For unaligned offset pread() may return EINVAL even if the offset pass
EOF, although it should not, since otherwise there is no abiliity to
rely on read() == 0 is EOF (with pread() loop).
Here is a reproducer for the problem on 4.9.0-12-amd64:
$ head -c27 /dev/urandom > /tmp/pread.issue
$ xfs_io
xfs_io> open -d /tmp/pread.issue
xfs_io> pread 1000 4096
pread: Invalid argument
And this is how it should work:
xfs_io> pread 29 4096
read 0/4096 bytes at offset 29
Note, here I use interactive mode since we had old xfs_io that does not
allow to execute multiple commands at once, and to avoid EMFILE issue
Here is some history of a patches that affects this behaviour in the
linux kernel:
- the issue had been introduced in
torvalds/linux@9fe55eea7e v3.14
("Fix race when checking i_size on direct i/o read")
- an attempt to fix it had been made in
torvalds/linux@74cedf9b6c v4.4
("direct-io: Fix negative return from dio read beyond eof")
- but this wasn't enough, since alignment check was earlier, so
eventually fixed in
torvalds/linux@41b21af388 v5.10
("direct-io: defer alignment check until after the EOF check")
Someone may ask why CI does not shows the issue, since:
- it had 4.19 kernel when CI was in yandex
- now it has 5.4 when CI is in AWS
Since both of those kernels does not have the last patch.
But, this bug requires the following conditions to met:
- index_granularity_bytes=0
- min_merge_bytes_to_use_direct_io=1
Which was not covered by CI yet.
