[Bug nptl/11588] New: pthread condvars are not priority inheritance aware

[Bug nptl/11588] New: pthread condvars are not priority inheritance aware

[dvhltc at us dot ibm dot com]

When using a PTHREAD_PRIO_INHERIT mutex with a condvar, the pthread_cond* calls 
can still cause an unbounded priority inversion via the internal condvar lock. 
The POSIX specification doesn't provide a mechanism to specify the protocol of 
the condvar. We would like to do this at runtime, but unfortunately it is legal 
to call pthread_cond_signal() or pthread_cond_broadcast() without first waiting 
on the lock, so the mutex type may not be known the first time the condvar is 
used. A new API, pthread_condattr_setprotocol_np() and 
pthread_condattr_getprotocol_np(), would allow the user to create a 
PTHREAD_PRIO_INHERIT condvar. This would use a PTHREAD_PRIO_INHERIT mutex for 
the internal condvar lock, eliminating the potential for hitting an unbounded 
priority inversion on that lock.

This topic was initially discussed here:
http://sources.redhat.com/ml/libc-alpha/2010-01/msg00011.html

And again here:
http://sources.redhat.com/ml/libc-alpha/2010-02/msg00089.html

Test cases and patches for evaluation will be attached.

-- 
           Summary: pthread condvars are not priority inheritance aware
           Product: glibc
           Version: 2.12
            Status: NEW
          Severity: normal
          Priority: P2
         Component: nptl
        AssignedTo: drepper at redhat dot com
        ReportedBy: dvhltc at us dot ibm dot com
                CC: dvhltc at us dot ibm dot com,glibc-bugs at sources dot
                    redhat dot com


http://sourceware.org/bugzilla/show_bug.cgi?id=11588

------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

[Bug nptl/11588] pthread condvars are not priority inheritance aware

[dvhltc at us dot ibm dot com]

------- Additional Comments From dvhltc at us dot ibm dot com  2010-05-11 18:48 -------
Created an attachment (id=4779)
 --> (http://sourceware.org/bugzilla/attachment.cgi?id=4779&action=view)
Unbounded priority inversion testcase pthread_cond_hang.c

Using vanilla cond vars results in a pthread_cond_wait priority inversion
deadlock. With PI cond vars it works fine. This test uses dlsym to use the new
API if it exists.

-- 


http://sourceware.org/bugzilla/show_bug.cgi?id=11588

------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

[Bug nptl/11588] pthread condvars are not priority inheritance aware

[dvhltc at us dot ibm dot com]

------- Additional Comments From dvhltc at us dot ibm dot com  2010-05-11 18:50 -------
Created an attachment (id=4780)
 --> (http://sourceware.org/bugzilla/attachment.cgi?id=4780&action=view)
pi condvar: pthread_condattr_setprotocol_np 2.12 port


-- 


http://sourceware.org/bugzilla/show_bug.cgi?id=11588

------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

[Bug nptl/11588] pthread condvars are not priority inheritance aware

[dvhltc at us dot ibm dot com]

------- Additional Comments From dvhltc at us dot ibm dot com  2010-05-11 18:50 -------
Created an attachment (id=4781)
 --> (http://sourceware.org/bugzilla/attachment.cgi?id=4781&action=view)
pi condvar: add tst-condpi1.c basic API test


-- 


http://sourceware.org/bugzilla/show_bug.cgi?id=11588

------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

[Bug nptl/11588] pthread condvars are not priority inheritance aware

[dvhltc at us dot ibm dot com]

------- Additional Comments From dvhltc at us dot ibm dot com  2010-05-11 18:55 -------
Created an attachment (id=4782)
 --> (http://sourceware.org/bugzilla/attachment.cgi?id=4782&action=view)
pi condvar patch impact on normal path testcase

This test is designed to test the efficiency of the pthread_cond_wait() and
pthread_cond_signal() code paths within glibc. Ulrich has expressed concern
about these patches negatively impacting the non-pi condvar paths. I felt the
changes to the common case (when PI is not involved) were pretty minimal.
Certainly any significant regression there would be unacceptable.

I built three versions of glibc 2.11.1 from git:
1)	  git: unmodified git sources
2)     c_only: pthread_cond*.S files deleted
3) pi_condvar: same as c_only with the pi_condvar patches applied

Comparing #3 against #2 allows us to eliminate any gains #1 would have solely
from the hand written asm. 3 will eventually contain hand written asm, but
until the non-posix API is agreed upon, it doesn't make sense to expend the
effort of writing the asm code in my opinion.

I then ran 10 runs of 10M iterations each at SCHED_FIFO 1 priority on each of
the three glibcs, the results (following) suggest no significant change in the
non PI condvar performance, sitting right at ~270k (avg) cycles/sec for each
glibc.

build-x86_64-2.11.1-git
Cycles/Second: 279831.187500
Cycles/Second: 261911.421875
Cycles/Second: 277664.125000
Cycles/Second: 284847.718750
Cycles/Second: 285067.281250
Cycles/Second: 267918.718750
Cycles/Second: 284785.656250
Cycles/Second: 277402.843750
Cycles/Second: 202379.703125
Cycles/Second: 266421.718750

Min:  202379.703125  us
Max:  285067.28125  us
Avg:  268823.0375  us


build-x86_64-2.11.1-c_only
Cycles/Second: 277931.781250
Cycles/Second: 275614.093750
Cycles/Second: 271194.125000
Cycles/Second: 280155.093750
Cycles/Second: 284708.156250
Cycles/Second: 190936.031250
Cycles/Second: 264253.468750
Cycles/Second: 281354.281250
Cycles/Second: 290366.218750
Cycles/Second: 279962.000000

Min:  190936.03125  us
Max:  290366.21875  us
Avg:  269647.525  us


build-x86_64-2.11.1-pi_condvar
Cycles/Second: 263975.937500
Cycles/Second: 279577.281250
Cycles/Second: 276504.531250
Cycles/Second: 266163.562500
Cycles/Second: 262115.796875
Cycles/Second: 279219.406250
Cycles/Second: 265263.812500
Cycles/Second: 262226.468750
Cycles/Second: 284592.687500
Cycles/Second: 278975.875000

Min:  262115.796875  us
Max:  284592.6875  us
Avg:  271861.535938  us


This is only the cond_signal case, and doesn't account for cond_timedwait or
cond_broadcast, but I wouldn't expect those to experience any additional impact
from this patch. Are there scenarios you can think of that are likely to suffer
greater impact that are not covered by this rather simple test case?

-- 


http://sourceware.org/bugzilla/show_bug.cgi?id=11588

------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

[Bug nptl/11588] pthread condvars are not priority inheritance aware

[dino at in dot ibm dot com]

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |dino at in dot ibm dot com


http://sourceware.org/bugzilla/show_bug.cgi?id=11588

------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

[Bug nptl/11588] pthread condvars are not priority inheritance aware

[pasky at suse dot cz]

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |pasky at suse dot cz


http://sourceware.org/bugzilla/show_bug.cgi?id=11588

------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

[Bug nptl/11588] pthread condvars are not priority inheritance aware

[dvhltc at us dot ibm dot com]

------- Additional Comments From dvhltc at us dot ibm dot com  2010-05-13 01:55 -------
While working up tst-condpi2.c (make check version of pthread_cond_hang.c), I 
found something went wrong with my port attached as "0001-pi-condvar-
pthread_condattr_setprotocol_np-2.12-port.patch" and the test case still hangs 
when affined to CPU 0. I'll look into why and get an updated patch posted as soon 
as possible.

-- 


http://sourceware.org/bugzilla/show_bug.cgi?id=11588

------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

[Bug nptl/11588] pthread condvars are not priority inheritance aware

[dvhltc at us dot ibm dot com]

------- Additional Comments From dvhltc at us dot ibm dot com  2010-05-13 08:02 -------
Turns out the trouble in comment 5 was due to my forgetting to remove the 
nptl/pthread_cond*.S files before building the library this time around. The 
original 0001 patch did indeed function correctly. I have reworked the patches, 
improved the patch headers and naming as well as added a new tst-condpi2.c 
testcase which tests for priority inversion avoidance. A follow-on patch detects 
an unbounded inversion and exits gracefully. Without the last patch, the glibc 
test framework will kill the test after the 2 second timeout - which may be the 
preferred approach. I'll attach the updated patches now.

-- 


http://sourceware.org/bugzilla/show_bug.cgi?id=11588

------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

[Bug nptl/11588] pthread condvars are not priority inheritance aware

[dvhltc at us dot ibm dot com]

------- Additional Comments From dvhltc at us dot ibm dot com  2010-05-13 08:05 -------
Created an attachment (id=4785)
 --> (http://sourceware.org/bugzilla/attachment.cgi?id=4785&action=view)
Unbounded priority inversion test case

The first version (attachment 4779) was missing the "_np" in the dl lookup.
This fixes that.

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
Attachment #4779 is|0                           |1
           obsolete|                            |


http://sourceware.org/bugzilla/show_bug.cgi?id=11588

------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

[Bug nptl/11588] pthread condvars are not priority inheritance aware

[dvhltc at us dot ibm dot com]

------- Additional Comments From dvhltc at us dot ibm dot com  2010-05-13 08:07 -------
Created an attachment (id=4786)
 --> (http://sourceware.org/bugzilla/attachment.cgi?id=4786&action=view)
[PATCH 1/4 V2] pi-condvars: add protocol support to pthread_condattr_t 

Updated patch header. Functionally equivalent to attachment 4780.

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
Attachment #4780 is|0                           |1
           obsolete|                            |


http://sourceware.org/bugzilla/show_bug.cgi?id=11588

------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

[Bug nptl/11588] pthread condvars are not priority inheritance aware

[dvhltc at us dot ibm dot com]

------- Additional Comments From dvhltc at us dot ibm dot com  2010-05-13 08:08 -------
Created an attachment (id=4787)
 --> (http://sourceware.org/bugzilla/attachment.cgi?id=4787&action=view)
[PATCH 2/4 V2] pi-condvars: add tst-condpi1.c, basic API test 

Updated patch header.

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
Attachment #4781 is|0                           |1
           obsolete|                            |


http://sourceware.org/bugzilla/show_bug.cgi?id=11588

------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

[Bug nptl/11588] pthread condvars are not priority inheritance aware

[dvhltc at us dot ibm dot com]

------- Additional Comments From dvhltc at us dot ibm dot com  2010-05-13 08:10 -------
Created an attachment (id=4788)
 --> (http://sourceware.org/bugzilla/attachment.cgi?id=4788&action=view)
[PATCH 3/4 V2] pi-condvars: add tst-condpi2.c, priority inversion avoidance
test 

New "make check" test, tst-condpi2.c, to test priority inversion avoidance.

-- 


http://sourceware.org/bugzilla/show_bug.cgi?id=11588

------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

[Bug nptl/11588] pthread condvars are not priority inheritance aware

[dvhltc at us dot ibm dot com]

------- Additional Comments From dvhltc at us dot ibm dot com  2010-05-13 08:13 -------
Created an attachment (id=4789)
 --> (http://sourceware.org/bugzilla/attachment.cgi?id=4789&action=view)
[PATCH 4/4 V2] pi-condvars: tst-condpi2.c, detect priority inversion 

Optional patch. This goes beyond relying on the glibc 2 second timeout to abort
the test. Instead, it detects the unbounded priority inversion and exits
gracefully. Depending on the preferred approach, this patch can be ignored
without a problem.

-- 


http://sourceware.org/bugzilla/show_bug.cgi?id=11588

------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

[Bug nptl/11588] pthread condvars are not priority inheritance aware

[dvhltc at us dot ibm dot com]

------- Additional Comments From dvhltc at us dot ibm dot com  2010-05-13 18:44 -------
It was brought to my attention that the Signed-off-by's in the patches attached 
and sent to libc-alpha are not appropriate for glibc development. This was just 
an artifact of my git usage, and I will correct in subsequent versions. May we 
discuss these as is, as we our foremost concern is agreeing on the proposed new 
condattr protocol APIs?

-- 


http://sourceware.org/bugzilla/show_bug.cgi?id=11588

------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

[Bug nptl/11588] pthread condvars are not priority inheritance aware

[dvhltc at us dot ibm dot com]

------- Additional Comments From dvhltc at us dot ibm dot com  2010-05-27 23:18 -------
I have posted V3 of the patches to libc-alpha. Rather than bog bugzilla down with 
additional attachments, I'll wait until an agreement has been made on libc-alpha. 
These patches have proper ChangeLogs and have removed the requeue-pi C 
implementation which the previous versions were carrying along.

-- 


http://sourceware.org/bugzilla/show_bug.cgi?id=11588

------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

[Bug nptl/11588] pthread condvars are not priority inheritance aware

[dvhltc at us dot ibm dot com]

------- Additional Comments From dvhltc at us dot ibm dot com  2010-05-28 20:34 -------
Created an attachment (id=4821)
 --> (http://sourceware.org/bugzilla/attachment.cgi?id=4821&action=view)
pi condvar patch impact on normal path testcase

Updated condvar_perf testcase which runs the previous test multiple times and
calculates various statistics to aid in determining the impact of the
pi_condvar patches.

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
Attachment #4782 is|0                           |1
           obsolete|                            |


http://sourceware.org/bugzilla/show_bug.cgi?id=11588

------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

[Bug nptl/11588] pthread condvars are not priority inheritance aware

[dvhltc at us dot ibm dot com]

------- Additional Comments From dvhltc at us dot ibm dot com  2010-06-17 20:59 -------
Ulrich, before I spend more time on the test cases and possibly trying to 
optimize the current implementation, can you comment on if you have any objection 
to the proposed API changes themselves?

-- 


http://sourceware.org/bugzilla/show_bug.cgi?id=11588

------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

[Bug nptl/11588] pthread condvars are not priority inheritance aware

[tglx at linutronix dot de]

------- Additional Comments From tglx at linutronix dot de  2010-06-17 22:17 -------
(In reply to comment #15)
> Ulrich, before I spend more time on the test cases and possibly trying to 
> optimize the current implementation, can you comment on if you have any objection 
> to the proposed API changes themselves?

Is there any chance that this issue is going to make any progress ? It affects
real users and we have a working solution for it.

Are there any issues with the approach itself ? If not, what needs to be done to
get this resolved ?

Thanks,

      tglx



-- 


http://sourceware.org/bugzilla/show_bug.cgi?id=11588

------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.