virtio net: fix oops on interface-up
authorChristian Borntraeger <borntraeger@de.ibm.com>
Wed, 6 Feb 2008 07:50:11 +0000 (08:50 +0100)
committerJeff Garzik <jeff@garzik.org>
Wed, 6 Feb 2008 11:42:30 +0000 (06:42 -0500)
I got the following oops during interface ifup. Unfortunately its not
easily reproducable so I cant say for sure that my fix fixes this
problem, but I am confident and I think its correct anyway:

   <2>kernel BUG at /space/kvm/drivers/virtio/virtio_ring.c:234!
    <4>illegal operation: 0001 [#1] PREEMPT SMP
    <4>Modules linked in:
    <4>CPU: 0 Not tainted 2.6.24zlive-guest-07293-gf1ca151-dirty #91
    <4>Process swapper (pid: 0, task: 0000000000800938, ksp: 000000000084ddb8)
    <4>Krnl PSW : 0404300180000000 0000000000466374 (vring_disable_cb+0x30/0x34)
    <4>           R:0 T:1 IO:0 EX:0 Key:0 M:1 W:0 P:0 AS:0 CC:3 PM:0 EA:3
    <4>Krnl GPRS: 0000000000000001 0000000000000001 0000000010003800 0000000000466344
    <4>           000000000e980900 00000000008848b0 000000000084e748 0000000000000000
    <4>           000000000087b300 0000000000001237 0000000000001237 000000000f85bdd8
    <4>           000000000e980920 00000000001137c0 0000000000464754 000000000f85bdd8
    <4>Krnl Code: 0000000000466368e3b0b0700004        lg      %r11,112(%r11)
    <4>           000000000046636e: 07fe                bcr     15,%r14
    <4>           0000000000466370a7f40001            brc     15,466372
    <4>          >0000000000466374a7f4fff6            brc     15,466360
    <4>           0000000000466378eb7ff0500024        stmg    %r7,%r15,80(%r15)
    <4>           000000000046637ea7f13e00            tmll    %r15,15872
    <4>           0000000000466382b90400ef            lgr     %r14,%r15
    <4>           0000000000466386a7840001            brc     8,466388
    <4>Call Trace:
    <4>([<000201500f85c000>] 0x201500f85c000)
    <4> [<0000000000466556>] vring_interrupt+0x72/0x88
    <4> [<00000000004801a0>] kvm_extint_handler+0x34/0x44
    <4> [<000000000010d22c>] do_extint+0xbc/0xf8
    <4> [<0000000000113f98>] ext_no_vtime+0x16/0x1a
    <4> [<000000000010a182>] cpu_idle+0x216/0x238
    <4>([<000000000010a162>] cpu_idle+0x1f6/0x238)
    <4> [<0000000000568656>] rest_init+0xaa/0xb8
    <4> [<000000000084ee2c>] start_kernel+0x3fc/0x490
    <4> [<0000000000100020>] _stext+0x20/0x80
    <4>
    <4> <0>Kernel panic - not syncing: Fatal exception in interrupt
    <4>

After looking at the code and the dump I think the following scenario
happened: Ifup was running on cpu2 and the interrupt arrived on cpu0.
Now virtnet_open on cpu 2 managed to execute napi_enable and disable_cb
but did not execute rx_schedule. Meanwhile on cpu 0 skb_recv_done was
called by vring_interrupt, executed netif_rx_schedule_prep, which
succeeded and therefore called disable_cb. This triggered the BUG_ON,
as interrupts were already disabled by cpu 2.

I think the proper solution is to make the call to disable_cb depend on
the atomic update of NAPI_STATE_SCHED by using netif_rx_schedule_prep
in the same way as skb_recv_done.

Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
Acked-by: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
drivers/net/virtio_net.c

index e66de0c12fc10e1455e4137f8f68271ca44505b5..fdc23678117bdd31157a13c6d428d2c1951dc778 100644 (file)
@@ -302,10 +302,12 @@ static int virtnet_open(struct net_device *dev)
 
        /* If all buffers were filled by other side before we napi_enabled, we
         * won't get another interrupt, so process any outstanding packets
-        * now.  virtnet_poll wants re-enable the queue, so we disable here. */
-       vi->rvq->vq_ops->disable_cb(vi->rvq);
-       netif_rx_schedule(vi->dev, &vi->napi);
-
+        * now.  virtnet_poll wants re-enable the queue, so we disable here.
+        * We synchronize against interrupts via NAPI_STATE_SCHED */
+       if (netif_rx_schedule_prep(dev, &vi->napi)) {
+               vi->rvq->vq_ops->disable_cb(vi->rvq);
+               __netif_rx_schedule(dev, &vi->napi);
+       }
        return 0;
 }