diff options
Diffstat (limited to 'erts/emulator/beam/erl_process.c')
-rw-r--r-- | erts/emulator/beam/erl_process.c | 506 |
1 files changed, 311 insertions, 195 deletions
diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c index f8997f3c07..5ceb4ce9a8 100644 --- a/erts/emulator/beam/erl_process.c +++ b/erts/emulator/beam/erl_process.c @@ -134,15 +134,15 @@ int erts_disable_proc_not_running_opt; #ifndef DEBUG #define ERTS_SCHDLR_SSPND_CHNG_SET(VAL, OLD_VAL) \ - erts_smp_atomic32_set(&schdlr_sspnd.changing, (VAL)) + erts_smp_atomic32_set_nob(&schdlr_sspnd.changing, (VAL)) #else #define ERTS_SCHDLR_SSPND_CHNG_SET(VAL, OLD_VAL) \ do { \ erts_aint32_t old_val__; \ - old_val__ = erts_smp_atomic32_xchg(&schdlr_sspnd.changing, \ - (VAL)); \ + old_val__ = erts_smp_atomic32_xchg_nob(&schdlr_sspnd.changing, \ + (VAL)); \ ASSERT(old_val__ == (OLD_VAL)); \ } while (0) @@ -158,7 +158,7 @@ static struct { erts_smp_atomic32_t changing; erts_smp_atomic32_t active; struct { - erts_smp_atomic32_t ongoing; + int ongoing; long wait_active; ErtsProcList *procs; } msb; /* Multi Scheduling Block */ @@ -166,9 +166,8 @@ static struct { static struct { erts_smp_mtx_t update_mtx; - erts_smp_atomic32_t active_runqs; + erts_smp_atomic32_t no_runqs; int last_active_runqs; - erts_smp_atomic32_t used_runqs; int forced_check_balance; erts_smp_atomic32_t checking_balance; int halftime; @@ -411,7 +410,7 @@ erts_init_process(int ncpu) init_proclist_alloc(); - erts_smp_atomic32_init(&process_count, 0); + erts_smp_atomic32_init_nob(&process_count, 0); if (erts_use_r9_pids_ports) { proc_bits = ERTS_R9_PROC_BITS; @@ -695,8 +694,8 @@ erts_smp_schedule_misc_aux_work(int ignore_self, erts_smp_mtx_unlock(&misc_aux_work_queues[ix].data.mtx); ssi = ERTS_SCHED_SLEEP_INFO_IX(ix); - aux_work = erts_smp_atomic32_bor(&ssi->aux_work, - ERTS_SSI_AUX_WORK_MISC); + aux_work = erts_smp_atomic32_read_bor_nob(&ssi->aux_work, + ERTS_SSI_AUX_WORK_MISC); if ((aux_work & ERTS_SSI_AUX_WORK_MISC) == 0) erts_sched_poke(ssi); } @@ -712,8 +711,8 @@ erts_smp_notify_check_children_needed(void) erts_aint32_t aux_work; ErtsSchedulerSleepInfo *ssi; ssi = ERTS_SCHED_SLEEP_INFO_IX(i); - aux_work = erts_smp_atomic32_bor(&ssi->aux_work, - ERTS_SSI_AUX_WORK_CHECK_CHILDREN); + aux_work = erts_smp_atomic32_read_bor_nob(&ssi->aux_work, + ERTS_SSI_AUX_WORK_CHECK_CHILDREN); if (!(aux_work & ERTS_SSI_AUX_WORK_CHECK_CHILDREN)) erts_sched_poke(ssi); } @@ -728,15 +727,15 @@ blockable_aux_work(ErtsSchedulerData *esdp, { if (aux_work & ERTS_SSI_BLOCKABLE_AUX_WORK_MASK) { if (aux_work & ERTS_SSI_AUX_WORK_MISC) { - aux_work = erts_smp_atomic32_band(&ssi->aux_work, - ~ERTS_SSI_AUX_WORK_MISC); + aux_work = erts_smp_atomic32_read_band_nob(&ssi->aux_work, + ~ERTS_SSI_AUX_WORK_MISC); aux_work &= ~ERTS_SSI_AUX_WORK_MISC; handle_misc_aux_work(esdp); } #ifdef ERTS_SMP_SCHEDULERS_NEED_TO_CHECK_CHILDREN if (aux_work & ERTS_SSI_AUX_WORK_CHECK_CHILDREN) { - aux_work = erts_smp_atomic32_band(&ssi->aux_work, - ~ERTS_SSI_AUX_WORK_CHECK_CHILDREN); + aux_work = erts_smp_atomic32_band_nob(&ssi->aux_work, + ~ERTS_SSI_AUX_WORK_CHECK_CHILDREN); aux_work &= ~ERTS_SSI_AUX_WORK_CHECK_CHILDREN; erts_check_children(); } @@ -816,7 +815,7 @@ erts_active_schedulers(void) static ERTS_INLINE void clear_sys_scheduling(void) { - erts_smp_atomic32_set_relb(&doing_sys_schedule, 0); + erts_smp_atomic32_set_mb(&doing_sys_schedule, 0); } static ERTS_INLINE int @@ -883,42 +882,43 @@ sched_active(Uint no, ErtsRunQueue *rq) static int ERTS_INLINE ongoing_multi_scheduling_block(void) { - return erts_smp_atomic32_read(&schdlr_sspnd.msb.ongoing) != 0; + ERTS_SMP_LC_ASSERT(erts_lc_mtx_is_locked(&schdlr_sspnd.mtx)); + return schdlr_sspnd.msb.ongoing; } static ERTS_INLINE void empty_runq(ErtsRunQueue *rq) { - erts_aint32_t oifls = erts_smp_atomic32_band(&rq->info_flags, - ~ERTS_RUNQ_IFLG_NONEMPTY); + erts_aint32_t oifls = erts_smp_atomic32_read_band_nob(&rq->info_flags, + ~ERTS_RUNQ_IFLG_NONEMPTY); if (oifls & ERTS_RUNQ_IFLG_NONEMPTY) { #ifdef DEBUG - erts_aint32_t empty = erts_smp_atomic32_read(&no_empty_run_queues); + erts_aint32_t empty = erts_smp_atomic32_read_nob(&no_empty_run_queues); /* * For a short period of time no_empty_run_queues may have * been increased twice for a specific run queue. */ ASSERT(0 <= empty && empty < 2*erts_no_run_queues); #endif - erts_smp_atomic32_inc(&no_empty_run_queues); + erts_smp_atomic32_inc_relb(&no_empty_run_queues); } } static ERTS_INLINE void non_empty_runq(ErtsRunQueue *rq) { - erts_aint32_t oifls = erts_smp_atomic32_bor(&rq->info_flags, - ERTS_RUNQ_IFLG_NONEMPTY); + erts_aint32_t oifls = erts_smp_atomic32_read_bor_nob(&rq->info_flags, + ERTS_RUNQ_IFLG_NONEMPTY); if (!(oifls & ERTS_RUNQ_IFLG_NONEMPTY)) { #ifdef DEBUG - erts_aint32_t empty = erts_smp_atomic32_read(&no_empty_run_queues); + erts_aint32_t empty = erts_smp_atomic32_read_nob(&no_empty_run_queues); /* * For a short period of time no_empty_run_queues may have * been increased twice for a specific run queue. */ ASSERT(0 < empty && empty <= 2*erts_no_run_queues); #endif - erts_smp_atomic32_dec(&no_empty_run_queues); + erts_smp_atomic32_dec_relb(&no_empty_run_queues); } } @@ -931,7 +931,7 @@ sched_prep_spin_wait(ErtsSchedulerSleepInfo *ssi) erts_aint32_t xflgs = 0; do { - oflgs = erts_smp_atomic32_cmpxchg(&ssi->flags, nflgs, xflgs); + oflgs = erts_smp_atomic32_cmpxchg_acqb(&ssi->flags, nflgs, xflgs); if (oflgs == xflgs) return nflgs; xflgs = oflgs; @@ -948,7 +948,7 @@ sched_prep_cont_spin_wait(ErtsSchedulerSleepInfo *ssi) erts_aint32_t xflgs = ERTS_SSI_FLG_WAITING; do { - oflgs = erts_smp_atomic32_cmpxchg(&ssi->flags, nflgs, xflgs); + oflgs = erts_smp_atomic32_cmpxchg_acqb(&ssi->flags, nflgs, xflgs); if (oflgs == xflgs) return nflgs; xflgs = oflgs; @@ -965,7 +965,7 @@ sched_spin_wait(ErtsSchedulerSleepInfo *ssi, int spincount) erts_aint32_t flgs; do { - flgs = erts_smp_atomic32_read(&ssi->flags); + flgs = erts_smp_atomic32_read_acqb(&ssi->flags); if ((flgs & (ERTS_SSI_FLG_SLEEPING|ERTS_SSI_FLG_WAITING)) != (ERTS_SSI_FLG_SLEEPING|ERTS_SSI_FLG_WAITING)) { break; @@ -990,7 +990,7 @@ sched_set_sleeptype(ErtsSchedulerSleepInfo *ssi, erts_aint32_t sleep_type) erts_tse_reset(ssi->event); while (1) { - oflgs = erts_smp_atomic32_cmpxchg(&ssi->flags, nflgs, xflgs); + oflgs = erts_smp_atomic32_cmpxchg_acqb(&ssi->flags, nflgs, xflgs); if (oflgs == xflgs) return nflgs; if ((oflgs & (ERTS_SSI_FLG_SLEEPING|ERTS_SSI_FLG_WAITING)) @@ -1050,7 +1050,7 @@ scheduler_wait(int *fcalls, ErtsSchedulerData *esdp, ErtsRunQueue *rq) tse_wait: #ifdef ERTS_SCHED_NEED_BLOCKABLE_AUX_WORK - aux_work = erts_smp_atomic32_read(&ssi->aux_work); + aux_work = erts_smp_atomic32_read_nob(&ssi->aux_work); tse_blockable_aux_work: aux_work = blockable_aux_work(esdp, ssi, aux_work); #endif @@ -1060,7 +1060,7 @@ scheduler_wait(int *fcalls, ErtsSchedulerData *esdp, ErtsRunQueue *rq) #ifdef ERTS_SCHED_NEED_NONBLOCKABLE_AUX_WORK #ifndef ERTS_SCHED_NEED_BLOCKABLE_AUX_WORK - aux_work = erts_smp_atomic32_read(&ssi->aux_work); + aux_work = erts_smp_atomic32_read_nob(&ssi->aux_work); #endif nonblockable_aux_work(esdp, ssi, aux_work); #endif @@ -1093,7 +1093,7 @@ scheduler_wait(int *fcalls, ErtsSchedulerData *esdp, ErtsRunQueue *rq) } #ifdef ERTS_SCHED_NEED_BLOCKABLE_AUX_WORK - aux_work = erts_smp_atomic32_read(&ssi->aux_work); + aux_work = erts_smp_atomic32_read_nob(&ssi->aux_work); if (aux_work & ERTS_SSI_BLOCKABLE_AUX_WORK_MASK) { erts_smp_activity_end(ERTS_ACTIVITY_WAIT, NULL, NULL, NULL); goto tse_blockable_aux_work; @@ -1105,7 +1105,7 @@ scheduler_wait(int *fcalls, ErtsSchedulerData *esdp, ErtsRunQueue *rq) erts_smp_activity_end(ERTS_ACTIVITY_WAIT, NULL, NULL, NULL); if (flgs & ~ERTS_SSI_FLG_SUSPENDED) - erts_smp_atomic32_band(&ssi->flags, ERTS_SSI_FLG_SUSPENDED); + erts_smp_atomic32_read_band_nob(&ssi->flags, ERTS_SSI_FLG_SUSPENDED); erts_smp_runq_lock(rq); sched_active(esdp->no, rq); @@ -1114,7 +1114,7 @@ scheduler_wait(int *fcalls, ErtsSchedulerData *esdp, ErtsRunQueue *rq) else { erts_aint_t dt; - erts_smp_atomic32_set(&function_calls, 0); + erts_smp_atomic32_set_relb(&function_calls, 0); *fcalls = 0; sched_waiting_sys(esdp->no, rq); @@ -1137,28 +1137,21 @@ scheduler_wait(int *fcalls, ErtsSchedulerData *esdp, ErtsRunQueue *rq) sys_aux_work: #ifdef ERTS_SCHED_NEED_BLOCKABLE_AUX_WORK - aux_work = erts_smp_atomic32_read(&ssi->aux_work); + aux_work = erts_smp_atomic32_read_nob(&ssi->aux_work); aux_work = blockable_aux_work(esdp, ssi, aux_work); #endif #ifdef ERTS_SCHED_NEED_NONBLOCKABLE_AUX_WORK #ifndef ERTS_SCHED_NEED_BLOCKABLE_AUX_WORK - aux_work = erts_smp_atomic32_read(&ssi->aux_work); + aux_work = erts_smp_atomic32_read_nob(&ssi->aux_work); #endif nonblockable_aux_work(esdp, ssi, aux_work); #endif - flgs = erts_smp_atomic32_read(&ssi->flags); + flgs = erts_smp_atomic32_read_acqb(&ssi->flags); if (!(flgs & ERTS_SSI_FLG_WAITING)) { ASSERT(!(flgs & ERTS_SSI_FLG_SLEEPING)); goto sys_woken; } - if (!(flgs & ERTS_SSI_FLG_SLEEPING)) { - flgs = sched_prep_cont_spin_wait(ssi); - if (!(flgs & ERTS_SSI_FLG_WAITING)) { - ASSERT(!(flgs & ERTS_SSI_FLG_SLEEPING)); - goto sys_woken; - } - } /* * If we got new I/O tasks we aren't allowed to @@ -1240,7 +1233,7 @@ scheduler_wait(int *fcalls, ErtsSchedulerData *esdp, ErtsRunQueue *rq) sys_locked_woken: clear_sys_scheduling(); if (flgs & ~ERTS_SSI_FLG_SUSPENDED) - erts_smp_atomic32_band(&ssi->flags, ERTS_SSI_FLG_SUSPENDED); + erts_smp_atomic32_read_band_nob(&ssi->flags, ERTS_SSI_FLG_SUSPENDED); sched_active_sys(esdp->no, rq); } } @@ -1256,7 +1249,7 @@ ssi_flags_set_wake(ErtsSchedulerSleepInfo *ssi) erts_aint32_t nflgs = 0; erts_aint32_t xflgs = ERTS_SSI_FLG_SLEEPING|ERTS_SSI_FLG_WAITING; while (1) { - oflgs = erts_smp_atomic32_cmpxchg(&ssi->flags, nflgs, xflgs); + oflgs = erts_smp_atomic32_cmpxchg_relb(&ssi->flags, nflgs, xflgs); if (oflgs == xflgs) return oflgs; nflgs = oflgs & ERTS_SSI_FLG_SUSPENDED; @@ -1267,7 +1260,6 @@ ssi_flags_set_wake(ErtsSchedulerSleepInfo *ssi) static void wake_scheduler(ErtsRunQueue *rq, int incq, int one) { - int res; ErtsSchedulerSleepInfo *ssi; ErtsSchedulerSleepList *sl; @@ -1298,10 +1290,8 @@ wake_scheduler(ErtsRunQueue *rq, int incq, int one) if (ssi->next) ssi->next->prev = ssi->prev; - res = sl->list != NULL; erts_smp_spin_unlock(&sl->lock); - ERTS_THR_MEMORY_BARRIER; flgs = ssi_flags_set_wake(ssi); erts_sched_finish_poke(ssi, flgs); @@ -1335,6 +1325,79 @@ wake_all_schedulers(void) } } +#define ERTS_NO_USED_RUNQS_SHIFT 16 +#define ERTS_NO_RUNQS_MASK 0xffff + +#if ERTS_MAX_NO_OF_SCHEDULERS > ERTS_NO_RUNQS_MASK +# error "Too large amount of schedulers allowed" +#endif + +static ERTS_INLINE void +init_no_runqs(int active, int used) +{ + erts_aint32_t no_runqs = (erts_aint32_t) (active & ERTS_NO_RUNQS_MASK); + no_runqs |= (erts_aint32_t) ((used & ERTS_NO_RUNQS_MASK) << ERTS_NO_USED_RUNQS_SHIFT); + erts_smp_atomic32_init_nob(&balance_info.no_runqs, no_runqs); +} + +static ERTS_INLINE void +get_no_runqs(int *active, int *used) +{ + erts_aint32_t no_runqs = erts_smp_atomic32_read_nob(&balance_info.no_runqs); + if (active) + *active = (int) (no_runqs & ERTS_NO_RUNQS_MASK); + if (used) + *used = (int) ((no_runqs >> ERTS_NO_USED_RUNQS_SHIFT) & ERTS_NO_RUNQS_MASK); +} + +static ERTS_INLINE void +set_no_used_runqs(int used) +{ + erts_aint32_t exp = erts_smp_atomic32_read_nob(&balance_info.no_runqs); + while (1) { + erts_aint32_t act, new; + new = (used & ERTS_NO_RUNQS_MASK) << ERTS_NO_USED_RUNQS_SHIFT; + new |= exp & ERTS_NO_RUNQS_MASK; + act = erts_smp_atomic32_cmpxchg_nob(&balance_info.no_runqs, new, exp); + if (act == exp) + break; + exp = act; + } +} + +static ERTS_INLINE void +set_no_active_runqs(int active) +{ + erts_aint32_t exp = erts_smp_atomic32_read_nob(&balance_info.no_runqs); + while (1) { + erts_aint32_t act, new; + new = exp & (ERTS_NO_RUNQS_MASK << ERTS_NO_USED_RUNQS_SHIFT); + new |= active & ERTS_NO_RUNQS_MASK; + act = erts_smp_atomic32_cmpxchg_nob(&balance_info.no_runqs, new, exp); + if (act == exp) + break; + exp = act; + } +} + +static ERTS_INLINE int +try_inc_no_active_runqs(int active) +{ + erts_aint32_t exp = erts_smp_atomic32_read_nob(&balance_info.no_runqs); + if (((exp >> ERTS_NO_USED_RUNQS_SHIFT) & ERTS_NO_RUNQS_MASK) < active) + return 0; + if ((exp & ERTS_NO_RUNQS_MASK) + 1 == active) { + erts_aint32_t new, act; + new = exp & (ERTS_NO_RUNQS_MASK << ERTS_NO_USED_RUNQS_SHIFT); + new |= active & ERTS_NO_RUNQS_MASK; + act = erts_smp_atomic32_cmpxchg_nob(&balance_info.no_runqs, new, exp); + if (act == exp) + return 1; + } + return 0; +} + + static ERTS_INLINE int chk_wake_sched(ErtsRunQueue *crq, int ix, int activate) { @@ -1343,12 +1406,10 @@ chk_wake_sched(ErtsRunQueue *crq, int ix, int activate) if (crq->ix == ix) return 0; wrq = ERTS_RUNQ_IX(ix); - iflgs = erts_smp_atomic32_read(&wrq->info_flags); + iflgs = erts_smp_atomic32_read_nob(&wrq->info_flags); if (!(iflgs & (ERTS_RUNQ_IFLG_SUSPENDED|ERTS_RUNQ_IFLG_NONEMPTY))) { if (activate) { - if (ix == erts_smp_atomic32_cmpxchg(&balance_info.active_runqs, - ix+1, - ix)) { + if (try_inc_no_active_runqs(ix+1)) { erts_smp_xrunq_lock(crq, wrq); wrq->flags &= ~ERTS_RUNQ_FLG_INACTIVE; erts_smp_xrunq_unlock(crq, wrq); @@ -1365,8 +1426,9 @@ wake_scheduler_on_empty_runq(ErtsRunQueue *crq) { int ix = crq->ix; int stop_ix = ix; - int active_ix = erts_smp_atomic32_read(&balance_info.active_runqs); - int balance_ix = erts_smp_atomic32_read(&balance_info.used_runqs); + int active_ix, balance_ix; + + get_no_runqs(&active_ix, &balance_ix); if (active_ix > balance_ix) active_ix = balance_ix; @@ -1418,7 +1480,7 @@ erts_sched_notify_check_cpu_bind(void) int ix; if (erts_common_run_queue) { for (ix = 0; ix < erts_no_schedulers; ix++) - erts_smp_atomic32_set(&ERTS_SCHEDULER_IX(ix)->chk_cpu_bind, 1); + erts_smp_atomic32_set_relb(&ERTS_SCHEDULER_IX(ix)->chk_cpu_bind, 1); wake_all_schedulers(); } else { @@ -1586,15 +1648,15 @@ evacuate_run_queue(ErtsRunQueue *evac_rq, ErtsRunQueue *rq) erts_smp_runq_lock(evac_rq); - erts_smp_atomic32_bor(&evac_rq->scheduler->ssi->flags, - ERTS_SSI_FLG_SUSPENDED); + erts_smp_atomic32_read_bor_nob(&evac_rq->scheduler->ssi->flags, + ERTS_SSI_FLG_SUSPENDED); evac_rq->flags &= ~ERTS_RUNQ_FLGS_IMMIGRATE_QMASK; evac_rq->flags |= (ERTS_RUNQ_FLGS_EMIGRATE_QMASK | ERTS_RUNQ_FLGS_EVACUATE_QMASK | ERTS_RUNQ_FLG_SUSPENDED); - erts_smp_atomic32_bor(&evac_rq->info_flags, ERTS_RUNQ_IFLG_SUSPENDED); + erts_smp_atomic32_read_bor_nob(&evac_rq->info_flags, ERTS_RUNQ_IFLG_SUSPENDED); /* * Need to set up evacuation paths first since we * may release the run queue lock on evac_rq @@ -1843,7 +1905,7 @@ static ERTS_INLINE int check_possible_steal_victim(ErtsRunQueue *rq, int *rq_lockedp, int vix) { ErtsRunQueue *vrq = ERTS_RUNQ_IX(vix); - erts_aint32_t iflgs = erts_smp_atomic32_read(&vrq->info_flags); + erts_aint32_t iflgs = erts_smp_atomic32_read_nob(&vrq->info_flags); if (iflgs & ERTS_RUNQ_IFLG_NONEMPTY) return try_steal_task_from_victim(rq, rq_lockedp, vrq); else @@ -1873,8 +1935,7 @@ try_steal_task(ErtsRunQueue *rq) ERTS_SMP_LC_CHK_RUNQ_LOCK(rq, rq_locked); - active_rqs = erts_smp_atomic32_read(&balance_info.active_runqs); - blnc_rqs = erts_smp_atomic32_read(&balance_info.used_runqs); + get_no_runqs(&active_rqs, &blnc_rqs); if (active_rqs > blnc_rqs) active_rqs = blnc_rqs; @@ -1885,7 +1946,7 @@ try_steal_task(ErtsRunQueue *rq) if (active_rqs < blnc_rqs) { int no = blnc_rqs - active_rqs; int stop_ix = vix = active_rqs + rq->ix % no; - while (erts_smp_atomic32_read(&no_empty_run_queues) < blnc_rqs) { + while (erts_smp_atomic32_read_acqb(&no_empty_run_queues) < blnc_rqs) { res = check_possible_steal_victim(rq, &rq_locked, vix); if (res) goto done; @@ -1900,7 +1961,7 @@ try_steal_task(ErtsRunQueue *rq) vix = rq->ix; /* ... then try to steal a job from another active queue... */ - while (erts_smp_atomic32_read(&no_empty_run_queues) < blnc_rqs) { + while (erts_smp_atomic32_read_acqb(&no_empty_run_queues) < blnc_rqs) { vix++; if (vix >= active_rqs) vix = 0; @@ -1996,15 +2057,15 @@ check_balance(ErtsRunQueue *c_rq) int forced, active, current_active, oowc, half_full_scheds, full_scheds, mmax_len, blnc_no_rqs, qix, pix, freds_hist_ix; - if (erts_smp_atomic32_xchg(&balance_info.checking_balance, 1)) { + if (erts_smp_atomic32_xchg_nob(&balance_info.checking_balance, 1)) { c_rq->check_balance_reds = INT_MAX; return; } - blnc_no_rqs = (int) erts_smp_atomic32_read(&balance_info.used_runqs); + get_no_runqs(NULL, &blnc_no_rqs); if (blnc_no_rqs == 1) { c_rq->check_balance_reds = INT_MAX; - erts_smp_atomic32_set(&balance_info.checking_balance, 0); + erts_smp_atomic32_set_nob(&balance_info.checking_balance, 0); return; } @@ -2012,7 +2073,7 @@ check_balance(ErtsRunQueue *c_rq) if (balance_info.halftime) { balance_info.halftime = 0; - erts_smp_atomic32_set(&balance_info.checking_balance, 0); + erts_smp_atomic32_set_nob(&balance_info.checking_balance, 0); ERTS_FOREACH_RUNQ(rq, { if (rq->waiting) @@ -2040,12 +2101,13 @@ check_balance(ErtsRunQueue *c_rq) forced = balance_info.forced_check_balance; balance_info.forced_check_balance = 0; - blnc_no_rqs = (int) erts_smp_atomic32_read(&balance_info.used_runqs); + get_no_runqs(¤t_active, &blnc_no_rqs); + if (blnc_no_rqs == 1) { erts_smp_mtx_unlock(&balance_info.update_mtx); erts_smp_runq_lock(c_rq); c_rq->check_balance_reds = INT_MAX; - erts_smp_atomic32_set(&balance_info.checking_balance, 0); + erts_smp_atomic32_set_nob(&balance_info.checking_balance, 0); return; } @@ -2054,8 +2116,6 @@ check_balance(ErtsRunQueue *c_rq) if (balance_info.full_reds_history_index >= ERTS_FULL_REDS_HISTORY_SIZE) balance_info.full_reds_history_index = 0; - current_active = erts_smp_atomic32_read(&balance_info.active_runqs); - /* Read balance information for all run queues */ for (qix = 0; qix < blnc_no_rqs; qix++) { ErtsRunQueue *rq = ERTS_RUNQ_IX(qix); @@ -2389,10 +2449,10 @@ erts_fprintf(stderr, "--------------------------------\n"); } balance_info.last_active_runqs = active; - erts_smp_atomic32_set(&balance_info.active_runqs, active); + set_no_active_runqs(active); balance_info.halftime = 1; - erts_smp_atomic32_set(&balance_info.checking_balance, 0); + erts_smp_atomic32_set_nob(&balance_info.checking_balance, 0); /* Write migration paths and reset balance statistics in all queues */ for (qix = 0; qix < blnc_no_rqs; qix++) { @@ -2534,7 +2594,7 @@ erts_init_scheduling(int mrq, int no_schedulers, int no_schedulers_online) erts_alloc_permanent_cache_aligned(ERTS_ALC_T_RUNQS, sizeof(ErtsAlignedRunQueue) * n); #ifdef ERTS_SMP - erts_smp_atomic32_init(&no_empty_run_queues, 0); + erts_smp_atomic32_init_nob(&no_empty_run_queues, 0); #endif erts_no_run_queues = n; @@ -2544,7 +2604,7 @@ erts_init_scheduling(int mrq, int no_schedulers, int no_schedulers_online) ErtsRunQueue *rq = ERTS_RUNQ_IX(ix); rq->ix = ix; - erts_smp_atomic32_init(&rq->info_flags, ERTS_RUNQ_IFLG_NONEMPTY); + erts_smp_atomic32_init_nob(&rq->info_flags, ERTS_RUNQ_IFLG_NONEMPTY); /* make sure that the "extra" id correponds to the schedulers * id if the esdp->no <-> ix+1 mapping change. @@ -2637,9 +2697,9 @@ erts_init_scheduling(int mrq, int no_schedulers, int no_schedulers_online) ssi->next = NULL; ssi->prev = NULL; #endif - erts_smp_atomic32_init(&ssi->flags, 0); + erts_smp_atomic32_init_nob(&ssi->flags, 0); ssi->event = NULL; /* initialized in sched_thread_func */ - erts_smp_atomic32_init(&ssi->aux_work, 0); + erts_smp_atomic32_init_nob(&ssi->aux_work, 0); } #endif @@ -2656,11 +2716,14 @@ erts_init_scheduling(int mrq, int no_schedulers, int no_schedulers_online) esdp->match_pseudo_process = NULL; esdp->ssi = ERTS_SCHED_SLEEP_INFO_IX(ix); esdp->free_process = NULL; -#if HALFWORD_HEAP - /* Registers need to be heap allocated (correct memory range) for tracing to work */ - esdp->save_reg = erts_alloc(ERTS_ALC_T_BEAM_REGISTER, ERTS_X_REGS_ALLOCATED * sizeof(Eterm)); -#endif #endif + esdp->x_reg_array = + erts_alloc_permanent_cache_aligned(ERTS_ALC_T_BEAM_REGISTER, + ERTS_X_REGS_ALLOCATED * + sizeof(Eterm)); + esdp->f_reg_array = + erts_alloc_permanent_cache_aligned(ERTS_ALC_T_BEAM_REGISTER, + MAX_REG * sizeof(FloatDef)); #if !HEAP_ON_C_STACK esdp->num_tmp_heap_used = 0; #endif @@ -2683,7 +2746,7 @@ erts_init_scheduling(int mrq, int no_schedulers, int no_schedulers_online) } #ifdef ERTS_SMP - erts_smp_atomic32_init(&esdp->chk_cpu_bind, 0); + erts_smp_atomic32_init_nob(&esdp->chk_cpu_bind, 0); #endif } @@ -2691,21 +2754,20 @@ erts_init_scheduling(int mrq, int no_schedulers, int no_schedulers_online) erts_smp_mtx_init(&schdlr_sspnd.mtx, "schdlr_sspnd"); erts_smp_cnd_init(&schdlr_sspnd.cnd); - erts_smp_atomic32_init(&schdlr_sspnd.changing, 0); + erts_smp_atomic32_init_nob(&schdlr_sspnd.changing, 0); schdlr_sspnd.online = no_schedulers_online; schdlr_sspnd.curr_online = no_schedulers; - erts_smp_atomic32_init(&schdlr_sspnd.msb.ongoing, 0); - erts_smp_atomic32_init(&schdlr_sspnd.active, no_schedulers); + schdlr_sspnd.msb.ongoing = 0; + erts_smp_atomic32_init_nob(&schdlr_sspnd.active, no_schedulers); schdlr_sspnd.msb.procs = NULL; - erts_smp_atomic32_set(&balance_info.used_runqs, - erts_common_run_queue ? 1 : no_schedulers_online); - erts_smp_atomic32_init(&balance_info.active_runqs, no_schedulers); + init_no_runqs(no_schedulers, + erts_common_run_queue ? 1 : no_schedulers_online); balance_info.last_active_runqs = no_schedulers; erts_smp_mtx_init(&balance_info.update_mtx, "migration_info_update"); balance_info.forced_check_balance = 0; balance_info.halftime = 1; balance_info.full_reds_history_index = 0; - erts_smp_atomic32_init(&balance_info.checking_balance, 0); + erts_smp_atomic32_init_nob(&balance_info.checking_balance, 0); balance_info.prev_rise.active_runqs = 0; balance_info.prev_rise.max_len = 0; balance_info.prev_rise.reds = 0; @@ -2714,8 +2776,8 @@ erts_init_scheduling(int mrq, int no_schedulers, int no_schedulers_online) if (no_schedulers_online < no_schedulers) { if (erts_common_run_queue) { for (ix = no_schedulers_online; ix < no_schedulers; ix++) - erts_smp_atomic32_bor(&ERTS_SCHED_SLEEP_INFO_IX(ix)->flags, - ERTS_SSI_FLG_SUSPENDED); + erts_smp_atomic32_read_bor_nob(&ERTS_SCHED_SLEEP_INFO_IX(ix)->flags, + ERTS_SSI_FLG_SUSPENDED); } else { for (ix = no_schedulers_online; ix < erts_no_run_queues; ix++) @@ -2729,7 +2791,7 @@ erts_init_scheduling(int mrq, int no_schedulers, int no_schedulers_online) ERTS_SCHDLR_SSPND_CHNG_SET((ERTS_SCHDLR_SSPND_CHNG_ONLN | ERTS_SCHDLR_SSPND_CHNG_WAITER), 0); - erts_smp_atomic32_init(&doing_sys_schedule, 0); + erts_smp_atomic32_init_nob(&doing_sys_schedule, 0); init_misc_aux_work(); @@ -2745,7 +2807,7 @@ erts_init_scheduling(int mrq, int no_schedulers, int no_schedulers_online) erts_no_schedulers = 1; #endif - erts_smp_atomic32_init(&function_calls, 0); + erts_smp_atomic32_init_nob(&function_calls, 0); /* init port tasks */ erts_port_task_init(); @@ -2857,7 +2919,6 @@ resume_process(Process *p) return; switch(p->rstatus) { case P_RUNABLE: - *statusp = P_WAITING; /* make erts_add_to_runq work */ erts_add_to_runq(p); break; case P_WAITING: @@ -2873,10 +2934,10 @@ int erts_get_max_no_executing_schedulers(void) { #ifdef ERTS_SMP - if (erts_smp_atomic32_read(&schdlr_sspnd.changing)) + if (erts_smp_atomic32_read_nob(&schdlr_sspnd.changing)) return (int) erts_no_schedulers; ERTS_THR_MEMORY_BARRIER; - return (int) erts_smp_atomic32_read(&schdlr_sspnd.active); + return (int) erts_smp_atomic32_read_nob(&schdlr_sspnd.active); #else return 1; #endif @@ -2906,7 +2967,7 @@ scheduler_ix_resume_wake(Uint ix) | ERTS_SSI_FLG_SUSPENDED); erts_aint32_t oflgs; do { - oflgs = erts_smp_atomic32_cmpxchg(&ssi->flags, 0, xflgs); + oflgs = erts_smp_atomic32_cmpxchg_relb(&ssi->flags, 0, xflgs); if (oflgs == xflgs) { erts_sched_finish_poke(ssi, oflgs); break; @@ -2925,7 +2986,7 @@ sched_prep_spin_suspended(ErtsSchedulerSleepInfo *ssi, erts_aint32_t xpct) erts_aint32_t xflgs = xpct; do { - oflgs = erts_smp_atomic32_cmpxchg(&ssi->flags, nflgs, xflgs); + oflgs = erts_smp_atomic32_cmpxchg_acqb(&ssi->flags, nflgs, xflgs); if (oflgs == xflgs) return nflgs; xflgs = oflgs; @@ -2942,7 +3003,7 @@ sched_spin_suspended(ErtsSchedulerSleepInfo *ssi, int spincount) erts_aint32_t flgs; do { - flgs = erts_smp_atomic32_read(&ssi->flags); + flgs = erts_smp_atomic32_read_acqb(&ssi->flags); if ((flgs & (ERTS_SSI_FLG_SLEEPING | ERTS_SSI_FLG_WAITING | ERTS_SSI_FLG_SUSPENDED)) @@ -2975,7 +3036,7 @@ sched_set_suspended_sleeptype(ErtsSchedulerSleepInfo *ssi) erts_tse_reset(ssi->event); while (1) { - oflgs = erts_smp_atomic32_cmpxchg(&ssi->flags, nflgs, xflgs); + oflgs = erts_smp_atomic32_cmpxchg_acqb(&ssi->flags, nflgs, xflgs); if (oflgs == xflgs) return nflgs; if ((oflgs & (ERTS_SSI_FLG_SLEEPING @@ -3030,15 +3091,15 @@ suspend_scheduler(ErtsSchedulerData *esdp) flgs = sched_prep_spin_suspended(ssi, ERTS_SSI_FLG_SUSPENDED); if (flgs & ERTS_SSI_FLG_SUSPENDED) { - active_schedulers = erts_smp_atomic32_dectest(&schdlr_sspnd.active); + active_schedulers = erts_smp_atomic32_dec_read_nob(&schdlr_sspnd.active); ASSERT(active_schedulers >= 1); - changing = erts_smp_atomic32_read(&schdlr_sspnd.changing); + changing = erts_smp_atomic32_read_nob(&schdlr_sspnd.changing); if (changing & ERTS_SCHDLR_SSPND_CHNG_MSB) { if (active_schedulers == schdlr_sspnd.msb.wait_active) wake = 1; if (active_schedulers == 1) { - changing = erts_smp_atomic32_band(&schdlr_sspnd.changing, - ~ERTS_SCHDLR_SSPND_CHNG_MSB); + changing = erts_smp_atomic32_read_band_nob(&schdlr_sspnd.changing, + ~ERTS_SCHDLR_SSPND_CHNG_MSB); changing &= ~ERTS_SCHDLR_SSPND_CHNG_MSB; } } @@ -3060,8 +3121,8 @@ suspend_scheduler(ErtsSchedulerData *esdp) && schdlr_sspnd.curr_online == schdlr_sspnd.wait_curr_online) wake = 1; if (schdlr_sspnd.online == schdlr_sspnd.curr_online) { - changing = erts_smp_atomic32_band(&schdlr_sspnd.changing, - ~ERTS_SCHDLR_SSPND_CHNG_ONLN); + changing = erts_smp_atomic32_read_band_nob(&schdlr_sspnd.changing, + ~ERTS_SCHDLR_SSPND_CHNG_ONLN); changing &= ~ERTS_SCHDLR_SSPND_CHNG_ONLN; } } @@ -3071,14 +3132,14 @@ suspend_scheduler(ErtsSchedulerData *esdp) wake = 0; } - flgs = erts_smp_atomic32_read(&ssi->flags); + flgs = erts_smp_atomic32_read_acqb(&ssi->flags); if (!(flgs & ERTS_SSI_FLG_SUSPENDED)) break; erts_smp_mtx_unlock(&schdlr_sspnd.mtx); #ifdef ERTS_SCHED_NEED_BLOCKABLE_AUX_WORK - aux_work = erts_smp_atomic32_read(&ssi->aux_work); + aux_work = erts_smp_atomic32_read_nob(&ssi->aux_work); blockable_aux_work: blockable_aux_work(esdp, ssi, aux_work); #endif @@ -3114,13 +3175,13 @@ suspend_scheduler(ErtsSchedulerData *esdp) | ERTS_SSI_FLG_SUSPENDED)); if (!(flgs & ERTS_SSI_FLG_SUSPENDED)) break; - changing = erts_smp_atomic32_read(&schdlr_sspnd.changing); + changing = erts_smp_atomic32_read_nob(&schdlr_sspnd.changing); if (changing & ~ERTS_SCHDLR_SSPND_CHNG_WAITER) break; #ifdef ERTS_SCHED_NEED_BLOCKABLE_AUX_WORK - aux_work = erts_smp_atomic32_read(&ssi->aux_work); + aux_work = erts_smp_atomic32_read_nob(&ssi->aux_work); if (aux_work & ERTS_SSI_BLOCKABLE_AUX_WORK_MASK) { erts_smp_activity_end(ERTS_ACTIVITY_WAIT, NULL, NULL, NULL); goto blockable_aux_work; @@ -3132,19 +3193,19 @@ suspend_scheduler(ErtsSchedulerData *esdp) erts_smp_activity_end(ERTS_ACTIVITY_WAIT, NULL, NULL, NULL); erts_smp_mtx_lock(&schdlr_sspnd.mtx); - changing = erts_smp_atomic32_read(&schdlr_sspnd.changing); + changing = erts_smp_atomic32_read_nob(&schdlr_sspnd.changing); } - active_schedulers = erts_smp_atomic32_inctest(&schdlr_sspnd.active); - changing = erts_smp_atomic32_read(&schdlr_sspnd.changing); + active_schedulers = erts_smp_atomic32_inc_read_nob(&schdlr_sspnd.active); + changing = erts_smp_atomic32_read_nob(&schdlr_sspnd.changing); if ((changing & ERTS_SCHDLR_SSPND_CHNG_MSB) && schdlr_sspnd.online == active_schedulers) { - erts_smp_atomic32_band(&schdlr_sspnd.changing, - ~ERTS_SCHDLR_SSPND_CHNG_MSB); + erts_smp_atomic32_read_band_nob(&schdlr_sspnd.changing, + ~ERTS_SCHDLR_SSPND_CHNG_MSB); } ASSERT(no <= schdlr_sspnd.online); - ASSERT(!erts_smp_atomic32_read(&schdlr_sspnd.msb.ongoing)); + ASSERT(!ongoing_multi_scheduling_block()); } @@ -3173,7 +3234,7 @@ do { \ (RQ)->flags |= (ERTS_RUNQ_FLG_OUT_OF_WORK \ | ERTS_RUNQ_FLG_HALFTIME_OUT_OF_WORK); \ (RQ)->check_balance_reds = ERTS_RUNQ_CALL_CHECK_BALANCE_REDS; \ - erts_smp_atomic32_band(&(RQ)->info_flags, ~ERTS_RUNQ_IFLG_SUSPENDED);\ + erts_smp_atomic32_read_band_nob(&(RQ)->info_flags, ~ERTS_RUNQ_IFLG_SUSPENDED);\ for (pix__ = 0; pix__ < ERTS_NO_PROC_PRIO_LEVELS; pix__++) { \ (RQ)->procs.prio_info[pix__].max_len = 0; \ (RQ)->procs.prio_info[pix__].reds = 0; \ @@ -3217,7 +3278,7 @@ erts_schedulers_state(Uint *total, int res; erts_aint32_t changing; erts_smp_mtx_lock(&schdlr_sspnd.mtx); - changing = erts_smp_atomic32_read(&schdlr_sspnd.changing); + changing = erts_smp_atomic32_read_nob(&schdlr_sspnd.changing); if (yield_allowed && (changing & ~ERTS_SCHDLR_SSPND_CHNG_WAITER)) res = ERTS_SCHDLR_SSPND_YIELD_RESTART; else { @@ -3248,7 +3309,7 @@ erts_set_schedulers_online(Process *p, have_unlocked_plocks = 0; no = (int) new_no; - changing = erts_smp_atomic32_read(&schdlr_sspnd.changing); + changing = erts_smp_atomic32_read_nob(&schdlr_sspnd.changing); if (changing) { res = ERTS_SCHDLR_SSPND_YIELD_RESTART; } @@ -3295,7 +3356,7 @@ erts_set_schedulers_online(Process *p, ErtsRunQueue *to_rq = ERTS_RUNQ_IX(ix % no); evacuate_run_queue(from_rq, to_rq); } - erts_smp_atomic32_set(&balance_info.used_runqs, no); + set_no_used_runqs(no); erts_smp_mtx_unlock(&balance_info.update_mtx); erts_smp_mtx_lock(&schdlr_sspnd.mtx); } @@ -3323,8 +3384,8 @@ erts_set_schedulers_online(Process *p, for (ix = no; ix < online; ix++) { ErtsSchedulerSleepInfo *ssi; ssi = ERTS_SCHED_SLEEP_INFO_IX(ix); - erts_smp_atomic32_bor(&ssi->flags, - ERTS_SSI_FLG_SUSPENDED); + erts_smp_atomic32_read_bor_nob(&ssi->flags, + ERTS_SSI_FLG_SUSPENDED); } wake_all_schedulers(); } @@ -3349,7 +3410,7 @@ erts_set_schedulers_online(Process *p, for (ix = erts_no_run_queues-1; ix >= no; ix--) evacuate_run_queue(ERTS_RUNQ_IX(ix), ERTS_RUNQ_IX(ix % no)); - erts_smp_atomic32_set(&balance_info.used_runqs, no); + set_no_used_runqs(no); erts_smp_mtx_unlock(&balance_info.update_mtx); erts_smp_mtx_lock(&schdlr_sspnd.mtx); for (ix = no; ix < online; ix++) { @@ -3371,11 +3432,11 @@ erts_set_schedulers_online(Process *p, NULL); ASSERT(res != ERTS_SCHDLR_SSPND_DONE ? (ERTS_SCHDLR_SSPND_CHNG_WAITER - & erts_smp_atomic32_read(&schdlr_sspnd.changing)) + & erts_smp_atomic32_read_nob(&schdlr_sspnd.changing)) : (ERTS_SCHDLR_SSPND_CHNG_WAITER - == erts_smp_atomic32_read(&schdlr_sspnd.changing))); - erts_smp_atomic32_band(&schdlr_sspnd.changing, - ~ERTS_SCHDLR_SSPND_CHNG_WAITER); + == erts_smp_atomic32_read_nob(&schdlr_sspnd.changing))); + erts_smp_atomic32_read_band_nob(&schdlr_sspnd.changing, + ~ERTS_SCHDLR_SSPND_CHNG_WAITER); } } @@ -3394,7 +3455,7 @@ erts_block_multi_scheduling(Process *p, ErtsProcLocks plocks, int on, int all) ErtsProcList *plp; erts_smp_mtx_lock(&schdlr_sspnd.mtx); - changing = erts_smp_atomic32_read(&schdlr_sspnd.changing); + changing = erts_smp_atomic32_read_nob(&schdlr_sspnd.changing); if (changing) { res = ERTS_SCHDLR_SSPND_YIELD_RESTART; /* Yield */ } @@ -3404,7 +3465,7 @@ erts_block_multi_scheduling(Process *p, ErtsProcLocks plocks, int on, int all) plp->next = schdlr_sspnd.msb.procs; schdlr_sspnd.msb.procs = plp; p->flags |= F_HAVE_BLCKD_MSCHED; - ASSERT(erts_smp_atomic32_read(&schdlr_sspnd.active) == 1); + ASSERT(erts_smp_atomic32_read_nob(&schdlr_sspnd.active) == 1); ASSERT(p->scheduler_data->no == 1); res = ERTS_SCHDLR_SSPND_DONE_MSCHED_BLOCKED; } @@ -3415,11 +3476,11 @@ erts_block_multi_scheduling(Process *p, ErtsProcLocks plocks, int on, int all) have_unlocked_plocks = 1; erts_smp_proc_unlock(p, plocks); } - ASSERT(0 == erts_smp_atomic32_read(&schdlr_sspnd.msb.ongoing)); - erts_smp_atomic32_set(&schdlr_sspnd.msb.ongoing, 1); + ASSERT(!ongoing_multi_scheduling_block()); + schdlr_sspnd.msb.ongoing = 1; if (online == 1) { res = ERTS_SCHDLR_SSPND_DONE_MSCHED_BLOCKED; - ASSERT(erts_smp_atomic32_read(&schdlr_sspnd.active) == 1); + ASSERT(erts_smp_atomic32_read_nob(&schdlr_sspnd.active) == 1); ASSERT(p->scheduler_data->no == 1); } else { @@ -3439,14 +3500,14 @@ erts_block_multi_scheduling(Process *p, ErtsProcLocks plocks, int on, int all) } if (erts_common_run_queue) { for (ix = 1; ix < online; ix++) - erts_smp_atomic32_bor(&ERTS_SCHED_SLEEP_INFO_IX(ix)->flags, - ERTS_SSI_FLG_SUSPENDED); + erts_smp_atomic32_read_bor_nob(&ERTS_SCHED_SLEEP_INFO_IX(ix)->flags, + ERTS_SSI_FLG_SUSPENDED); wake_all_schedulers(); } else { erts_smp_mtx_unlock(&schdlr_sspnd.mtx); erts_smp_mtx_lock(&balance_info.update_mtx); - erts_smp_atomic32_set(&balance_info.used_runqs, 1); + set_no_used_runqs(1); for (ix = 0; ix < online; ix++) { ErtsRunQueue *rq = ERTS_RUNQ_IX(ix); erts_smp_runq_lock(rq); @@ -3468,7 +3529,7 @@ erts_block_multi_scheduling(Process *p, ErtsProcLocks plocks, int on, int all) susp_sched_prep_block, susp_sched_resume_block, NULL); - while (erts_smp_atomic32_read(&schdlr_sspnd.active) + while (erts_smp_atomic32_read_nob(&schdlr_sspnd.active) != schdlr_sspnd.msb.wait_active) erts_smp_cnd_wait(&schdlr_sspnd.cnd, &schdlr_sspnd.mtx); erts_smp_activity_end(ERTS_ACTIVITY_WAIT, @@ -3477,11 +3538,11 @@ erts_block_multi_scheduling(Process *p, ErtsProcLocks plocks, int on, int all) NULL); ASSERT(res != ERTS_SCHDLR_SSPND_DONE_MSCHED_BLOCKED ? (ERTS_SCHDLR_SSPND_CHNG_WAITER - & erts_smp_atomic32_read(&schdlr_sspnd.changing)) + & erts_smp_atomic32_read_nob(&schdlr_sspnd.changing)) : (ERTS_SCHDLR_SSPND_CHNG_WAITER - == erts_smp_atomic32_read(&schdlr_sspnd.changing))); - erts_smp_atomic32_band(&schdlr_sspnd.changing, - ~ERTS_SCHDLR_SSPND_CHNG_WAITER); + == erts_smp_atomic32_read_nob(&schdlr_sspnd.changing))); + erts_smp_atomic32_read_band_nob(&schdlr_sspnd.changing, + ~ERTS_SCHDLR_SSPND_CHNG_WAITER); } plp = proclist_create(p); plp->next = schdlr_sspnd.msb.procs; @@ -3548,16 +3609,16 @@ erts_block_multi_scheduling(Process *p, ErtsProcLocks plocks, int on, int all) }); #endif p->flags &= ~F_HAVE_BLCKD_MSCHED; - erts_smp_atomic32_set(&schdlr_sspnd.msb.ongoing, 0); + schdlr_sspnd.msb.ongoing = 0; if (schdlr_sspnd.online == 1) { /* No schedulers to resume */ - ASSERT(erts_smp_atomic32_read(&schdlr_sspnd.active) == 1); + ASSERT(erts_smp_atomic32_read_nob(&schdlr_sspnd.active) == 1); ERTS_SCHDLR_SSPND_CHNG_SET(0, ERTS_SCHDLR_SSPND_CHNG_MSB); } else if (erts_common_run_queue) { for (ix = 1; ix < schdlr_sspnd.online; ix++) - erts_smp_atomic32_band(&ERTS_SCHED_SLEEP_INFO_IX(ix)->flags, - ~ERTS_SSI_FLG_SUSPENDED); + erts_smp_atomic32_read_band_nob(&ERTS_SCHED_SLEEP_INFO_IX(ix)->flags, + ~ERTS_SSI_FLG_SUSPENDED); wake_all_schedulers(); } else { @@ -3583,7 +3644,7 @@ erts_block_multi_scheduling(Process *p, ErtsProcLocks plocks, int on, int all) evacuate_run_queue(ERTS_RUNQ_IX(ix), ERTS_RUNQ_IX(ix % online)); - erts_smp_atomic32_set(&balance_info.used_runqs, online); + set_no_used_runqs(online); /* Make sure that we balance soon... */ balance_info.forced_check_balance = 1; erts_smp_runq_lock(ERTS_RUNQ_IX(0)); @@ -3607,7 +3668,7 @@ void erts_dbg_multi_scheduling_return_trap(Process *p, Eterm return_value) { if (return_value == am_blocked) { - erts_aint32_t active = erts_smp_atomic32_read(&schdlr_sspnd.active); + erts_aint32_t active = erts_smp_atomic32_read_nob(&schdlr_sspnd.active); ASSERT(1 <= active && active <= 2); ASSERT(ERTS_PROC_GET_SCHDATA(p)->no == 1); } @@ -3668,7 +3729,7 @@ sched_thread_func(void *vesdp) #ifdef ERTS_ENABLE_LOCK_CHECK { char buf[31]; - erts_snprintf(&buf[0], 31, "scheduler %bpu", no); + erts_snprintf(&buf[0], 31, "scheduler %beu", no); erts_lc_set_thread_name(&buf[0]); } #endif @@ -3690,12 +3751,12 @@ sched_thread_func(void *vesdp) erts_thread_init_float(); erts_smp_mtx_lock(&schdlr_sspnd.mtx); - ASSERT(erts_smp_atomic32_read(&schdlr_sspnd.changing) + ASSERT(erts_smp_atomic32_read_nob(&schdlr_sspnd.changing) & ERTS_SCHDLR_SSPND_CHNG_ONLN); if (--schdlr_sspnd.curr_online == schdlr_sspnd.wait_curr_online) { - erts_smp_atomic32_band(&schdlr_sspnd.changing, - ~ERTS_SCHDLR_SSPND_CHNG_ONLN); + erts_smp_atomic32_read_band_nob(&schdlr_sspnd.changing, + ~ERTS_SCHDLR_SSPND_CHNG_ONLN); if (((ErtsSchedulerData *) vesdp)->no != 1) erts_smp_cnd_signal(&schdlr_sspnd.cnd); } @@ -3726,7 +3787,7 @@ sched_thread_func(void *vesdp) process_main(); /* No schedulers should *ever* terminate */ - erl_exit(ERTS_ABORT_EXIT, "Scheduler thread number %bpu terminated\n", + erl_exit(ERTS_ABORT_EXIT, "Scheduler thread number %beu terminated\n", ((ErtsSchedulerData *) vesdp)->no); return NULL; } @@ -3775,8 +3836,8 @@ erts_start_schedulers(void) erts_dsprintf_buf_t *dsbufp = erts_create_logger_dsbuf(); ASSERT(actual != wanted_no_schedulers); erts_dsprintf(dsbufp, - "Failed to create %bpu scheduler-threads (%s:%d); " - "only %bpu scheduler-thread%s created.\n", + "Failed to create %beu scheduler-threads (%s:%d); " + "only %beu scheduler-thread%s created.\n", wanted_no_schedulers, erl_errno_id(res), res, actual, actual == 1 ? " was" : "s were"); erts_send_error_to_logger_nogl(dsbufp); @@ -3890,21 +3951,9 @@ handle_pend_sync_suspend(Process *suspendee, } } -/* - * Like erts_pid2proc() but: - * - * * At least ERTS_PROC_LOCK_MAIN have to be held on c_p. - * * At least ERTS_PROC_LOCK_MAIN have to be taken on pid. - * * It also waits for proc to be in a state != running and garbing. - * * If ERTS_PROC_LOCK_BUSY is returned, the calling process has to - * yield (ERTS_BIF_YIELD[0-3]()). c_p might in this case have been - * suspended. - */ - - -Process * -erts_pid2proc_not_running(Process *c_p, ErtsProcLocks c_p_locks, - Eterm pid, ErtsProcLocks pid_locks) +static Process * +pid2proc_not_running(Process *c_p, ErtsProcLocks c_p_locks, + Eterm pid, ErtsProcLocks pid_locks, int suspend) { Process *rp; int unlock_c_p_status; @@ -3931,7 +3980,7 @@ erts_pid2proc_not_running(Process *c_p, ErtsProcLocks c_p_locks, c_p->suspendee = NIL; ASSERT(c_p->flags & F_P2PNR_RESCHED); c_p->flags &= ~F_P2PNR_RESCHED; - if (rp) + if (!suspend && rp) resume_process(rp); } else { @@ -3995,6 +4044,8 @@ erts_pid2proc_not_running(Process *c_p, ErtsProcLocks c_p_locks, } /* rp is not running and we got the locks we want... */ + if (suspend) + suspend_process(rp_rq, rp); } erts_smp_runqs_unlock(cp_rq, rp_rq); } @@ -4007,6 +4058,35 @@ erts_pid2proc_not_running(Process *c_p, ErtsProcLocks c_p_locks, return rp; } + +/* + * Like erts_pid2proc() but: + * + * * At least ERTS_PROC_LOCK_MAIN have to be held on c_p. + * * At least ERTS_PROC_LOCK_MAIN have to be taken on pid. + * * It also waits for proc to be in a state != running and garbing. + * * If ERTS_PROC_LOCK_BUSY is returned, the calling process has to + * yield (ERTS_BIF_YIELD[0-3]()). c_p might in this case have been + * suspended. + */ +Process * +erts_pid2proc_not_running(Process *c_p, ErtsProcLocks c_p_locks, + Eterm pid, ErtsProcLocks pid_locks) +{ + return pid2proc_not_running(c_p, c_p_locks, pid, pid_locks, 0); +} + +/* + * Like erts_pid2proc_not_running(), but hands over the process + * in a suspended state unless (c_p is looked up). + */ +Process * +erts_pid2proc_suspend(Process *c_p, ErtsProcLocks c_p_locks, + Eterm pid, ErtsProcLocks pid_locks) +{ + return pid2proc_not_running(c_p, c_p_locks, pid, pid_locks, 1); +} + /* * erts_pid2proc_nropt() is normally the same as * erts_pid2proc_not_running(). However it is only @@ -4120,6 +4200,21 @@ handle_pend_bif_async_suspend(Process *suspendee, } } +#else + +/* + * Non-smp version of erts_pid2proc_suspend(). + */ +Process * +erts_pid2proc_suspend(Process *c_p, ErtsProcLocks c_p_locks, + Eterm pid, ErtsProcLocks pid_locks) +{ + Process *rp = erts_pid2proc(c_p, c_p_locks, pid, pid_locks); + if (rp) + erts_suspend(rp, pid_locks, NULL); + return rp; +} + #endif /* ERTS_SMP */ /* @@ -4653,7 +4748,7 @@ internal_add_to_runq(ErtsRunQueue *runq, Process *p) if (p->status_flags & ERTS_PROC_SFLG_INRUNQ) return NULL; else if (p->runq_flags & ERTS_PROC_RUNQ_FLG_RUNNING) { - ASSERT(p->status != P_SUSPENDED); + ASSERT(ERTS_PROC_IS_EXITING(p) || p->rcount == 0); ERTS_DBG_CHK_PROCS_RUNQ_NOPROC(runq, p); p->status_flags |= ERTS_PROC_SFLG_PENDADD2SCHEDQ; return NULL; @@ -4664,9 +4759,8 @@ internal_add_to_runq(ErtsRunQueue *runq, Process *p) ERTS_DBG_CHK_PROCS_RUNQ_NOPROC(runq, p); #ifndef ERTS_SMP /* Never schedule a suspended process (ok in smp case) */ - ASSERT(p->status != P_SUSPENDED); + ASSERT(ERTS_PROC_IS_EXITING(p) || p->rcount == 0); add_runq = runq; - #else ASSERT(!p->bound_runq || p->bound_runq == p->run_queue); if (p->bound_runq) { @@ -5101,7 +5195,7 @@ Process *schedule(Process *p, int calls) esdp = erts_get_scheduler_data(); rq = erts_get_runq_current(esdp); ASSERT(esdp); - fcalls = (int) erts_smp_atomic32_read(&function_calls); + fcalls = (int) erts_smp_atomic32_read_acqb(&function_calls); actual_reds = reds = 0; erts_smp_runq_lock(rq); } else { @@ -5119,7 +5213,7 @@ Process *schedule(Process *p, int calls) reds = ERTS_PROC_MIN_CONTEXT_SWITCH_REDS_COST; esdp->virtual_reds = 0; - fcalls = (int) erts_smp_atomic32_addtest(&function_calls, reds); + fcalls = (int) erts_smp_atomic32_add_read_acqb(&function_calls, reds); ASSERT(esdp && esdp == erts_get_scheduler_data()); rq = erts_get_runq_current(esdp); @@ -5166,7 +5260,7 @@ Process *schedule(Process *p, int calls) handle_pending_suspend(p, ERTS_PROC_LOCK_MAIN|ERTS_PROC_LOCK_STATUS); ASSERT(!(p->status_flags & ERTS_PROC_SFLG_PENDADD2SCHEDQ) - || p->status != P_SUSPENDED); + || p->rcount == 0); } #endif erts_smp_runq_lock(rq); @@ -5252,14 +5346,14 @@ Process *schedule(Process *p, int calls) | ERTS_RUNQ_FLG_CHK_CPU_BIND | ERTS_RUNQ_FLG_SUSPENDED)) { if ((rq->flags & ERTS_RUNQ_FLG_SUSPENDED) - || (erts_smp_atomic32_read(&esdp->ssi->flags) + || (erts_smp_atomic32_read_acqb(&esdp->ssi->flags) & ERTS_SSI_FLG_SUSPENDED)) { - ASSERT(erts_smp_atomic32_read(&esdp->ssi->flags) + ASSERT(erts_smp_atomic32_read_nob(&esdp->ssi->flags) & ERTS_SSI_FLG_SUSPENDED); suspend_scheduler(esdp); } if ((rq->flags & ERTS_RUNQ_FLG_CHK_CPU_BIND) - || erts_smp_atomic32_read(&esdp->chk_cpu_bind)) { + || erts_smp_atomic32_read_acqb(&esdp->chk_cpu_bind)) { erts_sched_check_cpu_bind(esdp); } } @@ -5268,7 +5362,7 @@ Process *schedule(Process *p, int calls) || defined(ERTS_SCHED_NEED_NONBLOCKABLE_AUX_WORK) { ErtsSchedulerSleepInfo *ssi = esdp->ssi; - erts_aint32_t aux_work = erts_smp_atomic32_read(&ssi->aux_work); + erts_aint32_t aux_work = erts_smp_atomic32_read_nob(&ssi->aux_work); if (aux_work) { erts_smp_runq_unlock(rq); #ifdef ERTS_SCHED_NEED_BLOCKABLE_AUX_WORK @@ -5310,9 +5404,9 @@ Process *schedule(Process *p, int calls) if (rq->flags & (ERTS_RUNQ_FLG_SHARED_RUNQ | ERTS_RUNQ_FLG_SUSPENDED)) { if ((rq->flags & ERTS_RUNQ_FLG_SUSPENDED) - || (erts_smp_atomic32_read(&esdp->ssi->flags) + || (erts_smp_atomic32_read_acqb(&esdp->ssi->flags) & ERTS_SSI_FLG_SUSPENDED)) { - ASSERT(erts_smp_atomic32_read(&esdp->ssi->flags) + ASSERT(erts_smp_atomic32_read_nob(&esdp->ssi->flags) & ERTS_SSI_FLG_SUSPENDED); non_empty_runq(rq); goto continue_check_activities_to_run; @@ -5354,7 +5448,7 @@ Process *schedule(Process *p, int calls) * Schedule system-level activities. */ - erts_smp_atomic32_set(&function_calls, 0); + erts_smp_atomic32_set_relb(&function_calls, 0); fcalls = 0; ASSERT(!erts_port_task_have_outstanding_io_tasks()); @@ -5396,7 +5490,7 @@ Process *schedule(Process *p, int calls) if (erts_common_run_queue->waiting) wake_scheduler(erts_common_run_queue, 0, 1); } - else if (erts_smp_atomic32_read(&no_empty_run_queues) != 0) { + else if (erts_smp_atomic32_read_acqb(&no_empty_run_queues) != 0) { wake_scheduler_on_empty_runq(rq); rq->wakeup_other = 0; } @@ -5687,10 +5781,13 @@ erts_sched_stat_term(Process *p, int total) void erts_schedule_misc_op(void (*func)(void *), void *arg) { - ErtsRunQueue *rq = erts_get_runq_current(NULL); + ErtsRunQueue *rq; ErtsMiscOpList *molp = misc_op_list_alloc(); + ErtsSchedulerData *esdp = erts_get_scheduler_data(); - if (!rq) { + if (esdp) { + rq = esdp->run_queue; + } else { /* * This can only happen when the sys msg dispatcher * thread schedules misc ops (this happens *very* @@ -5853,7 +5950,7 @@ erts_test_next_pid(int set, Uint next) Uint erts_process_count(void) { - erts_aint32_t res = erts_smp_atomic32_read(&process_count); + erts_aint32_t res = erts_smp_atomic32_read_nob(&process_count); ASSERT(res >= 0); return (Uint) res; } @@ -5902,7 +5999,7 @@ alloc_process(void) ASSERT(!process_tab[p_next]); process_tab[p_next] = p; - erts_smp_atomic32_inc(&process_count); + erts_smp_atomic32_inc_nob(&process_count); p->id = make_internal_pid(p_serial << p_serial_shift | p_next); if (p->id == ERTS_INVALID_PID) { /* Do not use the invalid pid; change serial */ @@ -6028,7 +6125,7 @@ erl_create_process(Process* parent, /* Parent of process (default group leader). p->min_heap_size = H_MIN_SIZE; p->min_vheap_size = BIN_VH_MIN_SIZE; p->prio = PRIORITY_NORMAL; - p->max_gen_gcs = (Uint16) erts_smp_atomic32_read(&erts_max_gen_gcs); + p->max_gen_gcs = (Uint16) erts_smp_atomic32_read_nob(&erts_max_gen_gcs); } p->skipped = 0; ASSERT(p->min_heap_size == erts_next_heap_size(p->min_heap_size, 0)); @@ -7485,8 +7582,8 @@ continue_exit_process(Process *p p->status_flags = 0; #endif process_tab[pix] = NULL; /* Time of death! */ - ASSERT(erts_smp_atomic32_read(&process_count) > 0); - erts_smp_atomic32_dec(&process_count); + ASSERT(erts_smp_atomic32_read_nob(&process_count) > 0); + erts_smp_atomic32_dec_nob(&process_count); #ifdef ERTS_SMP erts_pix_unlock(pix_lock); @@ -7607,14 +7704,33 @@ continue_exit_process(Process *p static void timeout_proc(Process* p) { - p->i = *((BeamInstr **) (UWord) p->def_arg_reg); + BeamInstr** pi = (BeamInstr **) p->def_arg_reg; + p->i = *pi; p->flags |= F_TIMO; p->flags &= ~F_INSLPQUEUE; - if (p->status == P_WAITING) - erts_add_to_runq(p); - if (p->status == P_SUSPENDED) + switch (p->status) { + case P_GARBING: + switch (p->gcstatus) { + case P_SUSPENDED: + goto suspended; + case P_WAITING: + goto waiting; + default: + break; + } + break; + case P_WAITING: + waiting: + erts_add_to_runq(p); + break; + case P_SUSPENDED: + suspended: p->rstatus = P_RUNABLE; /* MUST set resume status to runnable */ + break; + default: + break; + } } |