On 12/01/2016 12:06 AM, Dmitry Monakhov wrote:

Maxim Patlasov <mpatla...@virtuozzo.com> writes:


You're right. And while composing the patch I well understood that it's
possible to rework fuse_sync_writes() using a counter instead of
negative bias. But the problem with flush_mtime still exists anyway.
Think about it: we firstly acquire local mtime from local inode, then
fill and submit mtime-update-request. Since then, we don't know when
exactly fuse daemon will apply that new mtime to its metadata
structures. If another mtime-update is generated in-between (e.g. "touch
-d <date> file", or even simplier -- just a single direct write
implicitly updating mtime), we wouldn't know which of those two
mtime-update-requests are processed by fused first. That comes from a
general FUSE protocol limitation: when kernel fuse queues request A,
then request B, it cannot be sure if they will be processed by userspace
as <A, then B> or <B, then A>.

The big advantage of the patch I sent is that it's very simple,
straightforward and presumably will remove 99% of contention between
fsync and io_submit (assuming we spend most of time waiting for
userspace ACK for FUSE_FSYNC request. There are actually three questions
to answer:

1) Do we really must honor a crazy app who mixes a lot of fsyncs with a
lot of io_submits? The goal of fsync is to ensure that some state is
actually went to platters. An app who races io_submit-s with fsync-s
actually doesn't care which state will come to platters. I'm not sure
that it's reasonable to work very hard to achieve the best possible
performance for such a marginal app.
Obiously any filesystem behave like this.
Task A(mail-server) may perform write/fsync, task B(mysql) do a lot of 
All that io may happens in parallel, fs guarantee only that metadata
will be serialized. So all that concurent IO flowa to blockdevice which
does no have i_mutex so all IO indeed happen concurrently.

Looks as you're comparing an app doing POSIX open/read/write/fsync/close with fs doing submit_bio. This is a stretch. But OK, there is a similarity. But I don't think this rather vague similarity proves something.

But when we dealt with fs-in-file (loop/ploop/qemu-nbd) we face i_mutex
on file.

This really makes sense. If an app inside a VM loops over ordinary direct writes, while another app (in the same VM) does fsync, it's not fair to suspend the first app for long while just because fuse holds i_mutex for long somewhere deep in fuse_fsync.

For general filesystem (xfs/ext4)

What makes xfs/ext4 general? They are *local* fs having almost direct access to underlying HDD. Have you explored how CEPH implements fsync?

we grab i_mutex only on write
path, fsync is lockless.

What about the following part of ext4_sync_file:

    if (!journal) {
        ret = generic_file_fsync(file, start, end, datasync);
? Is it lockless?

But int case of fuse we artificially introduce
i_mutex inside fsync

exactly the same as generic_file_fsync()

which basically kill concurrency for upper FS.
As result we have SMP scalability as we have in Linux-v2.2 with single
mutex in VFS.

You can't derive "kill concurrency" from "introduce i_mutex" as simple as that. It really does matter for how long the mutex is held and how often it contends with writes.

BTW: I'm wondering why do we care about mtime at all. for fs-in-file
we can relax that, for example flush mtime only on fsync, and not for fdatasync.

Good catch, makes sense!

2) Will the patch (in the form I sent it) break something? I think no.
If you know some usecase that can be broken, let's discuss it in more

3) Should we expect some noticeable (or significant) improvement in
performance comparing fuse_fsync with no locking at all vs. the locking
we have with that patch applied? I tend to think that the answer is "no"
because handling FUSE_FSYNC is notoriously heavy-weight operation. If
you disagree, let's firstly measure that difference in performance
(simply commenting out lock/unlock(i_mutex) in fuse_fsync) and then
start to think if it's really worthy to fully re-work locking scheme to
preserve flush_mtime correctness w/o i_mutex.



On 11/30/2016 05:09 AM, Alexey Kuznetsov wrote:
Sorry, missed that pair fuse_set_nowrite/fuse_release_writes
can be done only under i_mutex.

IMHO it is only due to bad implementation.
If fuse_set_nowrite would be done with separate
count instead of adding negative bias, it would
be possible.

On Wed, Nov 30, 2016 at 3:47 PM, Alexey Kuznetsov <kuz...@virtuozzo.com> wrote:

I do not think you got it right.

i_mutex in fsync is not about some atomicity,
it is about stopping data feed while fsync is executed
to prevent livelock.

I cannot tell anything about mtime update, it is just some voodoo
magic for me.

What's about fsync semantics, I see two different ways:


1. Remove useless write_inode_now. Its work is done
      by filemap_write_and_wait_range(), there is no need to repeat it
under mutex.
2. move mutex_lock _after_  fuse_sync_writes(), which is essentially
      fuse continuation forfilemap_write_and_wait_range().
3. i_mutex is preserved only around fsync call.

1. Remove  write_inode_now as well.
2. Remove i_mutex _completely_. (No idea about mtime voodo though)
2. Replace fuse_sync_writes() with fuse_set_nowrite()
      and add release after call to FSYNC.

Both prevent livelock. B is obviosly optimal.

But A preserves historic fuse protocol semantics.
F.e. I have no idea would user space survive truncate
racing with fsync. pstorage should survice, though this
path was never tested.

On Wed, Nov 30, 2016 at 4:02 AM, Maxim Patlasov <mpatla...@virtuozzo.com> wrote:
fuse_fsync_common() does need i_mutex for fuse_sync_writes() and
fuse_flush_mtime(). But when those operations are done, it's actually
doesn't matter whether to hold the lock over fuse_request_send(FUSE_FSYNC)
or not: we ensured that all relevant data were already seen by
userspace fuse daemon, and so it will sync them (by handling FUSE_FSYNC)
anyway; if the user screws up by leaking new data updates in-between, it
is up to the user and doesn't violate fsync(2) semantics.


Signed-off-by: Maxim Patlasov <mpatla...@virtuozzo.com>
   fs/fuse/file.c |    3 +++
   1 file changed, 3 insertions(+)

diff --git a/fs/fuse/file.c b/fs/fuse/file.c
index 464b2f5..559dfd9 100644
--- a/fs/fuse/file.c
+++ b/fs/fuse/file.c
@@ -697,6 +697,8 @@ int fuse_fsync_common(struct file *file, loff_t start, 
loff_t end,
                  goto out;

+       mutex_unlock(&inode->i_mutex);
          memset(&inarg, 0, sizeof(inarg));
          inarg.fh = ff->fh;
          inarg.fsync_flags = datasync ? 1 : 0;
@@ -715,6 +717,7 @@ int fuse_fsync_common(struct file *file, loff_t start, 
loff_t end,
                          fc->no_fsync = 1;
                  err = 0;
+       return err;
          return err;

Devel mailing list

Reply via email to