We force notification when the ring is full, even if the host has
indicated it doesn't want to know. This seemed like a good idea at
the time: if we fill the transmit ring, we should tell the host
immediately.
Unfortunately this logic also applies to the receiving ring, which is
refilled constantly. We should introduce real notification thesholds
to replace this logic. Meanwhile, removing the logic altogether breaks
the heuristics which KVM uses, so we use a hack: only notify if there are
outgoing parts of the new buffer.
Here are the number of exits with lguest's crappy network implementation:
Before:
network xmit 7859051 recv 236420
After:
network xmit 7858610 recv 118136
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
virtio allows drivers to suppress callbacks (ie. interrupts) for
efficiency (no locking, it's just an optimization).
There's a similar mechanism for the host to suppress notifications
coming from the guest: in that case, we ignore the suppression if the
ring is completely full.
It turns out that life is simpler if the host similarly ignores
callback suppression when the ring is completely empty: the network
driver wants to free up old packets in a timely manner, and otherwise
has to use a timer to poll.
We have to remove the code which ignores interrupts when the driver
has disabled them (again, it had no locking and hence was unreliable
anyway).
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Hello Rusty,
seems that we still have a problem with virtio_net and the enable_cb callback.
During a long running network stress tests with virtio and got the following
oops:
------------[ cut here ]------------
kernel BUG at drivers/virtio/virtio_ring.c:230!
illegal operation: 0001 [#1] SMP
Modules linked in:
CPU: 0 Not tainted 2.6.26-rc2-kvm-00436-gc94c08b-dirty #34
Process netserver (pid: 2582, task: 000000000fbc4c68, ksp: 000000000f42b990)
Krnl PSW : 0704c00180000000 00000000002d0ec8 (vring_enable_cb+0x1c/0x60)
R:0 T:1 IO:1 EX:1 Key:0 M:1 W:0 P:0 AS:3 CC:0 PM:0 EA:3
Krnl GPRS: 0000000000000000 0000000000000000 000000000ef3d000 0000000010009800
0000000000000000 0000000000419ce0 0000000000000080 000000000000007b
000000000adb5538 000000000ef40900 000000000ef40000 000000000ef40920
0000000000000000 0000000000000005 000000000029c1b0 000000000fea7d18
Krnl Code: 00000000002d0ebc: a7110001 tmll %r1,1
00000000002d0ec0: a7740004 brc 7,2d0ec8
00000000002d0ec4: a7f40001 brc 15,2d0ec6
>00000000002d0ec8: a517fffe nill %r1,65534
00000000002d0ecc: 40103000 sth %r1,0(%r3)
00000000002d0ed0: 07f0 bcr 15,%r0
00000000002d0ed2: e31020380004 lg %r1,56(%r2)
00000000002d0ed8: a7480000 lhi %r4,0
Call Trace:
([<000000000029c0fc>] virtnet_poll+0x290/0x3b8)
[<0000000000333fb8>] net_rx_action+0x9c/0x1b8
[<00000000001394bc>] __do_softirq+0x74/0x108
[<000000000010d16a>] do_softirq+0x92/0xac
[<0000000000139826>] irq_exit+0x72/0xc8
[<000000000010a7b6>] do_extint+0xe2/0x104
[<0000000000110508>] ext_no_vtime+0x16/0x1a
Last Breaking-Event-Address:
[<00000000002d0ec4>] vring_enable_cb+0x18/0x60
I looked into the virtio_net code for some time and I think the following
scenario happened. Please look at virtnet_poll:
[...]
/* Out of packets? */
if (received < budget) {
netif_rx_complete(vi->dev, napi);
if (unlikely(!vi->rvq->vq_ops->enable_cb(vi->rvq))
&& napi_schedule_prep(napi)) {
vi->rvq->vq_ops->disable_cb(vi->rvq);
__netif_rx_schedule(vi->dev, napi);
goto again;
}
}
If an interrupt arrives after netif_rx_complete, a second poll routine can run
on a different cpu. The second check for napi_schedule_prep would prevent any
harm in the network stack, but we have called enable_cb possibly after the
disable_cb in skb_recv_done.
static void skb_recv_done(struct virtqueue *rvq)
{
struct virtnet_info *vi = rvq->vdev->priv;
/* Schedule NAPI, Suppress further interrupts if successful. */
if (netif_rx_schedule_prep(vi->dev, &vi->napi)) {
rvq->vq_ops->disable_cb(rvq);
__netif_rx_schedule(vi->dev, &vi->napi);
}
}
That means that the second poll routine runs with interrupts enabled, which is
ok, since we can handle additional interrupts. The problem is now that the
second poll routine might also call enable_cb, triggering the BUG.
The only solution I can come up with, is to remove the BUG statement in
enable_cb - similar to disable_cb. Opinions or better ideas where the oops
could come from?
Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
A corrupt virtqueue (caused by the other end screwing up) can have
strange results such as a driver spinning: just bail when we try to
get a buffer from a known-broken queue.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
The 'disable_cb' callback is designed as an optimization to tell the host
we don't need callbacks now. As it is not reliable, the debug check is
overzealous: it can happen on two CPUs at the same time. Document this.
Even if it were reliable, the virtio_net driver doesn't disable
callbacks on transmit so the START_USE/END_USE debugging reentrance
protection can be easily tripped even on UP.
Thanks to Balaji Rao for the bug report and testing.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
CC: Balaji Rao <balajirrao@gmail.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
There is a race in virtio_net, dealing with disabling/enabling the callback.
I saw the following oops:
kernel BUG at /space/kvm/drivers/virtio/virtio_ring.c:218!
illegal operation: 0001 [#1] SMP
Modules linked in: sunrpc dm_mod
CPU: 2 Not tainted 2.6.25-rc1zlive-host-10623-gd358142-dirty #99
Process swapper (pid: 0, task: 000000000f85a610, ksp: 000000000f873c60)
Krnl PSW : 0404300180000000 00000000002b81a6 (vring_disable_cb+0x16/0x20)
R:0 T:1 IO:0 EX:0 Key:0 M:1 W:0 P:0 AS:0 CC:3 PM:0 EA:3
Krnl GPRS: 0000000000000001 0000000000000001 0000000010005800 0000000000000001
000000000f3a0900 000000000f85a610 0000000000000000 0000000000000000
0000000000000000 000000000f870000 0000000000000000 0000000000001237
000000000f3a0920 000000000010ff74 00000000002846f6 000000000fa0bcd8
Krnl Code: 00000000002b819a: a7110001 tmll %r1,1
00000000002b819e: a7840004 brc 8,2b81a6
00000000002b81a2: a7f40001 brc 15,2b81a4
>00000000002b81a6: a51b0001 oill %r1,1
00000000002b81aa: 40102000 sth %r1,0(%r2)
00000000002b81ae: 07fe bcr 15,%r14
00000000002b81b0: eb7ff0380024 stmg %r7,%r15,56(%r15)
00000000002b81b6: a7f13e00 tmll %r15,15872
Call Trace:
([<000000000fa0bcd0>] 0xfa0bcd0)
[<00000000002b8350>] vring_interrupt+0x5c/0x6c
[<000000000010ab08>] do_extint+0xb8/0xf0
[<0000000000110716>] ext_no_vtime+0x16/0x1a
[<0000000000107e72>] cpu_idle+0x1c2/0x1e0
The problem can be triggered with a high amount of host->guest traffic.
I think its the following race:
poll says netif_rx_complete
poll calls enable_cb
enable_cb opens the interrupt mask
a new packet comes, an interrupt is triggered----\
enable_cb sees that there is more work |
enable_cb disables the interrupt |
. V
. interrupt is delivered
. skb_recv_done does atomic napi test, ok
some waiting disable_cb is called->check fails->bang!
.
poll would do napi check
poll would do disable_cb
The fix is to let enable_cb not disable the interrupt again, but expect the
caller to do the cleanup if it returns false. In that case, the interrupt is
only disabled, if the napi test_set_bit was successful.
Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au> (cleaned up doco)
This is needed for the virtio PCI device to be compiled as a module.
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Anthony Liguori found double interrupt suppression in the virtio_net
driver, triggered by two skb_recv_done's in a row. This is because
virtio_ring's interrupt suppression is a best-effort optimization: it
contains no synchronization so the host can miss it and still send
interrupts.
But it's certainly nicer for virtio users if calling disable_cb
actually disables callbacks, so we check for the race in the interrupt
routine.
Note: SMP guests might require syncronization here, but since
disable_cb is actually called from interrupt context, there has to be
some form of synchronization before the next same interrupt handler is
called (Linux guarantees that the same device's irq handler will never
run simultanously on multiple CPUs).
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
A reset function solves three problems:
1) It allows us to renegotiate features, eg. if we want to upgrade a
guest driver without rebooting the guest.
2) It gives us a clean way of shutting down virtqueues: after a reset,
we know that the buffers won't be used by the host, and
3) It helps the guest recover from messed-up drivers.
So we remove the ->shutdown hook, and the only way we now remove
feature bits is via reset.
We leave it to the driver to do the reset before it deletes queues:
the balloon driver, for example, needs to chat to the host in its
remove function.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
The other side (host) can set the NO_NOTIFY flag as an optimization,
to say "no need to kick me when you add things". Make it clear that
this is advisory only; especially that we should always notify when
the ring is full.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
It seems that virtio_net wants to disable callbacks (interrupts) before
calling netif_rx_schedule(), so we can't use the return value to do so.
Rename "restart" to "cb_enable" and introduce "cb_disable" hook: callback
now returns void, rather than a boolean.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
The virtio descriptor rings of size N-1 were nicely set up to be
aligned to an N-byte boundary. But as Anthony Liguori points out, the
free-running indices used by virtio require that the sizes be a power
of 2, otherwise we get problems on wrap (demonstrated with lguest).
So we replace the clever "2^n-1" scheme with a simple "align to page
boundary" scheme: this means that all virtio rings take at least two
pages, but it's safer than guessing cache alignment.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
The more_used() function compares the vq->vring.used->idx with last_used_idx.
Since vq->vring.used->idx is a 16-bit integer, and last_used_idx is an
unsigned int, this results in unpredictable behavior when vq->vring.used->idx
wraps around.
This patch corrects this by changing last_used_idx to the correct type.
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
These helper routines supply most of the virtqueue_ops for hypervisors
which want to use a ring for virtio. Unlike the previous lguest
implementation:
1) The rings are variable sized (2^n-1 elements).
2) They have an unfortunate limit of 65535 bytes per sg element.
3) The page numbers are always 64 bit (PAE anyone?)
4) They no longer place used[] on a separate page, just a separate
cacheline.
5) We do a modulo on a variable. We could be tricky if we cared.
6) Interrupts and notifies are suppressed using flags within the rings.
Users need only get the ring pages and provide a notify hook (KVM
wants the guest to allocate the rings, lguest does it sanely).
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Cc: Dor Laor <dor.laor@qumranet.com>
