[Bug 57388] New: Major perf loss with blocking ReadFile() & OVERLAPPED

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=57388

Bug ID: 57388
           Summary: Major perf loss with blocking ReadFile() & OVERLAPPED
           Product: Wine
           Version: 9.20
          Hardware: x86-64
                OS: Linux
            Status: UNCONFIRMED
          Severity: minor
          Priority: P2
         Component: kernel32
          Assignee: [e-mail]
          Reporter: [e-mail]
      Distribution: ---

When using usual blocking file IO & File pointer - based reads (Passing NULL to
ReadFile(lpOverlapped)), the performance is relatively similar to native
`read()` syscall on the host.
However, when passing an OVERLAPPED structure to ReadFile() on a file opened in
blocking mode (For pread-like behavior), the performance drops significantly
compared to a native pread() syscall on the host. It would be a good thing to
fix this, as positioned IO is pretty useful for a lot of apps. Some performance
measures:

Native Linux pread() test
512b reads: 1239998.666753 IOPS, 605.468099 MiB/s
4k reads: 1246816.786350 IOPS, 4870.378072 MiB/s
1M reads: 20914.414868 IOPS, 20914.414868 MiB/s

Wine 9.20 on the same host, blocking ReadFile() with OVERLAPPED file position
512b reads: 76785.134791 IOPS, 37.492742 MiB/s
4k reads: 74402.319471 IOPS, 290.634060 MiB/s
1M reads: 14141.221307 IOPS, 14141.221307 MiB/s

In both cases file fully resides in pagecache and no actual random IO is
performed, the mere fact that we specify a file position from OVERLAPPED
structure affects performance a lot

[Bug 57388] Major perf loss with blocking ReadFile() & OVERLAPPED

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=57388

Zeb Figura  changed:

What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |[e-mail]

--- Comment #1 from Zeb Figura  ---
Certain types of completion signalling are potentially cross-process, and need
to be managed by a central authority (the wineserver). This means RPC to the
wineserver process, which is slow.

This will happen if:

* an event is passed;

* an APC routine is passed;

* the operation might queue a packet to a completion port. (Note that only the
server knows if there is in fact a completion port associated with this file.)

The last one is presumably the culprit here, assuming you're not passing a
valid event. If no OVERLAPPED structure is provided, completion is never
queued, but if a valid OVERLAPPED structure is provided, it will be.

On the other hand, a non-overlapped file handle can't be associated with a
port, so we can do a little better, and avoid trying to queue completion in
that case. I've sent
https://gitlab.winehq.org/wine/wine/-/merge_requests/6768.

[Bug 57388] Major perf loss with blocking ReadFile() & OVERLAPPED

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=57388

--- Comment #2 from LekKit  ---
Thank you very much, it's really a huge improvement. Here are new results:

Wine 9.21 on the same host, synchronous ReadFile() with OVERLAPPED file
position

512b reads: 634344.851175 IOPS, 309.738697 MB/s
4k reads: 628872.834744 IOPS, 2456.534511 MB/s
1M reads: 18620.033840 IOPS, 18620.033840 MB/s

It's still slightly slower than a native version, but according to `strace`
it's just doing another lseek() to conform to WinAPI properly
(ReadFile()/WriteFile() seeks after buffer even with synchronous OVERLAPPED
mode). That explains why it's ~2x slower for 512b/4k reads, as the syscall
entry latency dominates the execution time (Mitigations enabled).

A pretty good result overall, improving it further would require moving the
file pointer into userspace which I'm not sure is doable for Wine at all.

Best regards

[Bug 57388] Major perf loss with blocking ReadFile() & OVERLAPPED

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=57388

LekKit  changed:

What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |RESOLVED
         Resolution|---                         |FIXED

--- Comment #3 from LekKit  ---
Closing this as FIXED.

[Bug 57388] Major perf loss with blocking ReadFile() & OVERLAPPED

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=57388

Alexandre Julliard  changed:

What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|RESOLVED                    |CLOSED

--- Comment #4 from Alexandre Julliard  ---
Closing bugs fixed in 9.22.