Edivaldo reports a problem that the array of NetClientState in NICState is too
large - MAX_QUEUE_NUM(1024) which will wastes memory even if multiqueue is not
used.
Instead of static arrays, solving this issue by allocating the queues on demand
for both the NetClientState array in NICState and VirtIONetQueue array in
VirtIONet.
Tested by myself, with single virtio-net-pci device. The memory allocation is
almost the same as when multiqueue is not merged.
Cc: Edivaldo de Araujo Pereira <edivaldoapereira@yahoo.com.br>
Cc: qemu-stable@nongnu.org
Signed-off-by: Jason Wang <jasowang@redhat.com>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
historically the kernel queues packets two times. once
at the device and second in qdisc. this is believed to cause
interface stalls if one of these queues overruns.
setting IFF_ONE_QUEUE is the default in kernels >= 3.8. the
flag is ignored since then. see kernel commit
5d097109257c03a71845729f8db6b5770c4bbedc
Signed-off-by: Peter Lieven <pl@kamp.de>
Acked-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Obviously, hub does not support multiqueue tap. So this patch forbids creating
multiple queue tap when hub is used to prevent the crash when command line such
as "-net tap,queues=2" is used.
Cc: qemu-stable@nongnu.org
Signed-off-by: Jason Wang <jasowang@redhat.com>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
In the current implementation of qemu, running without a network
backend will cause the queue to grow unbounded when the guest is
transmitting traffic.
This patch fixes the problem by implementing bounded size NetQueue,
used with an arbitrary limit of 10000 packets, and dropping packets
when the queue is full _and_ the sender does not pass a callback.
The second condition makes sure that we never drop packets that
contains a callback (which would be tricky, because the producer
expects the callback to be run when all previous packets have been
consumed; so we cannot run it when the packet is dropped).
If documentation is correct, producers that submit a callback should
stop sending when their packet is queued, so there is no real risk
that the queue exceeds the max size by large values.
Signed-off-by: Luigi Rizzo <rizzo@iet.unipi.it>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
When frontend and backend are connected through a hub as below
(showing only one direction), and the frontend (or in general, all
output ports of the hub) cannot accept more traffic, the backend
queues packets in queue-A.
When the frontend (or in general, one output port) becomes ready again,
quemu tries to flush packets from queue-B, which is unfortunately empty.
e1000.0 <--[queue B]-- hub0port0(hub)hub0port1 <--[queue A]-- tap.0
To fix this i propose to introduce a new function net_hub_flush()
which is called when trying to flush a queue connected to a hub.
Signed-off-by: Luigi Rizzo <rizzo@iet.unipi.it>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
The pSeries machine and some other devices don't supply a cleanup
callback. Revert part of 1ceef9f273 that
started calling it unconditionally.
Cc: Jason Wang <jasowang@redhat.com>
Signed-off-by: Andreas Färber <afaerber@suse.de>
Message-id: 1360707366-9271-1-git-send-email-afaerber@suse.de
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
1ceef9f273 added handling for cleaning
up multiple queues in qemu_del_nic() for cases where multiqueue is in
use. To determine the number of queues it looks at nic->conf->queues,
then iterates through all the queues to cleanup the associated
NetClientStates. If no queues are found, no NetClientStates are deleted.
However, nic->conf->queues is only set when a peer is created via
-netdev or netdev_add, and is otherwise 0. This causes us to spin in
net_cleanup() if we attempt to shut down qemu before adding a host
device.
Since qemu_new_nic() unconditionally creates at least 1
queue/NetClientState at queue idx 0, make qemu_del_nic() always attempt
to clean it up.
Signed-off-by: Michael Roth <mdroth@linux.vnet.ibm.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
The multiqueue patch series broke -netdev tap,fd=X which manifests
as libvirt not being able to start a guest. This was because it
passed NULL for the netdev name which results in an anonymous netdev
device regardless of what the user specified.
Cc: Jason Wang <jasowang@redhat.com>
Cc: Bruce Rogers <brogers@suse.com>
Reported-by: Bruce Rogers <brogers@suse.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
Recently, linux support multiqueue tap which could let userspace call TUNSETIFF
for a signle device many times to create multiple file descriptors as
independent queues. User could also enable/disabe a specific queue through
TUNSETQUEUE.
The patch adds the generic infrastructure to create multiqueue taps. To achieve
this a new parameter "queues" were introduced to specify how many queues were
expected to be created for tap by qemu itself. Alternatively, management could
also pass multiple pre-created tap file descriptors separated with ':' through a
new parameter fds like -netdev tap,id=hn0,fds="X:Y:..:Z". Multiple vhost file
descriptors could also be passed in this way.
Each TAPState were still associated to a tap fd, which mean multiple TAPStates
were created when user needs multiqueue taps. Since each TAPState contains one
NetClientState, with the multiqueue nic support, an N peers of NetClientState
were built up.
A new parameter, mq_required were introduce in tap_open() to create multiqueue
tap fds.
Signed-off-by: Jason Wang <jasowang@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
This patch introduces a helper tap_get_ifname() to get the device name of tap
device. This is needed when ifname is unspecified in the command line and qemu
were asked to create tap device by itself. In this situation, the name were
allocated by kernel, so if multiqueue is asked, we need to fetch its name after
creating the first queue.
Only linux has this support since it's the only platform that supports
multiqueue tap.
Signed-off-by: Jason Wang <jasowang@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
This patch introduce a new bit - enabled in TAPState which tracks whether a
specific queue/fd is enabled. The tap/fd is enabled during initialization and
could be enabled/disabled by tap_enalbe() and tap_disable() which calls platform
specific helpers to do the real work. Polling of a tap fd can only done when
the tap was enabled.
Signed-off-by: Jason Wang <jasowang@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
This patch add basic multiqueue support for Linux. When multiqueue is needed, we
will first check whether kernel support multiqueue tap before creating more
queues. Two new functions tap_fd_enable() and tap_fd_disable() were introduced
to enable and disable a specific queue. Since the multiqueue is only supported
in Linux, return error on other platforms.
Signed-off-by: Jason Wang <jasowang@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
This patch factors out the common initialization of tap into a new helper
net_init_tap_one(). This will be used by multiqueue tap patches.
Signed-off-by: Jason Wang <jasowang@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
Import multiqueue constants from if_tun.h from 3.8-rc3. A new ifr flag
IFF_MULTI_QUEUE were introduced to create a multiqueue backend by calling
TUNSETIFF with the this flag and with the same interface name many times.
A new ioctl TUNSETQUEUE were introduced. When doing this ioctl with
IFF_DETACH_QUEUE, the queue were disabled in the linux kernel. When doing this
ioctl with IFF_ATTACH_QUEUE, the queue were enabled in the linux kernel.
Signed-off-by: Jason Wang <jasowang@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
This patch adds basic multiqueue support for qemu. The idea is simple, an array
of NetClientStates were introduced in NICState, parse_netdev() were extended to
find and match all NetClientStates belongs to the backend and place their
pointers in NICConf. Then qemu_new_nic can setup a N:N mapping between NICStates
that belongs to a nic and NICStates belongs to the netdev. And a queue_index
were introduced in NetClientState to track its index. After this, each peers of
a NICState were abstracted as a queue.
After this change, all NetClientState that belongs to the same backend/nic has
the same id. When use want to change the link status, all NetClientStates that
belongs to the same backend/nic will be also changed. When user want to delete
a device or netdev, all NetClientStates that belongs to the same backend/nic
will be deleted also. Changing or deleting an specific queue is not allowed.
Signed-off-by: Jason Wang <jasowang@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
To allow allocating an array of NetClientState and free it once, this patch
introduces destructor of NetClientState. Which could do type specific free,
which could be used by multiqueue to free the array once.
Signed-off-by: Jason Wang <jasowang@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
This patch separates the setup of NetClientState from its allocation, this will
allow allocating an arrays of NetClientState and does the initialization one by
one which is what multiqueue needs.
Signed-off-by: Jason Wang <jasowang@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
In multiqueue, all NetClientState that belongs to the same netdev or nic has the
same id. So this patches introduces an helper qemu_find_net_clients_except()
which finds all NetClientState with the same id. This will be used by multiqueue
networking.
Signed-off-by: Jason Wang <jasowang@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
To support multiqueue nic, this patch separate the nic destructor from
qemu_del_net_client() to a new helper qemu_del_nic() since the mapping bettween
NiCState and NetClientState were not 1:1 in multiqueue. The following patches
would refactor this function to support multiqueue nic.
Signed-off-by: Jason Wang <jasowang@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
To support multiqueue, this patch introduces a helper qemu_get_nic() to get
NICState from a NetClientState. The following patches would refactor this helper
to support multiqueue.
Signed-off-by: Jason Wang <jasowang@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
To support multiqueue, the patch introduce a helper qemu_get_queue()
which is used to get the NetClientState of a device. The following patches would
refactor this helper to support multiqueue.
Signed-off-by: Jason Wang <jasowang@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
This patch change all info call back function to take
additional QDict * parameter, which allow those command
take parameter. Now it is set to NULL at default case.
Signed-off-by: Wenchao Xia <xiawenc@linux.vnet.ibm.com>
Signed-off-by: Luiz Capitulino <lcapitulino@redhat.com>
To fix building error:
CC net/vde.o
net/vde.c: In function ‘vde_cleanup’:
net/vde.c:65:5: error: implicit declaration of function ‘qemu_set_fd_handler’ [-Werror=implicit-function-declaration]
net/vde.c:65:5: error: nested extern declaration of ‘qemu_set_fd_handler’ [-Werror=nested-externs]
cc1: all warnings being treated as errors
Signed-off-by: Liming Wang <walimisdev@gmail.com>
Signed-off-by: Blue Swirl <blauwirbel@gmail.com>
These and some more compiler warnings were caused by a recent commit:
net/tap-win32.c:724: warning: no previous prototype for ‘tap_has_ufo’
net/tap-win32.c:729: warning: no previous prototype for ‘tap_has_vnet_hdr’
...
Signed-off-by: Stefan Weil <sw@weilnetz.de>
Signed-off-by: Blue Swirl <blauwirbel@gmail.com>
* bonzini/header-dirs: (45 commits)
janitor: move remaining public headers to include/
hw: move executable format header files to hw/
fpu: move public header file to include/fpu
softmmu: move remaining include files to include/ subdirectories
softmmu: move include files to include/sysemu/
misc: move include files to include/qemu/
qom: move include files to include/qom/
migration: move include files to include/migration/
monitor: move include files to include/monitor/
exec: move include files to include/exec/
block: move include files to include/block/
qapi: move include files to include/qobject/
janitor: add guards to headers
qapi: make struct Visitor opaque
qapi: remove qapi/qapi-types-core.h
qapi: move inclusions of qemu-common.h from headers to .c files
ui: move files to ui/ and include/ui/
qemu-ga: move qemu-ga files to qga/
net: reorganize headers
net: move net.c to net/
...
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
Move public headers to include/net, and leave private headers in net/.
Put the virtio headers in include/net/tap.h, removing the multiple copies
that existed. Leave include/net/tap.h as the interface for NICs, and
net/tap_int.h as the interface for OS-specific parts of the tap backend.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Touching char/char.h basically causes the whole of QEMU to
be rebuilt. Avoid this, it is usually unnecessary.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Remove some redundant blanks in the comments of
net_hub_id_for_client().
Signed-off-by: Zhi Yong Wu <wuzhy@linux.vnet.ibm.com>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
For tap, we currently assume the vnet header size is 10
(the default value) but that might not be the case
if tap is persistent and has been used by qemu previously.
To fix, set vnet header size correctly on open.
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
For tap, we currently assume the vnet header size is 10
(the default value) but that might not be the case
if tap is persistent and has been used by qemu previously.
To fix, set host header size in tap device on open.
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Tested-by: Alexander Graf <agraf@suse.de>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
This patch will allow the user to include the domain-search option in
replies from the built-in DHCP server. The domain suffixes can be
specified by adding dnssearch= entries to the "-net user" parameter.
[Jan: tiny style adjustments]
Signed-off-by: Klaus Stengel <Klaus.Stengel@asamnet.de>
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>
Fix the problem that can not delete the udp socket.
It's caused by passing "udp" model to net_socket_udp_init,
but we do not have "udp" model in our model list.
Pass the right model "socket" to init function.
https://bugs.launchpad.net/qemu/+bug/1073585?comments=all
Signed-off-by: Lei Li <lilei@linux.vnet.ibm.com>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Add missing stubs to win32 to fix link failure.
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Reported-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Blue Swirl <blauwirbel@gmail.com>
The include file for net_init_tap was missing:
net/tap-win32.c:703:
warning: no previous prototype for ‘net_init_tap’
Signed-off-by: Stefan Weil <sw@weilnetz.de>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
This patch doesn't seem much useful alone, I must admit. However,
it makes sense as part of the upcoming directory reorganization,
where I want to have include/net/tap.h as the net<->hw interface
for tap. Then having both net/tap.h and include/net/tap.h does
not work. "Fixed" by moving all the init functions to a single
header file net/clients.h.
The patch also adopts a uniform style for including net/*.h files
from net/*.c, without the net/ path.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Stefan Hajnoczi <stefanha@gmail.com>
Commit 213fd5087e removed a type cast
which is needed for MinGW:
net/socket.c:136: warning:
pointer targets in passing argument 2 of ‘sendto’ differ in signedness
/usr/lib/gcc/amd64-mingw32msvc/4.4.4/../../../../amd64-mingw32msvc/include/winsock2.h:1313: note:
expected ‘const char *’ but argument is of type ‘const uint8_t *’
Add a 'qemu_sendto' macro which provides that type cast where needed
and use the new macro instead of 'sendto'.
Signed-off-by: Stefan Weil <sw@weilnetz.de>
Signed-off-by: Stefan Hajnoczi <stefanha@gmail.com>
Replace spinning send_all() with a proper non-blocking send. When the
socket write buffer limit is reached, we should stop trying to send and
wait for the socket to become writable again.
Non-blocking TCP sockets can return in two different ways when the write
buffer limit is reached:
1. ret = -1 and errno = EAGAIN/EWOULDBLOCK. No data has been written.
2. ret < total_size. Short write, only part of the message was
transmitted.
Handle both cases and keep track of how many bytes have been written in
s->send_index. (This includes the 'length' header before the actual
payload buffer.)
Signed-off-by: Stefan Hajnoczi <stefanha@linux.vnet.ibm.com>
Implement asynchronous send for UDP (or other SOCK_DGRAM) sockets. If
send fails with EAGAIN we wait for the socket to become writable again.
Signed-off-by: Stefan Hajnoczi <stefanha@linux.vnet.ibm.com>
The net/socket.c net client is not truly asynchronous. This patch
borrows the qemu_set_fd_handler2() code from net/tap.c as the basis for
proper asynchronous send/receive.
Only read packets from the socket when the peer is able to receive.
This avoids needless queuing.
Later patches implement asynchronous send.
Signed-off-by: Stefan Hajnoczi <stefanha@linux.vnet.ibm.com>
