Does this add a constraint in libbsd that it only works with SMP for schedulers that support thread pinning? Is that a problem at all? Last I checked, not all the SMP schedulers have pinning. Actually, I think only the EDF SMP scheduler supports pinning.
On Fri, Jun 11, 2021 at 12:46 AM Sebastian Huber <[email protected]> wrote: > > Since the epoch call callbacks may lock/unlock a mutex the callbacks must be > called from within thread context with thread dispatching enabled. Use thread > pinning to ensure that the calling thread stays on the right processor. Use > the interrupt server to ensure the thread context for the epoch drain. > --- > rtemsbsd/rtems/rtems-kernel-epoch.c | 39 ++++++++++++++++++++++++++--- > 1 file changed, 35 insertions(+), 4 deletions(-) > > diff --git a/rtemsbsd/rtems/rtems-kernel-epoch.c > b/rtemsbsd/rtems/rtems-kernel-epoch.c > index 0c070bd8..642b5854 100644 > --- a/rtemsbsd/rtems/rtems-kernel-epoch.c > +++ b/rtemsbsd/rtems/rtems-kernel-epoch.c > @@ -349,15 +349,29 @@ epoch_call(epoch_t epoch, epoch_context_t ctx, > void (*callback) (epoch_context_t)) > { > Per_CPU_Control *cpu_self; > - struct epoch_record *er; > struct epoch_pcpu *epcpu; > + struct epoch_record *er; > +#ifdef RTEMS_SMP > + ISR_Level level; > + Thread_Control *executing; > + > + _ISR_Local_disable(level); > + cpu_self = _Per_CPU_Get(); > + executing = _Per_CPU_Get_executing(cpu_self); > + _Thread_Pin(executing); > + _ISR_Local_enable(level); > +#endif > > - cpu_self = _Thread_Dispatch_disable(); > epcpu = PER_CPU_DATA_GET(cpu_self, struct epoch_pcpu, epoch); > epcpu->cb_count += 1; > er = EPOCH_GET_RECORD(cpu_self, epoch); > ck_epoch_call(&er->er_record, ctx, callback); > + > +#ifdef RTEMS_SMP > + cpu_self = _Thread_Dispatch_disable(); > + _Thread_Unpin(executing, cpu_self); > _Thread_Dispatch_enable(cpu_self); > +#endif > } > > #ifdef INVARIANTS > @@ -411,7 +425,7 @@ epoch_call_drain_cb(void *arg) > struct epoch_record *er; > > epoch = arg; > - cpu = _Per_CPU_Get(); > + cpu = _Per_CPU_Get_snapshot(); > er = EPOCH_GET_RECORD(cpu, epoch); > epoch_call(epoch, &er->er_drain_ctx, epoch_drain_cb); > } > @@ -425,6 +439,7 @@ epoch_drain_callbacks(epoch_t epoch) > uint32_t cpu_max; > rtems_id id; > rtems_status_code sc; > + rtems_interrupt_server_request req[CPU_MAXIMUM_PROCESSORS]; > #else > struct epoch_record *er; > #endif > @@ -433,6 +448,7 @@ epoch_drain_callbacks(epoch_t epoch) > mtx_lock(&epoch->e_drain_mtx); > > #ifdef RTEMS_SMP > + memset(&req, 0, sizeof(req)); > cpu_max = rtems_scheduler_get_processor_maximum(); > > for (cpu_index = 0; cpu_index <= cpu_max; ++cpu_index) { > @@ -445,8 +461,15 @@ epoch_drain_callbacks(epoch_t epoch) > for (cpu_index = 0; cpu_index <= cpu_max; ++cpu_index) { > sc = rtems_scheduler_ident_by_processor(cpu_index, &id); > if (sc == RTEMS_SUCCESSFUL) { > - _SMP_Unicast_action(cpu_index, epoch_call_drain_cb, > + sc = rtems_interrupt_server_request_initialize( > + cpu_index, &req[cpu_index], epoch_call_drain_cb, > epoch); > + if (sc == RTEMS_SUCCESSFUL) { > + rtems_interrupt_server_request_submit( > + &req[cpu_index]); > + } else { > + panic("no interrupt server for epoch drain"); > + } > } > } > #else > @@ -461,4 +484,12 @@ epoch_drain_callbacks(epoch_t epoch) > > mtx_unlock(&epoch->e_drain_mtx); > sx_xunlock(&epoch->e_drain_sx); > + > +#ifdef RTEMS_SMP > + for (cpu_index = 0; cpu_index <= cpu_max; ++cpu_index) { > + if (req[cpu_index].action.handler != NULL) { > + > rtems_interrupt_server_request_destroy(&req[cpu_index]); > + } > + } > +#endif > } > -- > 2.26.2 > > _______________________________________________ > devel mailing list > [email protected] > http://lists.rtems.org/mailman/listinfo/devel _______________________________________________ devel mailing list [email protected] http://lists.rtems.org/mailman/listinfo/devel
