[Bug 19982] New: SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

Summary: SetThreadAffinityMask() seems not to work correctly
           Product: Wine
           Version: 1.1.29
          Platform: PC
               URL: http://rh-software.com/
        OS/Version: Linux
            Status: UNCONFIRMED
          Severity: normal
          Priority: P2
         Component: kernel32
        AssignedTo: [e-mail]
        ReportedBy: [e-mail]

Created an attachment (id=23510)
 --> (http://bugs.winehq.org/attachment.cgi?id=23510)
SIV log shouing pseuso random APICs

When my SIV application (http://rh-software.com/) on under Wine the reported
processor topography is often incorrect.  This is caused by the APIC returned
by the CPUID 1 instruction not being the APIC for the CPU I have set affinity
to.  This makes me believe that SetThreadAffinityMask() is not working
correctly.  This can easily be shown using me SIV utility with the command
line:
SIV -SAVE=[processors]=stdout -nodriver | find "CPU-"
which when run under Wine typically reports:
                    MPS  MCP  HTT          APIC  G  N  Affinity       Speed
Processors  CPU-0   0    3    0    (core)  0003  0  0  0x00000001   3.00GHz
            CPU-1   0    3    0    ...     0003  0  1  0x00000002   3.00GHz
            CPU-2   0    3    0    ...     0003  0  2  0x00000004   3.00GHz
            CPU-3   0    3    0    ...     0003  0  3  0x00000008   3.00GHz
if SetThreadAffinityMask() was woring as expected each CPU would have a
different APIC giving a report similar to
Processors  CPU-0   0    0    0               0000  0  0  0x00000001   3.20GHz
            CPU-1   0    1    0    (core)     0002  0  1  0x00000002   3.20GHz
            CPU-2   0    0    1    (logical)  0001  0  2  0x00000004   3.20GHz
            CPU-3   0    1    1    ...        0003  0  3  0x00000008   3.20GHz
this has been seen on at least 2 different systems.  Further the APIC returned
for a given CPU seems to vary from run-to-run as shown in the siv_apic.txt
attachment.

Looking through the bug reports this could also be what is fundamental cause of
issue http://bugs.winehq.org/show_bug.cgi?id=11674.  If WOW is using
SetThreadAffinityMask() it could happen that all the threads end up running on
the same CPU because of SetThreadAffinityMask() not working as it should.

The -NODRIVER qualifier is needed with SIV 4.00 and earlier and will not be
needed from SIV 4.01 onwards.

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

Boris  changed:

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

--- Comment #1 from Boris   2009-09-08 14:47:20 ---
I tried this on my system, so I confirm:
$ wine siv -save=[processors]=stdout -nodriver 2> /dev/null | egrep "(CPU-|APIC
 G)"
                    MPS  MCP  HTT          APIC  G  N  Affinity       Speed
Processors  CPU-0   0    1    0    (core)  0001  0  0  0x00000001   3.17GHz
            CPU-1   0    1    0    ...     0001  0  1  0x00000002   3.17GHz

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

Vitaliy Margolen  changed:

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

--- Comment #2 from Vitaliy Margolen   2009-09-08 20:46:12 ---
Duplicate

*** This bug has been marked as a duplicate of bug 7768 ***

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

Vitaliy Margolen  changed:

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

--- Comment #3 from Vitaliy Margolen   2009-09-08 20:46:55 ---
Closing duplicate.

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

Juan Lang  changed:

What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|CLOSED                      |UNCONFIRMED
         Resolution|DUPLICATE                   |

--- Comment #4 from Juan Lang   2009-09-08 21:46:03 ---
Actually, this isn't a dup.  Bug 7768 is about process affinity, and this is
about thread affinity.  They're related, but separate.  This one's probably
easier to fix than 7768, because in Linux, the affinity API applies to a
thread, so implementing this on top of the Linux API should be relatively
straightforward.

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

Juan Lang  changed:

What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
     Ever Confirmed|0                           |1

--- Comment #5 from Juan Lang   2009-09-08 21:47:20 ---
Reconfirming per comment 1.  (Reopening it didn't give me the option to mark it
confirmed.)

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #6 from Juan Lang   2009-09-09 12:39:00 ---
Created an attachment (id=23529)
 --> (http://bugs.winehq.org/attachment.cgi?id=23529)
Patch: Use kernel support for thread affinity when available

It turns out I'd already written a patch for thread affinity, but it was
rejected along with the process one, probably because the process one needed
work.  It also doesn't compile on MacOS, but as far as I know there isn't a
similar system call there anyway.  Does it address the issue for you?

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

Juan Lang  changed:

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

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #7 from Juan Lang   2009-09-09 12:52:50 ---
(In reply to comment #6)
> It also doesn't compile on MacOS, but as far as I know there isn't a
> similar system call there anyway.

Correcting myself:  there is a MacOS API for setting thread affinity[1].  Its
programming model is rather different than in Windows/Linux, so mapping between
the two would be a challenge.  Still, I'm curious whether the attached patch
works for you (under Linux.)

[1]
http://developer.apple.com/mac/library/releasenotes/Performance/RN-AffinityA...

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #8 from red-ray   2009-09-10 05:55:46 ---
Thank you for this.  I confirm that with this patch SIV reportts the correct
processor topography.  The output from SIV is now:

$ wine siv -save=[processors]=stdout -nodriver 2> /dev/null | egrep "(CPU-|APIC
 G)"
                    MPS  MCP  HTT          APIC  G  N  Affinity       Speed
Processors  CPU-0   0    0    0            0000  0  0  0x00000001   3.17GHz
            CPU-1   0    1    0    (core)  0001  0  1  0x00000002   3.17GHz

I hope the patch is accepted this time around.  Is there any change it will
make 1.1.30 ?

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #9 from Juan Lang   2009-09-10 10:08:40 ---
(In reply to comment #8)
> I hope the patch is accepted this time around.  Is there any change it will
> make 1.1.30 ?

There's a chance.  It depends on my being able to figure out the correct
configure magic to make this code be excluded on MacOSX, which has sched.h, but
not cpu_set_t nor sched_getaffinity/sched_setaffinity.

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

Juan Lang  changed:

What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |patch

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #10 from Juan Lang   2009-09-11 11:14:42 ---
Well, my patch didn't make the latest round of commits.  Alexandre tends to be
sensitive about changes to the server, so it's not surprising.  I don't know
the reason why.  I'll ask for feedback if it doesn't go in in a day or two.

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #11 from Alexandre Julliard   2009-09-11 12:52:51 ---
I don't think you can treat thread affinity independently from process
affinity. Some test cases would be in order. Also your patch is again fetching
the affinity on every thread startup.

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #12 from red-ray   2009-09-11 14:23:48 ---
(In reply to comment #11)
> I don't think you can treat thread affinity independently from process
> affinity. Some test cases would be in order. Also your patch is again fetching
> the affinity on every thread startup.

Looking at http://msdn.microsoft.com/en-us/library/ms686247(VS.85).aspx the
only requirement from the Windows specification is the thread affinity must be
a subset of the process affinity.

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #13 from Juan Lang   2009-09-11 14:37:24 ---
(In reply to comment #12)
> Looking at http://msdn.microsoft.com/en-us/library/ms686247(VS.85).aspx the
> only requirement from the Windows specification is the thread affinity must be
> a subset of the process affinity.

My patch doesn't do that.  Like Alexandre suggested, this needs tests, and
since I'm currently without a Windows box, that's a bit of a challenge for me.

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #14 from red-ray   2009-09-15 08:08:59 ---
(In reply to comment #8)
> Thank you for this.  I confirm that with this patch SIV reportts the correct
> processor topography.  The output from SIV is now:
> $ wine siv -save=[processors]=stdout -nodriver 2> /dev/null | egrep "(CPU-|APIC
>  G)"
>                     MPS  MCP  HTT          APIC  G  N  Affinity       Speed
> Processors  CPU-0   0    0    0            0000  0  0  0x00000001   3.17GHz
>             CPU-1   0    1    0    (core)  0001  0  1  0x00000002   3.17GHz
> I hope the patch is accepted this time around.  Is there any change it will
> make 1.1.30 ?

I have done more testing using the SIV 4.01 [Test APIC] page.  The patch
improves things but when lots of SetProcessAffinityMask() calls are made the
the threads end up on the incorrect CPU.  It's possible I did not apply the
patch correctly, but I also applied
http://bugs.winehq.org/show_bug.cgi?id=19964 which did work, so I am fairly
sure it went in correctly.

To run the test download http://rh-software.com/downloads/siz.zip, run SIV.exe,
select the [Test APIC] page (Right/Click tray Icon or System
Menu->Machine->Test APIC) and then press [Start].  If that does not fail then
press [Swap CPU] and it soon will.

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #15 from red-ray   2009-09-15 09:02:58 ---
Oops, should be  http://rh-software.com/downloads/siv.zip,

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #16 from Juan Lang   2009-09-28 11:50:34 ---
I sent an updated patch series to fix this, starting with this one:
http://www.winehq.org/pipermail/wine-patches/2009-September/079014.html

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #17 from [e-mail]  2009-09-30 04:43:08 ---
Created an attachment (id=23846)
 --> (http://bugs.winehq.org/attachment.cgi?id=23846)
Test Program Binary

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #18 from [e-mail]  2009-09-30 04:43:35 ---
Created an attachment (id=23847)
 --> (http://bugs.winehq.org/attachment.cgi?id=23847)
Test Program Source

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #19 from [e-mail]  2009-09-30 04:57:08 ---
(In reply to comment #16)
> I sent an updated patch series to fix this, starting with this one:
> http://www.winehq.org/pipermail/wine-patches/2009-September/079014.html

I have applied these pataches and my ART test program shows a logic error.  If
the Process Affinity is 0x01 setting a Thread afffinity of 0x02 does not fail
and it should do.
$ wine art 
ART - CPU-1   +++ SetThreadAffinityMask(  0x00000002 ) Test Fail - Enabled 
Mask 0x00000001

Further when I tested with SIV the returned APICs on SIV startup vary from
run-to-run and the [Test APIC] page detects incorrect APICs being returned
meaning that the threads are still not running on the appropiate processor.

I did these tests on an ubunto linux 2.6.28-15 host system with a two core CPU.

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #20 from [e-mail]  2009-09-30 05:12:37 ---
Created an attachment (id=23848)
 --> (http://bugs.winehq.org/attachment.cgi?id=23848)
Screen Caprure showing the [Test APIC] test failing

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #21 from Juan Lang   2009-09-30 11:11:36 ---
(In reply to comment #19)
> I have applied these pataches and my ART test program shows a logic error.  If
> the Process Affinity is 0x01 setting a Thread afffinity of 0x02 does not fail
> and it should do.

Thanks, you're right, I missed that in my tests and implementation.
> Further when I tested with SIV the returned APICs on SIV startup vary from
> run-to-run and the [Test APIC] page detects incorrect APICs being returned
> meaning that the threads are still not running on the appropiate processor.

Alexandre committed a variant of my patches today.  Would you mind retrying
with today's git?

If it still doesn't work for you, please check whether HAVE_SCHED_SETAFFINITY
is defined in include/config.h.  If it isn't defined, configure isn't finding
the function on your system.  If it is, then the Linux call isn't doing what
it's supposed to (and it's a Linux bug rather than a Wine one.)

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

Austin English  changed:

What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |download, testcase

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #22 from Juan Lang   2009-09-30 15:00:22 ---
(In reply to comment #19)
> I have applied these pataches and my ART test program shows a logic error.  If
> the Process Affinity is 0x01 setting a Thread afffinity of 0x02 does not fail
> and it should do.

I sent a patch for this issue:
http://www.winehq.org/pipermail/wine-patches/2009-September/079205.html

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #23 from [e-mail]  2009-09-30 15:15:01 ---
(In reply to comment #21)
> Would you mind retrying with today's git?

I do not know how to actually do this.  Is there a patch file ?

When you ran SIV [Test APIC] on your wine system did it pass and how many CPUs
does your system have ?

I have just deleted and regenerated include/config.h and when I searched for
HAVE_SCHED_SETAFFINITY could not find it at all.  there is "#define
HAVE_SCHED_H 1".  I searched /usr/include/sched.h and found definitions for
sched_setaffinity() and sched_getaffinity().
I would expect that if Wine is not able to support SetThreadAffinityMask() I
would get an error 50 (ERROR_NOT_SUPPORTED) return which does not seem to be
the case.

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #24 from [e-mail]  2009-09-30 15:26:22 ---
(In reply to comment #22)
> (In reply to comment #19)
> > I have applied these pataches and my ART test program shows a logic error.  If
> > the Process Affinity is 0x01 setting a Thread afffinity of 0x02 does not fail
> > and it should do.
>
> I sent a patch for this issue:
> http://www.winehq.org/pipermail/wine-patches/2009-September/079205.html

I have just looked at this patch and reading the code it looks incorrect to me.

void set_thread_affinity( struct thread *thread, affinity_t affinity )
 {
+    if (!(affinity & thread->process->affinity))
+    {
+        set_error( STATUS_INVALID_PARAMETER );
+        return;
+    }

If the process affinity is 0x05 and I try and set a thread affinity of 0x03
this should fail and this will not be the case with the above code.

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #25 from Juan Lang   2009-09-30 15:31:18 ---
(In reply to comment #24)
> If the process affinity is 0x05 and I try and set a thread affinity of 0x03
> this should fail and this will not be the case with the above code.

Oh right, okay.  Harder for me to test as I only have a 2-core system, but I'll
think about it and have another go.

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #26 from Juan Lang   2009-09-30 15:34:09 ---
(In reply to comment #23)
> (In reply to comment #21)
> > Would you mind retrying with today's git?
>
> I do not know how to actually do this.  Is there a patch file ?

Use:
git clone git://source.winehq.org/git/wine.git

The most recent source tree will then be in the wine subdirectory of your
current directory.  See http://wiki.winehq.org/GitWine if you want to learn
more about using git with wine.

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #27 from Boris   2009-09-30 15:38:43 ---
Created an attachment (id=23856)
 --> (http://bugs.winehq.org/attachment.cgi?id=23856)
Screenshot of Test APIC window on WinXP (Test case 1)

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #28 from Boris   2009-09-30 15:39:16 ---
Created an attachment (id=23857)
 --> (http://bugs.winehq.org/attachment.cgi?id=23857)
Screenshot of Test APIC window on Wine 1.1.30 + 5 patches (Test case 2)

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #29 from Boris   2009-09-30 15:40:04 ---
Created an attachment (id=23858)
 --> (http://bugs.winehq.org/attachment.cgi?id=23858)
Screenshot of Test APIC window on Wine 1.1.30 + patch of comment 19 (Test case
3)

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #30 from Boris   2009-09-30 15:40:20 ---
I performed some tests on a workstation with WinXP at work and on my home
computer. Here are the results.

Test cases:
1) WinXP. For reference.
2) Wine 1.1.30 with five patches applied (see comment #16).
3) Wine 1.1.30 from git (downloaded about 1.5 hours ago) + patch from comment
#19.

$ grep HAVE_SCHED_SETAFFINITY include/config.h
#define HAVE_SCHED_SETAFFINITY 1
This command was executed just after Wine (test case #3) installed with
./tools/wineinstall

Screenshots of 'Test APIC' window of SIV in 3 test cases attached:
1) Reference.
2) Failed.
3) Failed.

'Processors' output of SIV:
1) Reference:
Processors  CPU-0   0    0    0               0000  0  0  0x00000001   2.67GHz

CPU-1   0    1    0    (core)     0002  0  1  0x00000002   2.67GHz

CPU-2   0    2    0    ...        0004  0  2  0x00000004   2.67GHz

CPU-3   0    3    0    ...        0006  0  3  0x00000008   2.67GHz

CPU-4   0    0    1    (logical)  0001  0  4  0x00000010   2.67GHz

CPU-5   0    1    1    ...        0003  0  5  0x00000020   2.67GHz

CPU-6   0    2    1    ...        0005  0  6  0x00000040   2.67GHz

CPU-7   0    3    1    ...        0007  0  7  0x00000080   2.67GHz
2) Failed:
$ wine siv -save=[processors]=stdout -nodriver 2> /dev/null | egrep "(CPU-|APIC
 G)"
                    MPS  MCP  HTT          APIC  G  N  Affinity       Speed
Processors  CPU-0   0    1    0    (core)  0001  0  0  0x00000001   3.17GHz
            CPU-1   0    1    0    ...     0001  0  1  0x00000002   3.17GHz
3) Good:
$ wine siv -save=[processors]=stdout -nodriver 2> /dev/null | egrep "(CPU-|APIC
 G)"
                    MPS  MCP  HTT          APIC  G  N  Affinity       Speed
Processors  CPU-0   0    0    0            0000  0  0  0x00000001   3.17GHz
            CPU-1   0    1    0    (core)  0001  0  1  0x00000002   3.17GHz

ART test:
1) Good.
2) N/A (forget to do it).
3) Good:
$ wine art
ART - All Tests Passed !

Hope this helps.

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #31 from Boris   2009-09-30 15:44:39 ---
My bad, sorry. Test case 3 was with patch from comment #22

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #32 from [e-mail]  2009-09-30 15:53:18 ---
(In reply to comment #22)
I have also just spotted another very obscure bug:

proc_affinity = (1 << si.dwNumberOfProcessors) - 1;

This will typically generate:

00474CBD   push        1
00474CBF   pop         eax
00474CC0   mov         ecx,esi
00474CC2   shl         eax,cl
00474CC4   dec         eax

which for the case of 32 CPUs the after the shl eax,cl EAX will still contain 1
!

If you look in http://www.intel.com/Assets/PDF/manual/253667.pdf you will find:

IA-32 Architecture Compatibility
The 8086 does not mask the shift count. However, all other IA-32 processors
(starting with the Intel 286 processor) do mask the shift count to 5 bits,
resulting in a maximum count of 31. This masking is done in all operating modes
(including the virtual-8086 mode) to reduce the maximum execution time of the
instructions.

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #33 from [e-mail]  2009-09-30 15:57:28 ---
(In reply to comment #25)
> (In reply to comment #24)
> > If the process affinity is 0x05 and I try and set a thread affinity of 0x03
> > this should fail and this will not be the case with the above code.
>
> Oh right, okay.  Harder for me to test as I only have a 2-core system, but I'll
> think about it and have another go.

Try Process Affinity 0x01 and set Thread Affinity 0x03 if you only have 2
cores.

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #34 from Juan Lang   2009-09-30 16:05:51 ---
(In reply to comment #24)
> If the process affinity is 0x05 and I try and set a thread affinity of 0x03
> this should fail and this will not be the case with the above code.

I sent an improved patch that fixes this:
http://www.winehq.org/pipermail/wine-patches/2009-September/079213.html

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #35 from Juan Lang   2009-09-30 16:27:19 ---
(In reply to comment #30)
> Screenshots of 'Test APIC' window of SIV in 3 test cases attached:
> - Reference.
> - Failed.
> - Failed.

Please explain how they failed.  I looked at the screenshots, and I have no
idea what they mean.  Remember:  assume I've never used this program before. 
(In fact, I haven't.)  Explain what I should see, and what I see instead, and
how to verify it myself.

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #36 from Boris   2009-09-30 16:51:09 ---
> Please explain how they failed

At least, green&red bars in test cases 2 and 3 should have spaces just as green
bars. My english is bad, so I attached an image

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #37 from Boris   2009-09-30 16:51:32 ---
Created an attachment (id=23860)
 --> (http://bugs.winehq.org/attachment.cgi?id=23860)
Why tests 2 and 3 failed

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #38 from Boris   2009-09-30 17:00:27 ---
Steps to reproduce:
1) Download SIV: http://rh-software.com/downloads/siv_beta.zip
2) Extract it at any location
3) Run it with 'wine siv'
4) Right click on tray icon \ Machine \ CPU Usage
5) Press 'Test APIC' button at right of buttons row
6) Press 'Start' button
7) Now play with checkboxes at bottom of the window. Each pair of checkboxes
should affect corresponding CPU graph only

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #39 from Juan Lang   2009-09-30 17:04:27 ---
Okay, that helps, thanks.  Ray, how is that time calculated?  Are the incorrect
values related to bug 19960?

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #40 from Boris   2009-09-30 17:11:28 ---
> I sent an improved patch that fixes this:
> http://www.winehq.org/pipermail/wine-patches/2009-September/079213.html

I tried it. No changes since 'test case 3'. But I don't know if there are
should be any.

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #41 from [e-mail]  2009-09-30 17:14:20 ---
(In reply to comment #38)
> Steps to reproduce:
> - Download SIV: http://rh-software.com/downloads/siv_beta.zip
> - Extract it at any location
> - Run it with 'wine siv'
> - Right click on tray icon \ Machine \ CPU Usage
> - Press 'Test APIC' button at right of buttons row
> - Press 'Start' button
> - Now play with checkboxes at bottom of the window. Each pair of checkboxes
>
> should affect corresponding CPU graph only

The http://rh-software.com/downloads/siv_beta.zip should not have been posted
on this public board.  Can it be removed ?

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #42 from [e-mail]  2009-09-30 17:17:07 ---
(In reply to comment #39)
> Okay, that helps, thanks.  Ray, how is that time calculated?  Are the incorrect
> values related to bug 19960?

This problem is nothing to do with CPU times.  SIV 4.01 and later work around
that bug anyway.  The problem is that the threads are not being run on the CPU
they set affinity to:

What the [Test APIC] page is doing is checking
that the APIC of the CPU it had set affinity to is as expected.  The actual
code is a loop that does:

if( cib->dkc )                              // CPU Switching Mode ?
    {
      if( ( cpo += 1 ) == smb->cpu + smb->log_cpus )
      {
        cpo = smb->cpu;                         // cycle round all CPUs
      }

cpu_ssa_set( cth, cpo->ssa );             // switch to another
      cpu_ssa_set( cth, cpu->ssa );             // and back to my CPU
    }
    else                                        // no, so wait a while
    {
      ULONG volatile dly;

for( dly = 0; dly < 3000; dly++ ) {}      // so rate is similar
    }

if( cpu->apic == cpu_get_apic( smb ) )      // APIC correct ?
    {
      InterlockedIncrement( &cib->gar );        // Good APIC Rate
    }
    else
    {
      InterlockedIncrement( &cib->bac );        // Bad  APIC Count
    }

cpu_ssa_set() is a wrapper for SetThreadAffinityMask()
cpu_get_apic() reads the APIC of the curent CPU

As soon as the returned APIC is not as expected then FAIL is reported.

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #43 from [e-mail]  2009-09-30 17:31:07 ---
(In reply to comment #40)
> > I sent an improved patch that fixes this:
> > http://www.winehq.org/pipermail/wine-patches/2009-September/079213.html
>
> I tried it. No changes since 'test case 3'. But I don't know if there are
> should be any.

The patch is to detect programs trying to set in invalid thread affinity mask. 
It should have no effect of the operation of SIV as SIV does not do this.

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #44 from Juan Lang   2009-09-30 17:40:51 ---
That sounds like either the Linux kernel isn't scheduling on the thread you're
asking it to, or that the APIC-based CPU identification isn't doing what you
want.

I compiled and ran the test program here:
http://lkml.org/lkml/2007/5/10/578
and it reported 0, which indicates that after calling sched_setaffinity,
sched_getcpu reports the same CPU as sched_setaffinity requested.  Thus, either
sched_getcpu lying, or sched_setaffinity is doing what I want.  If it's lying,
it's a Linux kernel bug.  Otherwise, the CPU identification is flawed under
Wine, and likely a Wine bug.

In either case, _this_ bug will be fixed once my latest patch is committed,
except on >= 32 CPU systems.. but I'm willing to leave that one alone until
someone actually needs support for those.

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #45 from [e-mail]  2009-09-30 17:57:04 ---
(In reply to comment #44)
> That sounds like either the Linux kernel isn't scheduling on the thread you're
> asking it to, or that the APIC-based CPU identification isn't doing what you
> want.

The SIV APIC-based CPU identification works on real Windows systems as shown by
http://bugs.winehq.org/attachment.cgi?id=23856 but not on Wine.  What do I need
to do to prove beyond any reasonable doubt that this is a Wine/*nix issue ? 
Have you run SIV on your system ?  Does the failure get shown on your system
please ?

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #46 from Juan Lang   2009-09-30 18:04:27 ---
(In reply to comment #45)
> The SIV APIC-based CPU identification works on real Windows systems as shown by
> http://bugs.winehq.org/attachment.cgi?id=23856 but not on Wine.  What do I need
> to do to prove beyond any reasonable doubt that this is a Wine/*nix issue ?

You already have.  My point is, this is either a Linux kernel bug--we're making
the correct call to schedule the thread on a different CPU, and the kernel
isn't honoring that request--or the APIC identification is failing under Wine. 
If the former is true, you'd need to take this up with the Linux folks.  If the
latter is true, it's a different bug.

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #47 from [e-mail]  2009-09-30 18:25:53 ---
(In reply to comment #46)
> the APIC identification is failing under Wine.

I am unclear as to what you mean by this.  The APIC is returned  in bits
31...24 of EBX after issuing a CPUID 1 instruction.  Wine/*nix can not affect
the operation of the CPUID instruction, it's just like any other unprivileged
x86 instruction.

Have you run SIV on your system ?  Does the failure get shown on your system
please ?  I would like to get an idea if this is just a ubuntu Linux issue.  To
actually know it works at all with Wine would be a valuable input.

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #48 from Juan Lang   2009-09-30 18:32:39 ---
(In reply to comment #47)
> > the APIC identification is failing under Wine.
>
> I am unclear as to what you mean by this.  The APIC is returned  in bits
> 31...24 of EBX after issuing a CPUID 1 instruction.  Wine/*nix can not affect
> the operation of the CPUID instruction, it's just like any other unprivileged
> x86 instruction.

Okay.  I didn't remember that that's how you were doing it.
> Have you run SIV on your system ?  Does the failure get shown on your system
> please ?  I would like to get an idea if this is just a ubuntu Linux issue.  To
> actually know it works at all with Wine would be a valuable input.

Yes, and it fails in the same way here, if I understand the program well
enough.  (I confess to finding SIV pretty confusing.)  Nevertheless, the
sched_setaffinity calls succeed.  (I instrumented my own wineserver to verify
it.)  So I'm leaning toward this being a kernel bug.  Once we ask the kernel to
schedule a thread on a different CPU, and the kernel says, "righto," there's
nothing more we can do, is there?

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #49 from [e-mail]  2009-09-30 19:40:33 ---
(In reply to comment #48)
> Yes, and it fails in the same way here, if I understand the program well
> enough.  (I confess to finding SIV pretty confusing.)  Nevertheless, the
> sched_setaffinity calls succeed.  (I instrumented my own wineserver to verify
> it.)  So I'm leaning toward this being a kernel bug.

I think having a simple program to test this would be a good move.  Looking in
the Windows API I would like to use GetCurrentProcessorNumber() which returns
the zero based CPU number.  It's new in Vista and not in Wine.  Having this
would be very helpful, what is the best way to get it added to Wine ?  Is there
a Linux API that will do this ?
> Once we ask the kernel to
> schedule a thread on a different CPU, and the kernel says, "righto," there's
> nothing more we can do, is there?

If this is a Linux bug then can it we worked around ?  I am wondering if doing
a sched_yield() or similar after the affinity has been set would trigger a
switch to the appropriate CPU.

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #50 from Juan Lang   2009-09-30 19:50:27 ---
Actually, I take it back.  I wrote a test program that runs under Linux that
calls sched_setaffinity and cpuid 1 directly, and it does what I expect:  after
calling sched_setaffinity, the process's CPU, as reported by cpuid 1, changes. 
I modifed kernel32 to call cpuid 1 also, and after changing the affinity mask,
it does not change CPUs.  So this remains a Wine bug, not a kernel bug.

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #51 from Juan Lang   2009-09-30 20:00:10 ---
Now I'm just confused.  I added traces to wineserver, and it reports that
sched_setaffinity succeeds, and that sched_getaffinity returns the same
affinity mask that I set.  So perhaps it is a problem with the thread not
getting rescheduled.

One difference between the wine environment and my test program is that in
wine, the server is a different process than the program who's requesting an
affinity change.  So maybe an affinity change, while valid, doesn't take effect
immediately?

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #52 from Juan Lang   2009-09-30 20:03:58 ---
(In reply to comment #49)
> I think having a simple program to test this would be a good move.  Looking in
> the Windows API I would like to use GetCurrentProcessorNumber() which returns
> the zero based CPU number.  It's new in Vista and not in Wine.  Having this
> would be very helpful, what is the best way to get it added to Wine ?  Is there
> a Linux API that will do this ?

Implementing it, and sending a patch to wine-patches, is the best way to get it
added.  You could call sched_getcpu() if it's available.
> If this is a Linux bug then can it we worked around ?  I am wondering if doing
> a sched_yield() or similar after the affinity has been set would trigger a
> switch to the appropriate CPU.

Thanks for the suggestion.  I tried that, but no such luck.

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #53 from [e-mail]  2009-10-01 13:56:02 ---
Created an attachment (id=23869)
 --> (http://bugs.winehq.org/attachment.cgi?id=23869)
Test Version of SIV and Test Result files

I have extended SIV to have a have a command line test option to check the APIC
is as expected.  To run this new test the command is:
   wine siv -save=[wine-apic]=stdout
Within this .ZIP file apic-real.txt shows the output from a real windows
system, apic-wine.txt showing the results from my wine system which was built
from today's GIT and siv.exe is the test program

The general test loop is run 128K times and it uses SetThreadAffinityMask() to
change to a different CPU, read the CPU APIC and check it is as expected. The
1st time uses a work thread and the actual thread handle, the 2nd test also
uses a work thread but the Pseudo thread handle (0xFFFFFFFE) and the final test
uses the main thread.

When a work thread is used the test currently fails, but passes when the main
thread is used.

I hope this is of help and simple enough to use.  I feel once this is passing
then it would still be a good idea to run the SIV [Test APIC] test as this uses
a work thread per CPU and may highlight issues not shown by the [wine-apic]
test.

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

Juan Lang  changed:

What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|patch                       |

--- Comment #54 from Juan Lang   2009-10-03 11:50:45 ---
I wonder whether wine should be using pthread_setaffinity_np() instead of
sched_setaffinity().  Also, it's possible that I should be setting the affinity
directly in ntdll, rather than in wineserver.  I'll have another go when I have
a chance.

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

Juan Lang  changed:

What    |Removed                     |Added
----------------------------------------------------------------------------
  Attachment #23529|0                           |1
        is obsolete|                            |

--- Comment #55 from Juan Lang   2009-10-03 11:51:24 ---
(From update of attachment 23529)
This was already committed, though it doesn't fix the bug yet.

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #56 from [e-mail]  2009-10-04 07:30:01 ---
(In reply to comment #54)
> I wonder whether wine should be using pthread_setaffinity_np() instead of
> sched_setaffinity().

Looking at
http://www.kernel.org/doc/man-pages/online/pages/man2/sched_setaffinity.2.ht...
it says:

If you are using the POSIX threads API, then use pthread_setaffinity_np(3)
instead of sched_setaffinity().

Does Wine use the "POSIX threads API" ?

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #57 from [e-mail]  2009-10-04 08:07:41 ---
(In reply to comment #55)
> (From update of attachment 23529 [details])
> This was already committed, though it doesn't fix the bug yet.

I think currently it fixes SetThreadAffinityMask() for the Main Thread, but not
for any additional threads that are created with Windows API call
CreateThread().

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #58 from Juan Lang   2009-10-04 11:49:46 ---
(In reply to comment #56)
> Does Wine use the "POSIX threads API" ?

Yes.

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #59 from Juan Lang   2009-10-08 17:12:00 ---
I tried fixing this by using the pthread_setaffinity_np function instead, but
didn't succeed.  There were two issues:

1. It can only be done by the calling process, not by the wineserver, because a
pthread_t only has meaning within the current process, not in a separate
process.  This means it can't work in all cases in which it works in Windows.
2. pthread_setaffinity_np crashed when I tried to use it.

I can attach the patches I wrote if anyone's interested, but I'm currently at a
loss with this one, and don't think I'll be putting more time into it.

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #60 from Juan Lang   2009-10-12 17:26:55 ---
Created an attachment (id=24060)
 --> (http://bugs.winehq.org/attachment.cgi?id=24060)
Patch 1/3: Revert 474b2e4bfb46cf1bb8db0faa6ecfdb3fd157b117

I'm attaching my last attempt at fixing this in case someone else feels like
taking it up.

This first patch reverts using sched_setaffinity in wineserver.

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #61 from Juan Lang   2009-10-12 17:27:31 ---
Created an attachment (id=24061)
 --> (http://bugs.winehq.org/attachment.cgi?id=24061)
Patch 2/3: Use pthread id as thread id on Linux

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #62 from Juan Lang   2009-10-12 17:30:45 ---
Created an attachment (id=24062)
 --> (http://bugs.winehq.org/attachment.cgi?id=24062)
Patch 3/3: Use pthread_setaffinity_np to set thread affinity when available

This third patch adds the tid to the reply to the set_thread_info server
request.  The tid was set to the pthread id in patch 2 of the series.  Using
it, a process should be able to set the affinity of a thread within the same
process as it to the requested affinity.  Unfortunately pthread_setaffinity_np
crashed for me when I tried to use it, and I don't know the reason.

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #63 from Boris   2009-10-14 13:20:06 ---
Thank you for your work, Juan. Should I test your patches on my system? If
'yes', what should I do/run to perform the test, and what feedback do you need?

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #64 from Juan Lang   2009-10-14 13:29:57 ---
(In reply to comment #63)
> Should I test your patches on my system?

No, don't bother.  They don't work for me.  I posted them in case they're of
use to someone else who wants to take this on.  Thanks for the help testing my
patches.

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #65 from [e-mail]  2009-11-23 11:55:05 ---
(In reply to comment #64)
Hi Juan,  It looks like the patces for
http://bugs.winehq.org/show_bug.cgi?id=20602 have also fixed this bug.

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

Louis Lenders  changed:

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

--- Comment #66 from Louis Lenders   2009-11-23 12:02:26 ---
(In reply to comment #65)
> (In reply to comment #64)
> Hi Juan,  It looks like the patces for
> http://bugs.winehq.org/show_bug.cgi?id=20602 have also fixed this bug.

So can we clode this one , Ray?

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

Dmitry Timoshkov  changed:

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

--- Comment #67 from Dmitry Timoshkov   2009-11-24 02:14:12 ---
Reported fixed.

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

Alexandre Julliard  changed:

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

--- Comment #68 from Alexandre Julliard   2009-12-04 12:16:19 ---
Closing bugs fixed in 1.1.34.

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

[e-mail] changed:

What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|CLOSED                      |REOPENED
         Resolution|FIXED                       |

--- Comment #69 from [e-mail]  2010-07-04 16:28:14 ---
This was working in 1.2-rc5 and regressed in 1.2-rc6.

This was logged as http://bugs.winehq.org/show_bug.cgi?id=23509

I can't confirm, so I have re-opened my original bug report.

In addition to inforrect APICs being reported the Menu->Machine->Teat APIC page
is also not reporting failure, especially when [Swap CPU] mode is selected.

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

--- Comment #70 from [e-mail]  2010-07-04 16:41:57 ---
(In reply to comment #69)
In addition to inforrect APICs being reported the Menu->Machine->Test APIC page
is also now reporting failure, especially when [Swap CPU] mode is selected.

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

Alexandre Julliard  changed:

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

--- Comment #71 from Alexandre Julliard   2010-07-04 17:09:11 ---
This was fixed, and there's no reason to believe that the patch no longer
works. Please don't reopen. New problems should go into new bugs.

[Bug 19982] SetThreadAffinityMask() seems not to work correctly

http://bugs.winehq.org/show_bug.cgi?id=19982

Alexandre Julliard  changed:

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

--- Comment #72 from Alexandre Julliard   2010-07-04 17:09:28 ---
Closing.