[Bug 54933] New: Wine slow on certain fuse filesystems, with fix

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=54933

Bug ID: 54933
           Summary: Wine slow on certain fuse filesystems, with fix
           Product: Wine
           Version: 8.8
          Hardware: x86-64
                OS: Linux
            Status: UNCONFIRMED
          Severity: enhancement
          Priority: P2
         Component: ntdll
          Assignee: [e-mail]
          Reporter: [e-mail]
      Distribution: ---

Created attachment 74452
  --> https://bugs.winehq.org/attachment.cgi?id=74452
CIOPFS speed patch

I found a description of users experiencing slowdowns (even using explorer.exe)
on an MTP filesystem, and I experienced similar slowdowns on an s3ql
filesystem.  In short, even using explorer on certain filesystems is slow, and
opening programs that open up more than a handful of files is very slow off
these filesystems.:
https://forum.winehq.org/viewtopic.php?t=36400

I found the root cause courtesy of strace -- in dlls/ntdll/unix/file.c in
get_dir_case_sensitivity_stat (which is called very frequently, I think on any
file open or creation, among others), there is a fstatat system call, then a
check on that result to see if the filesystem is a FUSE filesystem, then if it
is it looks for a .ciopfs directory (because presumably CIOPFS -- Case
Insensitive On Purpose Filesystem -- places a .ciopfs in every directory
specifically so wine can do this test.)  It turns out the fstatat call gets
used and free inode count, used and free disk space, etc. along with filesystem
type, and on my large s3ql filesystem at least this takes *8/10ths* of a
second, versus the .ciopfs lookup being in dentry cache -- I don't recall if it
used 0.0001 or 0.00001 seconds, but either way extremely fast compared to the
other call.

Reversing the order of these calls (check for .ciopfs first, then if it's there
run fstatat and check result to see if it's a FUSE filesystem) should not cause
a performance regression on any filesystem (since it's usually going to be out
of the dentry cache) while avoiding this slowdown on a filesystems with slow
filesystem stats.

Just a note, I see the equivalent "is filesystem case sensitive?" check for
MacOS actually only runs any tests once then caches the results; using a
similar "case sensitivity" cache for Linux would also be effective (since it
would only run this piece of code once per wine run then.)

Finally, I wonder if CIOPFS is actually used? I see the last update on it is
from 2011, it may have turned out that Wine's case insensitivity support is
fast enough that CIOPFS did not prove useful.  If it's not actually being used
by anyone, removing the CIOPFS tests from file.c could be a reasonable thing to
do.

Please find attached a small patch to implement this change for Linux -- it
simply reverses the order of 3 lines of code, so it runs the (fast) .ciopfs
directory existence test first, then only runs the (on some filesystems much
slower) fstatat call and "is it FUSE filesystem?" test if the.ciopfs directory
actually exists.

Thanks, and keep up the good work!!!
--Henry

[Bug 54933] Wine slow on certain fuse filesystems, with fix

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=54933

--- Comment #1 from Henry Wertz  ---
Just to add, I recently upgrades to s3ql-5.0.0 and it has a statfs cache (it
stores the results for 30 seconds).  The comment in the code says VSCode
repeatedly makes statfs calls and it was put in to speed that up; but it's
equally effective at speeding up wine's continuous use of fstatat.