On 09/15/2016 08:23 AM, Fabrice Bacchella wrote:
fio uses posix_fadvise, but can only take two values :
if (td->o.invalidate_cache && file_invalidate_cache(td, f))
goto err;
if (td->o.fadvise_hint &&
(f->filetype == FIO_TYPE_BD || f->filetype == FIO_TYPE_FILE)) {
int flags;
if (td_random(td))
flags = POSIX_FADV_RANDOM;
else
flags = POSIX_FADV_SEQUENTIAL;
if (posix_fadvise(f->fd, f->file_offset, f->io_size, flags) < 0) {
td_verror(td, errno, "fadvise");
goto err;
}
}
But it can take many values, from linux/fadvise.h:
#define POSIX_FADV_NORMAL 0 /* No further special treatment. */
#define POSIX_FADV_RANDOM 1 /* Expect random page references. */
#define POSIX_FADV_SEQUENTIAL 2 /* Expect sequential page references. */
#define POSIX_FADV_WILLNEED 3 /* Will need these pages. */
...
#define POSIX_FADV_DONTNEED 4 /* Don't need these pages. */
#define POSIX_FADV_NOREUSE 5 /* Data will be accessed once. */
POSIX_FADV_NOREUSE have no effect on current kernel
(https://github.com/torvalds/linux/blob/master/mm/fadvise.c#L115) but
DONTNEED is interesting as it allow to flush page cache after use It
allows to do bench with machines with huge memory without having to
bench huge files, and without using directio that is a strange beast.
But it needs to be call right after use.
It's used in file_invalidate_cache but this function seems to be call
once, when file is open for the first time.
Is there any plan to support it ?
What kind of use case did you have in mind? Dropping pages after writing
them? That basically boils down to being a non data integrity fdatasync.
You would get basically the same effect by using fdatasync=1 in your
workload. The difference would mainly be that DONTNEED would factor in
the range and offset. So if you have multiple writers on the same file,
then there would be a difference in behavior.
That said, I don't mind adding support for using DONTNEED specifically.
Currently it's just used to set a permanent hint, which is why it's only
done once. For DONTNEED, you'd want to call it after each write instead.
--
Jens Axboe
--
To unsubscribe from this list: send the line "unsubscribe fio" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
