mmapped netlink has a number of unresolved issues:
- TX zerocopy support had to be disabled more than a year ago via
commit 4682a03586 ("netlink: Always copy on mmap TX.")
because the content of the mmapped area can change after netlink
attribute validation but before message processing.
- RX support was implemented mainly to speed up nfqueue dumping packet
payload to userspace. However, since commit ae08ce0021
("netfilter: nfnetlink_queue: zero copy support") we avoid one copy
with the socket-based interface too (via the skb_zerocopy helper).
The other problem is that skbs attached to mmaped netlink socket
behave different from normal skbs:
- they don't have a shinfo area, so all functions that use skb_shinfo()
(e.g. skb_clone) cannot be used.
- reserving headroom prevents userspace from seeing the content as
it expects message to start at skb->head.
See for instance
commit aa3a022094 ("netlink: not trim skb for mmaped socket when dump").
- skbs handed e.g. to netlink_ack must have non-NULL skb->sk, else we
crash because it needs the sk to check if a tx ring is attached.
Also not obvious, leads to non-intuitive bug fixes such as 7c7bdf359
("netfilter: nfnetlink: use original skbuff when acking batches").
mmaped netlink also didn't play nicely with the skb_zerocopy helper
used by nfqueue and openvswitch. Daniel Borkmann fixed this via
commit 6bb0fef489 ("netlink, mmap: fix edge-case leakages in nf queue
zero-copy")' but at the cost of also needing to provide remaining
length to the allocation function.
nfqueue also has problems when used with mmaped rx netlink:
- mmaped netlink doesn't allow use of nfqueue batch verdict messages.
Problem is that in the mmap case, the allocation time also determines
the ordering in which the frame will be seen by userspace (A
allocating before B means that A is located in earlier ring slot,
but this also means that B might get a lower sequence number then A
since seqno is decided later. To fix this we would need to extend the
spinlocked region to also cover the allocation and message setup which
isn't desirable.
- nfqueue can now be configured to queue large (GSO) skbs to userspace.
Queing GSO packets is faster than having to force a software segmentation
in the kernel, so this is a desirable option. However, with a mmap based
ring one has to use 64kb per ring slot element, else mmap has to fall back
to the socket path (NL_MMAP_STATUS_COPY) for all large packets.
To use the mmap interface, userspace not only has to probe for mmap netlink
support, it also has to implement a recv/socket receive path in order to
handle messages that exceed the size of an rx ring element.
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Ken-ichirou MATSUZAWA <chamaken@gmail.com>
Cc: Pablo Neira Ayuso <pablo@netfilter.org>
Cc: Patrick McHardy <kaber@trash.net>
Cc: Thomas Graf <tgraf@suug.ch>
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
On Mon, Sep 21, 2015 at 02:20:22PM -0400, Tejun Heo wrote:
>
> store_release and load_acquire are different from the usual memory
> barriers and can't be paired this way. You have to pair store_release
> and load_acquire. Besides, it isn't a particularly good idea to
OK I've decided to drop the acquire/release helpers as they don't
help us at all and simply pessimises the code by using full memory
barriers (on some architectures) where only a write or read barrier
is needed.
> depend on memory barriers embedded in other data structures like the
> above. Here, especially, rhashtable_insert() would have write barrier
> *before* the entry is hashed not necessarily *after*, which means that
> in the above case, a socket which appears to have set bound to a
> reader might not visible when the reader tries to look up the socket
> on the hashtable.
But you are right we do need an explicit write barrier here to
ensure that the hashing is visible.
> There's no reason to be overly smart here. This isn't a crazy hot
> path, write barriers tend to be very cheap, store_release more so.
> Please just do smp_store_release() and note what it's paired with.
It's not about being overly smart. It's about actually understanding
what's going on with the code. I've seen too many instances of
people simply sprinkling synchronisation primitives around without
any knowledge of what is happening underneath, which is just a recipe
for creating hard-to-debug races.
> > @@ -1539,7 +1546,7 @@ static int netlink_bind(struct socket *sock, struct sockaddr *addr,
> > }
> > }
> >
> > - if (!nlk->portid) {
> > + if (!nlk->bound) {
>
> I don't think you can skip load_acquire here just because this is the
> second deref of the variable. That doesn't change anything. Race
> condition could still happen between the first and second tests and
> skipping the second would lead to the same kind of bug.
The reason this one is OK is because we do not use nlk->portid or
try to get nlk from the hash table before we return to user-space.
However, there is a real bug here that none of these acquire/release
helpers discovered. The two bound tests here used to be a single
one. Now that they are separate it is entirely possible for another
thread to come in the middle and bind the socket. So we need to
repeat the portid check in order to maintain consistency.
> > @@ -1587,7 +1594,7 @@ static int netlink_connect(struct socket *sock, struct sockaddr *addr,
> > !netlink_allowed(sock, NL_CFG_F_NONROOT_SEND))
> > return -EPERM;
> >
> > - if (!nlk->portid)
> > + if (!nlk->bound)
>
> Don't we need load_acquire here too? Is this path holding a lock
> which makes that unnecessary?
Ditto.
---8<---
The commit 1f770c0a09 ("netlink:
Fix autobind race condition that leads to zero port ID") created
some new races that can occur due to inconcsistencies between the
two port IDs.
Tejun is right that a barrier is unavoidable. Therefore I am
reverting to the original patch that used a boolean to indicate
that a user netlink socket has been bound.
Barriers have been added where necessary to ensure that a valid
portid and the hashed socket is visible.
I have also changed netlink_insert to only return EBUSY if the
socket is bound to a portid different to the requested one. This
combined with only reading nlk->bound once in netlink_bind fixes
a race where two threads that bind the socket at the same time
with different port IDs may both succeed.
Fixes: 1f770c0a09 ("netlink: Fix autobind race condition that leads to zero port ID")
Reported-by: Tejun Heo <tj@kernel.org>
Reported-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Nacked-by: Tejun Heo <tj@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
The commit c0bb07df7d ("netlink:
Reset portid after netlink_insert failure") introduced a race
condition where if two threads try to autobind the same socket
one of them may end up with a zero port ID. This led to kernel
deadlocks that were observed by multiple people.
This patch reverts that commit and instead fixes it by introducing
a separte rhash_portid variable so that the real portid is only set
after the socket has been successfully hashed.
Fixes: c0bb07df7d ("netlink: Reset portid after netlink_insert failure")
Reported-by: Tejun Heo <tj@kernel.org>
Reported-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
Ken-ichirou reported that running netlink in mmap mode for receive in
combination with nlmon will throw a NULL pointer dereference in
__kfree_skb() on nlmon_xmit(), in my case I can also trigger an "unable
to handle kernel paging request". The problem is the skb_clone() in
__netlink_deliver_tap_skb() for skbs that are mmaped.
I.e. the cloned skb doesn't have a destructor, whereas the mmap netlink
skb has it pointed to netlink_skb_destructor(), set in the handler
netlink_ring_setup_skb(). There, skb->head is being set to NULL, so
that in such cases, __kfree_skb() doesn't perform a skb_release_data()
via skb_release_all(), where skb->head is possibly being freed through
kfree(head) into slab allocator, although netlink mmap skb->head points
to the mmap buffer. Similarly, the same has to be done also for large
netlink skbs where the data area is vmalloced. Therefore, as discussed,
make a copy for these rather rare cases for now. This fixes the issue
on my and Ken-ichirou's test-cases.
Reference: http://thread.gmane.org/gmane.linux.network/371129
Fixes: bcbde0d449 ("net: netlink: virtual tap device management")
Reported-by: Ken-ichirou MATSUZAWA <chamaken@gmail.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Tested-by: Ken-ichirou MATSUZAWA <chamaken@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Conflicts:
arch/arm/boot/dts/imx6sx-sdb.dts
net/sched/cls_bpf.c
Two simple sets of overlapping changes.
Signed-off-by: David S. Miller <davem@davemloft.net>
In addition to the problem Jeff Layton reported, I looked at the code
and reproduced the same warning by subscribing and removing the genl
family with a socket still open. This is a fairly tricky race which
originates in the fact that generic netlink allows the family to go
away while sockets are still open - unlike regular netlink which has
a module refcount for every open socket so in general this cannot be
triggered.
Trying to resolve this issue by the obvious locking isn't possible as
it will result in deadlocks between unregistration and group unbind
notification (which incidentally lockdep doesn't find due to the home
grown locking in the netlink table.)
To really resolve this, introduce a "closing socket" reference counter
(for generic netlink only, as it's the only affected family) in the
core netlink code and use that in generic netlink to wait for all the
sockets that are being closed at the same time as a generic netlink
family is removed.
This fixes the race that when a socket is closed, it will should call
the unbind, but if the family is removed at the same time the unbind
will not find it, leading to the warning. The real problem though is
that in this case the unbind could actually find a new family that is
registered to have a multicast group with the same ID, and call its
mcast_unbind() leading to confusing.
Also remove the warning since it would still trigger, but is now no
longer a problem.
This also moves the code in af_netlink.c to before unreferencing the
module to avoid having the same problem in the normal non-genl case.
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
As rhashtable_lookup_compare_insert() can guarantee the process
of search and insertion is atomic, it's safe to eliminate the
nl_sk_hash_lock. After this, object insertion or removal will
be protected with per bucket lock on write side while object
lookup is guarded with rcu read lock on read side.
Signed-off-by: Ying Xue <ying.xue@windriver.com>
Cc: Thomas Graf <tgraf@suug.ch>
Acked-by: Thomas Graf <tgraf@suug.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>
Defers the release of the socket reference using call_rcu() to
allow using an RCU read-side protected call to rhashtable_lookup()
This restores behaviour and performance gains as previously
introduced by e341694 ("netlink: Convert netlink_lookup() to use
RCU protected hash table") without the side effect of severely
delayed socket destruction.
Signed-off-by: Thomas Graf <tgraf@suug.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>
Netlink families can exist in multiple namespaces, and for the most
part multicast subscriptions are per network namespace. Thus it only
makes sense to have bind/unbind notifications per network namespace.
To achieve this, pass the network namespace of a given client socket
to the bind/unbind functions.
Also do this in generic netlink, and there also make sure that any
bind for multicast groups that only exist in init_net is rejected.
This isn't really a problem if it is accepted since a client in a
different namespace will never receive any notifications from such
a group, but it can confuse the family if not rejected (it's also
possible to silently (without telling the family) accept it, but it
would also have to be ignored on unbind so families that take any
kind of action on bind/unbind won't do unnecessary work for invalid
clients like that.
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Although RCU protection would be possible during diag dump, doing
so allows for concurrent table mutations which can render the
in-table offset between individual Netlink messages invalid and
thus cause legitimate sockets to be skipped in the dump.
Since the diag dump is relatively low volume and consistency is
more important than performance, the table mutex is held during
dump.
Reported-by: Andrey Wagin <avagin@gmail.com>
Signed-off-by: Thomas Graf <tgraf@suug.ch>
Fixes: e341694e3e ("netlink: Convert netlink_lookup() to use RCU protected hash table")
Signed-off-by: David S. Miller <davem@davemloft.net>
Heavy Netlink users such as Open vSwitch spend a considerable amount of
time in netlink_lookup() due to the read-lock on nl_table_lock. Use of
RCU relieves the lock contention.
Makes use of the new resizable hash table to avoid locking on the
lookup.
The hash table will grow if entries exceeds 75% of table size up to a
total table size of 64K. It will automatically shrink if usage falls
below 30%.
Also splits nl_table_lock into a separate mutex to protect hash table
mutations and allow synchronize_rcu() to sleep while waiting for readers
during expansion and shrinking.
Before:
9.16% kpktgend_0 [openvswitch] [k] masked_flow_lookup
6.42% kpktgend_0 [pktgen] [k] mod_cur_headers
6.26% kpktgend_0 [pktgen] [k] pktgen_thread_worker
6.23% kpktgend_0 [kernel.kallsyms] [k] memset
4.79% kpktgend_0 [kernel.kallsyms] [k] netlink_lookup
4.37% kpktgend_0 [kernel.kallsyms] [k] memcpy
3.60% kpktgend_0 [openvswitch] [k] ovs_flow_extract
2.69% kpktgend_0 [kernel.kallsyms] [k] jhash2
After:
15.26% kpktgend_0 [openvswitch] [k] masked_flow_lookup
8.12% kpktgend_0 [pktgen] [k] pktgen_thread_worker
7.92% kpktgend_0 [pktgen] [k] mod_cur_headers
5.11% kpktgend_0 [kernel.kallsyms] [k] memset
4.11% kpktgend_0 [openvswitch] [k] ovs_flow_extract
4.06% kpktgend_0 [kernel.kallsyms] [k] _raw_spin_lock
3.90% kpktgend_0 [kernel.kallsyms] [k] jhash2
[...]
0.67% kpktgend_0 [kernel.kallsyms] [k] netlink_lookup
Signed-off-by: Thomas Graf <tgraf@suug.ch>
Reviewed-by: Nikolay Aleksandrov <nikolay@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Have the netlink per-protocol optional bind function return an int error code
rather than void to signal a failure.
This will enable netlink protocols to perform extra checks including
capabilities and permissions verifications when updating memberships in
multicast groups.
In netlink_bind() and netlink_setsockopt() the call to the per-protocol bind
function was moved above the multicast group update to prevent any access to
the multicast socket groups before checking with the per-protocol bind
function. This will enable the per-protocol bind function to be used to check
permissions which could be denied before making them available, and to avoid
the messy job of undoing the addition should the per-protocol bind function
fail.
The netfilter subsystem seems to be the only one currently using the
per-protocol bind function.
Signed-off-by: Richard Guy Briggs <rgb@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
One known problem with netlink is the fact that NLMSG_GOODSIZE is
really small on PAGE_SIZE==4096 architectures, and it is difficult
to know in advance what buffer size is used by the application.
This patch adds an automatic learning of the size.
First netlink message will still be limited to ~4K, but if user used
bigger buffers, then following messages will be able to use up to 16KB.
This speedups dump() operations by a large factor and should be safe
for legacy applications.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Thomas Graf <tgraf@suug.ch>
Acked-by: Thomas Graf <tgraf@suug.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>
Following patch stores struct netlink_callback in netlink_sock
to avoid allocating and freeing it on every netlink dump msg.
Only one dump operation is allowed for a given socket at a time
therefore we can safely convert cb pointer to cb struct inside
netlink_sock.
Signed-off-by: Pravin B Shelar <pshelar@nicira.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
As we know, netlink sockets are private resource of
net namespace, they can communicate with each other
only when they in the same net namespace. this works
well until we try to add namespace support for other
subsystems which use netlink.
Don't like ipv4 and route table.., it is not suited to
make these subsytems belong to net namespace, Such as
audit and crypto subsystems,they are more suitable to
user namespace.
So we must have the ability to make the netlink sockets
in same user namespace can communicate with each other.
This patch adds a new function pointer "compare" for
netlink_table, we can decide if the netlink sockets can
communicate with each other through this netlink_table
self-defined compare function.
The behavior isn't changed if we don't provide the compare
function for netlink_table.
Signed-off-by: Gao feng <gaofeng@cn.fujitsu.com>
Acked-by: Serge E. Hallyn <serge.hallyn@ubuntu.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add support for mmap'ed RX and TX ring setup and teardown based on the
af_packet.c code. The following patches will use this to add the real
mmap'ed receive and transmit functionality.
Signed-off-by: Patrick McHardy <kaber@trash.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
Move a few declarations in a header.
Acked-by: Pavel Emelyanov <xemul@parallels.com>
Cc: "David S. Miller" <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Pablo Neira Ayuso <pablo@netfilter.org>
Cc: "Eric W. Biederman" <ebiederm@xmission.com>
Cc: Gao feng <gaofeng@cn.fujitsu.com>
Cc: Thomas Graf <tgraf@suug.ch>
Signed-off-by: Andrey Vagin <avagin@openvz.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Florian Westphal
Herbert Xu
Daniel Borkmann
David S. Miller
Johannes Berg
Ying Xue
Thomas Graf
Richard Guy Briggs
Eric Dumazet
Pravin B Shelar
Gao feng
Patrick McHardy
Andrey Vagin