On 8/8/19 3:32 AM, Stefan Hajnoczi wrote:
> On Mon, Aug 05, 2019 at 05:14:11PM +0100, Stefan Hajnoczi wrote:
>
> Sorry Jens, I had your old email address. Resent.
>
>> Hi Jens,
>> I've minimized one of the qemu-iotests failures and it looks like a
>> kernel bug. io_uring completes a request with EAGAIN on an O_DIRECT
>> file with Linux 5.3-rc1 on XFS + dm-thin + dm-crypt:
>>
>> $ qemu-img create -f raw t.img 6G
>> $ qemu-io -f raw -n -i io_uring -c "aio_write 0 1M" -c "aio_write 512 1M" t.img
>> $ qemu-io -f raw -n -i io_uring -c "aio_write $((1 * 1024 * 1024)) 1M" -c "aio_write $((1 * 1024 * 1024 + 512)) 1M" t.img
>> Formatting 'tests/qemu-iotests/scratch/t.img', fmt=raw size=6442450944
>> wrote 1048576/1048576 bytes at offset 0
>> 1 MiB, 1 ops; 00.01 sec (128.598 MiB/sec and 128.5985 ops/sec)
>> wrote 1048576/1048576 bytes at offset 512
>> 1 MiB, 1 ops; 00.00 sec (201.017 MiB/sec and 201.0169 ops/sec)
>> aio_write failed: Resource temporarily unavailable
>> wrote 1048576/1048576 bytes at offset 1049088
>> 1 MiB, 1 ops; 00.00 sec (391.620 MiB/sec and 391.6198 ops/sec)
>>
>> The third write request should not fail with EAGAIN ("Resource
>> temporarily available"). I think device-mapper is returning EAGAIN at
>> some point but I haven't gotten to the bottom of this yet.
>>
>> Interestingly it only happens when qemu-img create is run before the
>> test to create a new file. Rerun qemu-io on an existing raw image file
>> and it succeeds. This could be because the failing request is XFS a
>> metadata I/O request that only happens when writing the file for the
>> first time.
It's due to how XFS handles (or doesn't handle) REQ_NOWAIT. I've got
a patch for this that needs some testing, but I probably won't get to
this until Monday, currently out on vacation until Sunday.
--
Jens Axboe
