On 12/13/2017 05:05 AM, Eryu Guan wrote:
> On Thu, Dec 07, 2017 at 10:00:42AM -0600, Goldwyn Rodrigues wrote:
>> From: Goldwyn Rodrigues <rgoldwyn@xxxxxxxx>
>>
>> There are some parameters added with xfs_io. Check if the pwrite
>> parameters are available. For some cases, xfs_io now returns "command
>> -%c not supported", so added "not supported" to count as error.
>>
>> Signed-off-by: Goldwyn Rodrigues <rgoldwyn@xxxxxxxx>
>>
>> Changes since v3:
>> - pwrite -N check to add -V parameters to check on pwritev2()
>>
>> ---
>> common/rc | 16 +++++++++++++++-
>> 1 file changed, 15 insertions(+), 1 deletion(-)
>>
>> diff --git a/common/rc b/common/rc
>> index 4c053a53..4f4b9e10 100644
>> --- a/common/rc
>> +++ b/common/rc
>> @@ -2035,6 +2035,7 @@ _require_xfs_io_command()
>> shift
>> local param="$*"
>> local param_checked=0
>> + local opts=""
>>
>> testfile=$TEST_DIR/$$.xfs_io
>> case $command in
>> @@ -2079,6 +2080,17 @@ _require_xfs_io_command()
>> echo $testio | grep -q "invalid option" && \
>> _notrun "xfs_io $command support is missing"
>> ;;
>> + "pwrite")
>> + # -N (RWF_NOWAIT) only works with direct I/O writes
>> + local pwrite_opts=" "
>> + if [ "$param" == "-N" ]; then
>> + opts+=" -d"
>> + pwrite_opts+="-V 1 -b 4k"
>> + fi
>> + testio=`$XFS_IO_PROG -f $opts -c \
>> + "pwrite $pwrite_opts $param 0 4k" $testfile 2>&1`
>> + param_checked=1
>
> I think I found why xfs_io returned EOPNOTSUPP but strace showed EAGAIN
> for me. It's pwritev2 glibc wrapper that modifies the errno before
> return, please see glibc commit 852d63120783 ("posix: Set p{read,write}v2
> to return ENOTSUP (BZ#21780)").
>
Oh yes, that explains it.
> diff --git a/sysdeps/unix/sysv/linux/pwritev2.c b/sysdeps/unix/sysv/linux/pwritev2.c
> index 72f0471f969a..8e5032fe2fed 100644
> --- a/sysdeps/unix/sysv/linux/pwritev2.c
> +++ b/sysdeps/unix/sysv/linux/pwritev2.c
> @@ -28,7 +28,7 @@ pwritev2 (int fd, const struct iovec *vector, int count, off_t offset,
> # ifdef __NR_pwritev2
> ssize_t result = SYSCALL_CANCEL (pwritev2, fd, vector, count,
> LO_HI_LONG (offset), flags);
> - if (result >= 0 || errno != ENOSYS)
> + if (result >= 0)
> return result;
> # endif
> /* Trying to emulate the pwritev2 syscall flags is troublesome:
> @@ -42,7 +42,7 @@ pwritev2 (int fd, const struct iovec *vector, int count, off_t offset,
>
> if (flags != 0)
> {
> - __set_errno (EOPNOTSUPP);
> + __set_errno (ENOTSUP);
> return -1;
> }
> return pwritev (fd, vector, count, offset);
>
> So originally, pritev2 wrapper returned any error except ENOSYS, but now
> it translates all errors to ENOTSUP (EOPNOTSUPP). That's why kernel
> returned EAGAIN and strace captured EAGAIN but perror() in xfs_io still
> printed EOPNOTSUPP. It looks like a glibc bug to me.
>
> Apparently, my Fedora rawhide system is new enough to have this glibc
> version. And I did see generic/470 pass with 4.15-rc2 kernel with that
> glibc patch reverted.
>
> Anyway, with this additional "-V 1" in pwrite test, generic/470 would
> _notrun on buggy system.
>
> I'll queue this patchset for next release if there's no more comments
> from other reivewers.
Thanks for your efforts.
Just to let you know that generic/471 does not pass as yet. IOW, direct
I/O write() writes the data to the file _and_ returns -ENOSPC. I have
sent a patch upstream [1] which fixes the bug but it has not been
incorporated as yet. In any case, the write() should either write and
return number of bytes written or return -ENOSPC and not write anything
at all. Following semantics, it should be the former.
[1] https://patchwork.kernel.org/patch/10070399/
--
Goldwyn
--
To unsubscribe from this list: send the line "unsubscribe fstests" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
