liblttdvfs: put fadvise hint after sync writeback
authorMathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
Thu, 20 May 2010 02:06:46 +0000 (22:06 -0400)
committerMathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
Thu, 20 May 2010 02:06:46 +0000 (22:06 -0400)
commit825cbbaee78eb17f1381a6be2f36b8023ac433da
tree1ae8ef622448fd07f72dcbc6be91441a947f76fc
parentd39f4b57abf039991b1eb29c33b02e1d593f1137
liblttdvfs: put fadvise hint after sync writeback

On Wed, 19 May 2010, Mathieu Desnoyers wrote:
>
> A faced a small counter-intuitive fadvise behavior though.
>
>   posix_fadvise(fd, 0, 0, POSIX_FADV_DONTNEED);
>
> only seems to affect the parts of a file that already exist.

POSIX_FADV_DONTNEED does not have _any_ long-term behavior. So when you do
a

        posix_fadvise(fd, 0, 0, POSIX_FADV_DONTNEED);

it only affects the pages that are there right now, it has no effect on
any future actions.

> So after each splice() that appends to the file, I have to call fadvise
> again. I would have expected the "0" len parameter to tell the kernel to
> apply the hint to the whole file, even parts that will be added in the
> future.

It's not a hint about future at all. It's a "throw current pages away".

I would also suggest against doing that kind of thing in a streaming write
situation. The behavior for dirty page writeback is _not_ welldefined, and
if you do POSIX_FADV_DONTNEED, I would suggest you do it as part of that
writeback logic, ie you do it only on ranges that you have just waited on.

IOW, in my example, you'd couple the

        sync_file_range(fd, (index-1)*BUFSIZE, BUFSIZE,
+SYNC_FILE_RANGE_WAIT_BEFORE|SYNC_FILE_RANGE_WRITE|SYNC_FILE_RANGE_WAIT_AFTER);

with a

        posix_fadvise(fd, (index-1)*BUFSIZE, BUFSIZE, POSIX_FADV_DONTNEED);

afterwards to throw out the pages that you just waited for.

                Linus

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
liblttd/liblttdvfs.c
This page took 0.024521 seconds and 4 git commands to generate.