Writes permanently hang with very heavy I/O on virtio-scsi - worse on virtio-blk

[James Harvey]

Up to date Arch Linux on host and guest. linux 5.2.11. QEMU 4.1.0. Full command line at bottom.

Host gives QEMU two thin LVM volumes. The first is the root filesystem, and the second is for heavy I/O, on a Samsung 970 Evo 1TB.

When maxing out the I/O on the second virtual block device using virtio-blk, I often get a "lockup" in about an hour or two. From the advise of iggy in IRC, I switched over to virtio-scsi. It ran perfectly for a few days, but then "locked up" in the same way.

By "lockup", I mean writes to the second virtual block device permanently hang. I can read files from it, but even "touch foo" never times out, cannot be "kill -9"'ed, and is stuck in uninterruptible sleep.

When this happens, writes to the first virtual block device with the root filesystem are fine, so the O/S itself remains responsive.

The second virtual block device uses BTRFS. But, I have also tried XFS and reproduced the issue.

In guest, when this starts, it starts logging "task X blocked for more than Y seconds". Below is an example of one of these. At this point, anything that is or does in the future write to this block device gets stuck in uninterruptible sleep.

-----

INFO: task kcompactd:232 blocked for more than 860 seconds.
      Not tained 5.2.11-1 #1
"echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this messae.
kcompactd0 D 0 232 2 0x80004000
Call Trace:
 ? __schedule+0x27f/0x6d0
 schedule+0x3d/0xc0
 io_schedule+0x12/0x40
 __lock_page+0x14a/0x250
 ? add_to_page_cache_lru+0xe0/0xe0
 migrate_pages+0x803/0xb70
 ? isolate_migratepages_block+0x9f0/0x9f0
 ? __reset_isolation_suitable+0x110/0x110
 compact_zone+0x6a2/0xd30
 kcompactd_do_work+0x134/0x260
 ? kvm_clock_read+0x14/0x30
 ? kvm_sched_clock_read+0x5/0x10
 kcompactd+0xd3/0x220
 ? wait_woken+0x80/0x80
 kthread+0xfd/0x130
 ? kcompactd_do_work+0x260/0x260
 ? kthread_park+0x80/0x80
 ret_from_fork+0x35/0x40

-----

In guest, there are no other dmesg/journalctl entries other than "task...blocked".

On host, there are no dmesg/journalctl entries whatsoever. Everything else in host continues to work fine, including other QEMU VM's on the same underlying SSD (but obviously different lvm volumes.)

I understand there might not be enough to go on here, and I also understand it's possible this isn't a QEMU bug. Happy to run given commands or patches to help diagnose what's going on here.

I'm now running a custom compiled QEMU 4.1.0, with debug symbols, so I can get a meaningful backtrace from the host point of view.

I've only recently tried this level of I/O, so can't say if this is a new issue.

When writes are hanging, on host, I can connect to the monitor. Running "info block" shows nothing unusual.

-----

/usr/bin/qemu-system-x86_64
   -name arch,process=qemu:arch
   -no-user-config
   -nodefaults
   -nographic
   -uuid 0528162b-2371-41d5-b8da-233fe61b6458
   -pidfile /tmp/0528162b-2371-41d5-b8da-233fe61b6458.pid
   -machine q35,accel=kvm,vmport=off,dump-guest-core=off
   -cpu SandyBridge-IBRS
   -smp cpus=24,cores=12,threads=1,sockets=2
   -m 24G
   -drive if=pflash,format=raw,readonly,file=/usr/share/ovmf/x64/OVMF_CODE.fd
   -drive if=pflash,format=raw,readonly,file=/var/qemu/0528162b-2371-41d5-b8da-233fe61b6458.fd
   -monitor telnet:localhost:8000,server,nowait,nodelay
   -spice unix,addr=/tmp/0528162b-2371-41d5-b8da-233fe61b6458.sock,disable-ticketing
   -device ioh3420,id=pcie.1,bus=pcie.0,slot=0
   -device virtio-vga,bus=pcie.1,addr=0
   -usbdevice tablet
   -netdev bridge,id=network0,br=br0
   -device virtio-net-pci,netdev=network0,mac=02:37:de:79:19:09,bus=pcie.0,addr=3
   -device virtio-scsi-pci,id=scsi1
   -drive driver=raw,node-name=hd0,file=/dev/lvm/arch_root,if=none,discard=unmap
   -device scsi-hd,drive=hd0,bootindex=1
   -drive driver=raw,node-name=hd1,file=/dev/lvm/arch_nvme,if=none,discard=unmap
   -device scsi-hd,drive=hd1,bootindex=2

-----

[Stefan Hajnoczi (stefanha)]

On Thu, Sep 05, 2019 at 03:42:03AM -0000, James Harvey wrote:
> ** Description changed:
>
> Up to date Arch Linux on host and guest. linux 5.2.11. QEMU 4.1.0.
> Full command line at bottom.
>
> Host gives QEMU two thin LVM volumes. The first is the root filesystem,
> and the second is for heavy I/O, on a Samsung 970 Evo 1TB.
>
> When maxing out the I/O on the second virtual block device using virtio-
> blk, I often get a "lockup" in about an hour or two. From the advise of
> iggy in IRC, I switched over to virtio-scsi. It ran perfectly for a few
> days, but then "locked up" in the same way.
>
> By "lockup", I mean writes to the second virtual block device
> permanently hang. I can read files from it, but even "touch foo" never
> times out, cannot be "kill -9"'ed, and is stuck in uninterruptible
> sleep.
>
> When this happens, writes to the first virtual block device with the
> root filesystem are fine, so the O/S itself remains responsive.
>
> The second virtual block device uses BTRFS. But, I have also tried XFS
> and reproduced the issue.
>
> In guest, when this starts, it starts logging "task X blocked for more
> than Y seconds". Below is an example of one of these. At this point,
> anything that is or does in the future write to this block device gets
> stuck in uninterruptible sleep.
>
> -----
>
> INFO: task kcompactd:232 blocked for more than 860 seconds.
>       Not tained 5.2.11-1 #1
> "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this messae.
> kcompactd0 D 0 232 2 0x80004000
> Call Trace:
>  ? __schedule+0x27f/0x6d0
>  schedule+0x3d/0xc0
>  io_schedule+0x12/0x40
>  __lock_page+0x14a/0x250
>  ? add_to_page_cache_lru+0xe0/0xe0
>  migrate_pages+0x803/0xb70
>  ? isolate_migratepages_block+0x9f0/0x9f0
>  ? __reset_isolation_suitable+0x110/0x110
>  compact_zone+0x6a2/0xd30
>  kcompactd_do_work+0x134/0x260
>  ? kvm_clock_read+0x14/0x30
>  ? kvm_sched_clock_read+0x5/0x10
>  kcompactd+0xd3/0x220
>  ? wait_woken+0x80/0x80
>  kthread+0xfd/0x130
>  ? kcompactd_do_work+0x260/0x260
>  ? kthread_park+0x80/0x80
>  ret_from_fork+0x35/0x40
>
> -----
>
> In guest, there are no other dmesg/journalctl entries other than
> "task...blocked".
>
> On host, there are no dmesg/journalctl entries whatsoever. Everything
> else in host continues to work fine, including other QEMU VM's on the
> same underlying SSD (but obviously different lvm volumes.)
>
> I understand there might not be enough to go on here, and I also
> understand it's possible this isn't a QEMU bug. Happy to run given
> commands or patches to help diagnose what's going on here.
>
> I'm now running a custom compiled QEMU 4.1.0, with debug symbols, so I
> can get a meaningful backtrace from the host point of view.
>
> I've only recently tried this level of I/O, so can't say if this is a
> new issue.
>
> + When writes are hanging, on host, I can connect to the monitor. Running
> + "info block" shows nothing unusual.
> +
> -----
>
> /usr/bin/qemu-system-x86_64
>    -name arch,process=qemu:a...

[James Harvey (jamespharvey20)]

Apologies, it looks like I ran into two separate bugs, one with XFS, and one with BTRFS, that had the same symptom, initially making me to think this must be a QEMU issue.

Using blktrace, I was able to see within the VM, that the virtio block device wasn't getting the writes that were going into uninterruptible sleep.

So, this should be able to be closed. For some reason, virtio-blk seemed to trigger the bugs more rapidly, but at this point, I can't say there is anything at fault with it or virtio-scsi.

BTRFS issue was discussed and linked to here https://lore.kernel.org<email address hidden>/ and has been released. I've been able to run it for several days without a lockup, so it seems to have fixed the issue for me.

I just emailed the XFS list about the separate problems with it. No idea if it's an issue in more recent kernels than 5.1.15-5.1.16, which is what I was running at the time of the XFS errors. (Like the original report said, I was on 5.2.11 at that point.) See https://www.spinics.net/lists/linux-xfs/msg31927.html

[Stefan Hajnoczi (stefanha)]

Thanks for updating us on this issue, which turned out not to be a QEMU bug.