That certainly seems to be the upshot for now from yesterday's discussions. Whether this is something that will persist longer-term remains to be seen.
The net is that if you want to pause there a certain amount of choreography that the application needs to do to perform such role-switching in a graceful manner. A good reason, perhaps, to consider designing the application so that pauses are not needed. We still need to discuss how to deal with potentially cached work if we want to be able to support a more robust programming model in which threads/processes can fail without killing the entire application. On Thu, Jan 8, 2015 at 6:27 AM, Ciprian Barbu <[email protected]> wrote: > On Wed, Jan 7, 2015 at 9:41 PM, Mike Holmes <[email protected]> > wrote: > > I am unsure if I need to pay attention to this for 0.7.0 > > Still abit unclear after the discussion we had with Petri, but I think > we need to keep the behavior as it is, meaning applications need to > take care of the scheduling "cache", and consume everything after > issuing an odp_schedule_pause call. This would also mean my test case > behaves as expected. > > > > > On 7 January 2015 at 04:39, Ciprian Barbu <[email protected]> > wrote: > >> > >> On Tue, Jan 6, 2015 at 4:03 PM, Bill Fischofer > >> <[email protected]> wrote: > >> > I think it's something we need to discuss during the sync call. > >> > > >> > On Tue, Jan 6, 2015 at 7:48 AM, Mike Holmes <[email protected]> > >> > wrote: > >> >> > >> >> Should a bug be made to track a needed change or is it important for > >> >> 1.0 > >> >> and needs to be in the delta doc ? > >> >> > >> >> On 6 January 2015 at 08:40, Bill Fischofer < > [email protected]> > >> >> wrote: > >> >>> > >> >>> Caches should be transparent. While this may be needed here, it's a > >> >>> poor > >> >>> set of semantics to expose as part of the formal APIs. This is > >> >>> definitely > >> >>> something we need to address. My suggestion is that a > >> >>> odp_schedule_pause() > >> >>> should cause an implicit cache flush if the implementation is using > a > >> >>> scheduling cache. That way any cache being used is truly > transparent > >> >>> and > >> >>> moreover there won't be unnecessary delays in event processing since > >> >>> who > >> >>> knows how long a pause may last? Clearly it won't be brief since > >> >>> otherwise > >> >>> the application would not have bothered with a pause/resume in the > >> >>> first > >> >>> place. > >> > >> Sorry, I couldn't join you in the ODP call yesterday, mind if you give > >> a brief update on what was decided? > >> > >> >>> > >> >>> On Tue, Jan 6, 2015 at 7:17 AM, Ciprian Barbu > >> >>> <[email protected]> > >> >>> wrote: > >> >>>> > >> >>>> On Wed, Dec 17, 2014 at 5:09 PM, Jerin Jacob > >> >>>> <[email protected]> wrote: > >> >>>> > On Wed, Dec 17, 2014 at 03:10:11PM +0200, Ciprian Barbu wrote: > >> >>>> >> Signed-off-by: Ciprian Barbu <[email protected]> > >> >>>> >> --- > >> >>>> >> test/validation/odp_schedule.c | 63 > >> >>>> >> ++++++++++++++++++++++++++++++++++++++---- > >> >>>> >> 1 file changed, 58 insertions(+), 5 deletions(-) > >> >>>> >> > >> >>>> >> diff --git a/test/validation/odp_schedule.c > >> >>>> >> b/test/validation/odp_schedule.c > >> >>>> >> index 31be742..bdbcf77 100644 > >> >>>> >> --- a/test/validation/odp_schedule.c > >> >>>> >> +++ b/test/validation/odp_schedule.c > >> >>>> >> @@ -11,9 +11,11 @@ > >> >>>> >> #define MSG_POOL_SIZE (4*1024*1024) > >> >>>> >> #define QUEUES_PER_PRIO 16 > >> >>>> >> #define BUF_SIZE 64 > >> >>>> >> -#define TEST_NUM_BUFS 100 > >> >>>> >> +#define NUM_BUFS 100 > >> >>>> >> #define BURST_BUF_SIZE 4 > >> >>>> >> -#define TEST_NUM_BUFS_EXCL 10000 > >> >>>> >> +#define NUM_BUFS_EXCL 10000 > >> >>>> >> +#define NUM_BUFS_PAUSE 1000 > >> >>>> >> +#define NUM_BUFS_BEFORE_PAUSE 10 > >> >>>> >> > >> >>>> >> #define GLOBALS_SHM_NAME "test_globals" > >> >>>> >> #define MSG_POOL_NAME "msg_pool" > >> >>>> >> @@ -229,7 +231,7 @@ static void > >> >>>> >> schedule_common(odp_schedule_sync_t > >> >>>> >> sync, int num_queues, > >> >>>> >> args.sync = sync; > >> >>>> >> args.num_queues = num_queues; > >> >>>> >> args.num_prio = num_prio; > >> >>>> >> - args.num_bufs = TEST_NUM_BUFS; > >> >>>> >> + args.num_bufs = NUM_BUFS; > >> >>>> >> args.num_cores = 1; > >> >>>> >> args.enable_schd_multi = enable_schd_multi; > >> >>>> >> args.enable_excl_atomic = 0; /* Not needed with a > single > >> >>>> >> core */ > >> >>>> >> @@ -261,9 +263,9 @@ static void > >> >>>> >> parallel_execute(odp_schedule_sync_t > >> >>>> >> sync, int num_queues, > >> >>>> >> thr_args->num_queues = num_queues; > >> >>>> >> thr_args->num_prio = num_prio; > >> >>>> >> if (enable_excl_atomic) > >> >>>> >> - thr_args->num_bufs = TEST_NUM_BUFS_EXCL; > >> >>>> >> + thr_args->num_bufs = NUM_BUFS_EXCL; > >> >>>> >> else > >> >>>> >> - thr_args->num_bufs = TEST_NUM_BUFS; > >> >>>> >> + thr_args->num_bufs = NUM_BUFS; > >> >>>> >> thr_args->num_cores = globals->core_count; > >> >>>> >> thr_args->enable_schd_multi = enable_schd_multi; > >> >>>> >> thr_args->enable_excl_atomic = enable_excl_atomic; > >> >>>> >> @@ -459,6 +461,56 @@ static void > >> >>>> >> test_schedule_multi_1q_mt_a_excl(void) > >> >>>> >> ENABLE_EXCL_ATOMIC); > >> >>>> >> } > >> >>>> >> > >> >>>> >> +static void test_schedule_pause_resume(void) > >> >>>> >> +{ > >> >>>> >> + odp_queue_t queue; > >> >>>> >> + odp_buffer_t buf; > >> >>>> >> + odp_queue_t from; > >> >>>> >> + int i; > >> >>>> >> + int local_bufs = 0; > >> >>>> >> + > >> >>>> >> + queue = odp_queue_lookup("sched_0_0_n"); > >> >>>> >> + CU_ASSERT(queue != ODP_QUEUE_INVALID); > >> >>>> >> + > >> >>>> >> + pool = odp_buffer_pool_lookup(MSG_POOL_NAME); > >> >>>> >> + CU_ASSERT_FATAL(pool != ODP_BUFFER_POOL_INVALID); > >> >>>> >> + > >> >>>> >> + > >> >>>> >> + for (i = 0; i < NUM_BUFS_PAUSE; i++) { > >> >>>> >> + buf = odp_buffer_alloc(pool); > >> >>>> >> + CU_ASSERT(buf != ODP_BUFFER_INVALID); > >> >>>> >> + odp_queue_enq(queue, buf); > >> >>>> >> + } > >> >>>> >> + > >> >>>> >> + for (i = 0; i < NUM_BUFS_BEFORE_PAUSE; i++) { > >> >>>> >> + buf = odp_schedule(&from, ODP_SCHED_NO_WAIT); > >> >>>> >> + CU_ASSERT(from == queue); > >> >>>> >> + odp_buffer_free(buf); > >> >>>> >> + } > >> >>>> >> + > >> >>>> >> + odp_schedule_pause(); > >> >>>> >> + > >> >>>> >> + while (1) { > >> >>>> >> + buf = odp_schedule(&from, ODP_SCHED_NO_WAIT); > >> >>>> >> + if (buf == ODP_BUFFER_INVALID) > >> >>>> >> + break; > >> >>>> >> + > >> >>>> >> + CU_ASSERT(from == queue); > >> >>>> >> + odp_buffer_free(buf); > >> >>>> >> + local_bufs++; > >> >>>> >> + } > >> >>>> >> + > >> >>>> >> + CU_ASSERT(local_bufs < NUM_BUFS_PAUSE - > >> >>>> >> NUM_BUFS_BEFORE_PAUSE); > >> >>>> > > >> >>>> > Whats is the expected behavior here, Shouldn't it be > >> >>>> > CU_ASSERT(local_bufs == 0) ? > >> >>>> > meaning, the complete pause ? > >> >>>> > >> >>>> Sorry about the delay, I've been playing around with mutt and I > must > >> >>>> have accidentally marked this email as read. > >> >>>> The explanation here is that after pausing the scheduling, there > >> >>>> might > >> >>>> still be locally reserved buffers (see the odp_schedule_pause > >> >>>> documentation). For linux-generic for instance the scheduler > dequeues > >> >>>> buffers in bursts, odp_scheduler_pause only stops further dequeues, > >> >>>> buffers may still be in the 'reservoirs'. With that in mind, the > >> >>>> check > >> >>>> above makes sure that after pausing only a limited number of > packets > >> >>>> are still scheduled, or else said pausing seems to work, not all > >> >>>> packets being drained. > >> >>>> > >> >>>> > > >> >>>> >> + > >> >>>> >> + odp_schedule_resume(); > >> >>>> >> + > >> >>>> >> + for (i = local_bufs + NUM_BUFS_BEFORE_PAUSE; i < > >> >>>> >> NUM_BUFS_PAUSE; i++) { > >> >>>> >> + buf = odp_schedule(&from, ODP_SCHED_WAIT); > >> >>>> >> + CU_ASSERT(from == queue); > >> >>>> >> + odp_buffer_free(buf); > >> >>>> >> + } > >> >>>> >> +} > >> >>>> >> + > >> >>>> >> static int create_queues(void) > >> >>>> >> { > >> >>>> >> int i, j, prios; > >> >>>> >> @@ -594,6 +646,7 @@ struct CU_TestInfo test_odp_schedule[] = { > >> >>>> >> {"schedule_multi_mq_mt_prio_a", > >> >>>> >> test_schedule_multi_mq_mt_prio_a}, > >> >>>> >> {"schedule_multi_mq_mt_prio_o", > >> >>>> >> test_schedule_multi_mq_mt_prio_o}, > >> >>>> >> {"schedule_multi_1q_mt_a_excl", > >> >>>> >> test_schedule_multi_1q_mt_a_excl}, > >> >>>> >> + {"schedule_pause_resume", > test_schedule_pause_resume}, > >> >>>> >> CU_TEST_INFO_NULL, > >> >>>> >> }; > >> >>>> >> > >> >>>> >> -- > >> >>>> >> 1.8.3.2 > >> >>>> >> > >> >>>> >> > >> >>>> >> _______________________________________________ > >> >>>> >> lng-odp mailing list > >> >>>> >> [email protected] > >> >>>> >> http://lists.linaro.org/mailman/listinfo/lng-odp > >> >>>> > >> >>>> _______________________________________________ > >> >>>> lng-odp mailing list > >> >>>> [email protected] > >> >>>> http://lists.linaro.org/mailman/listinfo/lng-odp > >> >>> > >> >>> > >> >>> > >> >>> _______________________________________________ > >> >>> lng-odp mailing list > >> >>> [email protected] > >> >>> http://lists.linaro.org/mailman/listinfo/lng-odp > >> >>> > >> >> > >> >> > >> >> > >> >> -- > >> >> Mike Holmes > >> >> Linaro Sr Technical Manager > >> >> LNG - ODP > >> > > >> > > > > > > > > > > > -- > > Mike Holmes > > Linaro Sr Technical Manager > > LNG - ODP >
_______________________________________________ lng-odp mailing list [email protected] http://lists.linaro.org/mailman/listinfo/lng-odp
