2006-11-02 17:19:21 +01:00
|
|
|
/******************************************************************************
|
|
|
|
*******************************************************************************
|
|
|
|
**
|
|
|
|
** Copyright (C) Sistina Software, Inc. 1997-2003 All rights reserved.
|
|
|
|
** Copyright (C) 2004-2006 Red Hat, Inc. All rights reserved.
|
|
|
|
**
|
|
|
|
** This copyrighted material is made available to anyone wishing to use,
|
|
|
|
** modify, copy, or redistribute it subject to the terms and conditions
|
|
|
|
** of the GNU General Public License v.2.
|
|
|
|
**
|
|
|
|
*******************************************************************************
|
|
|
|
******************************************************************************/
|
|
|
|
|
|
|
|
/*
|
|
|
|
* lowcomms.c
|
|
|
|
*
|
|
|
|
* This is the "low-level" comms layer.
|
|
|
|
*
|
|
|
|
* It is responsible for sending/receiving messages
|
|
|
|
* from other nodes in the cluster.
|
|
|
|
*
|
|
|
|
* Cluster nodes are referred to by their nodeids. nodeids are
|
|
|
|
* simply 32 bit numbers to the locking module - if they need to
|
|
|
|
* be expanded for the cluster infrastructure then that is it's
|
|
|
|
* responsibility. It is this layer's
|
|
|
|
* responsibility to resolve these into IP address or
|
|
|
|
* whatever it needs for inter-node communication.
|
|
|
|
*
|
|
|
|
* The comms level is two kernel threads that deal mainly with
|
|
|
|
* the receiving of messages from other nodes and passing them
|
|
|
|
* up to the mid-level comms layer (which understands the
|
|
|
|
* message format) for execution by the locking core, and
|
|
|
|
* a send thread which does all the setting up of connections
|
|
|
|
* to remote nodes and the sending of data. Threads are not allowed
|
|
|
|
* to send their own data because it may cause them to wait in times
|
|
|
|
* of high load. Also, this way, the sending thread can collect together
|
|
|
|
* messages bound for one node and send them in one block.
|
|
|
|
*
|
|
|
|
* I don't see any problem with the recv thread executing the locking
|
|
|
|
* code on behalf of remote processes as the locking code is
|
|
|
|
* short, efficient and never waits.
|
|
|
|
*
|
|
|
|
*/
|
|
|
|
|
|
|
|
|
|
|
|
#include <asm/ioctls.h>
|
|
|
|
#include <net/sock.h>
|
|
|
|
#include <net/tcp.h>
|
|
|
|
#include <linux/pagemap.h>
|
|
|
|
|
|
|
|
#include "dlm_internal.h"
|
|
|
|
#include "lowcomms.h"
|
|
|
|
#include "midcomms.h"
|
|
|
|
#include "config.h"
|
|
|
|
|
|
|
|
struct cbuf {
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
unsigned int base;
|
|
|
|
unsigned int len;
|
|
|
|
unsigned int mask;
|
2006-11-02 17:19:21 +01:00
|
|
|
};
|
|
|
|
|
|
|
|
#define NODE_INCREMENT 32
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
static void cbuf_add(struct cbuf *cb, int n)
|
|
|
|
{
|
|
|
|
cb->len += n;
|
|
|
|
}
|
2006-11-02 17:19:21 +01:00
|
|
|
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
static int cbuf_data(struct cbuf *cb)
|
|
|
|
{
|
|
|
|
return ((cb->base + cb->len) & cb->mask);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void cbuf_init(struct cbuf *cb, int size)
|
|
|
|
{
|
|
|
|
cb->base = cb->len = 0;
|
|
|
|
cb->mask = size-1;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void cbuf_eat(struct cbuf *cb, int n)
|
|
|
|
{
|
|
|
|
cb->len -= n;
|
|
|
|
cb->base += n;
|
|
|
|
cb->base &= cb->mask;
|
|
|
|
}
|
|
|
|
|
|
|
|
static bool cbuf_empty(struct cbuf *cb)
|
|
|
|
{
|
|
|
|
return cb->len == 0;
|
|
|
|
}
|
2006-11-02 17:19:21 +01:00
|
|
|
|
|
|
|
/* Maximum number of incoming messages to process before
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
doing a cond_resched()
|
2006-11-02 17:19:21 +01:00
|
|
|
*/
|
|
|
|
#define MAX_RX_MSG_COUNT 25
|
|
|
|
|
|
|
|
struct connection {
|
|
|
|
struct socket *sock; /* NULL if not connected */
|
|
|
|
uint32_t nodeid; /* So we know who we are in the list */
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
struct rw_semaphore sock_sem; /* Stop connect races */
|
|
|
|
struct list_head read_list; /* On this list when ready for reading */
|
|
|
|
struct list_head write_list; /* On this list when ready for writing */
|
|
|
|
struct list_head state_list; /* On this list when ready to connect */
|
2006-11-02 17:19:21 +01:00
|
|
|
unsigned long flags; /* bit 1,2 = We are on the read/write lists */
|
|
|
|
#define CF_READ_PENDING 1
|
|
|
|
#define CF_WRITE_PENDING 2
|
|
|
|
#define CF_CONNECT_PENDING 3
|
|
|
|
#define CF_IS_OTHERCON 4
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
struct list_head writequeue; /* List of outgoing writequeue_entries */
|
|
|
|
struct list_head listenlist; /* List of allocated listening sockets */
|
2006-11-02 17:19:21 +01:00
|
|
|
spinlock_t writequeue_lock;
|
|
|
|
int (*rx_action) (struct connection *); /* What to do when active */
|
|
|
|
struct page *rx_page;
|
|
|
|
struct cbuf cb;
|
|
|
|
int retries;
|
|
|
|
atomic_t waiting_requests;
|
|
|
|
#define MAX_CONNECT_RETRIES 3
|
|
|
|
struct connection *othercon;
|
|
|
|
};
|
|
|
|
#define sock2con(x) ((struct connection *)(x)->sk_user_data)
|
|
|
|
|
|
|
|
/* An entry waiting to be sent */
|
|
|
|
struct writequeue_entry {
|
|
|
|
struct list_head list;
|
|
|
|
struct page *page;
|
|
|
|
int offset;
|
|
|
|
int len;
|
|
|
|
int end;
|
|
|
|
int users;
|
|
|
|
struct connection *con;
|
|
|
|
};
|
|
|
|
|
|
|
|
static struct sockaddr_storage dlm_local_addr;
|
|
|
|
|
|
|
|
/* Manage daemons */
|
|
|
|
static struct task_struct *recv_task;
|
|
|
|
static struct task_struct *send_task;
|
|
|
|
|
|
|
|
static wait_queue_t lowcomms_send_waitq_head;
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
static DECLARE_WAIT_QUEUE_HEAD(lowcomms_send_waitq);
|
2006-11-02 17:19:21 +01:00
|
|
|
static wait_queue_t lowcomms_recv_waitq_head;
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
static DECLARE_WAIT_QUEUE_HEAD(lowcomms_recv_waitq);
|
2006-11-02 17:19:21 +01:00
|
|
|
|
|
|
|
/* An array of pointers to connections, indexed by NODEID */
|
|
|
|
static struct connection **connections;
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
static DECLARE_MUTEX(connections_lock);
|
2006-11-02 17:19:21 +01:00
|
|
|
static kmem_cache_t *con_cache;
|
|
|
|
static int conn_array_size;
|
|
|
|
|
|
|
|
/* List of sockets that have reads pending */
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
static LIST_HEAD(read_sockets);
|
|
|
|
static DEFINE_SPINLOCK(read_sockets_lock);
|
2006-11-02 17:19:21 +01:00
|
|
|
|
|
|
|
/* List of sockets which have writes pending */
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
static LIST_HEAD(write_sockets);
|
|
|
|
static DEFINE_SPINLOCK(write_sockets_lock);
|
2006-11-02 17:19:21 +01:00
|
|
|
|
|
|
|
/* List of sockets which have connects pending */
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
static LIST_HEAD(state_sockets);
|
|
|
|
static DEFINE_SPINLOCK(state_sockets_lock);
|
2006-11-02 17:19:21 +01:00
|
|
|
|
|
|
|
static struct connection *nodeid2con(int nodeid, gfp_t allocation)
|
|
|
|
{
|
|
|
|
struct connection *con = NULL;
|
|
|
|
|
|
|
|
down(&connections_lock);
|
|
|
|
if (nodeid >= conn_array_size) {
|
|
|
|
int new_size = nodeid + NODE_INCREMENT;
|
|
|
|
struct connection **new_conns;
|
|
|
|
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
new_conns = kzalloc(sizeof(struct connection *) *
|
2006-11-02 17:19:21 +01:00
|
|
|
new_size, allocation);
|
|
|
|
if (!new_conns)
|
|
|
|
goto finish;
|
|
|
|
|
|
|
|
memcpy(new_conns, connections, sizeof(struct connection *) * conn_array_size);
|
|
|
|
conn_array_size = new_size;
|
|
|
|
kfree(connections);
|
|
|
|
connections = new_conns;
|
|
|
|
|
|
|
|
}
|
|
|
|
|
|
|
|
con = connections[nodeid];
|
|
|
|
if (con == NULL && allocation) {
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
con = kmem_cache_zalloc(con_cache, allocation);
|
2006-11-02 17:19:21 +01:00
|
|
|
if (!con)
|
|
|
|
goto finish;
|
|
|
|
|
|
|
|
con->nodeid = nodeid;
|
|
|
|
init_rwsem(&con->sock_sem);
|
|
|
|
INIT_LIST_HEAD(&con->writequeue);
|
|
|
|
spin_lock_init(&con->writequeue_lock);
|
|
|
|
|
|
|
|
connections[nodeid] = con;
|
|
|
|
}
|
|
|
|
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
finish:
|
2006-11-02 17:19:21 +01:00
|
|
|
up(&connections_lock);
|
|
|
|
return con;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Data available on socket or listen socket received a connect */
|
|
|
|
static void lowcomms_data_ready(struct sock *sk, int count_unused)
|
|
|
|
{
|
|
|
|
struct connection *con = sock2con(sk);
|
|
|
|
|
|
|
|
atomic_inc(&con->waiting_requests);
|
|
|
|
if (test_and_set_bit(CF_READ_PENDING, &con->flags))
|
|
|
|
return;
|
|
|
|
|
|
|
|
spin_lock_bh(&read_sockets_lock);
|
|
|
|
list_add_tail(&con->read_list, &read_sockets);
|
|
|
|
spin_unlock_bh(&read_sockets_lock);
|
|
|
|
|
|
|
|
wake_up_interruptible(&lowcomms_recv_waitq);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void lowcomms_write_space(struct sock *sk)
|
|
|
|
{
|
|
|
|
struct connection *con = sock2con(sk);
|
|
|
|
|
|
|
|
if (test_and_set_bit(CF_WRITE_PENDING, &con->flags))
|
|
|
|
return;
|
|
|
|
|
|
|
|
spin_lock_bh(&write_sockets_lock);
|
|
|
|
list_add_tail(&con->write_list, &write_sockets);
|
|
|
|
spin_unlock_bh(&write_sockets_lock);
|
|
|
|
|
|
|
|
wake_up_interruptible(&lowcomms_send_waitq);
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline void lowcomms_connect_sock(struct connection *con)
|
|
|
|
{
|
|
|
|
if (test_and_set_bit(CF_CONNECT_PENDING, &con->flags))
|
|
|
|
return;
|
|
|
|
|
|
|
|
spin_lock_bh(&state_sockets_lock);
|
|
|
|
list_add_tail(&con->state_list, &state_sockets);
|
|
|
|
spin_unlock_bh(&state_sockets_lock);
|
|
|
|
|
|
|
|
wake_up_interruptible(&lowcomms_send_waitq);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void lowcomms_state_change(struct sock *sk)
|
|
|
|
{
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
if (sk->sk_state == TCP_ESTABLISHED)
|
2006-11-02 17:19:21 +01:00
|
|
|
lowcomms_write_space(sk);
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Make a socket active */
|
|
|
|
static int add_sock(struct socket *sock, struct connection *con)
|
|
|
|
{
|
|
|
|
con->sock = sock;
|
|
|
|
|
|
|
|
/* Install a data_ready callback */
|
|
|
|
con->sock->sk->sk_data_ready = lowcomms_data_ready;
|
|
|
|
con->sock->sk->sk_write_space = lowcomms_write_space;
|
|
|
|
con->sock->sk->sk_state_change = lowcomms_state_change;
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Add the port number to an IP6 or 4 sockaddr and return the address
|
|
|
|
length */
|
|
|
|
static void make_sockaddr(struct sockaddr_storage *saddr, uint16_t port,
|
|
|
|
int *addr_len)
|
|
|
|
{
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
saddr->ss_family = dlm_local_addr.ss_family;
|
|
|
|
if (saddr->ss_family == AF_INET) {
|
2006-11-02 17:19:21 +01:00
|
|
|
struct sockaddr_in *in4_addr = (struct sockaddr_in *)saddr;
|
|
|
|
in4_addr->sin_port = cpu_to_be16(port);
|
|
|
|
*addr_len = sizeof(struct sockaddr_in);
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
} else {
|
2006-11-02 17:19:21 +01:00
|
|
|
struct sockaddr_in6 *in6_addr = (struct sockaddr_in6 *)saddr;
|
|
|
|
in6_addr->sin6_port = cpu_to_be16(port);
|
|
|
|
*addr_len = sizeof(struct sockaddr_in6);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Close a remote connection and tidy up */
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
static void close_connection(struct connection *con, bool and_other)
|
2006-11-02 17:19:21 +01:00
|
|
|
{
|
|
|
|
down_write(&con->sock_sem);
|
|
|
|
|
|
|
|
if (con->sock) {
|
|
|
|
sock_release(con->sock);
|
|
|
|
con->sock = NULL;
|
|
|
|
}
|
|
|
|
if (con->othercon && and_other) {
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
/* Will only re-enter once. */
|
|
|
|
close_connection(con->othercon, false);
|
2006-11-02 17:19:21 +01:00
|
|
|
}
|
|
|
|
if (con->rx_page) {
|
|
|
|
__free_page(con->rx_page);
|
|
|
|
con->rx_page = NULL;
|
|
|
|
}
|
|
|
|
con->retries = 0;
|
|
|
|
up_write(&con->sock_sem);
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Data received from remote end */
|
|
|
|
static int receive_from_sock(struct connection *con)
|
|
|
|
{
|
|
|
|
int ret = 0;
|
|
|
|
struct msghdr msg;
|
|
|
|
struct iovec iov[2];
|
|
|
|
mm_segment_t fs;
|
|
|
|
unsigned len;
|
|
|
|
int r;
|
|
|
|
int call_again_soon = 0;
|
|
|
|
|
|
|
|
down_read(&con->sock_sem);
|
|
|
|
|
|
|
|
if (con->sock == NULL)
|
|
|
|
goto out;
|
|
|
|
if (con->rx_page == NULL) {
|
|
|
|
/*
|
|
|
|
* This doesn't need to be atomic, but I think it should
|
|
|
|
* improve performance if it is.
|
|
|
|
*/
|
|
|
|
con->rx_page = alloc_page(GFP_ATOMIC);
|
|
|
|
if (con->rx_page == NULL)
|
|
|
|
goto out_resched;
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
cbuf_init(&con->cb, PAGE_CACHE_SIZE);
|
2006-11-02 17:19:21 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
msg.msg_control = NULL;
|
|
|
|
msg.msg_controllen = 0;
|
|
|
|
msg.msg_iovlen = 1;
|
|
|
|
msg.msg_iov = iov;
|
|
|
|
msg.msg_name = NULL;
|
|
|
|
msg.msg_namelen = 0;
|
|
|
|
msg.msg_flags = 0;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* iov[0] is the bit of the circular buffer between the current end
|
|
|
|
* point (cb.base + cb.len) and the end of the buffer.
|
|
|
|
*/
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
iov[0].iov_len = con->cb.base - cbuf_data(&con->cb);
|
|
|
|
iov[0].iov_base = page_address(con->rx_page) + cbuf_data(&con->cb);
|
2006-11-02 17:19:21 +01:00
|
|
|
iov[1].iov_len = 0;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* iov[1] is the bit of the circular buffer between the start of the
|
|
|
|
* buffer and the start of the currently used section (cb.base)
|
|
|
|
*/
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
if (cbuf_data(&con->cb) >= con->cb.base) {
|
|
|
|
iov[0].iov_len = PAGE_CACHE_SIZE - cbuf_data(&con->cb);
|
2006-11-02 17:19:21 +01:00
|
|
|
iov[1].iov_len = con->cb.base;
|
|
|
|
iov[1].iov_base = page_address(con->rx_page);
|
|
|
|
msg.msg_iovlen = 2;
|
|
|
|
}
|
|
|
|
len = iov[0].iov_len + iov[1].iov_len;
|
|
|
|
|
|
|
|
fs = get_fs();
|
|
|
|
set_fs(get_ds());
|
|
|
|
r = ret = sock_recvmsg(con->sock, &msg, len,
|
|
|
|
MSG_DONTWAIT | MSG_NOSIGNAL);
|
|
|
|
set_fs(fs);
|
|
|
|
|
|
|
|
if (ret <= 0)
|
|
|
|
goto out_close;
|
|
|
|
if (ret == len)
|
|
|
|
call_again_soon = 1;
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
cbuf_add(&con->cb, ret);
|
2006-11-02 17:19:21 +01:00
|
|
|
ret = dlm_process_incoming_buffer(con->nodeid,
|
|
|
|
page_address(con->rx_page),
|
|
|
|
con->cb.base, con->cb.len,
|
|
|
|
PAGE_CACHE_SIZE);
|
|
|
|
if (ret == -EBADMSG) {
|
|
|
|
printk(KERN_INFO "dlm: lowcomms: addr=%p, base=%u, len=%u, "
|
|
|
|
"iov_len=%u, iov_base[0]=%p, read=%d\n",
|
|
|
|
page_address(con->rx_page), con->cb.base, con->cb.len,
|
|
|
|
len, iov[0].iov_base, r);
|
|
|
|
}
|
|
|
|
if (ret < 0)
|
|
|
|
goto out_close;
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
cbuf_eat(&con->cb, ret);
|
2006-11-02 17:19:21 +01:00
|
|
|
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
if (cbuf_empty(&con->cb) && !call_again_soon) {
|
2006-11-02 17:19:21 +01:00
|
|
|
__free_page(con->rx_page);
|
|
|
|
con->rx_page = NULL;
|
|
|
|
}
|
|
|
|
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
out:
|
2006-11-02 17:19:21 +01:00
|
|
|
if (call_again_soon)
|
|
|
|
goto out_resched;
|
|
|
|
up_read(&con->sock_sem);
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
return 0;
|
2006-11-02 17:19:21 +01:00
|
|
|
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
out_resched:
|
2006-11-02 17:19:21 +01:00
|
|
|
lowcomms_data_ready(con->sock->sk, 0);
|
|
|
|
up_read(&con->sock_sem);
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
cond_resched();
|
|
|
|
return 0;
|
2006-11-02 17:19:21 +01:00
|
|
|
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
out_close:
|
2006-11-02 17:19:21 +01:00
|
|
|
up_read(&con->sock_sem);
|
|
|
|
if (ret != -EAGAIN && !test_bit(CF_IS_OTHERCON, &con->flags)) {
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
close_connection(con, false);
|
2006-11-02 17:19:21 +01:00
|
|
|
/* Reconnect when there is something to send */
|
|
|
|
}
|
|
|
|
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Listening socket is busy, accept a connection */
|
|
|
|
static int accept_from_sock(struct connection *con)
|
|
|
|
{
|
|
|
|
int result;
|
|
|
|
struct sockaddr_storage peeraddr;
|
|
|
|
struct socket *newsock;
|
|
|
|
int len;
|
|
|
|
int nodeid;
|
|
|
|
struct connection *newcon;
|
|
|
|
|
|
|
|
memset(&peeraddr, 0, sizeof(peeraddr));
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
result = sock_create_kern(dlm_local_addr.ss_family, SOCK_STREAM,
|
|
|
|
IPPROTO_TCP, &newsock);
|
2006-11-02 17:19:21 +01:00
|
|
|
if (result < 0)
|
|
|
|
return -ENOMEM;
|
|
|
|
|
|
|
|
down_read(&con->sock_sem);
|
|
|
|
|
|
|
|
result = -ENOTCONN;
|
|
|
|
if (con->sock == NULL)
|
|
|
|
goto accept_err;
|
|
|
|
|
|
|
|
newsock->type = con->sock->type;
|
|
|
|
newsock->ops = con->sock->ops;
|
|
|
|
|
|
|
|
result = con->sock->ops->accept(con->sock, newsock, O_NONBLOCK);
|
|
|
|
if (result < 0)
|
|
|
|
goto accept_err;
|
|
|
|
|
|
|
|
/* Get the connected socket's peer */
|
|
|
|
memset(&peeraddr, 0, sizeof(peeraddr));
|
|
|
|
if (newsock->ops->getname(newsock, (struct sockaddr *)&peeraddr,
|
|
|
|
&len, 2)) {
|
|
|
|
result = -ECONNABORTED;
|
|
|
|
goto accept_err;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Get the new node's NODEID */
|
|
|
|
make_sockaddr(&peeraddr, 0, &len);
|
|
|
|
if (dlm_addr_to_nodeid(&peeraddr, &nodeid)) {
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
printk("dlm: connect from non cluster node\n");
|
2006-11-02 17:19:21 +01:00
|
|
|
sock_release(newsock);
|
|
|
|
up_read(&con->sock_sem);
|
|
|
|
return -1;
|
|
|
|
}
|
|
|
|
|
|
|
|
log_print("got connection from %d", nodeid);
|
|
|
|
|
|
|
|
/* Check to see if we already have a connection to this node. This
|
|
|
|
* could happen if the two nodes initiate a connection at roughly
|
|
|
|
* the same time and the connections cross on the wire.
|
|
|
|
* TEMPORARY FIX:
|
|
|
|
* In this case we store the incoming one in "othercon"
|
|
|
|
*/
|
|
|
|
newcon = nodeid2con(nodeid, GFP_KERNEL);
|
|
|
|
if (!newcon) {
|
|
|
|
result = -ENOMEM;
|
|
|
|
goto accept_err;
|
|
|
|
}
|
|
|
|
down_write(&newcon->sock_sem);
|
|
|
|
if (newcon->sock) {
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
struct connection *othercon = newcon->othercon;
|
2006-11-02 17:19:21 +01:00
|
|
|
|
|
|
|
if (!othercon) {
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
othercon = kmem_cache_zalloc(con_cache, GFP_KERNEL);
|
2006-11-02 17:19:21 +01:00
|
|
|
if (!othercon) {
|
|
|
|
printk("dlm: failed to allocate incoming socket\n");
|
|
|
|
up_write(&newcon->sock_sem);
|
|
|
|
result = -ENOMEM;
|
|
|
|
goto accept_err;
|
|
|
|
}
|
|
|
|
othercon->nodeid = nodeid;
|
|
|
|
othercon->rx_action = receive_from_sock;
|
|
|
|
init_rwsem(&othercon->sock_sem);
|
|
|
|
set_bit(CF_IS_OTHERCON, &othercon->flags);
|
|
|
|
newcon->othercon = othercon;
|
|
|
|
}
|
|
|
|
othercon->sock = newsock;
|
|
|
|
newsock->sk->sk_user_data = othercon;
|
|
|
|
add_sock(newsock, othercon);
|
|
|
|
}
|
|
|
|
else {
|
|
|
|
newsock->sk->sk_user_data = newcon;
|
|
|
|
newcon->rx_action = receive_from_sock;
|
|
|
|
add_sock(newsock, newcon);
|
|
|
|
|
|
|
|
}
|
|
|
|
|
|
|
|
up_write(&newcon->sock_sem);
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Add it to the active queue in case we got data
|
|
|
|
* beween processing the accept adding the socket
|
|
|
|
* to the read_sockets list
|
|
|
|
*/
|
|
|
|
lowcomms_data_ready(newsock->sk, 0);
|
|
|
|
up_read(&con->sock_sem);
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
accept_err:
|
2006-11-02 17:19:21 +01:00
|
|
|
up_read(&con->sock_sem);
|
|
|
|
sock_release(newsock);
|
|
|
|
|
|
|
|
if (result != -EAGAIN)
|
|
|
|
printk("dlm: error accepting connection from node: %d\n", result);
|
|
|
|
return result;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Connect a new socket to its peer */
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
static void connect_to_sock(struct connection *con)
|
2006-11-02 17:19:21 +01:00
|
|
|
{
|
|
|
|
int result = -EHOSTUNREACH;
|
|
|
|
struct sockaddr_storage saddr;
|
|
|
|
int addr_len;
|
|
|
|
struct socket *sock;
|
|
|
|
|
|
|
|
if (con->nodeid == 0) {
|
|
|
|
log_print("attempt to connect sock 0 foiled");
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
return;
|
2006-11-02 17:19:21 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
down_write(&con->sock_sem);
|
|
|
|
if (con->retries++ > MAX_CONNECT_RETRIES)
|
|
|
|
goto out;
|
|
|
|
|
|
|
|
/* Some odd races can cause double-connects, ignore them */
|
|
|
|
if (con->sock) {
|
|
|
|
result = 0;
|
|
|
|
goto out;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Create a socket to communicate with */
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
result = sock_create_kern(dlm_local_addr.ss_family, SOCK_STREAM,
|
|
|
|
IPPROTO_TCP, &sock);
|
2006-11-02 17:19:21 +01:00
|
|
|
if (result < 0)
|
|
|
|
goto out_err;
|
|
|
|
|
|
|
|
memset(&saddr, 0, sizeof(saddr));
|
|
|
|
if (dlm_nodeid_to_addr(con->nodeid, &saddr))
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
goto out_err;
|
2006-11-02 17:19:21 +01:00
|
|
|
|
|
|
|
sock->sk->sk_user_data = con;
|
|
|
|
con->rx_action = receive_from_sock;
|
|
|
|
|
|
|
|
make_sockaddr(&saddr, dlm_config.tcp_port, &addr_len);
|
|
|
|
|
|
|
|
add_sock(sock, con);
|
|
|
|
|
|
|
|
log_print("connecting to %d", con->nodeid);
|
|
|
|
result =
|
|
|
|
sock->ops->connect(sock, (struct sockaddr *)&saddr, addr_len,
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
O_NONBLOCK);
|
2006-11-02 17:19:21 +01:00
|
|
|
if (result == -EINPROGRESS)
|
|
|
|
result = 0;
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
if (result == 0)
|
|
|
|
goto out;
|
2006-11-02 17:19:21 +01:00
|
|
|
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
out_err:
|
2006-11-02 17:19:21 +01:00
|
|
|
if (con->sock) {
|
|
|
|
sock_release(con->sock);
|
|
|
|
con->sock = NULL;
|
|
|
|
}
|
|
|
|
/*
|
|
|
|
* Some errors are fatal and this list might need adjusting. For other
|
|
|
|
* errors we try again until the max number of retries is reached.
|
|
|
|
*/
|
|
|
|
if (result != -EHOSTUNREACH && result != -ENETUNREACH &&
|
|
|
|
result != -ENETDOWN && result != EINVAL
|
|
|
|
&& result != -EPROTONOSUPPORT) {
|
|
|
|
lowcomms_connect_sock(con);
|
|
|
|
result = 0;
|
|
|
|
}
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
out:
|
|
|
|
up_write(&con->sock_sem);
|
|
|
|
return;
|
2006-11-02 17:19:21 +01:00
|
|
|
}
|
|
|
|
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
static struct socket *create_listen_sock(struct connection *con,
|
|
|
|
struct sockaddr_storage *saddr)
|
2006-11-02 17:19:21 +01:00
|
|
|
{
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
struct socket *sock = NULL;
|
2006-11-02 17:19:21 +01:00
|
|
|
mm_segment_t fs;
|
|
|
|
int result = 0;
|
|
|
|
int one = 1;
|
|
|
|
int addr_len;
|
|
|
|
|
|
|
|
if (dlm_local_addr.ss_family == AF_INET)
|
|
|
|
addr_len = sizeof(struct sockaddr_in);
|
|
|
|
else
|
|
|
|
addr_len = sizeof(struct sockaddr_in6);
|
|
|
|
|
|
|
|
/* Create a socket to communicate with */
|
|
|
|
result = sock_create_kern(dlm_local_addr.ss_family, SOCK_STREAM, IPPROTO_TCP, &sock);
|
|
|
|
if (result < 0) {
|
|
|
|
printk("dlm: Can't create listening comms socket\n");
|
|
|
|
goto create_out;
|
|
|
|
}
|
|
|
|
|
|
|
|
fs = get_fs();
|
|
|
|
set_fs(get_ds());
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
result = sock_setsockopt(sock, SOL_SOCKET, SO_REUSEADDR,
|
|
|
|
(char *)&one, sizeof(one));
|
2006-11-02 17:19:21 +01:00
|
|
|
set_fs(fs);
|
|
|
|
if (result < 0) {
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
printk("dlm: Failed to set SO_REUSEADDR on socket: result=%d\n",
|
|
|
|
result);
|
2006-11-02 17:19:21 +01:00
|
|
|
}
|
|
|
|
sock->sk->sk_user_data = con;
|
|
|
|
con->rx_action = accept_from_sock;
|
|
|
|
con->sock = sock;
|
|
|
|
|
|
|
|
/* Bind to our port */
|
|
|
|
make_sockaddr(saddr, dlm_config.tcp_port, &addr_len);
|
|
|
|
result = sock->ops->bind(sock, (struct sockaddr *) saddr, addr_len);
|
|
|
|
if (result < 0) {
|
|
|
|
printk("dlm: Can't bind to port %d\n", dlm_config.tcp_port);
|
|
|
|
sock_release(sock);
|
|
|
|
sock = NULL;
|
|
|
|
con->sock = NULL;
|
|
|
|
goto create_out;
|
|
|
|
}
|
|
|
|
|
|
|
|
fs = get_fs();
|
|
|
|
set_fs(get_ds());
|
|
|
|
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
result = sock_setsockopt(sock, SOL_SOCKET, SO_KEEPALIVE,
|
|
|
|
(char *)&one, sizeof(one));
|
2006-11-02 17:19:21 +01:00
|
|
|
set_fs(fs);
|
|
|
|
if (result < 0) {
|
|
|
|
printk("dlm: Set keepalive failed: %d\n", result);
|
|
|
|
}
|
|
|
|
|
|
|
|
result = sock->ops->listen(sock, 5);
|
|
|
|
if (result < 0) {
|
|
|
|
printk("dlm: Can't listen on port %d\n", dlm_config.tcp_port);
|
|
|
|
sock_release(sock);
|
|
|
|
sock = NULL;
|
|
|
|
goto create_out;
|
|
|
|
}
|
|
|
|
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
create_out:
|
2006-11-02 17:19:21 +01:00
|
|
|
return sock;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
/* Listen on all interfaces */
|
|
|
|
static int listen_for_all(void)
|
|
|
|
{
|
|
|
|
struct socket *sock = NULL;
|
|
|
|
struct connection *con = nodeid2con(0, GFP_KERNEL);
|
|
|
|
int result = -EINVAL;
|
|
|
|
|
|
|
|
/* We don't support multi-homed hosts */
|
|
|
|
set_bit(CF_IS_OTHERCON, &con->flags);
|
|
|
|
|
|
|
|
sock = create_listen_sock(con, &dlm_local_addr);
|
|
|
|
if (sock) {
|
|
|
|
add_sock(sock, con);
|
|
|
|
result = 0;
|
|
|
|
}
|
|
|
|
else {
|
|
|
|
result = -EADDRINUSE;
|
|
|
|
}
|
|
|
|
|
|
|
|
return result;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
static struct writequeue_entry *new_writequeue_entry(struct connection *con,
|
|
|
|
gfp_t allocation)
|
|
|
|
{
|
|
|
|
struct writequeue_entry *entry;
|
|
|
|
|
|
|
|
entry = kmalloc(sizeof(struct writequeue_entry), allocation);
|
|
|
|
if (!entry)
|
|
|
|
return NULL;
|
|
|
|
|
|
|
|
entry->page = alloc_page(allocation);
|
|
|
|
if (!entry->page) {
|
|
|
|
kfree(entry);
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
|
|
|
entry->offset = 0;
|
|
|
|
entry->len = 0;
|
|
|
|
entry->end = 0;
|
|
|
|
entry->users = 0;
|
|
|
|
entry->con = con;
|
|
|
|
|
|
|
|
return entry;
|
|
|
|
}
|
|
|
|
|
|
|
|
void *dlm_lowcomms_get_buffer(int nodeid, int len,
|
|
|
|
gfp_t allocation, char **ppc)
|
|
|
|
{
|
|
|
|
struct connection *con;
|
|
|
|
struct writequeue_entry *e;
|
|
|
|
int offset = 0;
|
|
|
|
int users = 0;
|
|
|
|
|
|
|
|
con = nodeid2con(nodeid, allocation);
|
|
|
|
if (!con)
|
|
|
|
return NULL;
|
|
|
|
|
|
|
|
e = list_entry(con->writequeue.prev, struct writequeue_entry, list);
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
if ((&e->list == &con->writequeue) ||
|
2006-11-02 17:19:21 +01:00
|
|
|
(PAGE_CACHE_SIZE - e->end < len)) {
|
|
|
|
e = NULL;
|
|
|
|
} else {
|
|
|
|
offset = e->end;
|
|
|
|
e->end += len;
|
|
|
|
users = e->users++;
|
|
|
|
}
|
|
|
|
spin_unlock(&con->writequeue_lock);
|
|
|
|
|
|
|
|
if (e) {
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
got_one:
|
2006-11-02 17:19:21 +01:00
|
|
|
if (users == 0)
|
|
|
|
kmap(e->page);
|
|
|
|
*ppc = page_address(e->page) + offset;
|
|
|
|
return e;
|
|
|
|
}
|
|
|
|
|
|
|
|
e = new_writequeue_entry(con, allocation);
|
|
|
|
if (e) {
|
|
|
|
spin_lock(&con->writequeue_lock);
|
|
|
|
offset = e->end;
|
|
|
|
e->end += len;
|
|
|
|
users = e->users++;
|
|
|
|
list_add_tail(&e->list, &con->writequeue);
|
|
|
|
spin_unlock(&con->writequeue_lock);
|
|
|
|
goto got_one;
|
|
|
|
}
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
|
|
|
void dlm_lowcomms_commit_buffer(void *mh)
|
|
|
|
{
|
|
|
|
struct writequeue_entry *e = (struct writequeue_entry *)mh;
|
|
|
|
struct connection *con = e->con;
|
|
|
|
int users;
|
|
|
|
|
|
|
|
users = --e->users;
|
|
|
|
if (users)
|
|
|
|
goto out;
|
|
|
|
e->len = e->end - e->offset;
|
|
|
|
kunmap(e->page);
|
|
|
|
spin_unlock(&con->writequeue_lock);
|
|
|
|
|
|
|
|
if (test_and_set_bit(CF_WRITE_PENDING, &con->flags) == 0) {
|
|
|
|
spin_lock_bh(&write_sockets_lock);
|
|
|
|
list_add_tail(&con->write_list, &write_sockets);
|
|
|
|
spin_unlock_bh(&write_sockets_lock);
|
|
|
|
|
|
|
|
wake_up_interruptible(&lowcomms_send_waitq);
|
|
|
|
}
|
|
|
|
return;
|
|
|
|
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
out:
|
2006-11-02 17:19:21 +01:00
|
|
|
spin_unlock(&con->writequeue_lock);
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void free_entry(struct writequeue_entry *e)
|
|
|
|
{
|
|
|
|
__free_page(e->page);
|
|
|
|
kfree(e);
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Send a message */
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
static void send_to_sock(struct connection *con)
|
2006-11-02 17:19:21 +01:00
|
|
|
{
|
|
|
|
int ret = 0;
|
|
|
|
ssize_t(*sendpage) (struct socket *, struct page *, int, size_t, int);
|
|
|
|
const int msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL;
|
|
|
|
struct writequeue_entry *e;
|
|
|
|
int len, offset;
|
|
|
|
|
|
|
|
down_read(&con->sock_sem);
|
|
|
|
if (con->sock == NULL)
|
|
|
|
goto out_connect;
|
|
|
|
|
|
|
|
sendpage = con->sock->ops->sendpage;
|
|
|
|
|
|
|
|
spin_lock(&con->writequeue_lock);
|
|
|
|
for (;;) {
|
|
|
|
e = list_entry(con->writequeue.next, struct writequeue_entry,
|
|
|
|
list);
|
|
|
|
if ((struct list_head *) e == &con->writequeue)
|
|
|
|
break;
|
|
|
|
|
|
|
|
len = e->len;
|
|
|
|
offset = e->offset;
|
|
|
|
BUG_ON(len == 0 && e->users == 0);
|
|
|
|
spin_unlock(&con->writequeue_lock);
|
|
|
|
|
|
|
|
ret = 0;
|
|
|
|
if (len) {
|
|
|
|
ret = sendpage(con->sock, e->page, offset, len,
|
|
|
|
msg_flags);
|
|
|
|
if (ret == -EAGAIN || ret == 0)
|
|
|
|
goto out;
|
|
|
|
if (ret <= 0)
|
|
|
|
goto send_error;
|
|
|
|
}
|
|
|
|
else {
|
|
|
|
/* Don't starve people filling buffers */
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
cond_resched();
|
2006-11-02 17:19:21 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
spin_lock(&con->writequeue_lock);
|
|
|
|
e->offset += ret;
|
|
|
|
e->len -= ret;
|
|
|
|
|
|
|
|
if (e->len == 0 && e->users == 0) {
|
|
|
|
list_del(&e->list);
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
kunmap(e->page);
|
2006-11-02 17:19:21 +01:00
|
|
|
free_entry(e);
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
spin_unlock(&con->writequeue_lock);
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
out:
|
2006-11-02 17:19:21 +01:00
|
|
|
up_read(&con->sock_sem);
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
return;
|
2006-11-02 17:19:21 +01:00
|
|
|
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
send_error:
|
2006-11-02 17:19:21 +01:00
|
|
|
up_read(&con->sock_sem);
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
close_connection(con, false);
|
2006-11-02 17:19:21 +01:00
|
|
|
lowcomms_connect_sock(con);
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
return;
|
2006-11-02 17:19:21 +01:00
|
|
|
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
out_connect:
|
2006-11-02 17:19:21 +01:00
|
|
|
up_read(&con->sock_sem);
|
|
|
|
lowcomms_connect_sock(con);
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
return;
|
2006-11-02 17:19:21 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
static void clean_one_writequeue(struct connection *con)
|
|
|
|
{
|
|
|
|
struct list_head *list;
|
|
|
|
struct list_head *temp;
|
|
|
|
|
|
|
|
spin_lock(&con->writequeue_lock);
|
|
|
|
list_for_each_safe(list, temp, &con->writequeue) {
|
|
|
|
struct writequeue_entry *e =
|
|
|
|
list_entry(list, struct writequeue_entry, list);
|
|
|
|
list_del(&e->list);
|
|
|
|
free_entry(e);
|
|
|
|
}
|
|
|
|
spin_unlock(&con->writequeue_lock);
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Called from recovery when it knows that a node has
|
|
|
|
left the cluster */
|
|
|
|
int dlm_lowcomms_close(int nodeid)
|
|
|
|
{
|
|
|
|
struct connection *con;
|
|
|
|
|
|
|
|
if (!connections)
|
|
|
|
goto out;
|
|
|
|
|
|
|
|
log_print("closing connection to node %d", nodeid);
|
|
|
|
con = nodeid2con(nodeid, 0);
|
|
|
|
if (con) {
|
|
|
|
clean_one_writequeue(con);
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
close_connection(con, true);
|
2006-11-02 17:19:21 +01:00
|
|
|
atomic_set(&con->waiting_requests, 0);
|
|
|
|
}
|
|
|
|
return 0;
|
|
|
|
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
out:
|
2006-11-02 17:19:21 +01:00
|
|
|
return -1;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* API send message call, may queue the request */
|
|
|
|
/* N.B. This is the old interface - use the new one for new calls */
|
|
|
|
int lowcomms_send_message(int nodeid, char *buf, int len, gfp_t allocation)
|
|
|
|
{
|
|
|
|
struct writequeue_entry *e;
|
|
|
|
char *b;
|
|
|
|
|
|
|
|
e = dlm_lowcomms_get_buffer(nodeid, len, allocation, &b);
|
|
|
|
if (e) {
|
|
|
|
memcpy(b, buf, len);
|
|
|
|
dlm_lowcomms_commit_buffer(e);
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
return -ENOBUFS;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Look for activity on active sockets */
|
|
|
|
static void process_sockets(void)
|
|
|
|
{
|
|
|
|
struct list_head *list;
|
|
|
|
struct list_head *temp;
|
|
|
|
int count = 0;
|
|
|
|
|
|
|
|
spin_lock_bh(&read_sockets_lock);
|
|
|
|
list_for_each_safe(list, temp, &read_sockets) {
|
|
|
|
|
|
|
|
struct connection *con =
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
list_entry(list, struct connection, read_list);
|
2006-11-02 17:19:21 +01:00
|
|
|
list_del(&con->read_list);
|
|
|
|
clear_bit(CF_READ_PENDING, &con->flags);
|
|
|
|
|
|
|
|
spin_unlock_bh(&read_sockets_lock);
|
|
|
|
|
|
|
|
/* This can reach zero if we are processing requests
|
|
|
|
* as they come in.
|
|
|
|
*/
|
|
|
|
if (atomic_read(&con->waiting_requests) == 0) {
|
|
|
|
spin_lock_bh(&read_sockets_lock);
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
do {
|
|
|
|
con->rx_action(con);
|
|
|
|
|
|
|
|
/* Don't starve out everyone else */
|
|
|
|
if (++count >= MAX_RX_MSG_COUNT) {
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
cond_resched();
|
2006-11-02 17:19:21 +01:00
|
|
|
count = 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
} while (!atomic_dec_and_test(&con->waiting_requests) &&
|
|
|
|
!kthread_should_stop());
|
|
|
|
|
|
|
|
spin_lock_bh(&read_sockets_lock);
|
|
|
|
}
|
|
|
|
spin_unlock_bh(&read_sockets_lock);
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Try to send any messages that are pending
|
|
|
|
*/
|
|
|
|
static void process_output_queue(void)
|
|
|
|
{
|
|
|
|
struct list_head *list;
|
|
|
|
struct list_head *temp;
|
|
|
|
|
|
|
|
spin_lock_bh(&write_sockets_lock);
|
|
|
|
list_for_each_safe(list, temp, &write_sockets) {
|
|
|
|
struct connection *con =
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
list_entry(list, struct connection, write_list);
|
2006-11-02 17:19:21 +01:00
|
|
|
clear_bit(CF_WRITE_PENDING, &con->flags);
|
|
|
|
list_del(&con->write_list);
|
|
|
|
|
|
|
|
spin_unlock_bh(&write_sockets_lock);
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
send_to_sock(con);
|
2006-11-02 17:19:21 +01:00
|
|
|
spin_lock_bh(&write_sockets_lock);
|
|
|
|
}
|
|
|
|
spin_unlock_bh(&write_sockets_lock);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void process_state_queue(void)
|
|
|
|
{
|
|
|
|
struct list_head *list;
|
|
|
|
struct list_head *temp;
|
|
|
|
|
|
|
|
spin_lock_bh(&state_sockets_lock);
|
|
|
|
list_for_each_safe(list, temp, &state_sockets) {
|
|
|
|
struct connection *con =
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
list_entry(list, struct connection, state_list);
|
2006-11-02 17:19:21 +01:00
|
|
|
list_del(&con->state_list);
|
|
|
|
clear_bit(CF_CONNECT_PENDING, &con->flags);
|
|
|
|
spin_unlock_bh(&state_sockets_lock);
|
|
|
|
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
connect_to_sock(con);
|
2006-11-02 17:19:21 +01:00
|
|
|
spin_lock_bh(&state_sockets_lock);
|
|
|
|
}
|
|
|
|
spin_unlock_bh(&state_sockets_lock);
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
/* Discard all entries on the write queues */
|
|
|
|
static void clean_writequeues(void)
|
|
|
|
{
|
|
|
|
int nodeid;
|
|
|
|
|
|
|
|
for (nodeid = 1; nodeid < conn_array_size; nodeid++) {
|
|
|
|
struct connection *con = nodeid2con(nodeid, 0);
|
|
|
|
|
|
|
|
if (con)
|
|
|
|
clean_one_writequeue(con);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
static int read_list_empty(void)
|
|
|
|
{
|
|
|
|
int status;
|
|
|
|
|
|
|
|
spin_lock_bh(&read_sockets_lock);
|
|
|
|
status = list_empty(&read_sockets);
|
|
|
|
spin_unlock_bh(&read_sockets_lock);
|
|
|
|
|
|
|
|
return status;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* DLM Transport comms receive daemon */
|
|
|
|
static int dlm_recvd(void *data)
|
|
|
|
{
|
|
|
|
init_waitqueue_entry(&lowcomms_recv_waitq_head, current);
|
|
|
|
add_wait_queue(&lowcomms_recv_waitq, &lowcomms_recv_waitq_head);
|
|
|
|
|
|
|
|
while (!kthread_should_stop()) {
|
|
|
|
set_current_state(TASK_INTERRUPTIBLE);
|
|
|
|
if (read_list_empty())
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
cond_resched();
|
2006-11-02 17:19:21 +01:00
|
|
|
set_current_state(TASK_RUNNING);
|
|
|
|
|
|
|
|
process_sockets();
|
|
|
|
}
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
static int write_and_state_lists_empty(void)
|
|
|
|
{
|
|
|
|
int status;
|
|
|
|
|
|
|
|
spin_lock_bh(&write_sockets_lock);
|
|
|
|
status = list_empty(&write_sockets);
|
|
|
|
spin_unlock_bh(&write_sockets_lock);
|
|
|
|
|
|
|
|
spin_lock_bh(&state_sockets_lock);
|
|
|
|
if (list_empty(&state_sockets) == 0)
|
|
|
|
status = 0;
|
|
|
|
spin_unlock_bh(&state_sockets_lock);
|
|
|
|
|
|
|
|
return status;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* DLM Transport send daemon */
|
|
|
|
static int dlm_sendd(void *data)
|
|
|
|
{
|
|
|
|
init_waitqueue_entry(&lowcomms_send_waitq_head, current);
|
|
|
|
add_wait_queue(&lowcomms_send_waitq, &lowcomms_send_waitq_head);
|
|
|
|
|
|
|
|
while (!kthread_should_stop()) {
|
|
|
|
set_current_state(TASK_INTERRUPTIBLE);
|
|
|
|
if (write_and_state_lists_empty())
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
cond_resched();
|
2006-11-02 17:19:21 +01:00
|
|
|
set_current_state(TASK_RUNNING);
|
|
|
|
|
|
|
|
process_state_queue();
|
|
|
|
process_output_queue();
|
|
|
|
}
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void daemons_stop(void)
|
|
|
|
{
|
|
|
|
kthread_stop(recv_task);
|
|
|
|
kthread_stop(send_task);
|
|
|
|
}
|
|
|
|
|
|
|
|
static int daemons_start(void)
|
|
|
|
{
|
|
|
|
struct task_struct *p;
|
|
|
|
int error;
|
|
|
|
|
|
|
|
p = kthread_run(dlm_recvd, NULL, "dlm_recvd");
|
|
|
|
error = IS_ERR(p);
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
if (error) {
|
2006-11-02 17:19:21 +01:00
|
|
|
log_print("can't start dlm_recvd %d", error);
|
|
|
|
return error;
|
|
|
|
}
|
|
|
|
recv_task = p;
|
|
|
|
|
|
|
|
p = kthread_run(dlm_sendd, NULL, "dlm_sendd");
|
|
|
|
error = IS_ERR(p);
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
if (error) {
|
2006-11-02 17:19:21 +01:00
|
|
|
log_print("can't start dlm_sendd %d", error);
|
|
|
|
kthread_stop(recv_task);
|
|
|
|
return error;
|
|
|
|
}
|
|
|
|
send_task = p;
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Return the largest buffer size we can cope with.
|
|
|
|
*/
|
|
|
|
int lowcomms_max_buffer_size(void)
|
|
|
|
{
|
|
|
|
return PAGE_CACHE_SIZE;
|
|
|
|
}
|
|
|
|
|
|
|
|
void dlm_lowcomms_stop(void)
|
|
|
|
{
|
|
|
|
int i;
|
|
|
|
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
/* Set all the flags to prevent any
|
2006-11-02 17:19:21 +01:00
|
|
|
socket activity.
|
|
|
|
*/
|
|
|
|
for (i = 0; i < conn_array_size; i++) {
|
|
|
|
if (connections[i])
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
connections[i]->flags |= 0xFF;
|
2006-11-02 17:19:21 +01:00
|
|
|
}
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
|
2006-11-02 17:19:21 +01:00
|
|
|
daemons_stop();
|
|
|
|
clean_writequeues();
|
|
|
|
|
|
|
|
for (i = 0; i < conn_array_size; i++) {
|
|
|
|
if (connections[i]) {
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
close_connection(connections[i], true);
|
2006-11-02 17:19:21 +01:00
|
|
|
if (connections[i]->othercon)
|
|
|
|
kmem_cache_free(con_cache, connections[i]->othercon);
|
|
|
|
kmem_cache_free(con_cache, connections[i]);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
kfree(connections);
|
|
|
|
connections = NULL;
|
|
|
|
|
|
|
|
kmem_cache_destroy(con_cache);
|
|
|
|
}
|
|
|
|
|
|
|
|
/* This is quite likely to sleep... */
|
|
|
|
int dlm_lowcomms_start(void)
|
|
|
|
{
|
|
|
|
int error = 0;
|
|
|
|
|
|
|
|
error = -ENOMEM;
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
connections = kzalloc(sizeof(struct connection *) *
|
2006-11-02 17:19:21 +01:00
|
|
|
NODE_INCREMENT, GFP_KERNEL);
|
|
|
|
if (!connections)
|
|
|
|
goto out;
|
|
|
|
|
|
|
|
conn_array_size = NODE_INCREMENT;
|
|
|
|
|
|
|
|
if (dlm_our_addr(&dlm_local_addr, 0)) {
|
|
|
|
log_print("no local IP address has been set");
|
|
|
|
goto fail_free_conn;
|
|
|
|
}
|
|
|
|
if (!dlm_our_addr(&dlm_local_addr, 1)) {
|
|
|
|
log_print("This dlm comms module does not support multi-homed clustering");
|
|
|
|
goto fail_free_conn;
|
|
|
|
}
|
|
|
|
|
|
|
|
con_cache = kmem_cache_create("dlm_conn", sizeof(struct connection),
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
__alignof__(struct connection), 0,
|
|
|
|
NULL, NULL);
|
2006-11-02 17:19:21 +01:00
|
|
|
if (!con_cache)
|
|
|
|
goto fail_free_conn;
|
|
|
|
|
|
|
|
|
|
|
|
/* Start listening */
|
|
|
|
error = listen_for_all();
|
|
|
|
if (error)
|
|
|
|
goto fail_unlisten;
|
|
|
|
|
|
|
|
error = daemons_start();
|
|
|
|
if (error)
|
|
|
|
goto fail_unlisten;
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
fail_unlisten:
|
|
|
|
close_connection(connections[0], false);
|
2006-11-02 17:19:21 +01:00
|
|
|
kmem_cache_free(con_cache, connections[0]);
|
|
|
|
kmem_cache_destroy(con_cache);
|
|
|
|
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
fail_free_conn:
|
2006-11-02 17:19:21 +01:00
|
|
|
kfree(connections);
|
|
|
|
|
[DLM] Clean up lowcomms
This fixes up most of the things pointed out by akpm and Pavel Machek
with comments below indicating why some things have been left:
Andrew Morton wrote:
>
>> +static struct nodeinfo *nodeid2nodeinfo(int nodeid, gfp_t alloc)
>> +{
>> + struct nodeinfo *ni;
>> + int r;
>> + int n;
>> +
>> + down_read(&nodeinfo_lock);
>
> Given that this function can sleep, I wonder if `alloc' is useful.
>
> I see lots of callers passing in a literal "0" for `alloc'. That's in fact
> a secret (GFP_ATOMIC & ~__GFP_HIGH). I doubt if that's what you really
> meant. Particularly as the code could at least have used __GFP_WAIT (aka
> GFP_NOIO) which is much, much more reliable than "0". In fact "0" is the
> least reliable mode possible.
>
> IOW, this is all bollixed up.
When 0 is passed into nodeid2nodeinfo the function does not try to allocate a
new structure at all. it's an indication that the caller only wants the nodeinfo
struct for that nodeid if there actually is one in existance.
I've tidied the function itself so it's more obvious, (and tidier!)
>> +/* Data received from remote end */
>> +static int receive_from_sock(void)
>> +{
>> + int ret = 0;
>> + struct msghdr msg;
>> + struct kvec iov[2];
>> + unsigned len;
>> + int r;
>> + struct sctp_sndrcvinfo *sinfo;
>> + struct cmsghdr *cmsg;
>> + struct nodeinfo *ni;
>> +
>> + /* These two are marginally too big for stack allocation, but this
>> + * function is (currently) only called by dlm_recvd so static should be
>> + * OK.
>> + */
>> + static struct sockaddr_storage msgname;
>> + static char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> whoa. This is globally singly-threaded code??
Yes. it is only ever run in the context of dlm_recvd.
>>
>> +static void initiate_association(int nodeid)
>> +{
>> + struct sockaddr_storage rem_addr;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Another static buffer to worry about. Globally singly-threaded code?
Yes. Only ever called by dlm_sendd.
>> +
>> +/* Send a message */
>> +static int send_to_sock(struct nodeinfo *ni)
>> +{
>> + int ret = 0;
>> + struct writequeue_entry *e;
>> + int len, offset;
>> + struct msghdr outmsg;
>> + static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
>
> Singly-threaded?
Yep.
>>
>> +static void dealloc_nodeinfo(void)
>> +{
>> + int i;
>> +
>> + for (i=1; i<=max_nodeid; i++) {
>> + struct nodeinfo *ni = nodeid2nodeinfo(i, 0);
>> + if (ni) {
>> + idr_remove(&nodeinfo_idr, i);
>
> Didn't that need locking?
Not. it's only ever called at DLM shutdown after all the other threads
have been stopped.
>>
>> +static int write_list_empty(void)
>> +{
>> + int status;
>> +
>> + spin_lock_bh(&write_nodes_lock);
>> + status = list_empty(&write_nodes);
>> + spin_unlock_bh(&write_nodes_lock);
>> +
>> + return status;
>> +}
>
> This function's return value is meaningless. As soon as the lock gets
> dropped, the return value can get out of sync with reality.
>
> Looking at the caller, this _might_ happen to be OK, but it's a nasty and
> dangerous thing. Really the locking should be moved into the caller.
It's just an optimisation to allow the caller to schedule if there is no work
to do. if something arrives immediately afterwards then it will get picked up
when the process re-awakes (and it will be woken by that arrival).
The 'accepting' atomic has gone completely. as Andrew pointed out it didn't
really achieve much anyway. I suspect it was a plaster over some other
startup or shutdown bug to be honest.
Signed-off-by: Patrick Caulfield <pcaulfie@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Pavel Machek <pavel@ucw.cz>
2006-12-06 16:10:37 +01:00
|
|
|
out:
|
2006-11-02 17:19:21 +01:00
|
|
|
return error;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Overrides for Emacs so that we follow Linus's tabbing style.
|
|
|
|
* Emacs will notice this stuff at the end of the file and automatically
|
|
|
|
* adjust the settings for this buffer only. This must remain at the end
|
|
|
|
* of the file.
|
|
|
|
* ---------------------------------------------------------------------------
|
|
|
|
* Local variables:
|
|
|
|
* c-file-style: "linux"
|
|
|
|
* End:
|
|
|
|
*/
|