Mailinglist Archive: opensuse (1239 mails)

< Previous Next >
Re: [opensuse] Re: XFS and openSUSE 12.1
Greg Freemyer wrote:
On Wed, Jun 12, 2013 at 6:38 PM, Linda Walsh <suse@xxxxxxxxx> wrote:
Greg Freemyer wrote:
open()
write()
fsync()
posix_fadvise(fd, 0, len, POSIX_FADV_DONTNEED)
close()
----
FWIW -- the calls are *advisory*, meaning the kernel may still
not do what you want, versus, O_DIRECT you are avoiding the membuffer
and its associated "garbage collection", altogether. So if it works
for you, yeay!

Asking if it works seems like an easy question for the xfs list, but
it seems like it should be filesystem independent to me.
----
I will likely work, but I used that call to fix xfs_fsr to not eat my
memory and submitted the patch. The bit where it reads files from
the old position and writes to the new -- it used kernel buff reads
and I would notice my buffers get thrashed. So I tried that and
it worked great.

The xfs people didn't really want to add a posix call as there was
no reason to write through the kernel in the first place, so later
one of them came out with a similar patch that used O_direct for
the file reading as well as the file writing(which due to how xfs_fsr
reorganizes the disk was already O_DIRECT).

So their preference was to use O_DIRECT which is probably better
for this use case -- known, fixed-writes occurring at regular
intervals. IF the write call took any real time to perform (which
would likely be negligible), it could be done in a separate thread
in background (roll-yur-own AIO -- or it could be done in linux AID
calls, but I doubt either would be necessary the usage). Since
there's no reason to buffer the data in ram, there doesn't seem to be
much benefit to writing through the kernel file-buffers. Anyway,
in their use case, O_DIRECT was a more efficient choice. I was just
tossing out other options cuz it sounded like he didn't like
O_DIRECT.

Meh.
;-)
--
To unsubscribe, e-mail: opensuse+unsubscribe@xxxxxxxxxxxx
To contact the owner, e-mail: opensuse+owner@xxxxxxxxxxxx

< Previous Next >
This Thread