While trying to use posix_fadvise(...POSIX_FADVISE_WILLNEED) to
implement userspace read-ahead (with a bigger read-ahead window than
the kernel default) I found that if the underlying filesystem is
fuse/sshfs posix_fadvise is no longer doing asynchronous reads:
strace -tt with mnt/testfile on sshfs and server with very slow
upstream (ADSL):
5345 00:00:17.334209 open("mnt/testfile", O_RDONLY|O_LARGEFILE) = 3
5345 00:00:18.011383 _llseek(3, 0, [3544379], SEEK_END) = 0
5345 00:00:18.011626 _llseek(3, 0, [0], SEEK_SET) = 0
5345 00:00:18.012393 fadvise64_64(3, 0, 1048576, POSIX_FADV_WILLNEED) = 0
5345 00:01:02.438097 write(1, "[file] File size is 3544379 byte"..., 34) = 34
Note that fadvise takes 40 seconds...
Is this expected behaviour?
I would have expected that fadvise is always asynchronous and I
could rely on the call to return almost immediately.
--
Tobias PGP: http://8ef7ddba.uguu.de
On Wed, 11 Nov 2009, Tobias Diedrich wrote:
> While trying to use posix_fadvise(...POSIX_FADVISE_WILLNEED) to
> implement userspace read-ahead (with a bigger read-ahead window than
> the kernel default) I found that if the underlying filesystem is
> fuse/sshfs posix_fadvise is no longer doing asynchronous reads:
>
> strace -tt with mnt/testfile on sshfs and server with very slow
> upstream (ADSL):
> 5345 00:00:17.334209 open("mnt/testfile", O_RDONLY|O_LARGEFILE) = 3
> 5345 00:00:18.011383 _llseek(3, 0, [3544379], SEEK_END) = 0
> 5345 00:00:18.011626 _llseek(3, 0, [0], SEEK_SET) = 0
> 5345 00:00:18.012393 fadvise64_64(3, 0, 1048576, POSIX_FADV_WILLNEED) = 0
> 5345 00:01:02.438097 write(1, "[file] File size is 3544379 byte"..., 34) = 34
>
> Note that fadvise takes 40 seconds...
> Is this expected behaviour?
I think this is because fuse limits the number of outstanding
requests.
In 2.6.31 you can raise these limits in
/sys/fs/fuse/connections/N/congestion_threshold
/sys/fs/fuse/connections/N/max_background
Thanks,
Miklos