On Sun, Dec 22, 2019 at 9:57 AM Michael S. Tsirkin <mst@xxxxxxxxxx> wrote: > > On Sun, Dec 22, 2019 at 09:21:43AM -0500, Willem de Bruijn wrote: > > On Sun, Dec 22, 2019 at 8:11 AM Michael S. Tsirkin <mst@xxxxxxxxxx> wrote: > > > > > > On Fri, Dec 20, 2019 at 10:08:41PM -0500, Willem de Bruijn wrote: > > > > On Fri, Dec 20, 2019 at 4:22 PM Alistair Delva <adelva@xxxxxxxxxx> wrote: > > > > > > > > > > On devices without control virtqueue support, such as the virtio_net > > > > > implementation in crosvm[1], attempting to configure LRO will panic the > > > > > kernel: > > > > > > > > > > kernel BUG at drivers/net/virtio_net.c:1591! > > > > > invalid opcode: 0000 [#1] PREEMPT SMP PTI > > > > > CPU: 1 PID: 483 Comm: Binder:330_1 Not tainted 5.4.5-01326-g19463e9acaac #1 > > > > > Hardware name: ChromiumOS crosvm, BIOS 0 > > > > > RIP: 0010:virtnet_send_command+0x15d/0x170 [virtio_net] > > > > > Code: d8 00 00 00 80 78 02 00 0f 94 c0 65 48 8b 0c 25 28 00 00 00 48 3b 4c 24 70 75 11 48 8d 65 d8 5b 41 5c 41 5d 41 5e 41 5f 5d c3 <0f> 0b e8 ec a4 12 c8 66 90 66 2e 0f 1f 84 00 00 00 00 00 55 48 89 > > > > > RSP: 0018:ffffb97940e7bb50 EFLAGS: 00010246 > > > > > RAX: ffffffffc0596020 RBX: ffffa0e1fc8ea840 RCX: 0000000000000017 > > > > > RDX: ffffffffc0596110 RSI: 0000000000000011 RDI: 000000000000000d > > > > > RBP: ffffb97940e7bbf8 R08: ffffa0e1fc8ea0b0 R09: ffffa0e1fc8ea0b0 > > > > > R10: ffffffffffffffff R11: ffffffffc0590940 R12: 0000000000000005 > > > > > R13: ffffa0e1ffad2c00 R14: ffffb97940e7bc08 R15: 0000000000000000 > > > > > FS: 0000000000000000(0000) GS:ffffa0e1fd100000(006b) knlGS:00000000e5ef7494 > > > > > CS: 0010 DS: 002b ES: 002b CR0: 0000000080050033 > > > > > CR2: 00000000e5eeb82c CR3: 0000000079b06001 CR4: 0000000000360ee0 > > > > > DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 > > > > > DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 > > > > > Call Trace: > > > > > ? preempt_count_add+0x58/0xb0 > > > > > ? _raw_spin_lock_irqsave+0x36/0x70 > > > > > ? _raw_spin_unlock_irqrestore+0x1a/0x40 > > > > > ? __wake_up+0x70/0x190 > > > > > virtnet_set_features+0x90/0xf0 [virtio_net] > > > > > __netdev_update_features+0x271/0x980 > > > > > ? nlmsg_notify+0x5b/0xa0 > > > > > dev_disable_lro+0x2b/0x190 > > > > > ? inet_netconf_notify_devconf+0xe2/0x120 > > > > > devinet_sysctl_forward+0x176/0x1e0 > > > > > proc_sys_call_handler+0x1f0/0x250 > > > > > proc_sys_write+0xf/0x20 > > > > > __vfs_write+0x3e/0x190 > > > > > ? __sb_start_write+0x6d/0xd0 > > > > > vfs_write+0xd3/0x190 > > > > > ksys_write+0x68/0xd0 > > > > > __ia32_sys_write+0x14/0x20 > > > > > do_fast_syscall_32+0x86/0xe0 > > > > > entry_SYSENTER_compat+0x7c/0x8e > > > > > > > > > > This happens because virtio_set_features() does not check the presence > > > > > of the control virtqueue feature, which is sanity checked by a BUG_ON > > > > > in virtnet_send_command(). > > > > > > > > > > Fix this by skipping any feature processing if the control virtqueue is > > > > > missing. This should be OK for any future feature that is added, as > > > > > presumably all of them would require control virtqueue support to notify > > > > > the endpoint that offload etc. should begin. > > > > > > > > > > [1] https://chromium.googlesource.com/chromiumos/platform/crosvm/ > > > > > > > > > > Fixes: a02e8964eaf9 ("virtio-net: ethtool configurable LRO") > > > > > Cc: stable@xxxxxxxxxxxxxxx [4.20+] > > > > > Cc: Michael S. Tsirkin <mst@xxxxxxxxxx> > > > > > Cc: Jason Wang <jasowang@xxxxxxxxxx> > > > > > Cc: David S. Miller <davem@xxxxxxxxxxxxx> > > > > > Cc: kernel-team@xxxxxxxxxxx > > > > > Cc: virtualization@xxxxxxxxxxxxxxxxxxxxxxxxxx > > > > > Cc: linux-kernel@xxxxxxxxxxxxxxx > > > > > Signed-off-by: Alistair Delva <adelva@xxxxxxxxxx> > > > > > > > > Thanks for debugging this, Alistair. > > > > > > > > > --- > > > > > drivers/net/virtio_net.c | 3 +++ > > > > > 1 file changed, 3 insertions(+) > > > > > > > > > > diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c > > > > > index 4d7d5434cc5d..709bcd34e485 100644 > > > > > --- a/drivers/net/virtio_net.c > > > > > +++ b/drivers/net/virtio_net.c > > > > > @@ -2560,6 +2560,9 @@ static int virtnet_set_features(struct net_device *dev, > > > > > u64 offloads; > > > > > int err; > > > > > > > > > > + if (!vi->has_cvq) > > > > > + return 0; > > > > > + > > > > > > > > Instead of checking for this in virtnet_set_features, how about we > > > > make configurability contingent on cvq in virtnet_probe: > > > > > > > > - if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS)) > > > > + if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS) && > > > > + virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ)) > > > > dev->hw_features |= NETIF_F_LRO; > > > > > > > > Based on this logic a little below in the same function > > > > > > > > if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ)) > > > > vi->has_cvq = true; > > > > > > > > > This would be a regression on old hypervisors which didn't have > > > CTL VQ - suddenly they will lose offloads. > > > > dev->features still correctly displays whether offloads are enabled. > > Removing it from dev->hw_features just renders it non-configurable. > > Oh you are right. I confused it with dev->features. > > > Note that before the patch that is being fixed the offloads were > > enabled, but ethtool would show them as off. > > So the bug is in spec, it should have said > VIRTIO_NET_F_CTRL_GUEST_OFFLOADS depends on VIRTIO_NET_F_CTRL_VQ, but we > missed that part. We can and I guess should add this as a recommendation > but it's too late to make it a MUST. > > Meanwhile I would say it's cleanest to work around > this in virtnet_validate by clearing VIRTIO_NET_F_CTRL_GUEST_OFFLOADS > if VIRTIO_NET_F_CTRL_VQ is off, with a big comment explaining > it's a spec bug. Wouldn't that cause precisely the regression you were concerned about? Workloads may now depend on LRO for cycle efficiency. Reverting to behavior before this patch (though now displaying the offload state correctly) is more conservative in that regard.
