Re: [Qemu-devel] [PATCH] block: workaround for unaligned byte range in fallocate()

[Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>]

22.08.2019 21:55, Vladimir Sementsov-Ogievskiy wrote:
> 22.08.2019 21:31, Andrey Shinkevich wrote:
>> Revert the commit 118f99442d 'block/io.c: fix for the allocation failure'
>> and make better error handling for the file systems that do not support
>> fallocate() for the unaligned byte range. Allow falling back to pwrite
>> in case fallocate() returns EINVAL.
>>
>> Suggested-by: Kevin Wolf <kwolf@redhat.com>
>> Suggested-by: Eric Blake <eblake@redhat.com>
>> Signed-off-by: Andrey Shinkevich <andrey.shinkevich@virtuozzo.com>
>> ---
>> Discussed in email thread with the message ID
>> <1554474244-553661-1-git-send-email-andrey.shinkevich@virtuozzo.com>
>>
>>   block/file-posix.c | 7 +++++++
>>   block/io.c         | 2 +-
>>   2 files changed, 8 insertions(+), 1 deletion(-)
>>
>> diff --git a/block/file-posix.c b/block/file-posix.c
>> index fbeb006..2c254ff 100644
>> --- a/block/file-posix.c
>> +++ b/block/file-posix.c
>> @@ -1588,6 +1588,13 @@ static int j(void *opaque)
>>       if (s->has_write_zeroes) {
>>           int ret = do_fallocate(s->fd, FALLOC_FL_ZERO_RANGE,
>>                                  aiocb->aio_offset, aiocb->aio_nbytes);
>> +        if (ret == -EINVAL) {
>> +            /*
>> +             * Allow falling back to pwrite for file systems that
>> +             * do not support fallocate() for unaligned byte range.
>> +             */
>> +            return -ENOTSUP;
>> +        }
>>           if (ret == 0 || ret != -ENOTSUP) {
>>               return ret;
>>           }
> 
> Hmm stop, you've done exactly what Den was afraid of:
> 
> the next line
>    s->has_write_zeroes = false;
> 
> will disable write_zeroes forever.
> 
> Something like
> 
> --- a/block/file-posix.c
> +++ b/block/file-posix.c
> @@ -1588,10 +1588,12 @@ static int handle_aiocb_write_zeroes(void *opaque)
>       if (s->has_write_zeroes) {
>           int ret = do_fallocate(s->fd, FALLOC_FL_ZERO_RANGE,
>                                  aiocb->aio_offset, aiocb->aio_nbytes);
> -        if (ret == 0 || ret != -ENOTSUP) {
> +        if (ret == 0 || (ret != -ENOTSUP && ret != -EINVAL)) {
>               return ret;
>           }
> -        s->has_write_zeroes = false;
> +        if (ret == -ENOTSUP) {
> +            s->has_write_zeroes = false;
> +        }
>       }
>   #endif
> 
> 
> will work better. So, handle ENOTSUP as "disable write_zeros forever", and EINVAL as
> "don't disable, but fallback to writing zeros". And we need same handling for following do_fallocate() calls
> too (otherwise they again fails with EINVAL which will break the whole thing).
> 

Oops, sorry, I misread your patch, it's OK.

Still we may want to handle other do_fallocate() calls in same manner, or may be just:

@@ -1558,7 +1558,13 @@ static int coroutine_fn bdrv_co_do_pwrite_zeroes(BlockDriverState *bs,
              assert(!bs->supported_zero_flags);
          }

-        if (ret < 0 && !(flags & BDRV_REQ_NO_FALLBACK)) {
+        /*
+         * We are sure that our arguments make sense, so consider "invalid
+         * argument" in same manner as "not supported".
+         */
+        if ((ret == -ENOTSUP || ret == -EINVAL) &&
+            !(flags & BDRV_REQ_NO_FALLBACK))
+        {
              /* Fall back to bounce buffer if write zeroes is unsupported */
              BdrvRequestFlags write_flags = flags & ~BDRV_REQ_ZERO_WRITE;




-- 
Best regards,
Vladimir