Re: fdatasync behaviour in macOS

Hi,

On Wed, 10 Dec 2025 at 06:07, Anil Sharma <[email protected]> wrote:
>
> I am using fio to benchmark serial fsync in macOS. From what I could understand, both fsync and fdatasync are linked to F_FULLSYNC in macOS. But the output is very different. fsync=1 can only get 110 tps where fdatasync=1 option can get 5.5K tps. Which seems like we are only writing to OS Page Cache.
>
> Question:
> Can we get a confirmation whether the fdatasync option is mapped to F_FULLSYNC?

fio's fdatasync call is currently (as fio fio 3.41) *not* mapped to
fcntl(..., F_FULLFSYNC) on macOS and if we read fio's source we can
confirm this:

If you look at https://github.com/axboe/fio/blob/fio-3.41/io_u.c#L2463-L2474
you can see what fio uses to do fsync depends on whether
CONFIG_FCNTL_SYNC is defined. In that block you can see the call to
fdatasync *only* depends on whether CONFIG_FDATASYNC is defined and it
will never be mapped to fcntl(..., F_FULLFSYNC). Over in configure you
can see the probe for fcntl(..., F_FULLFSYNC) in
https://github.com/axboe/fio/blob/fio-3.41/configure#L706-L723 and the
probe for fdatasync(...) over in
https://github.com/axboe/fio/blob/fio-3.41/configure#L1252-L1268 . If
you check the config.log on macOS you can see that "Compiling test
case fcntl(F_FULLFSYNC)" and "Compiling test case fdatasync" both
succeed leading to CONFIG_FCNTL_SYNC and CONFIG_FDATASYNC both being
defined (see https://github.com/axboe/fio/blob/fio-3.41/configure#L3075-L3077
and https://github.com/axboe/fio/blob/fio-3.41/configure#L3380-L3382
for where these symbols are output).

> Please let me know if this needs to go in github issues.

Hmm it's not something I personally will work on (although I can 't
speak for others) so it's one of those things where a well formed
patch/PR would be more useful but there's some debate as to whether
it's worth it. It's well known that fsync() alone on macOS doesn't
trigger *drive* cache flushes (unlike Linux) so I don't think it's a
surprise that fdatasync() has the same behaviour as fsync()...
However, I can see how fio's remapping of just fsync and not fdatasync
is surprising. I guess the question is: does it make sense to even use
fdatasync=1 with fio on macOS if it's just going to do the same
heavyweight fcntl(..., F_FULLFSYNC) work as the current fsync() ->
fcntl(..., F_FULLFSYNC) mapping or should it just be disabled
entirely?

-- 
Sitsofe