sched_football: Re-add crazy fans and other cleanups
cover sched_football: Re-add crazy fans and other cleanups
Just the other day (~18 years ago), I had implemented the sched_football test to validate the RT scheduling invariant that we always run the top NR_CPU priority tasks at any time.
That early version of the test included crazy fans, which disrupted the scheduling of the defensive threads. This was later removed from the test, likely due to occasional failures seen with it.
Recently, in testing the behavior of a scheduler patch I was working on I started tinkering with this test again, and found when I re-added the crazy fans, I could reproduce an actual scheduling problem with the vanilla kernel.
So I wanted to submit these changes to re-add the crazy fans along with some other cleanups and improvements to make it easier to use with tracing, so that the test can be used by others to reproduce this problem.
Feedback would be appreciated!
thanks
-john
John Stultz (6):
sched_football: Drop use of sched_yeild()
sched_football: Use atomic for ball
sched_football: Re-add the crazy fans to interrupt everyone
sched_football: Add a sleep before the game begins to get into steady
state
sched_football: Add prctrl calls to set thread comms
sched_football: Add trace_marker messages if we're tracing
.../func/sched_football/sched_football.c | 94 +++++++++++++++---- 1 file changed, 78 insertions(+), 16 deletions(-)
Signature · 5 lines
2.44.0.769.g3c40516874-goog -- Mailing list info: https://lists.linux.it/listinfo/ltp
PATCH 1/6 sched_football: Drop use of sched_yeild()
sched_yeild() just causes the SCHED_FIFO tasks to behave like SCHED_RR which makes analysis of the behavior more difficult to follow. So lets drop the calls.
Also fixes up some whitespace inconsistencies in the header comment.
--- .../realtime/func/sched_football/sched_football.c | 10 ++++------ 1 file changed, 4 insertions(+), 6 deletions(-)
testcases/realtime/func/sched_football/sched_football.c +4 -6
@@ -32,12 +32,12 @@* - Create a fixed number of offense threads (lower priority) * - Create a referee thread (highest priority) * - Once everyone is on the field, the offense thread increments the -* value of 'the_ball' and yields. The defense thread tries to block -* the ball by never letting the offense players get the CPU (it just -* does a sched_yield). +* value of 'the_ball'. The defense thread tries to block +* the ball by never letting the offense players get the CPU (it just +* spins). * - The refree threads wakes up regularly to check if the game is over :) * - In the end, if the value of 'the_ball' is >0, the test is considered -* to have failed. +* to have failed. * * USAGE: * Use run_auto.sh script in current directory to build and run test.
@@ -113,7 +113,6 @@ void *thread_defense(void *arg)atomic_inc (&players_ready);/*keep the ball from being moved */ while (1 ) { -sched_yield ();/* let other defenders run */ }return NULL ; }
@@ -124,7 +123,6 @@ void *thread_offense(void *arg)atomic_inc (&players_ready);while (1 ) { the_ball++;/* move the ball ahead one yard */ -sched_yield ();/* let other offensive players run */ }return NULL ; }
Signature · 5 lines
2.44.0.769.g3c40516874-goog -- Mailing list info: https://lists.linux.it/listinfo/ltp
PATCH 2/6 sched_football: Use atomic for ball
In reply to John Stultz via ltpUse atomic type for the ball value, as we don't have any locking going on.
--- testcases/realtime/func/sched_football/sched_football.c | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-)
testcases/realtime/func/sched_football/sched_football.c +4 -4
@@ -71,7 +71,7 @@#define DEF_GAME_LENGTH 5/* Here's the position of the ball */ -volatile int the_ball; +static atomic_t the_ball;static int players_per_team =0 ;static int game_length =DEF_GAME_LENGTH ;
@@ -122,7 +122,7 @@ void *thread_offense(void *arg) {atomic_inc (&players_ready);while (1 ) { - the_ball++;/* move the ball ahead one yard */ +atomic_inc (&the_ball);/* move the ball ahead one yard */ }return NULL ; }
@@ -138,16 +138,16 @@ int referee(int game_length) now = start;/* Start the game! */ - the_ball =0 ; +atomic_set (0 , &the_ball);/* Watch the game */ while ((now.tv_sec - start.tv_sec) < game_length) {sleep (1 );gettimeofday (&now,NULL ); } + final_ball =atomic_get (&the_ball);/* Blow the whistle */ printf ("Game Over!\n" ); - final_ball = the_ball;printf ("Final ball position: %d\n" , final_ball);return final_ball !=0 ; }
Signature · 5 lines
2.44.0.769.g3c40516874-goog -- Mailing list info: https://lists.linux.it/listinfo/ltp
Re: [LTP] [PATCH 2/6] sched_football: Use atomic for ball
Hi!
Wouldn't it make more sense to use the include/tst_atomic.h now instead of the functions from librttest.h. At least the version in the tst_atomic.h have proper memory barries for load, fallback in the case that compiler does not support __atomic_foo(), etc.
Signature · 5 lines
Cyril Hrubis [email protected] -- Mailing list info: https://lists.linux.it/listinfo/ltp
Re: [LTP] [PATCH 2/6] sched_football: Use atomic for ball
Quoted 6 lines · Cyril Hrubis
Hi!
Wouldn't it make more sense to use theinclude/tst_atomic.hnow instead of the functions fromlibrttest.h. At least the version in thetst_atomic.hhave proper memory barries for load, fallback in the case that compiler does not support__atomic_foo(), etc.
I don't object, but trying to include tst_atomic.h causes lots of redefinition warnings for PACKAGE, PACKAGE_*, and VERSION which are also set from librttest.h -> realtime_config.h
Suggestions on how to proceed there?
thanks
-john
Signature · 1 line
Mailing list info: https://lists.linux.it/listinfo/ltp
Re: [LTP] [PATCH 2/6] sched_football: Use atomic for ball
Hi!
Quoted 10 lines
Quoted 4 lines
Wouldn't it make more sense to use the
include/tst_atomic.hnow instead of the functions fromlibrttest.h. At least the version in thetst_atomic.hhave proper memory barries for load, fallback in the case that compiler does not support__atomic_foo(), etc.I don't object, but trying to include
tst_atomic.hcauses lots of redefinition warnings for PACKAGE, PACKAGE_*, and VERSION which are also set fromlibrttest.h->realtime_config.hSuggestions on how to proceed there?
The problem obviously is that we have two different config.h files that conflicts.
Given that there isn't much in the realtime/configure.ac and these checks does not seem to be relevant anymore it would make sense just to get rid of it.
What we check for:
PTHREAD_PRIO_INHERIT - likely required for LinuxThreads vs NPTL exp10 - seems to be supported in glibc since 2.1 __sync_add_and_fetch() - not needed when include/tst_atomic.h is used
Other than that the realtime/configure.ac contains stale data such as defunct email to the ltp sourceforge mailing list which should be removed anyways.
Signature · 5 lines
Cyril Hrubis [email protected] -- Mailing list info: https://lists.linux.it/listinfo/ltp
Re: [LTP] [PATCH 2/6] sched_football: Use atomic for ball
Hi!
FYI the patch that removes the autotools from the testcases/realtime has been merged, it should be possible to use the include/tst_atomic.h and other headers without conflicts now.
Signature · 5 lines
Cyril Hrubis [email protected] -- Mailing list info: https://lists.linux.it/listinfo/ltp
Re: [LTP] [PATCH 2/6] sched_football: Use atomic for ball
Quoted 5 lines · Cyril Hrubis
Hi!
FYI the patch that removes the autotools from the testcases/realtime has been merged, it should be possible to use theinclude/tst_atomic.hand other headers without conflicts now.
Hey Cyril!
I went to take a stab at this, but the master branch on
https://github.com/linux-test-project/ltp.git doesn't seem to build:
$ ./configure --with-realtime-testsuite $ make -j > /dev/null make[3]: *** No rule to make target 'Makefile.am', needed by 'Makefile.in'. Stop. make[2]: *** [../../include/mk/generic_trunk_target.inc:108: all] Error 2 make[1]: *** [../include/mk/generic_trunk_target.inc:108: all] Error 2 make: *** [Makefile:85: testcases-all] Error 2
Bisecting points to commit a05298ec4494 ("testcases: realtime: Get rid of autotools")
I'm unfortunately not savvy enough with build systems to have much of an idea of what is needed here.
Any suggestions?
thanks
-john
Signature · 1 line
Mailing list info: https://lists.linux.it/listinfo/ltp
Re: [LTP] [PATCH 2/6] sched_football: Use atomic for ball
Hi John,
Quoted 19 lines · John Stultz via ltp
Quoted 5 lines · Cyril Hrubis
Hi!
FYI the patch that removes the autotools from the testcases/realtime has been merged, it should be possible to use theinclude/tst_atomic.hand other headers without conflicts now.Hey Cyril!
I went to take a stab at this, but the master branch on
https://github.com/linux-test-project/ltp.git doesn't seem to build:$ ./configure --with-realtime-testsuite $ make -j > /dev/null make[3]: *** No rule to make target 'Makefile.am', needed by 'Makefile.in'. Stop. make[2]: *** [../../include/mk/generic_trunk_target.inc:108: all] Error 2 make[1]: *** [../include/mk/generic_trunk_target.inc:108: all] Error 2 make: *** [Makefile:85: testcases-all] Error 2
Quoted 8 lines
Bisecting points to commit
a05298ec4494("testcases: realtime: Get rid of autotools")I'm unfortunately not savvy enough with build systems to have much of an idea of what is needed here.
Any suggestions?
Looks like you were trying to configure realtime-testsuite in LTP build.
Try this:
# cd ltp dir/ # make autotools # ./configure # make -C testcases/realtime/ See: https://linux-test-project.readthedocs.io/en/latest/users/quick_start.html
Signature · 5 lines
Regards, Li Wang -- Mailing list info: https://lists.linux.it/listinfo/ltp
Re: [LTP] [PATCH 2/6] sched_football: Use atomic for ball
Quoted 16 lines · Li Wang
Quoted 15 lines · John Stultz via ltp
Hey Cyril!
I went to take a stab at this, but the master branch on
https://github.com/linux-test-project/ltp.git doesn't seem to build:$ ./configure --with-realtime-testsuite $ make -j > /dev/null make[3]: *** No rule to make target 'Makefile.am', needed by 'Makefile.in'. Stop. make[2]: *** [../../include/mk/generic_trunk_target.inc:108: all] Error 2 make[1]: *** [../include/mk/generic_trunk_target.inc:108: all] Error 2 make: *** [Makefile:85: testcases-all] Error 2Bisecting points to commit
a05298ec4494("testcases: realtime: Get rid of autotools")
Quoted 7 lines
Quoted 4 lines
I'm unfortunately not savvy enough with build systems to have much of an idea of what is needed here.
Any suggestions?
Looks like you were trying to configure realtime-testsuite in LTP build.
Yes, as previously that seemed to be required.
Quoted 6 lines
Try this:
# cd ltp dir/ # make autotools # ./configure # make -C testcases/realtime/
Thank you! That does indeed work.
Should the old --with-realtime-testsuite option be removed then?
thanks again!
-john
Signature · 1 line
Mailing list info: https://lists.linux.it/listinfo/ltp
Re: [LTP] [PATCH 2/6] sched_football: Use atomic for ball
Hi John,
Quoted 7 lines · John Stultz
Try this:
Quoted 5 lines
# cd ltp dir/ # make autotools # ./configure # make -C testcases/realtime/
Quoted 7 lines
Thank you! That does indeed work.
Quoted 1 line
See:
https://linux-test-project.readthedocs.io/en/latest/users/quick_start.html
Should the old --with-realtime-testsuite option be removed then?
No, we need it.
When configuring LTP without the --with-realtime-testsuite option, it will skip building the stuff under testcases/realtime/ by default.
Unless you build it manually like: make -C testcase/realtime/.
Signature · 5 lines
Regards, Li Wang -- Mailing list info: https://lists.linux.it/listinfo/ltp
Re: [LTP] [PATCH 2/6] sched_football: Use atomic for ball
Quoted 9 lines · Li Wang
Quoted 2 lines · John Stultz
Should the old --with-realtime-testsuite option be removed then?
No, we need it.
When configuring LTP without the --with-realtime-testsuite option, it will skip building the stuff under testcases/realtime/ by default.
Right. But as I mentioned earlier, building with it seems to be broken currently?
Quoted 1 line
Unless you build it manually like: make -C testcase/realtime/.
Should maybe the realtime tests just be included in the normal build now? Or is it still useful to keep it separate due to some of the additional dependencies it needs?
thanks
-john
Signature · 1 line
Mailing list info: https://lists.linux.it/listinfo/ltp
Re: [LTP] [PATCH 2/6] sched_football: Use atomic for ball
Hi John,
Quoted 6 lines · John Stultz
Quoted 5 lines · Li Wang
Quoted 2 lines · John Stultz
Should the old --with-realtime-testsuite option be removed then?
Quoted 9 lines
Quoted 3 lines
No, we need it.
When configuring LTP without the --with-realtime-testsuite option, it
will
Quoted 1 line
skip building the stuff under testcases/realtime/ by default.
Right. But as I mentioned earlier, building with it seems to be broken currently?
Not broken, that's because you didn't build in correct way. 'make autotools' is the first necessary step for the build.
Quoted 6 lines
Quoted 1 line
Unless you build it manually like: make -C testcase/realtime/.
Should maybe the realtime tests just be included in the normal build now? Or is it still useful to keep it separate due to some of the additional dependencies it needs?
Hmm, at least in the build.sh script, it enables in the configuration phase, if you build by that (./build.sh) it should be included.
But I'm not very sure if it's worth to enable by default. @Cryil, @Petr.
See: https://github.com/linux-test-project/ltp/blob/master/build.sh#L14
Signature · 5 lines
Regards, Li Wang -- Mailing list info: https://lists.linux.it/listinfo/ltp
Re: [LTP] [PATCH 2/6] sched_football: Use atomic for ball
Hi!
Quoted 4 lines
Hmm, at least in the
build.shscript, it enables in the configuration phase, if you build by that (./build.sh) it should be included.But I'm not very sure if it's worth to enable by default. @Cryil, @Petr.
I do not think that it actually makes a big sense to have the configure switch for something so small as the realtime testsuite. It's ~30 tests that are compiled in a few seconds and it does not add too much space to the installed LTP as well.
Signature · 5 lines
Cyril Hrubis [email protected] -- Mailing list info: https://lists.linux.it/listinfo/ltp
Re: [LTP] [PATCH 2/6] sched_football: Use atomic for ball
In reply to Li WangQuoted 12 lines · Li Wang
Quoted 11 lines · John Stultz
Quoted 9 lines · Li Wang
Quoted 2 lines · John Stultz
Should the old --with-realtime-testsuite option be removed then?
No, we need it.
When configuring LTP without the --with-realtime-testsuite option, it will skip building the stuff under testcases/realtime/ by default.
Quoted 7 lines
Quoted 2 lines
Right. But as I mentioned earlier, building with it seems to be broken currently?
Not broken, that's because you didn't build in correct way. 'make autotools' is the first necessary step for the build.
Ok, but even after doing that as you suggested, building with the following fails for me:
make autotools
./configure -C testcases/realtime/
make -j
...
make[3]: *** No rule to make target 'Makefile.am', needed by 'Makefile.in'. Stop.
make[2]: *** [../../include/mk/generic_trunk_target.inc:108: all] Error 2 make[1]: *** [../include/mk/generic_trunk_target.inc:108: all] Error 2
make: *** [Makefile:85: testcases-all] Error 2
So it does *seem* like something is broken. :)
thanks
-john
Signature · 1 line
Mailing list info: https://lists.linux.it/listinfo/ltp
Re: [LTP] [PATCH 2/6] sched_football: Use atomic for ball
Quoted 22 lines · John Stultz
Quoted 12 lines · Li Wang
Quoted 11 lines · John Stultz
Quoted 9 lines · Li Wang
Quoted 2 lines · John Stultz
Should the old --with-realtime-testsuite option be removed then?
No, we need it.
When configuring LTP without the --with-realtime-testsuite option, it will skip building the stuff under testcases/realtime/ by default.
Quoted 7 lines
Quoted 2 lines
Right. But as I mentioned earlier, building with it seems to be broken currently?
Not broken, that's because you didn't build in correct way. 'make autotools' is the first necessary step for the build.
Quoted 5 lines
Ok, but even after doing that as you suggested, building with the following fails for me:
make autotools
./configure -C testcases/realtime/
Oh, sorry, I cross wires typing this out. It was
./configure --with-realtime-testsuite
Quoted 12 lines
make -j
...
make[3]: *** No rule to make target 'Makefile.am', needed by 'Makefile.in'. Stop.
make[2]: *** [../../include/mk/generic_trunk_target.inc:108: all] Error 2 make[1]: *** [../include/mk/generic_trunk_target.inc:108: all] Error 2
make: *** [Makefile:85: testcases-all] Error 2So it does *seem* like something is broken. :)
thanks
-john
Signature · 1 line
Mailing list info: https://lists.linux.it/listinfo/ltp
Re: [LTP] [PATCH 2/6] sched_football: Use atomic for ball
In reply to John Stultz via ltpHi!
Quoted 13 lines
Ok, but even after doing that as you suggested, building with the following fails for me:
make autotools
./configure -C testcases/realtime/
make -j
...
make[3]: *** No rule to make target 'Makefile.am', needed by 'Makefile.in'. Stop.
make[2]: *** [../../include/mk/generic_trunk_target.inc:108: all] Error 2 make[1]: *** [../include/mk/generic_trunk_target.inc:108: all] Error 2
make: *** [Makefile:85: testcases-all] Error 2
Quoted 1 line
So it does *seem* like something is broken. :)
I can't reproduce that and it didn't show up on our CI that builds LTP on 16 different distributions.
And the error message you get does not make much sense either. The Makefile.am file is only referenced in the automake target. It shouldn't be referenced from the testcases-all target at all.
Signature · 5 lines
Cyril Hrubis [email protected] -- Mailing list info: https://lists.linux.it/listinfo/ltp
Re: [LTP] [PATCH 2/6] sched_football: Use atomic for ball
Quoted 13 lines · Cyril Hrubis
Quoted 13 lines
Ok, but even after doing that as you suggested, building with the following fails for me:
make autotools
./configure -C testcases/realtime/
make -j
...
make[3]: *** No rule to make target 'Makefile.am', needed by 'Makefile.in'. Stop.
make[2]: *** [../../include/mk/generic_trunk_target.inc:108: all] Error 2 make[1]: *** [../include/mk/generic_trunk_target.inc:108: all] Error 2
make: *** [Makefile:85: testcases-all] Error 2
Quoted 4 lines
Quoted 1 line
So it does *seem* like something is broken. :)
I can't reproduce that and it didn't show up on our CI that builds LTP on 16 different distributions.
Interesting. Thanks for checking on this! I appreciate your time in trying to reproduce it.
Quoted 3 lines
And the error message you get does not make much sense either. The Makefile.am file is only referenced in the automake target. It shouldn't be referenced from the testcases-all target at all.
So digging further, I think this has to do with some sort of state from prior to commit a05298ec4494 sticking around and causing problems in the build.
I had been using both make clean and make distclean and even git clean -f -d (as I noticed there were often git status leftover bits after the distclean), but even then I would trip the problem above.
But re-checking out a clean repo, I indeed don't hit the problem.
After distcleaning and git clean'ing both, the diffstat looks like:
* A bunch of confXXXXXX directories littering the old tree.
include/config.h.in~ | 656 testcases/kernel/syscalls/delete_module/.dummy_del_mod.o.d | 2 testcases/open_posix_testsuite/autom4te.cache/requests | 102 testcases/open_posix_testsuite/conformance/definitions/mqueue_h/null.o |binary testcases/open_posix_testsuite/functional/timers/timers/Makefile | 94 testcases/realtime/aclocal.m4 | 1165 - testcases/realtime/autom4te.cache/output.0 | 7058 ------ testcases/realtime/autom4te.cache/output.1 | 7059 ------ testcases/realtime/autom4te.cache/output.2 | 7058 ------ testcases/realtime/autom4te.cache/requests | 238 testcases/realtime/autom4te.cache/traces.0 | 1020 - testcases/realtime/autom4te.cache/traces.1 | 491 testcases/realtime/autom4te.cache/traces.2 | 491 testcases/realtime/compile | 348 testcases/realtime/config.guess | 1754 - testcases/realtime/config.log | 474 testcases/realtime/config.status | 1197 - testcases/realtime/config.sub | 1890 - testcases/realtime/configure | 7059 ------ testcases/realtime/include/realtime_config.h | 70 testcases/realtime/include/realtime_config.h.in | 69 testcases/realtime/include/stamp-h1 | 1 testcases/realtime/install-sh | 541 testcases/realtime/m4/Makefile.in | 411 testcases/realtime/missing | 215
So I'm guessing that leftover testcases/realtime/m4/Makefile.in is the issue. And I'm guessing the change where I saw things break removed the realtime dir from AUTOCONFED_SUBDIRS so it wouldn't get cleaned afterwards if it hadn't been before.
I guess in my bisection I should have done the distclean prior to switching branches, not first thing after. Or I guess I should have tried with git clean -x?
Well, apologies for the noise. I greatly appreciate the help resolving this!
thanks
-john
Signature · 1 line
Mailing list info: https://lists.linux.it/listinfo/ltp
PATCH 3/6 sched_football: Re-add the crazy fans to interrupt everyone
In reply to John Stultz via ltpAn idea that I had included long ago in this test case, but was later dropped was the "crazy fans", which added extra disruption to the scheduling behavior.
In my recent work, I've found that including them uncovers some unexpected behavior with current Linux kernels.
So re-add the high priroity tasks that wake up and run to disrupt the scheduling of defense threads. We run 2*NR_CPU "fan" threads and they spin for awhile, to force migrations.
--- .../func/sched_football/sched_football.c | 63 ++++++++++++++++--- 1 file changed, 55 insertions(+), 8 deletions(-)
testcases/realtime/func/sched_football/sched_football.c +55 -8
@@ -22,19 +22,23 @@* DESCRIPTION * This is a scheduler test that uses a football analogy. * The premise is that we want to make sure that lower priority threads -* (defensive team). The offense is trying to increment the balls position, -* while the defense is trying to block that from happening. +* don't run while we have runnable higher priority threads. +* The offense is trying to increment the balls position, while the +* defense is trying to block that from happening. * And the ref (highest priority thread) will blow the wistle if the * ball moves. Finally, we have crazy fans (higer prority) that try to * distract the defense by occasionally running onto the field. * * Steps: -* - Create a fixed number of offense threads (lower priority) +* - Create NR_CPU offense threads (lower priority) +* - Create NR_CPU defense threads (mid priority) +* - Create 2*NR_CPU fan threads (high priority) * - Create a referee thread (highest priority) -* - Once everyone is on the field, the offense thread increments the -* value of 'the_ball'. The defense thread tries to block -* the ball by never letting the offense players get the CPU (it just -* spins). +* - Once everyone is on the field, the offense thread spins incrementing +* the value of 'the_ball'. The defense thread tries to block the ball +* by never letting the offense players get the CPU (it just spins). +* The crazy fans sleep a bit, then jump the rail and run across the +* field, disrupting the players on the field. * - The refree threads wakes up regularly to check if the game is over :) * - In the end, if the value of 'the_ball' is >0, the test is considered * to have failed.
@@ -52,7 +56,7 @@* bugfixes and cleanups. -- Josh Triplett * 2009-06-23 Simplified atomic startup mechanism, avoiding thundering herd * scheduling at the beginning of the game. -- Darren Hart -* +* 2024-04-23 Re-add crazy fans! Along with minor cleanups -- John Stultz *****************************************************************************/ #include <stdio.h>
@@ -107,6 +111,38 @@ int parse_args(int c, char *v)return handled; } +#define NSEC_PER_SEC 1000000000ULL +unsigned long long ts_delta (struct timespec *start,struct timespec *stop) +{ +unsigned long long a, b; + + a = start->tv_sec *NSEC_PER_SEC + start->tv_nsec; + b = stop->tv_sec *NSEC_PER_SEC + stop->tv_nsec; +return b - a; +} + +#define SPIN_TIME_NS 200000000 +#define SLEEP_TIME_NS 50000000 +/* These are fans running across the field. They're trying to interrupt/distract everyone */ +void *thread_fan (void *arg) +{ +atomic_inc (&players_ready); +/*occasionally wake up and run across the field */ +while (1 ) { +int i; +struct timespec start, stop; + + start.tv_sec =0 ; + start.tv_nsec =SLEEP_TIME_NS ; +clock_nanosleep (CLOCK_MONOTONIC ,0 , &start,NULL ); +clock_gettime (CLOCK_MONOTONIC , &start); +clock_gettime (CLOCK_MONOTONIC , &stop); +while (ts_delta (&start, &stop) <SPIN_TIME_NS ) +clock_gettime (CLOCK_MONOTONIC , &stop); + } +return NULL ; +} +/* This is the defensive team. They're trying to block the offense */ void *thread_defense (void *arg) {
@@ -199,6 +235,17 @@ int main(int argc, char *argv[])while (atomic_get (&players_ready) < players_per_team *2 )usleep (100 ); +/* Start the crazy fans*/ + priority =50 ; +printf ("Starting %d fan threads at priority %d\n" , + players_per_team, priority); +for (i =0 ; i < players_per_team*2 ; i++) +create_fifo_thread (thread_fan,NULL , priority); + +/* Wait for the crazy fan threads to start */ +while (atomic_get (&players_ready) < players_per_team *4 ) +usleep (100 ); +/* Ok, everyone is on the field, bring out the ref */ printf ("Starting referee thread\n" ); result =referee (game_length);
Signature · 5 lines
2.44.0.769.g3c40516874-goog -- Mailing list info: https://lists.linux.it/listinfo/ltp
Re: [LTP] [PATCH 3/6] sched_football: Re-add the crazy fans to interrupt everyone
Hi!
Quoted 30 lines
* This is a scheduler test that uses a football analogy. * The premise is that we want to make sure that lower priority threads - * (defensive team). The offense is trying to increment the balls position, - * while the defense is trying to block that from happening. + * don't run while we have runnable higher priority threads. + * The offense is trying to increment the balls position, while the + * defense is trying to block that from happening. * And the ref (highest priority thread) will blow the wistle if the * ball moves. Finally, we have crazy fans (higer prority) that try to * distract the defense by occasionally running onto the field. * * Steps: - * - Create a fixed number of offense threads (lower priority) + * - Create NR_CPU offense threads (lower priority) + * - Create NR_CPU defense threads (mid priority) + * - Create 2*NR_CPU fan threads (high priority) * - Create a referee thread (highest priority) - * - Once everyone is on the field, the offense thread increments the - * value of 'the_ball'. The defense thread tries to block - * the ball by never letting the offense players get the CPU (it just - * spins). + * - Once everyone is on the field, the offense thread spins incrementing + * the value of 'the_ball'. The defense thread tries to block the ball + * by never letting the offense players get the CPU (it just spins). + * The crazy fans sleep a bit, then jump the rail and run across the + * field, disrupting the players on the field.* - The refree threads wakes up regularly to check if the game is over :)
* - In the end, if the value of 'the_ball' is >0, the test is considered
* to have failed.
testcases/realtime/func/sched_football/sched_football.c line 59
Quoted 5 lines
testcases/realtime/func/sched_football/sched_football.c +1
-1@@ -52,7 +56,7 @@* bugfixes and cleanups. -- Josh Triplett * 2009-06-23 Simplified atomic startup mechanism, avoiding thundering herd * scheduling at the beginning of the game. -- Darren Hart -* +* 2024-04-23 Re-add crazy fans! Along with minor cleanups -- John Stultz
We use git now, so it's kind of pointless to add more changelog lines into the file itself.
Quoted 6 lines
testcases/realtime/func/sched_football/sched_football.c +0
-0*****************************************************************************/ #include <stdio.h> @@ -107,6 +111,38 @@ int parse_args(int c, char *v)return handled; }
testcases/realtime/func/sched_football/sched_football.c int parse_args(int c, char *v) · line 122
Quoted 10 lines
testcases/realtime/func/sched_football/sched_football.c +9
-0+#define NSEC_PER_SEC 1000000000ULL +unsigned long long ts_delta (struct timespec *start,struct timespec *stop) +{ +unsigned long long a, b; + + a = start->tv_sec *NSEC_PER_SEC + start->tv_nsec; + b = stop->tv_sec *NSEC_PER_SEC + stop->tv_nsec; +return b - a; +}
This is tst_timespec_diff() from include/tst_timer.h
Quoted 49 lines
testcases/realtime/func/sched_football/sched_football.c +33
-0+#define SPIN_TIME_NS 200000000 +#define SLEEP_TIME_NS 50000000 +/* These are fans running across the field. They're trying to interrupt/distract everyone */ +void *thread_fan (void *arg) +{ +atomic_inc (&players_ready); +/*occasionally wake up and run across the field */ +while (1 ) { +int i; +struct timespec start, stop; + + start.tv_sec =0 ; + start.tv_nsec =SLEEP_TIME_NS ; +clock_nanosleep (CLOCK_MONOTONIC ,0 , &start,NULL ); +clock_gettime (CLOCK_MONOTONIC , &start); +clock_gettime (CLOCK_MONOTONIC , &stop); +while (ts_delta (&start, &stop) <SPIN_TIME_NS ) +clock_gettime (CLOCK_MONOTONIC , &stop); + } +return NULL ; +} +/* This is the defensive team. They're trying to block the offense */ void *thread_defense (void *arg) {@@ -199,6 +235,17 @@ int main(int argc, char *argv[])while (atomic_get (&players_ready) < players_per_team *2 )usleep (100 ); +/* Start the crazy fans*/ + priority =50 ; +printf ("Starting %d fan threads at priority %d\n" , + players_per_team, priority); +for (i =0 ; i < players_per_team*2 ; i++) +create_fifo_thread (thread_fan,NULL , priority); + +/* Wait for the crazy fan threads to start */ +while (atomic_get (&players_ready) < players_per_team *4 ) +usleep (100 ); +/* Ok, everyone is on the field, bring out the ref */ printf ("Starting referee thread\n" ); result =referee (game_length);Signature · 5 lines
2.44.0.769.g3c40516874-goog -- Mailing list info: https://lists.linux.it/listinfo/ltp
Signature · 5 lines
Cyril Hrubis [email protected] -- Mailing list info: https://lists.linux.it/listinfo/ltp
Re: [LTP] [PATCH 3/6] sched_football: Re-add the crazy fans to interrupt everyone
Quoted 32 lines · Cyril Hrubis
Hi!
Quoted 30 lines
* This is a scheduler test that uses a football analogy. * The premise is that we want to make sure that lower priority threads - * (defensive team). The offense is trying to increment the balls position, - * while the defense is trying to block that from happening. + * don't run while we have runnable higher priority threads. + * The offense is trying to increment the balls position, while the + * defense is trying to block that from happening. * And the ref (highest priority thread) will blow the wistle if the * ball moves. Finally, we have crazy fans (higer prority) that try to * distract the defense by occasionally running onto the field. * * Steps: - * - Create a fixed number of offense threads (lower priority) + * - Create NR_CPU offense threads (lower priority) + * - Create NR_CPU defense threads (mid priority) + * - Create 2*NR_CPU fan threads (high priority) * - Create a referee thread (highest priority) - * - Once everyone is on the field, the offense thread increments the - * value of 'the_ball'. The defense thread tries to block - * the ball by never letting the offense players get the CPU (it just - * spins). + * - Once everyone is on the field, the offense thread spins incrementing + * the value of 'the_ball'. The defense thread tries to block the ball + * by never letting the offense players get the CPU (it just spins). + * The crazy fans sleep a bit, then jump the rail and run across the + * field, disrupting the players on the field.* - The refree threads wakes up regularly to check if the game is over :)
* - In the end, if the value of 'the_ball' is >0, the test is considered
* to have failed.(quoted diff) +0
-0@@ -52,7 +56,7 @@
Quoted 8 lines
testcases/realtime/func/sched_football/sched_football.c line 59
Quoted 5 lines
testcases/realtime/func/sched_football/sched_football.c +1
-1* bugfixes and cleanups. -- Josh Triplett * 2009-06-23 Simplified atomic startup mechanism, avoiding thundering herd * scheduling at the beginning of the game. -- Darren Hart -* +* 2024-04-23 Re-add crazy fans! Along with minor cleanups -- John Stultz We use git now, so it's kind of pointless to add more changelog lines into the file itself.
Ack. Sounds good. Will drop.
Quoted 13 lines
testcases/realtime/func/sched_football/sched_football.c int parse_args(int c, char *v) · line 122
Quoted 10 lines
testcases/realtime/func/sched_football/sched_football.c +9
-0+#define NSEC_PER_SEC 1000000000ULL +unsigned long long ts_delta (struct timespec *start,struct timespec *stop) +{ +unsigned long long a, b; + + a = start->tv_sec *NSEC_PER_SEC + start->tv_nsec; + b = stop->tv_sec *NSEC_PER_SEC + stop->tv_nsec; +return b - a; +}This is
tst_timespec_diff()frominclude/tst_timer.h
Ah. I didn't even think to look, it's such a habit to hand write it in my own small tests.
Though this runs into the same issue of conflicting defines from the librttest.h, so your suggestions there would be appreciated.
I could probably utilize the ts_minus/ts_to_nsec logic as well in librttest.h as an alternative.
Thanks for the review feedback!
-john
Signature · 1 line
Mailing list info: https://lists.linux.it/listinfo/ltp
Re: [LTP] [PATCH 3/6] sched_football: Re-add the crazy fans to interrupt everyone
In reply to Cyril HrubisQuoted 5 lines · Cyril Hrubis
Hi!
Quoted 4 lines
testcases/realtime/func/sched_football/sched_football.c +0
-0@@ -107,6 +111,38 @@ int parse_args(int c, char *v)return handled; }
Quoted 11 lines
testcases/realtime/func/sched_football/sched_football.c int parse_args(int c, char *v) · line 122
Quoted 9 lines
testcases/realtime/func/sched_football/sched_football.c +9
-0+#define NSEC_PER_SEC 1000000000ULL +unsigned long long ts_delta (struct timespec *start,struct timespec *stop) +{ +unsigned long long a, b; + + a = start->tv_sec *NSEC_PER_SEC + start->tv_nsec; + b = stop->tv_sec *NSEC_PER_SEC + stop->tv_nsec; +return b - a; +}This is
tst_timespec_diff()frominclude/tst_timer.h
... which cannot be used in this test without porting it to new LTP API first. Otherwise you'll get undefined reference to tst_brk_().
Signature · 11 lines
Martin Doucha [email protected] SW Quality Engineer SUSE LINUX, s.r.o. CORSO IIa Krizikova 148/34 186 00 Prague 8 Czech Republic -- Mailing list info: https://lists.linux.it/listinfo/ltp
Re: [LTP] [PATCH 3/6] sched_football: Re-add the crazy fans to interrupt everyone
Hi!
Quoted 4 lines
Quoted 1 line
This is
tst_timespec_diff()frominclude/tst_timer.h... which cannot be used in this test without porting it to new LTP API first. Otherwise you'll get undefined reference to
tst_brk_().
Already send a patch to enable this. And I'm going to send v2 right now.
Signature · 5 lines
Cyril Hrubis [email protected] -- Mailing list info: https://lists.linux.it/listinfo/ltp
PATCH 4/6 sched_football: Add a sleep before the game begins to get into steady state
In reply to John Stultz via ltpBefore we begin the game, lets give a little extra time for any odd initial placement to get sorted before we start testing the RT scheduling invariant.
--- testcases/realtime/func/sched_football/sched_football.c | 2 ++ 1 file changed, 2 insertions(+)
testcases/realtime/func/sched_football/sched_football.c +2 -0
@@ -246,6 +246,8 @@ int main(int argc, char *argv[])while (atomic_get (&players_ready) < players_per_team *4 )usleep (100 ); +/* let things get into steady state */ +sleep (2 );/* Ok, everyone is on the field, bring out the ref */ printf ("Starting referee thread\n" ); result =referee (game_length);
Signature · 5 lines
2.44.0.769.g3c40516874-goog -- Mailing list info: https://lists.linux.it/listinfo/ltp
PATCH 5/6 sched_football: Add prctrl calls to set thread comms
In reply to John Stultz via ltpMake it easier to follow whats happening in traces by setting the thread comms to be more descriptive.
--- testcases/realtime/func/sched_football/sched_football.c | 5 +++++ 1 file changed, 5 insertions(+)
testcases/realtime/func/sched_football/sched_football.c +5 -0
@@ -69,6 +69,7 @@#include <errno.h> #include <sys/syscall.h> #include <unistd.h> +#include <sys/prctl.h> #include <sys/time.h> #include <librttest.h>
@@ -126,6 +127,7 @@ unsigned long long ts_delta(struct timespec *start, struct timespec *stop)/* These are fans running across the field. They're trying to interrupt/distract everyone */ void *thread_fan (void *arg) { +prctl (PR_SET_NAME ,"crazy_fan" ,0 ,0 ,0 );atomic_inc (&players_ready);/*occasionally wake up and run across the field */ while (1 ) {
@@ -146,6 +148,7 @@ void *thread_fan(void *arg)/* This is the defensive team. They're trying to block the offense */ void *thread_defense (void *arg) { +prctl (PR_SET_NAME ,"defense" ,0 ,0 ,0 );atomic_inc (&players_ready);/*keep the ball from being moved */ while (1 ) {
@@ -156,6 +159,7 @@ void *thread_defense(void *arg)/* This is the offensive team. They're trying to move the ball */ void *thread_offense (void *arg) { +prctl (PR_SET_NAME ,"offense" ,0 ,0 ,0 );atomic_inc (&players_ready);while (1 ) {atomic_inc (&the_ball);/* move the ball ahead one yard */
@@ -168,6 +172,7 @@ int referee(int game_length)struct timeval start, now;int final_ball; +prctl (PR_SET_NAME ,"referee" ,0 ,0 ,0 );printf ("Game On (%d seconds)!\n" , game_length);gettimeofday (&start,NULL );
Signature · 5 lines
2.44.0.769.g3c40516874-goog -- Mailing list info: https://lists.linux.it/listinfo/ltp
PATCH 6/6 sched_football: Add trace_marker messages if we're tracing
In reply to John Stultz via ltpTo further help with tracing, add trace_marker messages so we can see exactly when the game starts and ends in the tracelog.
--- .../realtime/func/sched_football/sched_football.c | 12 +++++++++++- 1 file changed, 11 insertions(+), 1 deletion(-)
testcases/realtime/func/sched_football/sched_football.c +11 -1
@@ -67,6 +67,7 @@#include <pthread.h> #include <sched.h> #include <errno.h> +#include <fcntl.h> #include <sys/syscall.h> #include <unistd.h> #include <sys/prctl.h>
@@ -167,10 +168,14 @@ void *thread_offense(void *arg)return NULL ; } +#define BUF_LEN 256int referee (int game_length) {struct timeval start, now; +char buf[BUF_LEN ];int final_ball; +int fd =open ("/sys/kernel/tracing/trace_marker" ,O_RDWR ,0 ); +int ret;prctl (PR_SET_NAME ,"referee" ,0 ,0 ,0 );printf ("Game On (%d seconds)!\n" , game_length);
@@ -178,14 +183,19 @@ int referee(int game_length)gettimeofday (&start,NULL ); now = start; +sprintf (buf,"I|%i|Game_Started!\n" ,getpid ());/* Start the game! */ atomic_set (0 , &the_ball); - +if (fd >0 ) + ret =write (fd, buf,strnlen (buf,BUF_LEN ));/* Watch the game */ while ((now.tv_sec - start.tv_sec) < game_length) {sleep (1 );gettimeofday (&now,NULL ); } +sprintf (buf,"I|%i|Game_Over!\n" ,getpid ()); +if (fd >0 ) + ret =write (fd, buf,strnlen (buf,BUF_LEN )); final_ball =atomic_get (&the_ball);/* Blow the whistle */ printf ("Game Over!\n" );
Signature · 5 lines
2.44.0.769.g3c40516874-goog -- Mailing list info: https://lists.linux.it/listinfo/ltp
Re: [LTP] [PATCH 6/6] sched_football: Add trace_marker messages if we're tracing
Hi!
Quoted 42 lines
--- .../realtime/func/sched_football/sched_football.c | 12 +++++++++++- 1 file changed, 11 insertions(+), 1 deletion(-)testcases/realtime/func/sched_football/sched_football.c +6
-1@@ -67,6 +67,7 @@#include <pthread.h> #include <sched.h> #include <errno.h> +#include <fcntl.h> #include <sys/syscall.h> #include <unistd.h> #include <sys/prctl.h> @@ -167,10 +168,14 @@ void *thread_offense(void *arg)return NULL ; } +#define BUF_LEN 256int referee (int game_length) {struct timeval start, now; +char buf[BUF_LEN ];int final_ball; +int fd =open ("/sys/kernel/tracing/trace_marker" ,O_RDWR ,0 ); +int ret;prctl (PR_SET_NAME ,"referee" ,0 ,0 ,0 );printf ("Game On (%d seconds)!\n" , game_length);@@ -178,14 +183,19 @@ int referee(int game_length)gettimeofday (&start,NULL ); now = start; +sprintf (buf,"I|%i|Game_Started!\n" ,getpid ());/* Start the game! */ atomic_set (0 , &the_ball); -
testcases/realtime/func/sched_football/sched_football.c int referee(int game_length) · line 198
Quoted 10 lines
testcases/realtime/func/sched_football/sched_football.c +5
-0+if (fd >0 ) + ret =write (fd, buf,strnlen (buf,BUF_LEN ));/* Watch the game */ while ((now.tv_sec - start.tv_sec) < game_length) {sleep (1 );gettimeofday (&now,NULL ); } +sprintf (buf,"I|%i|Game_Over!\n" ,getpid ()); +if (fd >0 ) + ret =write (fd, buf,strnlen (buf,BUF_LEN ));
I know that we havre only two usages now, but it would be a bit cleaner and readable to have a function to print the messages, e.g.
static void tracer_write(const char *msg);
And we would do:
tracer_write("Game_Started!");It would be a good idea to add the test name to the message as well, maybe as:
sprintf(buf, "sched_football|%i|%s\n", getpid(), msg);
Quoted 9 lines
testcases/realtime/func/sched_football/sched_football.c +0
-0final_ball =atomic_get (&the_ball);/* Blow the whistle */ printf ("Game Over!\n" );Signature · 5 lines
2.44.0.769.g3c40516874-goog -- Mailing list info: https://lists.linux.it/listinfo/ltp
Signature · 5 lines
Cyril Hrubis [email protected] -- Mailing list info: https://lists.linux.it/listinfo/ltp
Re: [LTP] [PATCH 6/6] sched_football: Add trace_marker messages if we're tracing
Quoted 4 lines · Cyril Hrubis
testcases/realtime/func/sched_football/sched_football.c int referee(int game_length) · line 198
Quoted 3 lines
testcases/realtime/func/sched_football/sched_football.c +3
-0+sprintf (buf,"I|%i|Game_Over!\n" ,getpid ()); +if (fd >0 ) + ret =write (fd, buf,strnlen (buf,BUF_LEN ));
Quoted 8 lines
I know that we havre only two usages now, but it would be a bit cleaner and readable to have a function to print the messages, e.g.
static void
tracer_write(const char *msg);And we would do:
tracer_write("Game_Started!");
Sure. For now I'll add it to the librttest code, but I'm happy to pull it out to more generic logic once there's a good approach to the config conflicts in the subsystem headers.
Quoted 5 lines
It would be a good idea to add the test name to the message as well, maybe as:
sprintf(buf, "sched_football|%i|%s\n", getpid(), msg);
How about sprintf(buf, "I|%i|%s: %s\n", getpid(), test_tag, msg) ?
This is mostly because I'm using the trace_marker format that perfetto understands, and can visualize.
thanks
-john
Signature · 1 line
Mailing list info: https://lists.linux.it/listinfo/ltp