On Wed, Aug 09, 2023 at 06:41:17AM -0700, Christoph Hellwig wrote:
> Please do not add a new ->direct_IO method. I'm currently working hard
> on removing it, just set FMODE_CAN_ODIRECT and handle the fallback in
> your read_iter/write_iter methods.
>
> But if we just start claiming direct I/O support for file systems that
> don't actually support it, I'm starting to seriously wonder why we
> bother with the flag at all and don't just allow O_DIRECT opens
> to always succeed..
I see it differently -- you can do byte-aligned directio to S_DAX files
on persistent memory, so I don't see why you can't do that for tmpfs
files too.
(I'm not advocating for letting *disk* based filesystems allow O_DIRECT
even if read and writes are always going to go through the page cache
and get flushed to disk. If programs wanted that, they'd use O_SYNC.)
/mnt is a pmem filesystem, /mnt/on/file has S_DAX set, and /mnt/off/file
does not:
# xfs_io -c statx /mnt/{on,off}/file
fd.path = "/mnt/on/file"
fd.flags = non-sync,non-direct,read-write
stat.ino = 132
stat.type = regular file
stat.size = 1048576
stat.blocks = 2048
fsxattr.xflags = 0x8002 [-p------------x--]
fsxattr.projid = 0
fsxattr.extsize = 0
fsxattr.cowextsize = 0
fsxattr.nextents = 1
fsxattr.naextents = 0
dioattr.mem = 0x200
dioattr.miniosz = 512
dioattr.maxiosz = 2147483136
fd.path = "/mnt/off/file"
fd.flags = non-sync,non-direct,read-write
stat.ino = 8388737
stat.type = regular file
stat.size = 1048576
stat.blocks = 2048
fsxattr.xflags = 0x2 [-p---------------]
fsxattr.projid = 0
fsxattr.extsize = 0
fsxattr.cowextsize = 0
fsxattr.nextents = 1
fsxattr.naextents = 0
dioattr.mem = 0x200
dioattr.miniosz = 512
dioattr.maxiosz = 2147483136
And now we try a byte-aligned direct write:
# xfs_io -d -c 'pwrite -S 0x58 47 1' /mnt/off/file
pwrite: Invalid argument
# xfs_io -d -c 'pwrite -S 0x58 47 1' /mnt/on/file
wrote 1/1 bytes at offset 47
1.000000 bytes, 1 ops; 0.0001 sec (5.194 KiB/sec and 5319.1489 ops/sec)
--D
