diff options
Diffstat (limited to 'erts/emulator/beam/erl_process.c')
| -rw-r--r-- | erts/emulator/beam/erl_process.c | 1498 |
1 files changed, 1005 insertions, 493 deletions
diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c index 9386f79b56..0a099e69bb 100644 --- a/erts/emulator/beam/erl_process.c +++ b/erts/emulator/beam/erl_process.c @@ -92,10 +92,6 @@ #undef HARDDEBUG #endif -#ifdef HARDDEBUG -#define HARDDEBUG_RUNQS -#endif - #ifdef HIPE #include "hipe_mode_switch.h" /* for hipe_init_process() */ #include "hipe_signal.h" /* for hipe_thread_signal_init() */ @@ -174,7 +170,6 @@ ErtsLcPSDLocks erts_psd_required_locks[ERTS_PSD_SIZE]; typedef struct { int aux_work; int tse; - int sys_schedule; } ErtsBusyWaitParams; static ErtsBusyWaitParams sched_busy_wait_params[ERTS_SCHED_TYPE_LAST + 1]; @@ -344,6 +339,10 @@ erts_sched_stat_t erts_sched_stat; static erts_tsd_key_t ERTS_WRITE_UNLIKELY(sched_data_key); +#if ERTS_POLL_USE_SCHEDULER_POLLING +static erts_atomic32_t function_calls; +static erts_atomic32_t doing_sys_schedule; +#endif static erts_atomic32_t no_empty_run_queues; long erts_runq_supervision_interval = 0; static ethr_event runq_supervision_event; @@ -736,12 +735,6 @@ erts_pre_init_process(void) #endif } -static void -release_process(void *vproc) -{ - erts_proc_dec_refc((Process *) vproc); -} - /* initialize the scheduler */ void erts_init_process(int ncpu, int proc_tab_size, int legacy_proc_tab) @@ -753,7 +746,7 @@ erts_init_process(int ncpu, int proc_tab_size, int legacy_proc_tab) erts_ptab_init_table(&erts_proc, ERTS_ALC_T_PROC_TABLE, - release_process, + NULL, (ErtsPTabElementCommon *) &erts_invalid_process.common, proc_tab_size, sizeof(Process), @@ -1477,7 +1470,10 @@ proclist_create(Process *p) { ErtsProcList *plp = proclist_alloc(); ensure_later_proc_interval(p->common.u.alive.started_interval); - plp->pid = p->common.id; + if (erts_atomic32_read_nob(&p->state) & ERTS_PSFLG_FREE) + plp->u.p = p; + else + plp->u.pid = p->common.id; plp->started_interval = p->common.u.alive.started_interval; return plp; } @@ -1486,7 +1482,7 @@ static ERTS_INLINE ErtsProcList * proclist_copy(ErtsProcList *plp0) { ErtsProcList *plp1 = proclist_alloc(); - plp1->pid = plp0->pid; + plp1->u.pid = plp0->u.pid; plp1->started_interval = plp0->started_interval; return plp1; } @@ -1521,7 +1517,10 @@ erts_proclist_dump(fmtfn_t to, void *to_arg, ErtsProcList *plp) ErtsProcList *first = plp; while (plp) { - erts_print(to, to_arg, "%T", plp->pid); + if (is_pid(plp->u.pid)) + erts_print(to, to_arg, "%T", plp->u.pid); + else + erts_print(to, to_arg, "%T", plp->u.p->common.id); plp = plp->next; if (plp == first) break; @@ -1646,7 +1645,7 @@ haw_thr_prgr_wakeup(ErtsAuxWorkData *awdp, ErtsThrPrgrVal val) awdp->latest_wakeup = val; haw_chk_later_cleanup_op_wakeup(awdp, val); } - erts_thr_progress_wakeup(awdp->esdp, val); + erts_thr_progress_wakeup(erts_thr_prgr_data(awdp->esdp), val); } } @@ -1656,7 +1655,7 @@ haw_thr_prgr_soft_wakeup(ErtsAuxWorkData *awdp, ErtsThrPrgrVal val) if (erts_thr_progress_cmp(val, awdp->latest_wakeup) > 0) { awdp->latest_wakeup = val; haw_chk_later_cleanup_op_wakeup(awdp, val); - erts_thr_progress_wakeup(awdp->esdp, val); + erts_thr_progress_wakeup(erts_thr_prgr_data(awdp->esdp), val); } } @@ -1670,7 +1669,7 @@ haw_thr_prgr_later_cleanup_op_wakeup(ErtsAuxWorkData *awdp, ErtsThrPrgrVal val, else { awdp->latest_wakeup = val; awdp->later_op.size = thr_prgr_later_cleanup_op_threshold; - erts_thr_progress_wakeup(awdp->esdp, val); + erts_thr_progress_wakeup(erts_thr_prgr_data(awdp->esdp), val); } } } @@ -2463,6 +2462,13 @@ handle_reap_ports(ErtsAuxWorkData *awdp, erts_aint32_t aux_work, int waiting) erts_port_lock(prt); + if (prt->common.u.alive.reg && + prt->common.u.alive.reg->name == am_heart_port) { + /* Leave heart port to not get killed before flushing is done*/ + erts_port_release(prt); + continue; + } + state = erts_atomic32_read_nob(&prt->state); if (!(state & (ERTS_PORT_SFLGS_INVALID_DRIVER_LOOKUP | ERTS_PORT_SFLG_HALT))) { @@ -3066,6 +3072,7 @@ aux_thread(void *unused) ErtsSchedulerSleepInfo *ssi = ERTS_SCHED_SLEEP_INFO_IX(-1); erts_aint32_t aux_work; ErtsThrPrgrCallbacks callbacks; + ErtsThrPrgrData *tpd; int thr_prgr_active = 1; ERTS_MSACC_DECLARE_CACHE(); @@ -3087,12 +3094,16 @@ aux_thread(void *unused) callbacks.wait = thr_prgr_wait; callbacks.finalize_wait = thr_prgr_fin_wait; - erts_thr_progress_register_managed_thread(NULL, &callbacks, 1); + tpd = erts_thr_progress_register_managed_thread(NULL, &callbacks, 1); init_aux_work_data(awdp, NULL, NULL); awdp->ssi = ssi; #if ERTS_POLL_USE_FALLBACK - ssi->psi = erts_create_pollset_thread(-1); +#if ERTS_POLL_USE_SCHEDULER_POLLING + ssi->psi = erts_create_pollset_thread(-2, tpd); +#else + ssi->psi = erts_create_pollset_thread(-1, tpd); +#endif #endif sched_prep_spin_wait(ssi); @@ -3105,11 +3116,11 @@ aux_thread(void *unused) aux_work = erts_atomic32_read_acqb(&ssi->aux_work); if (aux_work) { if (!thr_prgr_active) - erts_thr_progress_active(NULL, thr_prgr_active = 1); + erts_thr_progress_active(tpd, thr_prgr_active = 1); aux_work = handle_aux_work(awdp, aux_work, 1); ERTS_MSACC_UPDATE_CACHE(); - if (aux_work && erts_thr_progress_update(NULL)) - erts_thr_progress_leader_update(NULL); + if (aux_work && erts_thr_progress_update(tpd)) + erts_thr_progress_leader_update(tpd); } if (!aux_work) { @@ -3120,7 +3131,7 @@ aux_thread(void *unused) #endif if (thr_prgr_active) - erts_thr_progress_active(NULL, thr_prgr_active = 0); + erts_thr_progress_active(tpd, thr_prgr_active = 0); #if ERTS_POLL_USE_FALLBACK @@ -3132,11 +3143,11 @@ aux_thread(void *unused) if (flgs & ERTS_SSI_FLG_SLEEPING) { ASSERT(flgs & ERTS_SSI_FLG_POLL_SLEEPING); ASSERT(flgs & ERTS_SSI_FLG_WAITING); - erts_check_io(ssi->psi); + erts_check_io(ssi->psi, ERTS_POLL_INF_TIMEOUT); } } #else - erts_thr_progress_prepare_wait(NULL); + erts_thr_progress_prepare_wait(tpd); flgs = sched_spin_wait(ssi, 0); @@ -3153,7 +3164,7 @@ aux_thread(void *unused) ERTS_MSACC_SET_STATE_CACHED(ERTS_MSACC_STATE_OTHER); } } - erts_thr_progress_finalize_wait(NULL); + erts_thr_progress_finalize_wait(tpd); #endif } @@ -3171,7 +3182,8 @@ poll_thread(void *arg) erts_aint32_t aux_work; ErtsThrPrgrCallbacks callbacks; int thr_prgr_active = 1; - struct erts_poll_thread *psi = erts_create_pollset_thread(id); + struct erts_poll_thread *psi; + ErtsThrPrgrData *tpd; ERTS_MSACC_DECLARE_CACHE(); #ifdef ERTS_ENABLE_LOCK_CHECK @@ -3192,9 +3204,12 @@ poll_thread(void *arg) callbacks.wait = thr_prgr_wait; callbacks.finalize_wait = thr_prgr_fin_wait; - erts_thr_progress_register_managed_thread(NULL, &callbacks, 0); + tpd = erts_thr_progress_register_managed_thread(NULL, &callbacks, 0); init_aux_work_data(awdp, NULL, NULL); awdp->ssi = ssi; + + psi = erts_create_pollset_thread(id, tpd); + ssi->psi = psi; sched_prep_spin_wait(ssi); @@ -3207,16 +3222,16 @@ poll_thread(void *arg) aux_work = erts_atomic32_read_acqb(&ssi->aux_work); if (aux_work) { if (!thr_prgr_active) - erts_thr_progress_active(NULL, thr_prgr_active = 1); + erts_thr_progress_active(tpd, thr_prgr_active = 1); aux_work = handle_aux_work(awdp, aux_work, 1); ERTS_MSACC_UPDATE_CACHE(); - if (aux_work && erts_thr_progress_update(NULL)) - erts_thr_progress_leader_update(NULL); + if (aux_work && erts_thr_progress_update(tpd)) + erts_thr_progress_leader_update(tpd); } if (!aux_work) { if (thr_prgr_active) - erts_thr_progress_active(NULL, thr_prgr_active = 0); + erts_thr_progress_active(tpd, thr_prgr_active = 0); flgs = sched_spin_wait(ssi, 0); @@ -3226,7 +3241,7 @@ poll_thread(void *arg) if (flgs & ERTS_SSI_FLG_SLEEPING) { ASSERT(flgs & ERTS_SSI_FLG_POLL_SLEEPING); ASSERT(flgs & ERTS_SSI_FLG_WAITING); - erts_check_io(psi); + erts_check_io(psi, ERTS_POLL_INF_TIMEOUT); } } } @@ -3236,6 +3251,78 @@ poll_thread(void *arg) return NULL; } +#if ERTS_POLL_USE_SCHEDULER_POLLING +static ERTS_INLINE void +clear_sys_scheduling(void) +{ + erts_atomic32_set_relb(&function_calls, 0); + erts_atomic32_set_mb(&doing_sys_schedule, 0); +} + +static ERTS_INLINE int +try_set_sys_scheduling(void) +{ + return 0 == erts_atomic32_cmpxchg_acqb(&doing_sys_schedule, 1, 0); +} + + +static ERTS_INLINE int +prepare_for_sys_schedule(void) +{ + while (!erts_port_task_have_outstanding_io_tasks() + && try_set_sys_scheduling()) { + if (!erts_port_task_have_outstanding_io_tasks()) + return 1; + clear_sys_scheduling(); + } + return 0; +} + +#else +#define clear_sys_scheduling() +#define prepare_for_sys_schedule() 0 +#endif + +#ifdef HARDDEBUG +#define ERTS_HDBG_CHK_SLEEP_LIST(SL, L, F, FN) \ + check_sleepers_list((SL), (L), (F), (FN)) +static void check_sleepers_list(ErtsSchedulerSleepList *sl, + int lock, + ErtsSchedulerSleepInfo *find, + ErtsSchedulerSleepInfo *find_not) +{ + ErtsSchedulerSleepInfo *last_out; + int found = 0; + + if (lock) + erts_spin_lock(&sl->lock); + + ERTS_ASSERT(!find_not || (!find_not->next && !find_not->prev)); + + last_out = sl->list; + if (last_out) { + ErtsSchedulerSleepInfo *tmp = last_out; + do { + ERTS_ASSERT(tmp->next); + ERTS_ASSERT(tmp->prev); + ERTS_ASSERT(tmp->next->prev == tmp); + ERTS_ASSERT(tmp->prev->next == tmp); + ERTS_ASSERT(tmp != find_not); + if (tmp == find) + found = !0; + tmp = tmp->next; + + } while (tmp != last_out); + } + ERTS_ASSERT(!find || found); + + if (lock) + erts_spin_unlock(&sl->lock); +} +#else +#define ERTS_HDBG_CHK_SLEEP_LIST(SL, L, F, FN) ((void) 0) +#endif + static void scheduler_wait(int *fcalls, ErtsSchedulerData *esdp, ErtsRunQueue *rq) { @@ -3249,23 +3336,29 @@ scheduler_wait(int *fcalls, ErtsSchedulerData *esdp, ErtsRunQueue *rq) ERTS_LC_ASSERT(erts_lc_runq_is_locked(rq)); - if (ERTS_RUNQ_IX_IS_DIRTY(rq->ix)) - erts_spin_lock(&rq->sleepers.lock); flgs = sched_prep_spin_wait(ssi); if (flgs & ERTS_SSI_FLG_SUSPENDED) { /* Go suspend instead... */ - if (ERTS_RUNQ_IX_IS_DIRTY(rq->ix)) - erts_spin_unlock(&rq->sleepers.lock); return; } if (ERTS_RUNQ_IX_IS_DIRTY(rq->ix)) { - ssi->prev = NULL; - ssi->next = rq->sleepers.list; - if (rq->sleepers.list) - rq->sleepers.list->prev = ssi; - rq->sleepers.list = ssi; - erts_spin_unlock(&rq->sleepers.lock); + erts_spin_lock(&rq->sleepers.lock); + ERTS_HDBG_CHK_SLEEP_LIST(&rq->sleepers, 0, NULL, ssi); + ASSERT(!ssi->next); /* Not in sleepers list */ + ASSERT(!ssi->prev); + if (!rq->sleepers.list) { + ssi->next = ssi->prev = ssi; + rq->sleepers.list = ssi; + } + else { + ssi->prev = rq->sleepers.list; + ssi->next = rq->sleepers.list->next; + ssi->prev->next = ssi; + ssi->next->prev = ssi; + } + ERTS_HDBG_CHK_SLEEP_LIST(&rq->sleepers, 0, ssi, NULL); + erts_spin_unlock(&rq->sleepers.lock); dirty_active(esdp, -1); } @@ -3284,28 +3377,31 @@ scheduler_wait(int *fcalls, ErtsSchedulerData *esdp, ErtsRunQueue *rq) ErtsMonotonicTime current_time = 0; aux_work = erts_atomic32_read_acqb(&ssi->aux_work); - if (aux_work && !ERTS_SCHEDULER_IS_DIRTY(esdp)) { + + if (aux_work && ERTS_SCHEDULER_IS_DIRTY(esdp)) { + ERTS_INTERNAL_ERROR("Executing aux work on a dirty scheduler."); + } + + if (aux_work) { if (!thr_prgr_active) { - erts_thr_progress_active(esdp, thr_prgr_active = 1); + erts_thr_progress_active(erts_thr_prgr_data(esdp), thr_prgr_active = 1); sched_wall_time_change(esdp, 1); } aux_work = handle_aux_work(&esdp->aux_work_data, aux_work, 1); ERTS_MSACC_UPDATE_CACHE(); - if (aux_work && erts_thr_progress_update(esdp)) - erts_thr_progress_leader_update(esdp); + if (aux_work && erts_thr_progress_update(erts_thr_prgr_data(esdp))) + erts_thr_progress_leader_update(erts_thr_prgr_data(esdp)); } if (aux_work) { - if (!ERTS_SCHEDULER_IS_DIRTY(esdp)) { - flgs = erts_atomic32_read_acqb(&ssi->flags); - current_time = erts_get_monotonic_time(esdp); - if (current_time >= erts_next_timeout_time(esdp->next_tmo_ref)) { - if (!thr_prgr_active) { - erts_thr_progress_active(esdp, thr_prgr_active = 1); - sched_wall_time_change(esdp, 1); - } - erts_bump_timers(esdp->timer_wheel, current_time); + flgs = erts_atomic32_read_acqb(&ssi->flags); + current_time = erts_get_monotonic_time(esdp); + if (current_time >= erts_next_timeout_time(esdp->next_tmo_ref)) { + if (!thr_prgr_active) { + erts_thr_progress_active(erts_thr_prgr_data(esdp), thr_prgr_active = 1); + sched_wall_time_change(esdp, 1); } + erts_bump_timers(esdp->timer_wheel, current_time); } } else { @@ -3321,19 +3417,37 @@ scheduler_wait(int *fcalls, ErtsSchedulerData *esdp, ErtsRunQueue *rq) } if (do_timeout) { if (!thr_prgr_active) { - erts_thr_progress_active(esdp, thr_prgr_active = 1); + erts_thr_progress_active(erts_thr_prgr_data(esdp), thr_prgr_active = 1); sched_wall_time_change(esdp, 1); } } - else { + else if (!ERTS_SCHEDULER_IS_DIRTY(esdp) && prepare_for_sys_schedule()) { + /* We sleep in check_io, only for normal schedulers */ + if (thr_prgr_active) { + erts_thr_progress_active(erts_thr_prgr_data(esdp), thr_prgr_active = 0); + sched_wall_time_change(esdp, 0); + } + flgs = sched_spin_wait(ssi, 0); + if (flgs & ERTS_SSI_FLG_SLEEPING) { + ASSERT(flgs & ERTS_SSI_FLG_WAITING); + flgs = sched_set_sleeptype(ssi, ERTS_SSI_FLG_POLL_SLEEPING); + if (flgs & ERTS_SSI_FLG_SLEEPING) { + ASSERT(flgs & ERTS_SSI_FLG_POLL_SLEEPING); + ASSERT(flgs & ERTS_SSI_FLG_WAITING); + erts_check_io(ssi->psi, timeout_time); + current_time = erts_get_monotonic_time(esdp); + } + } + *fcalls = 0; + clear_sys_scheduling(); + } else { if (!ERTS_SCHEDULER_IS_DIRTY(esdp)) { if (thr_prgr_active) { - erts_thr_progress_active(esdp, thr_prgr_active = 0); + erts_thr_progress_active(erts_thr_prgr_data(esdp), thr_prgr_active = 0); sched_wall_time_change(esdp, 0); } - erts_thr_progress_prepare_wait(esdp); + erts_thr_progress_prepare_wait(erts_thr_prgr_data(esdp)); } - flgs = sched_spin_wait(ssi, spincount); if (flgs & ERTS_SSI_FLG_SLEEPING) { ASSERT(flgs & ERTS_SSI_FLG_WAITING); @@ -3363,7 +3477,7 @@ scheduler_wait(int *fcalls, ErtsSchedulerData *esdp, ErtsRunQueue *rq) } } if (!ERTS_SCHEDULER_IS_DIRTY(esdp)) - erts_thr_progress_finalize_wait(esdp); + erts_thr_progress_finalize_wait(erts_thr_prgr_data(esdp)); } if (!ERTS_SCHEDULER_IS_DIRTY(esdp) && current_time >= timeout_time) erts_bump_timers(esdp->timer_wheel, current_time); @@ -3389,10 +3503,30 @@ scheduler_wait(int *fcalls, ErtsSchedulerData *esdp, ErtsRunQueue *rq) (ERTS_SSI_FLG_SUSPENDED | ERTS_SSI_FLG_MSB_EXEC)); - if (ERTS_SCHEDULER_IS_DIRTY(esdp)) + if (ERTS_SCHEDULER_IS_DIRTY(esdp)) { dirty_sched_wall_time_change(esdp, working = 1); + erts_spin_lock(&rq->sleepers.lock); + ERTS_HDBG_CHK_SLEEP_LIST(&rq->sleepers, 0, ssi->next ? ssi : NULL, NULL); + if (ssi->next) { /* Still in list... */ + if (ssi->next == ssi) { + ASSERT(rq->sleepers.list == ssi); + ASSERT(ssi->prev == ssi); + rq->sleepers.list = NULL; + } + else { + ASSERT(ssi->prev != ssi); + if (rq->sleepers.list == ssi) + rq->sleepers.list = ssi->next; + ssi->prev->next = ssi->next; + ssi->next->prev = ssi->prev; + } + ssi->next = ssi->prev = NULL; + } + ERTS_HDBG_CHK_SLEEP_LIST(&rq->sleepers, 0, NULL, ssi); + erts_spin_unlock(&rq->sleepers.lock); + } else if (!thr_prgr_active) { - erts_thr_progress_active(esdp, thr_prgr_active = 1); + erts_thr_progress_active(erts_thr_prgr_data(esdp), thr_prgr_active = 1); sched_wall_time_change(esdp, 1); } @@ -3478,56 +3612,44 @@ wake_scheduler(ErtsRunQueue *rq) } static void -wake_dirty_schedulers(ErtsRunQueue *rq, int one) +wake_dirty_scheduler(ErtsRunQueue *rq) { - ErtsSchedulerSleepInfo *ssi; + ErtsSchedulerSleepInfo *lo_ssi, *fo_ssi; ErtsSchedulerSleepList *sl; ASSERT(ERTS_RUNQ_IX_IS_DIRTY(rq->ix)); sl = &rq->sleepers; erts_spin_lock(&sl->lock); - ssi = sl->list; - if (!ssi) { + ERTS_HDBG_CHK_SLEEP_LIST(&rq->sleepers, 0, NULL, NULL); + lo_ssi = sl->list; + if (!lo_ssi) { erts_spin_unlock(&sl->lock); - if (one) - wake_scheduler(rq); - } else if (one) { + wake_scheduler(rq); + } + else { erts_aint32_t flgs; - if (ssi->prev) - ssi->prev->next = ssi->next; - else { - ASSERT(sl->list == ssi); - sl->list = ssi->next; + fo_ssi = lo_ssi->next; + ASSERT(fo_ssi->prev == lo_ssi); + if (fo_ssi == lo_ssi) { + ASSERT(lo_ssi->prev == lo_ssi); + sl->list = NULL; + } + else { + ASSERT(lo_ssi->prev != lo_ssi); + lo_ssi->next = fo_ssi->next; + fo_ssi->next->prev = fo_ssi->prev; } - if (ssi->next) - ssi->next->prev = ssi->prev; - - erts_spin_unlock(&sl->lock); - - ERTS_THR_MEMORY_BARRIER; - flgs = ssi_flags_set_wake(ssi); - erts_sched_finish_poke(ssi, flgs); - } else { - sl->list = NULL; + fo_ssi->next = fo_ssi->prev = NULL; + ERTS_HDBG_CHK_SLEEP_LIST(&rq->sleepers, 0, NULL, fo_ssi); erts_spin_unlock(&sl->lock); ERTS_THR_MEMORY_BARRIER; - do { - ErtsSchedulerSleepInfo *wake_ssi = ssi; - ssi = ssi->next; - erts_sched_finish_poke(wake_ssi, ssi_flags_set_wake(wake_ssi)); - } while (ssi); + flgs = ssi_flags_set_wake(fo_ssi); + erts_sched_finish_poke(fo_ssi, flgs); } } -static void -wake_dirty_scheduler(ErtsRunQueue *rq) -{ - wake_dirty_schedulers(rq, 1); -} - - #define ERTS_NO_USED_RUNQS_SHIFT 16 #define ERTS_NO_RUNQS_MASK 0xffffU @@ -4022,9 +4144,7 @@ schedule_bound_processes(ErtsRunQueue *rq, static ERTS_INLINE void clear_proc_dirty_queue_bit(Process *p, ErtsRunQueue *rq, int prio_bit) { -#ifdef DEBUG erts_aint32_t old; -#endif erts_aint32_t qb = prio_bit; if (rq == ERTS_DIRTY_CPU_RUNQ) qb <<= ERTS_PDSFLGS_IN_CPU_PRQ_MASK_OFFSET; @@ -4032,13 +4152,8 @@ clear_proc_dirty_queue_bit(Process *p, ErtsRunQueue *rq, int prio_bit) ASSERT(rq == ERTS_DIRTY_IO_RUNQ); qb <<= ERTS_PDSFLGS_IN_IO_PRQ_MASK_OFFSET; } -#ifdef DEBUG - old = (int) -#else - (void) -#endif - erts_atomic32_read_band_mb(&p->dirty_state, ~qb); - ASSERT(old & qb); + old = (int) erts_atomic32_read_band_mb(&p->dirty_state, ~qb); + ASSERT(old & qb); (void)old; } @@ -5118,7 +5233,6 @@ erts_fprintf(stderr, "--------------------------------\n"); rq->out_of_work_count = 0; (void) ERTS_RUNQ_FLGS_READ_BSET(rq, ERTS_RUNQ_FLGS_MIGRATION_INFO, flags); - rq->max_len = erts_atomic32_read_dirty(&rq->len); for (pix = 0; pix < ERTS_NO_PRIO_LEVELS; pix++) { ErtsRunQueueInfo *rqi; @@ -5557,7 +5671,6 @@ erts_sched_set_busy_wait_threshold(ErtsSchedType sched_type, char *str) return EINVAL; } - params->sys_schedule = sys_sched; params->tse = sys_sched * ERTS_SCHED_TSE_SLEEP_SPINCOUNT_FACT; params->aux_work = sys_sched * aux_work_fact; @@ -5768,6 +5881,10 @@ erts_init_scheduling(int no_schedulers, int no_schedulers_online, int no_poll_th size_runqs = sizeof(ErtsAlignedRunQueue) * tot_rqs; erts_aligned_run_queues = erts_alloc_permanent_cache_aligned(ERTS_ALC_T_RUNQS, size_runqs); +#if ERTS_POLL_USE_SCHEDULER_POLLING + erts_atomic32_init_nob(&doing_sys_schedule, 0); + erts_atomic32_init_nob(&function_calls, 0); +#endif erts_atomic32_init_nob(&no_empty_run_queues, 0); erts_no_run_queues = n; @@ -5882,6 +5999,8 @@ erts_init_scheduling(int no_schedulers, int no_schedulers_online, int no_poll_th for (ix = 0; ix < no_dirty_cpu_schedulers; ix++) { ErtsSchedulerSleepInfo *ssi = &aligned_dirty_cpu_sched_sleep_info[ix].ssi; erts_atomic32_init_nob(&ssi->flags, 0); + ssi->next = NULL; + ssi->prev = NULL; ssi->event = NULL; /* initialized in sched_dirty_cpu_thread_func */ erts_atomic32_init_nob(&ssi->aux_work, 0); } @@ -5892,6 +6011,8 @@ erts_init_scheduling(int no_schedulers, int no_schedulers_online, int no_poll_th for (ix = 0; ix < no_dirty_io_schedulers; ix++) { ErtsSchedulerSleepInfo *ssi = &aligned_dirty_io_sched_sleep_info[ix].ssi; erts_atomic32_init_nob(&ssi->flags, 0); + ssi->next = NULL; + ssi->prev = NULL; ssi->event = NULL; /* initialized in sched_dirty_io_thread_func */ erts_atomic32_init_nob(&ssi->aux_work, 0); } @@ -6405,8 +6526,7 @@ schedule_out_process(ErtsRunQueue *c_rq, erts_aint32_t state, Process *p, n &= ~running_flgs; if ((!!(a & (ERTS_PSFLG_ACTIVE_SYS|ERTS_PSFLG_DIRTY_ACTIVE_SYS)) - | ((a & (ERTS_PSFLG_ACTIVE|ERTS_PSFLG_SUSPENDED)) == ERTS_PSFLG_ACTIVE)) - & !(a & ERTS_PSFLG_FREE)) { + | ((a & (ERTS_PSFLG_ACTIVE|ERTS_PSFLG_SUSPENDED)) == ERTS_PSFLG_ACTIVE))) { enqueue = check_enqueue_in_prio_queue(p, &enq_prio, &n, a); } a = erts_atomic32_cmpxchg_mb(&p->state, n, e); @@ -6441,7 +6561,6 @@ schedule_out_process(ErtsRunQueue *c_rq, erts_aint32_t state, Process *p, else { Process* sched_p; - ASSERT(!(n & ERTS_PSFLG_FREE)); ASSERT(!(n & ERTS_PSFLG_SUSPENDED) || (n & (ERTS_PSFLG_ACTIVE_SYS | ERTS_PSFLG_DIRTY_ACTIVE_SYS))); @@ -6571,8 +6690,8 @@ change_proc_schedule_state(Process *p, enqueue = ERTS_ENQUEUE_NOT; - if (a & ERTS_PSFLG_FREE) - break; /* We don't want to schedule free processes... */ + if ((a & (ERTS_PSFLG_FREE|ERTS_PSFLG_ACTIVE)) == ERTS_PSFLG_FREE) + break; /* If free and not active, do not schedule */ if (clear_state_flags) n &= ~clear_state_flags; @@ -7098,8 +7217,7 @@ schdlr_sspnd_resume_procs(ErtsSchedType sched_type, while (resume->msb.chngrs) { ErtsProcList *plp = resume->msb.chngrs; resume->msb.chngrs = plp->next; - schdlr_sspnd_resume_proc(sched_type, - plp->pid); + schdlr_sspnd_resume_proc(sched_type, plp->u.pid); proclist_destroy(plp); } } @@ -7175,9 +7293,7 @@ msb_scheduler_type_switch(ErtsSchedType sched_type, Uint32 nrml_prio, dcpu_prio, dio_prio; ErtsSchedType exec_type; ErtsRunQueue *exec_rq; -#ifdef DEBUG erts_aint32_t dbg_val; -#endif ASSERT(schdlr_sspnd.msb.ongoing); @@ -7292,16 +7408,12 @@ msb_scheduler_type_switch(ErtsSchedType sched_type, * Suspend this scheduler and wake up scheduler * number one of another type... */ -#ifdef DEBUG dbg_val = -#else - (void) -#endif erts_atomic32_read_bset_mb(&esdp->ssi->flags, (ERTS_SSI_FLG_SUSPENDED | ERTS_SSI_FLG_MSB_EXEC), ERTS_SSI_FLG_SUSPENDED); - ASSERT(dbg_val & ERTS_SSI_FLG_MSB_EXEC); + ASSERT(dbg_val & ERTS_SSI_FLG_MSB_EXEC); (void)dbg_val; switch (exec_type) { case ERTS_SCHED_NORMAL: @@ -7319,11 +7431,7 @@ msb_scheduler_type_switch(ErtsSchedType sched_type, break; } -#ifdef DEBUG dbg_val = -#else - (void) -#endif erts_atomic32_read_bset_mb(&exec_rq->scheduler->ssi->flags, (ERTS_SSI_FLG_SUSPENDED | ERTS_SSI_FLG_MSB_EXEC), @@ -7411,7 +7519,13 @@ suspend_scheduler(ErtsSchedulerData *esdp) return; } +#ifdef HARDDEBUG + if (sched_type != ERTS_SCHED_NORMAL) + ERTS_HDBG_CHK_SLEEP_LIST(&esdp->run_queue->sleepers, !0, NULL, ssi); +#endif + if (erts_atomic32_read_nob(&ssi->flags) & ERTS_SSI_FLG_MSB_EXEC) { + ASSERT(no == 1); if (!msb_scheduler_type_switch(sched_type, esdp, no)) return; @@ -7535,7 +7649,7 @@ suspend_scheduler(ErtsSchedulerData *esdp) else { schdlr_sspnd.changer = am_true; /* change right in transit */ /* resume process that is queued for next change... */ - resume.onln.nxt = plp->pid; + resume.onln.nxt = plp->u.pid; ASSERT(is_internal_pid(resume.onln.nxt)); } } @@ -7565,7 +7679,8 @@ suspend_scheduler(ErtsSchedulerData *esdp) if (aux_work|evacuate) { if (!thr_prgr_active) { - erts_thr_progress_active(esdp, thr_prgr_active = 1); + erts_thr_progress_active(erts_thr_prgr_data(esdp), + thr_prgr_active = 1); sched_wall_time_change(esdp, 1); } if (aux_work) @@ -7573,8 +7688,8 @@ suspend_scheduler(ErtsSchedulerData *esdp) aux_work, 1); - if (aux_work && erts_thr_progress_update(esdp)) - erts_thr_progress_leader_update(esdp); + if (aux_work && erts_thr_progress_update(erts_thr_prgr_data(esdp))) + erts_thr_progress_leader_update(erts_thr_prgr_data(esdp)); if (evacuate) { erts_runq_lock(esdp->run_queue); evacuate_run_queue(esdp->run_queue, &sbp); @@ -7593,18 +7708,18 @@ suspend_scheduler(ErtsSchedulerData *esdp) if (!aux_work && current_time < timeout_time) { /* go to sleep... */ if (thr_prgr_active) { - erts_thr_progress_active(esdp, thr_prgr_active = 0); + erts_thr_progress_active(erts_thr_prgr_data(esdp), thr_prgr_active = 0); sched_wall_time_change(esdp, 0); } - erts_thr_progress_prepare_wait(NULL); + erts_thr_progress_prepare_wait(erts_thr_prgr_data(NULL)); suspend_normal_scheduler_sleep(esdp); - erts_thr_progress_finalize_wait(NULL); + erts_thr_progress_finalize_wait(erts_thr_prgr_data(NULL)); current_time = erts_get_monotonic_time(esdp); } if (current_time >= timeout_time) { if (!thr_prgr_active) { - erts_thr_progress_active(esdp, thr_prgr_active = 1); + erts_thr_progress_active(erts_thr_prgr_data(esdp), thr_prgr_active = 1); sched_wall_time_change(esdp, 1); } erts_bump_timers(esdp->timer_wheel, current_time); @@ -7661,7 +7776,7 @@ suspend_scheduler(ErtsSchedulerData *esdp) profile_scheduler(make_small(esdp->no), am_active); if (!thr_prgr_active) { - erts_thr_progress_active(esdp, thr_prgr_active = 1); + erts_thr_progress_active(erts_thr_prgr_data(esdp), thr_prgr_active = 1); sched_wall_time_change(esdp, 1); } } @@ -7754,7 +7869,7 @@ abort_sched_onln_chng_waitq(Process *p) proclist_destroy(plp); plp = erts_proclist_peek_first(schdlr_sspnd.chngq); if (plp) - resume = plp->pid; + resume = plp->u.pid; else schdlr_sspnd.changer = am_false; } @@ -8020,7 +8135,8 @@ done: ErtsSchedSuspendResult erts_block_multi_scheduling(Process *p, ErtsProcLocks plocks, int on, int normal, int all) { - int resume_proc, ix, res, have_unlocked_plocks = 0; + ErtsSchedSuspendResult res; + int resume_proc, ix, have_unlocked_plocks = 0; ErtsProcList *plp; ErtsMultiSchedulingBlock *msbp; erts_aint32_t chng_flg; @@ -8253,10 +8369,10 @@ erts_multi_scheduling_blockers(Process *p, int normal) plp1; plp1 = erts_proclist_peek_next(msbp->blckrs, plp1)) { for (plp2 = erts_proclist_peek_first(msbp->blckrs); - plp2->pid != plp1->pid; + plp2->u.pid != plp1->u.pid; plp2 = erts_proclist_peek_next(msbp->blckrs, plp2)); if (plp2 == plp1) { - res = CONS(hp, plp1->pid, res); + res = CONS(hp, plp1->u.pid, res); hp += 2; } /* else: already in result list */ @@ -8296,6 +8412,11 @@ sched_thread_func(void *vesdp) erts_msacc_init_thread("scheduler", no, 1); erts_thr_progress_register_managed_thread(esdp, &callbacks, 0); + +#if ERTS_POLL_USE_SCHEDULER_POLLING + esdp->ssi->psi = erts_create_pollset_thread(-1, NULL); +#endif + erts_alloc_register_scheduler(vesdp); #ifdef ERTS_ENABLE_LOCK_CHECK { @@ -8888,11 +9009,8 @@ erts_suspend(Process* c_p, ErtsProcLocks c_p_locks, Port *busy_port) suspend = 1; if (suspend) { -#ifdef DEBUG - int res = -#endif - suspend_process(c_p, c_p); - ASSERT(res); + int res = suspend_process(c_p, c_p); + ASSERT(res); (void)res; } if (!(c_p_locks & ERTS_PROC_LOCK_STATUS)) @@ -8923,8 +9041,13 @@ erts_resume_processes(ErtsProcList *list) while (plp) { Process *proc; ErtsProcList *fplp; - ASSERT(is_internal_pid(plp->pid)); - proc = erts_pid2proc(NULL, 0, plp->pid, ERTS_PROC_LOCK_STATUS); + ASSERT(is_internal_pid(plp->u.pid) || is_CP((Eterm)plp->u.p)); + if (is_internal_pid(plp->u.pid)) + proc = erts_pid2proc(NULL, 0, plp->u.pid, ERTS_PROC_LOCK_STATUS); + else { + proc = plp->u.p; + erts_proc_lock(proc, ERTS_PROC_LOCK_STATUS); + } if (proc) { if (erts_proclist_same(plp, proc)) { resume_process(proc, ERTS_PROC_LOCK_STATUS); @@ -9079,6 +9202,9 @@ unlock_lock_rq(int pre_free, void *vrq) } +static void trace_schedule_in(Process *p, erts_aint32_t state); +static void trace_schedule_out(Process *p, erts_aint32_t state); + /* * schedule() is called from BEAM (process_main()) or HiPE * (hipe_mode_switch()) when the current process is to be @@ -9102,7 +9228,7 @@ Process *erts_schedule(ErtsSchedulerData *esdp, Process *p, int calls) Process *proxy_p = NULL; ErtsRunQueue *rq; int context_reds; - int fcalls; + int fcalls = 0; int actual_reds; int reds; Uint32 flags; @@ -9176,6 +9302,10 @@ Process *erts_schedule(ErtsSchedulerData *esdp, Process *p, int calls) reds = ERTS_PROC_MIN_CONTEXT_SWITCH_REDS_COST; esdp->virtual_reds = 0; +#if ERTS_POLL_USE_SCHEDULER_POLLING + fcalls = (int) erts_atomic32_add_read_acqb(&function_calls, reds); +#endif + ASSERT(esdp && esdp == erts_get_scheduler_data()); rq = erts_get_runq_current(esdp); @@ -9184,22 +9314,8 @@ Process *erts_schedule(ErtsSchedulerData *esdp, Process *p, int calls) state = erts_atomic32_read_nob(&p->state); - if (IS_TRACED(p)) { - if (IS_TRACED_FL(p, F_TRACE_CALLS) && !(state & ERTS_PSFLG_FREE)) - erts_schedule_time_break(p, ERTS_BP_CALL_TIME_SCHEDULE_OUT); - if ((state & (ERTS_PSFLG_FREE|ERTS_PSFLG_EXITING)) == ERTS_PSFLG_EXITING) { - if (ARE_TRACE_FLAGS_ON(p, F_TRACE_SCHED_EXIT)) - trace_sched(p, ERTS_PROC_LOCK_MAIN, - ((state & ERTS_PSFLG_FREE) - ? am_out_exited - : am_out_exiting)); - } - else { - if (ARE_TRACE_FLAGS_ON(p, F_TRACE_SCHED) || - ARE_TRACE_FLAGS_ON(p, F_TRACE_SCHED_PROCS)) - trace_sched(p, ERTS_PROC_LOCK_MAIN, am_out); - } - } + if (IS_TRACED(p)) + trace_schedule_out(p, state); erts_proc_lock(p, ERTS_PROC_LOCK_STATUS|ERTS_PROC_LOCK_TRACE); @@ -9324,12 +9440,12 @@ Process *erts_schedule(ErtsSchedulerData *esdp, Process *p, int calls) } } - leader_update = erts_thr_progress_update(esdp); + leader_update = erts_thr_progress_update(erts_thr_prgr_data(esdp)); aux_work = erts_atomic32_read_acqb(&esdp->ssi->aux_work); if (aux_work | leader_update) { erts_runq_unlock(rq); if (leader_update) - erts_thr_progress_leader_update(esdp); + erts_thr_progress_leader_update(erts_thr_prgr_data(esdp)); if (aux_work) handle_aux_work(&esdp->aux_work_data, aux_work, 0); erts_runq_lock(rq); @@ -9406,7 +9522,33 @@ Process *erts_schedule(ErtsSchedulerData *esdp, Process *p, int calls) non_empty_runq(rq); goto check_activities_to_run; - } + } else if (is_normal_sched && + fcalls > (2 * context_reds) && + prepare_for_sys_schedule()) { + ErtsMonotonicTime current_time; + /* + * Schedule system-level activities. + */ + + ERTS_MSACC_PUSH_STATE_CACHED_M(); + + erts_runq_unlock(rq); + + ERTS_MSACC_SET_STATE_CACHED_M(ERTS_MSACC_STATE_CHECK_IO); + LTTNG2(scheduler_poll, esdp->no, 1); + + erts_check_io(esdp->ssi->psi, ERTS_POLL_NO_TIMEOUT); + ERTS_MSACC_POP_STATE_M(); + + current_time = erts_get_monotonic_time(esdp); + if (current_time >= erts_next_timeout_time(esdp->next_tmo_ref)) + erts_bump_timers(esdp->timer_wheel, current_time); + + erts_runq_lock(rq); + fcalls = 0; + clear_sys_scheduling(); + goto continue_check_activities_to_run; + } if (flags & ERTS_RUNQ_FLG_MISC_OP) exec_misc_ops(rq); @@ -9497,7 +9639,8 @@ Process *erts_schedule(ErtsSchedulerData *esdp, Process *p, int calls) while (1) { erts_aint32_t exp, new; - int run_process; + int run_process, not_running, exiting_on_normal_sched, + not_suspended, not_exiting_on_dirty_sched; new = exp = state; new &= psflg_band_mask; /* @@ -9506,29 +9649,33 @@ Process *erts_schedule(ErtsSchedulerData *esdp, Process *p, int calls) * scheduler, and not suspended (and not in a * state where suspend should be ignored). */ - run_process = (((!(state & (ERTS_PSFLG_RUNNING - | ERTS_PSFLG_RUNNING_SYS - | ERTS_PSFLG_DIRTY_RUNNING - | ERTS_PSFLG_DIRTY_RUNNING_SYS - | ERTS_PSFLG_FREE))) - | (((state & (ERTS_PSFLG_RUNNING - - | ERTS_PSFLG_FREE - | ERTS_PSFLG_RUNNING_SYS - | ERTS_PSFLG_DIRTY_RUNNING_SYS - | ERTS_PSFLG_EXITING)) - == ERTS_PSFLG_EXITING) - & (!!is_normal_sched)) - ) - & ((state & (ERTS_PSFLG_SUSPENDED - | ERTS_PSFLG_EXITING - | ERTS_PSFLG_FREE - | ERTS_PSFLG_ACTIVE_SYS - | ERTS_PSFLG_DIRTY_ACTIVE_SYS)) - != ERTS_PSFLG_SUSPENDED) - & (!(state & ERTS_PSFLG_EXITING) - | (!!is_normal_sched)) - ); + not_running = !(state & (ERTS_PSFLG_RUNNING + | ERTS_PSFLG_RUNNING_SYS + | ERTS_PSFLG_DIRTY_RUNNING + | ERTS_PSFLG_DIRTY_RUNNING_SYS + | ERTS_PSFLG_FREE)); + exiting_on_normal_sched = + ((state & (ERTS_PSFLG_RUNNING + | ERTS_PSFLG_ACTIVE + | ERTS_PSFLG_RUNNING_SYS + | ERTS_PSFLG_DIRTY_RUNNING_SYS + | ERTS_PSFLG_EXITING)) + == (ERTS_PSFLG_EXITING|ERTS_PSFLG_ACTIVE)) + & (!!is_normal_sched); + + + not_suspended = ((state & (ERTS_PSFLG_SUSPENDED + | ERTS_PSFLG_EXITING + | ERTS_PSFLG_FREE + | ERTS_PSFLG_ACTIVE_SYS + | ERTS_PSFLG_DIRTY_ACTIVE_SYS)) + != ERTS_PSFLG_SUSPENDED); + + not_exiting_on_dirty_sched = !(state & ERTS_PSFLG_EXITING) | (!!is_normal_sched); + + run_process = (not_running | exiting_on_normal_sched) + & not_suspended + & not_exiting_on_dirty_sched; if (run_process) { if (state & (ERTS_PSFLG_ACTIVE_SYS @@ -9610,6 +9757,8 @@ Process *erts_schedule(ErtsSchedulerData *esdp, Process *p, int calls) /* Migrate to dirty scheduler... */ sunlock_sched_out_proc: erts_proc_unlock(p, ERTS_PROC_LOCK_STATUS); + if (IS_TRACED(p)) + trace_schedule_in(p, state); goto sched_out_proc; } } @@ -9643,29 +9792,14 @@ Process *erts_schedule(ErtsSchedulerData *esdp, Process *p, int calls) erts_proc_unlock(p, ERTS_PROC_LOCK_STATUS); - /* Clear tracer if it has been removed */ - if (IS_TRACED(p) && erts_is_tracer_proc_enabled( - p, ERTS_PROC_LOCK_MAIN, &p->common)) { - - if (state & ERTS_PSFLG_EXITING) { - if (ARE_TRACE_FLAGS_ON(p, F_TRACE_SCHED_EXIT)) - trace_sched(p, ERTS_PROC_LOCK_MAIN, am_in_exiting); - } - else { - if (ARE_TRACE_FLAGS_ON(p, F_TRACE_SCHED) || - ARE_TRACE_FLAGS_ON(p, F_TRACE_SCHED_PROCS)) - trace_sched(p, ERTS_PROC_LOCK_MAIN, am_in); - } - if (IS_TRACED_FL(p, F_TRACE_CALLS)) { - erts_schedule_time_break(p, ERTS_BP_CALL_TIME_SCHEDULE_IN); - } - } + if (IS_TRACED(p)) + trace_schedule_in(p, state); if (is_normal_sched) { if (state & ERTS_PSFLG_RUNNING_SYS) { if (state & (ERTS_PSFLG_SIG_Q|ERTS_PSFLG_SIG_IN_Q)) { int local_only = (!!(p->flags & F_LOCAL_SIGS_ONLY) - & !(state & ERTS_PSFLG_SUSPENDED)); + & !(state & (ERTS_PSFLG_SUSPENDED|ERTS_PSFLGS_DIRTY_WORK))); if (!local_only | !!(state & ERTS_PSFLG_SIG_Q)) { int sig_reds; /* @@ -9823,6 +9957,50 @@ Process *erts_schedule(ErtsSchedulerData *esdp, Process *p, int calls) } } +static void +trace_schedule_in(Process *p, erts_aint32_t state) +{ + ASSERT(IS_TRACED(p)); + ERTS_LC_ASSERT(erts_proc_lc_my_proc_locks(p) == ERTS_PROC_LOCK_MAIN); + + /* Clear tracer if it has been removed */ + if (erts_is_tracer_proc_enabled(p, ERTS_PROC_LOCK_MAIN, &p->common)) { + + if (state & ERTS_PSFLG_EXITING && p->u.terminate) { + if (ARE_TRACE_FLAGS_ON(p, F_TRACE_SCHED_EXIT)) + trace_sched(p, ERTS_PROC_LOCK_MAIN, am_in_exiting); + } + else { + if (ARE_TRACE_FLAGS_ON(p, F_TRACE_SCHED) || + ARE_TRACE_FLAGS_ON(p, F_TRACE_SCHED_PROCS)) + trace_sched(p, ERTS_PROC_LOCK_MAIN, am_in); + } + if (IS_TRACED_FL(p, F_TRACE_CALLS)) + erts_schedule_time_break(p, ERTS_BP_CALL_TIME_SCHEDULE_IN); + } + +} + +static void +trace_schedule_out(Process *p, erts_aint32_t state) +{ + ASSERT(IS_TRACED(p)); + ERTS_LC_ASSERT(erts_proc_lc_my_proc_locks(p) == ERTS_PROC_LOCK_MAIN); + + if (IS_TRACED_FL(p, F_TRACE_CALLS) && !(state & ERTS_PSFLG_FREE)) + erts_schedule_time_break(p, ERTS_BP_CALL_TIME_SCHEDULE_OUT); + + if (state & ERTS_PSFLG_EXITING && p->u.terminate) { + if (ARE_TRACE_FLAGS_ON(p, F_TRACE_SCHED_EXIT)) + trace_sched(p, ERTS_PROC_LOCK_MAIN, am_out_exiting); + } + else { + if (ARE_TRACE_FLAGS_ON(p, F_TRACE_SCHED) || + ARE_TRACE_FLAGS_ON(p, F_TRACE_SCHED_PROCS)) + trace_sched(p, ERTS_PROC_LOCK_MAIN, am_out); + } +} + static int notify_sys_task_executed(Process *c_p, ErtsProcSysTask *st, Eterm st_result, int normal_sched) @@ -11889,12 +12067,91 @@ erts_set_self_exiting(Process *c_p, Eterm reason) add2runq(enqueue, enq_prio, c_p, state, NULL); } -void -erts_proc_exit_handle_monitor(ErtsMonitor *mon, void *vctxt) +static int +erts_proc_exit_handle_dist_monitor(ErtsMonitor *mon, void *vctxt, Sint reds) +{ + ErtsProcExitContext *ctxt = (ErtsProcExitContext *) vctxt; + Process *c_p = ctxt->c_p; + Eterm reason = ctxt->reason; + int code; + ErtsDSigSendContext ctx; + ErtsMonLnkDist *dist; + DistEntry *dep; + Eterm watcher; + ErtsMonitorData *mdp = NULL; + Eterm watched; + + ASSERT(erts_monitor_is_target(mon) && mon->type == ERTS_MON_TYPE_DIST_PROC); + + mdp = erts_monitor_to_data(mon); + + if (mon->flags & ERTS_ML_FLG_NAME) + watched = ((ErtsMonitorDataExtended *) mdp)->u.name; + else + watched = c_p->common.id; + ASSERT(is_internal_pid(watched) || is_atom(watched)); + + watcher = mon->other.item; + ASSERT(is_external_pid(watcher)); + dep = external_pid_dist_entry(watcher); + ASSERT(dep); + dist = ((ErtsMonitorDataExtended *) mdp)->dist; + ASSERT(dist); + + code = erts_dsig_prepare(&ctx, dep, c_p, ERTS_PROC_LOCK_MAIN, + ERTS_DSP_NO_LOCK, 0, 0, 1); + + ctx.reds = (Sint) (reds * TERM_TO_BINARY_LOOP_FACTOR); + + switch (code) { + case ERTS_DSIG_PREP_NOT_ALIVE: + case ERTS_DSIG_PREP_NOT_CONNECTED: + break; + case ERTS_DSIG_PREP_PENDING: + case ERTS_DSIG_PREP_CONNECTED: + if (dist->connection_id != ctx.connection_id) + break; + code = erts_dsig_send_m_exit(&ctx, + watcher, + watched, + mdp->ref, + reason); + switch (code) { + case ERTS_DSIG_SEND_CONTINUE: + erts_set_gc_state(c_p, 0); + ctxt->dist_state = erts_dsend_export_trap_context(c_p, &ctx); + /* fall-through */ + case ERTS_DSIG_SEND_YIELD: + break; + case ERTS_DSIG_SEND_OK: + break; + case ERTS_DSIG_SEND_TOO_LRG: + erts_set_gc_state(c_p, 1); + break; + default: + ASSERT(! "Invalid dsig send exit monitor result"); + break; + } + break; + default: + ASSERT(! "Invalid dsig prep exit monitor result"); + break; + } + if (!erts_monitor_dist_delete(&mdp->origin)) + erts_monitor_release(mon); + else + erts_monitor_release_both(mdp); + return reds - (ctx.reds / TERM_TO_BINARY_LOOP_FACTOR); +} + +int +erts_proc_exit_handle_monitor(ErtsMonitor *mon, void *vctxt, Sint reds) { - Process *c_p = ((ErtsProcExitContext *) vctxt)->c_p; - Eterm reason = ((ErtsProcExitContext *) vctxt)->reason; + ErtsProcExitContext *ctxt = (ErtsProcExitContext *) vctxt; + Process *c_p = ctxt->c_p; + Eterm reason = ctxt->reason; ErtsMonitorData *mdp = NULL; + int res = 1; if (erts_monitor_is_target(mon)) { /* We are being watched... */ @@ -11924,43 +12181,48 @@ erts_proc_exit_handle_monitor(ErtsMonitor *mon, void *vctxt) case ERTS_MON_TYPE_DIST_PROC: { ErtsMonLnkDist *dist; DistEntry *dep; - ErtsDSigData dsd; + ErtsDSigSendContext ctx; int code; Eterm watcher; Eterm watched; - mdp = erts_monitor_to_data(mon); - - if (mon->flags & ERTS_ML_FLG_NAME) - watched = ((ErtsMonitorDataExtended *) mdp)->u.name; - else - watched = c_p->common.id; - ASSERT(is_internal_pid(watched) || is_atom(watched)); + if (is_immed(reason)) { + mdp = erts_monitor_to_data(mon); - watcher = mon->other.item; - ASSERT(is_external_pid(watcher)); - dep = external_pid_dist_entry(watcher); - ASSERT(dep); - dist = ((ErtsMonitorDataExtended *) mdp)->dist; - ASSERT(dist); - code = erts_dsig_prepare(&dsd, dep, NULL, 0, - ERTS_DSP_NO_LOCK, 0, 0); - switch (code) { - case ERTS_DSIG_PREP_CONNECTED: - case ERTS_DSIG_PREP_PENDING: - if (dist->connection_id == dsd.connection_id) { - code = erts_dsig_send_m_exit(&dsd, - watcher, - watched, - mdp->ref, - reason); - ASSERT(code == ERTS_DSIG_SEND_OK); + if (mon->flags & ERTS_ML_FLG_NAME) + watched = ((ErtsMonitorDataExtended *) mdp)->u.name; + else + watched = c_p->common.id; + ASSERT(is_internal_pid(watched) || is_atom(watched)); + + watcher = mon->other.item; + ASSERT(is_external_pid(watcher)); + dep = external_pid_dist_entry(watcher); + ASSERT(dep); + dist = ((ErtsMonitorDataExtended *) mdp)->dist; + ASSERT(dist); + code = erts_dsig_prepare(&ctx, dep, NULL, 0, + ERTS_DSP_NO_LOCK, 1, 1, 0); + switch (code) { + case ERTS_DSIG_PREP_CONNECTED: + case ERTS_DSIG_PREP_PENDING: + if (dist->connection_id == ctx.connection_id) { + code = erts_dsig_send_m_exit(&ctx, + watcher, + watched, + mdp->ref, + reason); + ASSERT(code == ERTS_DSIG_SEND_OK); + } + default: + break; } - default: - break; + if (!erts_monitor_dist_delete(&mdp->origin)) + mdp = NULL; + } else { + erts_monitor_tree_insert(&ctxt->dist_monitors, mon); + return 1; } - if (!erts_monitor_dist_delete(&mdp->origin)) - mdp = NULL; break; } default: @@ -12002,7 +12264,7 @@ erts_proc_exit_handle_monitor(ErtsMonitor *mon, void *vctxt) case ERTS_MON_TYPE_DIST_PROC: { ErtsMonLnkDist *dist; DistEntry *dep; - ErtsDSigData dsd; + ErtsDSigSendContext ctx; int code; Eterm watched; @@ -12019,17 +12281,16 @@ erts_proc_exit_handle_monitor(ErtsMonitor *mon, void *vctxt) ASSERT(is_external_pid(watched)); dep = external_pid_dist_entry(watched); } - code = erts_dsig_prepare(&dsd, dep, NULL, 0, - ERTS_DSP_NO_LOCK, 0, 0); + code = erts_dsig_prepare(&ctx, dep, NULL, 0, + ERTS_DSP_NO_LOCK, 1, 1, 0); switch (code) { case ERTS_DSIG_PREP_CONNECTED: case ERTS_DSIG_PREP_PENDING: - if (dist->connection_id == dsd.connection_id) { - code = erts_dsig_send_demonitor(&dsd, + if (dist->connection_id == ctx.connection_id) { + code = erts_dsig_send_demonitor(&ctx, c_p->common.id, watched, - mdp->ref, - 1); + mdp->ref); ASSERT(code == ERTS_DSIG_SEND_OK); } default: @@ -12037,6 +12298,7 @@ erts_proc_exit_handle_monitor(ErtsMonitor *mon, void *vctxt) } if (!erts_monitor_dist_delete(&mdp->target)) mdp = NULL; + res = 100; break; } default: @@ -12049,11 +12311,84 @@ erts_proc_exit_handle_monitor(ErtsMonitor *mon, void *vctxt) erts_monitor_release_both(mdp); else if (mon) erts_monitor_release(mon); + return res; } -void -erts_proc_exit_handle_link(ErtsLink *lnk, void *vctxt) +static int +erts_proc_exit_handle_dist_link(ErtsLink *lnk, void *vctxt, Sint reds) +{ + ErtsProcExitContext *ctxt = (ErtsProcExitContext *) vctxt; + Process *c_p = ctxt->c_p; + Eterm reason = ctxt->reason; + int code; + ErtsDSigSendContext ctx; + ErtsMonLnkDist *dist; + DistEntry *dep; + ErtsLink *dlnk; + ErtsLinkData *ldp = NULL; + + ASSERT(lnk->type == ERTS_LNK_TYPE_DIST_PROC); + dlnk = erts_link_to_other(lnk, &ldp); + dist = ((ErtsLinkDataExtended *) ldp)->dist; + + ASSERT(is_external_pid(lnk->other.item)); + dep = external_pid_dist_entry(lnk->other.item); + + ASSERT(dep != erts_this_dist_entry); + + if (!erts_link_dist_delete(dlnk)) + ldp = NULL; + + code = erts_dsig_prepare(&ctx, dep, c_p, ERTS_PROC_LOCK_MAIN, + ERTS_DSP_NO_LOCK, 0, 0, 0); + + ctx.reds = (Sint) (reds * TERM_TO_BINARY_LOOP_FACTOR); + + switch (code) { + case ERTS_DSIG_PREP_NOT_ALIVE: + case ERTS_DSIG_PREP_NOT_CONNECTED: + break; + case ERTS_DSIG_PREP_PENDING: + case ERTS_DSIG_PREP_CONNECTED: + if (dist->connection_id != ctx.connection_id) + break; + code = erts_dsig_send_exit_tt(&ctx, + c_p->common.id, + lnk->other.item, + reason, + SEQ_TRACE_TOKEN(c_p)); + switch (code) { + case ERTS_DSIG_SEND_CONTINUE: + erts_set_gc_state(c_p, 0); + ctxt->dist_state = erts_dsend_export_trap_context(c_p, &ctx); + /* fall-through */ + case ERTS_DSIG_SEND_YIELD: + break; + case ERTS_DSIG_SEND_OK: + break; + case ERTS_DSIG_SEND_TOO_LRG: + erts_set_gc_state(c_p, 1); + break; + default: + ASSERT(! "Invalid dsig send exit monitor result"); + break; + } + break; + default: + ASSERT(! "Invalid dsig prep exit monitor result"); + break; + } + if (ldp) + erts_link_release_both(ldp); + else if (lnk) + erts_link_release(lnk); + return reds - (ctx.reds / TERM_TO_BINARY_LOOP_FACTOR); +} + +int +erts_proc_exit_handle_link(ErtsLink *lnk, void *vctxt, Sint reds) { + ErtsProcExitContext *ctxt = (ErtsProcExitContext *) vctxt; Process *c_p = ((ErtsProcExitContext *) vctxt)->c_p; Eterm reason = ((ErtsProcExitContext *) vctxt)->reason; ErtsLinkData *ldp = NULL; @@ -12084,32 +12419,40 @@ erts_proc_exit_handle_link(ErtsLink *lnk, void *vctxt) DistEntry *dep; ErtsMonLnkDist *dist; ErtsLink *dlnk; - ErtsDSigData dsd; + ErtsDSigSendContext ctx; int code; - dlnk = erts_link_to_other(lnk, &ldp); - dist = ((ErtsLinkDataExtended *) ldp)->dist; + if (is_immed(reason)) { + dlnk = erts_link_to_other(lnk, &ldp); + dist = ((ErtsLinkDataExtended *) ldp)->dist; - ASSERT(is_external_pid(lnk->other.item)); - dep = external_pid_dist_entry(lnk->other.item); + ASSERT(is_external_pid(lnk->other.item)); + dep = external_pid_dist_entry(lnk->other.item); - ASSERT(dep != erts_this_dist_entry); + ASSERT(dep != erts_this_dist_entry); - if (!erts_link_dist_delete(dlnk)) - ldp = NULL; + if (!erts_link_dist_delete(dlnk)) + ldp = NULL; - code = erts_dsig_prepare(&dsd, dep, c_p, 0, ERTS_DSP_NO_LOCK, 0, 0); - switch (code) { - case ERTS_DSIG_PREP_CONNECTED: - case ERTS_DSIG_PREP_PENDING: - if (dist->connection_id == dsd.connection_id) { - code = erts_dsig_send_exit_tt(&dsd, - c_p->common.id, - lnk->other.item, - reason, - SEQ_TRACE_TOKEN(c_p)); - ASSERT(code == ERTS_DSIG_SEND_OK); + code = erts_dsig_prepare(&ctx, dep, c_p, 0, ERTS_DSP_NO_LOCK, 1, 1, 0); + switch (code) { + case ERTS_DSIG_PREP_CONNECTED: + case ERTS_DSIG_PREP_PENDING: + if (dist->connection_id == ctx.connection_id) { + code = erts_dsig_send_exit_tt(&ctx, + c_p->common.id, + lnk->other.item, + reason, + SEQ_TRACE_TOKEN(c_p)); + ASSERT(code == ERTS_DSIG_SEND_OK); + } + break; + default: + break; } + } else { + erts_link_tree_insert(&ctxt->dist_links, lnk); + return 1; } break; } @@ -12122,6 +12465,7 @@ erts_proc_exit_handle_link(ErtsLink *lnk, void *vctxt) erts_link_release_both(ldp); else if (lnk) erts_link_release(lnk); + return 1; } /* this function fishishes a process and propagates exit messages - called @@ -12155,11 +12499,8 @@ erts_do_exit_process(Process* p, Eterm reason) set_self_exiting(p, reason, NULL, NULL, NULL); - if (IS_TRACED(p)) { - if (IS_TRACED_FL(p, F_TRACE_CALLS)) - erts_schedule_time_break(p, ERTS_BP_CALL_TIME_SCHEDULE_EXITING); - - } + if (IS_TRACED_FL(p, F_TRACE_CALLS)) + erts_schedule_time_break(p, ERTS_BP_CALL_TIME_SCHEDULE_EXITING); erts_trace_check_exiting(p->common.id); @@ -12174,288 +12515,444 @@ erts_do_exit_process(Process* p, Eterm reason) erts_proc_unlock(p, ERTS_PROC_LOCKS_ALL_MINOR); - if (IS_TRACED_FL(p,F_TRACE_PROCS)) + if (IS_TRACED_FL(p, F_TRACE_PROCS)) trace_proc(p, ERTS_PROC_LOCK_MAIN, p, am_exit, reason); - /* * p->u.initial of this process can *not* be used anymore; * will be overwritten by misc termination data. */ p->u.terminate = NULL; + BUMP_REDS(p, 100); + erts_continue_exit_process(p); } -void -erts_continue_exit_process(Process *p) -{ +enum continue_exit_phase { + ERTS_CONTINUE_EXIT_TIMERS, + ERTS_CONTINUE_EXIT_BLCKD_MSHED, + ERTS_CONTINUE_EXIT_BLCKD_NMSHED, + ERTS_CONTINUE_EXIT_USING_DB, + ERTS_CONTINUE_EXIT_CLEAN_SYS_TASKS, + ERTS_CONTINUE_EXIT_FREE, + ERTS_CONTINUE_EXIT_CLEAN_SYS_TASKS_AFTER, + ERTS_CONTINUE_EXIT_LINKS, + ERTS_CONTINUE_EXIT_MONITORS, + ERTS_CONTINUE_EXIT_LT_MONITORS, + ERTS_CONTINUE_EXIT_HANDLE_PROC_SIG, + ERTS_CONTINUE_EXIT_DIST_LINKS, + ERTS_CONTINUE_EXIT_DIST_MONITORS, + ERTS_CONTINUE_EXIT_DONE, +}; + +struct continue_exit_state { + enum continue_exit_phase phase; ErtsLink *links; ErtsMonitor *monitors; ErtsMonitor *lt_monitors; + Eterm reason; + ErtsProcExitContext pectxt; + DistEntry *dep; + void *yield_state; +}; + +void +erts_continue_exit_process(Process *p) +{ + struct continue_exit_state static_state, *trap_state = &static_state; ErtsProcLocks curr_locks = ERTS_PROC_LOCK_MAIN; - Eterm reason = p->fvalue; - DistEntry *dep = NULL; erts_aint32_t state; int delay_del_proc = 0; - ErtsProcExitContext pectxt; - + Sint reds = ERTS_BIF_REDS_LEFT(p); #ifdef DEBUG int yield_allowed = 1; #endif + if (p->u.terminate) { + trap_state = p->u.terminate; + } else { + trap_state->phase = ERTS_CONTINUE_EXIT_TIMERS; + trap_state->reason = p->fvalue; + trap_state->dep = NULL; + trap_state->yield_state = NULL; + } + ERTS_LC_ASSERT(ERTS_PROC_LOCK_MAIN == erts_proc_lc_my_proc_locks(p)); ASSERT(ERTS_PROC_IS_EXITING(p)); ASSERT(erts_proc_read_refc(p) > 0); - if (p->bif_timers) { - if (erts_cancel_bif_timers(p, &p->bif_timers, &p->u.terminate)) { - ASSERT(erts_proc_read_refc(p) > 0); - goto yield; - } - ASSERT(erts_proc_read_refc(p) > 0); - p->bif_timers = NULL; - } - - if (p->flags & F_SCHDLR_ONLN_WAITQ) - abort_sched_onln_chng_waitq(p); - - if (p->flags & F_HAVE_BLCKD_MSCHED) { - ErtsSchedSuspendResult ssr; - ssr = erts_block_multi_scheduling(p, ERTS_PROC_LOCK_MAIN, 0, 0, 1); - switch (ssr) { - case ERTS_SCHDLR_SSPND_YIELD_RESTART: - goto yield; - case ERTS_SCHDLR_SSPND_DONE_MSCHED_BLOCKED: - case ERTS_SCHDLR_SSPND_DONE_NMSCHED_BLOCKED: - case ERTS_SCHDLR_SSPND_YIELD_DONE_MSCHED_BLOCKED: - case ERTS_SCHDLR_SSPND_YIELD_DONE_NMSCHED_BLOCKED: - case ERTS_SCHDLR_SSPND_DONE: - case ERTS_SCHDLR_SSPND_YIELD_DONE: - p->flags &= ~F_HAVE_BLCKD_MSCHED; - break; - case ERTS_SCHDLR_SSPND_EINVAL: - default: - erts_exit(ERTS_ABORT_EXIT, "%s:%d: Internal error: %d\n", - __FILE__, __LINE__, (int) ssr); - } - } - if (p->flags & F_HAVE_BLCKD_NMSCHED) { - ErtsSchedSuspendResult ssr; - ssr = erts_block_multi_scheduling(p, ERTS_PROC_LOCK_MAIN, 0, 1, 1); - switch (ssr) { - case ERTS_SCHDLR_SSPND_YIELD_RESTART: - goto yield; - case ERTS_SCHDLR_SSPND_DONE_MSCHED_BLOCKED: - case ERTS_SCHDLR_SSPND_DONE_NMSCHED_BLOCKED: - case ERTS_SCHDLR_SSPND_YIELD_DONE_MSCHED_BLOCKED: - case ERTS_SCHDLR_SSPND_YIELD_DONE_NMSCHED_BLOCKED: - case ERTS_SCHDLR_SSPND_DONE: - case ERTS_SCHDLR_SSPND_YIELD_DONE: - p->flags &= ~F_HAVE_BLCKD_MSCHED; - break; - case ERTS_SCHDLR_SSPND_EINVAL: - default: - erts_exit(ERTS_ABORT_EXIT, "%s:%d: Internal error: %d\n", - __FILE__, __LINE__, (int) ssr); - } - } +restart: + switch (trap_state->phase) { + case ERTS_CONTINUE_EXIT_TIMERS: + if (p->bif_timers) { + reds = erts_cancel_bif_timers(p, &p->bif_timers, &trap_state->yield_state, reds); + if (reds <= 0) goto yield; + p->bif_timers = NULL; + } - if (p->flags & F_USING_DB) { - if (erts_db_process_exiting(p, ERTS_PROC_LOCK_MAIN)) - goto yield; - p->flags &= ~F_USING_DB; - } + if (p->flags & F_SCHDLR_ONLN_WAITQ) { + abort_sched_onln_chng_waitq(p); + reds -= 100; + } - erts_set_gc_state(p, 1); - state = erts_atomic32_read_acqb(&p->state); - if ((state & ERTS_PSFLG_SYS_TASKS) || p->dirty_sys_tasks) { - if (cleanup_sys_tasks(p, state, CONTEXT_REDS) >= CONTEXT_REDS/2) - goto yield; - } + trap_state->phase = ERTS_CONTINUE_EXIT_BLCKD_MSHED; + if (reds <= 0) goto yield; + case ERTS_CONTINUE_EXIT_BLCKD_MSHED: + + if (p->flags & F_HAVE_BLCKD_MSCHED) { + ErtsSchedSuspendResult ssr; + ssr = erts_block_multi_scheduling(p, ERTS_PROC_LOCK_MAIN, 0, 0, 1); + switch (ssr) { + case ERTS_SCHDLR_SSPND_DONE: + case ERTS_SCHDLR_SSPND_DONE_MSCHED_BLOCKED: + case ERTS_SCHDLR_SSPND_DONE_NMSCHED_BLOCKED: + p->flags &= ~F_HAVE_BLCKD_MSCHED; + break; + default: + erts_exit(ERTS_ABORT_EXIT, "%s:%d: Internal error: %d\n", + __FILE__, __LINE__, (int) ssr); + } + reds -= 100; + } -#ifdef DEBUG - erts_proc_lock(p, ERTS_PROC_LOCK_STATUS); - ASSERT(ERTS_PROC_GET_DELAYED_GC_TASK_QS(p) == NULL); - ASSERT(p->dirty_sys_tasks == NULL); - erts_proc_unlock(p, ERTS_PROC_LOCK_STATUS); -#endif + trap_state->phase = ERTS_CONTINUE_EXIT_BLCKD_NMSHED; + if (reds <= 0) goto yield; + case ERTS_CONTINUE_EXIT_BLCKD_NMSHED: + + if (p->flags & F_HAVE_BLCKD_NMSCHED) { + ErtsSchedSuspendResult ssr; + ssr = erts_block_multi_scheduling(p, ERTS_PROC_LOCK_MAIN, 0, 1, 1); + switch (ssr) { + case ERTS_SCHDLR_SSPND_DONE: + case ERTS_SCHDLR_SSPND_DONE_MSCHED_BLOCKED: + case ERTS_SCHDLR_SSPND_DONE_NMSCHED_BLOCKED: + p->flags &= ~F_HAVE_BLCKD_MSCHED; + break; + default: + erts_exit(ERTS_ABORT_EXIT, "%s:%d: Internal error: %d\n", + __FILE__, __LINE__, (int) ssr); + } + reds -= 100; + } - if (p->flags & F_USING_DDLL) { - erts_ddll_proc_dead(p, ERTS_PROC_LOCK_MAIN); - p->flags &= ~F_USING_DDLL; - } + trap_state->yield_state = NULL; + trap_state->phase = ERTS_CONTINUE_EXIT_USING_DB; + if (reds <= 0) goto yield; + case ERTS_CONTINUE_EXIT_USING_DB: - /* - * The registered name *should* be the last "erlang resource" to - * cleanup. - */ - if (p->common.u.alive.reg) { - (void) erts_unregister_name(p, ERTS_PROC_LOCK_MAIN, NULL, THE_NON_VALUE); - ASSERT(!p->common.u.alive.reg); - } + if (p->flags & F_USING_DB) { + if (erts_db_process_exiting(p, ERTS_PROC_LOCK_MAIN, &trap_state->yield_state)) + goto yield; + p->flags &= ~F_USING_DB; + } - if (IS_TRACED_FL(p, F_TRACE_SCHED_EXIT)) - trace_sched(p, curr_locks, am_out_exited); + erts_set_gc_state(p, 1); - erts_proc_lock(p, ERTS_PROC_LOCKS_ALL_MINOR); - curr_locks = ERTS_PROC_LOCKS_ALL; + trap_state->phase = ERTS_CONTINUE_EXIT_CLEAN_SYS_TASKS; + case ERTS_CONTINUE_EXIT_CLEAN_SYS_TASKS: + + state = erts_atomic32_read_acqb(&p->state); + if ((state & ERTS_PSFLG_SYS_TASKS) || p->dirty_sys_tasks) { + reds -= cleanup_sys_tasks(p, state, reds); + if (reds <= 0) goto yield; + } + + trap_state->phase = ERTS_CONTINUE_EXIT_FREE; + case ERTS_CONTINUE_EXIT_FREE: - /* - * From this point on we are no longer allowed to yield - * this process. - */ #ifdef DEBUG - yield_allowed = 0; + erts_proc_lock(p, ERTS_PROC_LOCK_STATUS); + ASSERT(ERTS_PROC_GET_DELAYED_GC_TASK_QS(p) == NULL); + ASSERT(p->dirty_sys_tasks == NULL); + erts_proc_unlock(p, ERTS_PROC_LOCK_STATUS); #endif - /* - * Note! The monitor and link fields will be overwritten - * by erts_ptab_delete_element() below. - */ - links = ERTS_P_LINKS(p); - monitors = ERTS_P_MONITORS(p); - lt_monitors = ERTS_P_LT_MONITORS(p); + if (p->flags & F_USING_DDLL) { + erts_ddll_proc_dead(p, ERTS_PROC_LOCK_MAIN); + p->flags &= ~F_USING_DDLL; + } - { - /* Do *not* use erts_get_runq_proc() */ - ErtsRunQueue *rq; - rq = erts_get_runq_current(erts_proc_sched_data(p)); + /* + * The registered name *should* be the last "erlang resource" to + * cleanup. + */ + if (p->common.u.alive.reg) { + (void) erts_unregister_name(p, ERTS_PROC_LOCK_MAIN, NULL, THE_NON_VALUE); + ASSERT(!p->common.u.alive.reg); + } - erts_runq_lock(rq); + erts_proc_lock(p, ERTS_PROC_LOCKS_ALL_MINOR); + curr_locks = ERTS_PROC_LOCKS_ALL; - ASSERT(p->scheduler_data); - ASSERT(p->scheduler_data->current_process == p); - ASSERT(p->scheduler_data->free_process == NULL); + /* + * Note! The monitor and link fields will be overwritten + * by erts_ptab_delete_element() below. + */ + trap_state->links = ERTS_P_LINKS(p); + trap_state->monitors = ERTS_P_MONITORS(p); + trap_state->lt_monitors = ERTS_P_LT_MONITORS(p); - p->scheduler_data->current_process = NULL; - p->scheduler_data->free_process = p; + { + /* Do *not* use erts_get_runq_proc() */ + ErtsRunQueue *rq; + rq = erts_get_runq_current(erts_proc_sched_data(p)); - /* Time of death! */ - erts_ptab_delete_element(&erts_proc, &p->common); + erts_runq_lock(rq); - erts_runq_unlock(rq); - } + ASSERT(p->scheduler_data); + ASSERT(p->scheduler_data->current_process == p); + ASSERT(p->scheduler_data->free_process == NULL); - /* - * All "erlang resources" have to be deallocated before this point, - * e.g. registered name, so monitoring and linked processes can - * be sure that all interesting resources have been deallocated - * when the monitors and/or links hit. - */ + /* Time of death! */ + erts_ptab_delete_element(&erts_proc, &p->common); - { - /* Inactivate and notify free */ - erts_aint32_t n, e, a = erts_atomic32_read_nob(&p->state); - int refc_inced = 0; - while (1) { - n = e = a; - ASSERT(a & ERTS_PSFLG_EXITING); - n |= ERTS_PSFLG_FREE; - n &= ~(ERTS_PSFLG_ACTIVE - | ERTS_PSFLG_ACTIVE_SYS - | ERTS_PSFLG_DIRTY_ACTIVE_SYS); - if ((n & ERTS_PSFLG_IN_RUNQ) && !refc_inced) { - erts_proc_inc_refc(p); - refc_inced = 1; - } - a = erts_atomic32_cmpxchg_mb(&p->state, n, e); - if (a == e) - break; - } + erts_runq_unlock(rq); + } - if (a & (ERTS_PSFLG_DIRTY_RUNNING - | ERTS_PSFLG_DIRTY_RUNNING_SYS)) { - p->flags |= F_DELAYED_DEL_PROC; - delay_del_proc = 1; - /* - * The dirty scheduler decrease refc - * when done with the process... - */ - } + /* + * All "erlang resources" have to be deallocated before this point, + * e.g. registered name, so monitoring and linked processes can + * be sure that all interesting resources have been deallocated + * when the monitors and/or links hit. + */ - if (refc_inced && !(n & ERTS_PSFLG_IN_RUNQ)) - erts_proc_dec_refc(p); - } + { + /* Inactivate and notify free */ + erts_aint32_t n, e, a = erts_atomic32_read_nob(&p->state); + int refc_inced = 0; + while (1) { + n = e = a; + ASSERT(a & ERTS_PSFLG_EXITING); + n |= ERTS_PSFLG_FREE; + if ((n & ERTS_PSFLG_IN_RUNQ) && !refc_inced) { + erts_proc_inc_refc(p); + refc_inced = 1; + } + a = erts_atomic32_cmpxchg_mb(&p->state, n, e); + if (a == e) + break; + } - dep = ((p->flags & F_DISTRIBUTION) - ? ERTS_PROC_SET_DIST_ENTRY(p, NULL) - : NULL); + if (refc_inced && !(n & ERTS_PSFLG_IN_RUNQ)) + erts_proc_dec_refc(p); + } + trap_state->dep = ((p->flags & F_DISTRIBUTION) + ? ERTS_PROC_SET_DIST_ENTRY(p, NULL) + : NULL); + + reds -= 50; - /* - * It might show up signal prio elevation tasks until we - * have entered free state. Cleanup such tasks now. - */ - state = erts_atomic32_read_acqb(&p->state); - if (!(state & ERTS_PSFLG_SYS_TASKS)) - erts_proc_unlock(p, ERTS_PROC_LOCKS_ALL); - else { erts_proc_unlock(p, ERTS_PROC_LOCKS_ALL_MINOR); + curr_locks = ERTS_PROC_LOCK_MAIN; + trap_state->phase = ERTS_CONTINUE_EXIT_CLEAN_SYS_TASKS_AFTER; + case ERTS_CONTINUE_EXIT_CLEAN_SYS_TASKS_AFTER: + /* + * It might show up signal prio elevation tasks until we + * have entered free state. Cleanup such tasks now. + */ - do { - (void) cleanup_sys_tasks(p, state, CONTEXT_REDS); - state = erts_atomic32_read_acqb(&p->state); - } while (state & ERTS_PSFLG_SYS_TASKS); - - erts_proc_unlock(p, ERTS_PROC_LOCK_MAIN); - } + state = erts_atomic32_read_acqb(&p->state); + if ((state & ERTS_PSFLG_SYS_TASKS) || p->dirty_sys_tasks) { + reds -= cleanup_sys_tasks(p, state, reds); + if (reds <= 0) goto yield; + } + + /* Needs to be unlocked for erts_do_net_exits to work?!? */ + // erts_proc_unlock(p, ERTS_PROC_LOCK_MAIN); #ifdef DEBUG - erts_proc_lock(p, ERTS_PROC_LOCK_STATUS); - ASSERT(p->sys_task_qs == NULL); - erts_proc_unlock(p, ERTS_PROC_LOCK_STATUS); + erts_proc_lock(p, ERTS_PROC_LOCK_STATUS); + ASSERT(p->sys_task_qs == NULL); + erts_proc_unlock(p, ERTS_PROC_LOCK_STATUS); #endif - if (dep) { - erts_do_net_exits(dep, (reason == am_kill) ? am_killed : reason); - erts_deref_dist_entry(dep); - } + if (trap_state->dep) { + erts_do_net_exits(trap_state->dep, + (trap_state->reason == am_kill) ? am_killed : trap_state->reason); + erts_deref_dist_entry(trap_state->dep); + } - pectxt.c_p = p; - pectxt.reason = reason; + trap_state->pectxt.c_p = p; + trap_state->pectxt.reason = trap_state->reason; + trap_state->pectxt.dist_links = NULL; + trap_state->pectxt.dist_monitors = NULL; + trap_state->pectxt.dist_state = NIL; + + erts_proc_lock(p, ERTS_PROC_LOCK_MSGQ); + + erts_proc_sig_fetch(p); + + erts_proc_unlock(p, ERTS_PROC_LOCK_MSGQ); + + trap_state->yield_state = NULL; + trap_state->phase = ERTS_CONTINUE_EXIT_LINKS; + if (reds <= 0) goto yield; + case ERTS_CONTINUE_EXIT_LINKS: + + reds = erts_link_tree_foreach_delete_yielding( + &trap_state->links, + erts_proc_exit_handle_link, + (void *) &trap_state->pectxt, + &trap_state->yield_state, + reds); + if (reds <= 0) + goto yield; + + ASSERT(!trap_state->links); + trap_state->yield_state = NULL; + trap_state->phase = ERTS_CONTINUE_EXIT_MONITORS; + case ERTS_CONTINUE_EXIT_MONITORS: + + reds = erts_monitor_tree_foreach_delete_yielding( + &trap_state->monitors, + erts_proc_exit_handle_monitor, + (void *) &trap_state->pectxt, + &trap_state->yield_state, + reds); + if (reds <= 0) + goto yield; + + ASSERT(!trap_state->monitors); + trap_state->yield_state = NULL; + trap_state->phase = ERTS_CONTINUE_EXIT_LT_MONITORS; + case ERTS_CONTINUE_EXIT_LT_MONITORS: + + reds = erts_monitor_list_foreach_delete_yielding( + &trap_state->lt_monitors, + erts_proc_exit_handle_monitor, + (void *) &trap_state->pectxt, + &trap_state->yield_state, + reds); + if (reds <= 0) + goto yield; + + ASSERT(!trap_state->lt_monitors); + trap_state->phase = ERTS_CONTINUE_EXIT_HANDLE_PROC_SIG; + case ERTS_CONTINUE_EXIT_HANDLE_PROC_SIG: { + Sint r = reds; + + if (!erts_proc_sig_handle_exit(p, &r)) + goto yield; + + reds -= r; + + trap_state->phase = ERTS_CONTINUE_EXIT_DIST_LINKS; + } + case ERTS_CONTINUE_EXIT_DIST_LINKS: { + + continue_dist_send: + if (is_not_nil(trap_state->pectxt.dist_state)) { + Binary* bin = erts_magic_ref2bin(trap_state->pectxt.dist_state); + ErtsDSigSendContext* ctx = (ErtsDSigSendContext*) ERTS_MAGIC_BIN_DATA(bin); + Sint initial_reds = (Sint) (ERTS_BIF_REDS_LEFT(p) * TERM_TO_BINARY_LOOP_FACTOR); + int result; + + ctx->reds = initial_reds; + result = erts_dsig_send(ctx); + + /* erts_dsig_send bumps reductions on the process in the ctx */ + reds = ERTS_BIF_REDS_LEFT(p); + + switch (result) { + case ERTS_DSIG_SEND_OK: + case ERTS_DSIG_SEND_TOO_LRG: /*SEND_SYSTEM_LIMIT*/ + case ERTS_DSIG_SEND_YIELD: /*SEND_YIELD_RETURN*/ + break; + case ERTS_DSIG_SEND_CONTINUE: { /*SEND_YIELD_CONTINUE*/ + goto yield; + } + } + erts_set_gc_state(p, 1); + trap_state->pectxt.dist_state = NIL; + if (reds <= 0) + goto yield; + goto restart; + } - erts_proc_lock(p, ERTS_PROC_LOCK_MAIN|ERTS_PROC_LOCK_MSGQ); + reds = erts_link_tree_foreach_delete_yielding( + &trap_state->pectxt.dist_links, + erts_proc_exit_handle_dist_link, + (void *) &trap_state->pectxt, + &trap_state->yield_state, + reds); + if (reds <= 0 || is_not_nil(trap_state->pectxt.dist_state)) + goto yield; + trap_state->phase = ERTS_CONTINUE_EXIT_DIST_MONITORS; + } + case ERTS_CONTINUE_EXIT_DIST_MONITORS: { + + if (is_not_nil(trap_state->pectxt.dist_state)) + goto continue_dist_send; + + reds = erts_monitor_tree_foreach_delete_yielding( + &trap_state->pectxt.dist_monitors, + erts_proc_exit_handle_dist_monitor, + (void *) &trap_state->pectxt, + &trap_state->yield_state, + reds); + if (reds <= 0 || is_not_nil(trap_state->pectxt.dist_state)) + goto yield; + + trap_state->phase = ERTS_CONTINUE_EXIT_DONE; + } + case ERTS_CONTINUE_EXIT_DONE: { + erts_aint_t state; + /* + * From this point on we are no longer allowed to yield + * this process. + */ - erts_proc_sig_fetch(p); +#ifdef DEBUG + yield_allowed = 0; +#endif - erts_proc_unlock(p, ERTS_PROC_LOCK_MSGQ); + /* Set state to not active as we don't want this process + to be scheduled in again after this. */ + state = erts_atomic32_read_band_relb(&p->state, + ~(ERTS_PSFLG_ACTIVE + | ERTS_PSFLG_ACTIVE_SYS + | ERTS_PSFLG_DIRTY_ACTIVE_SYS)); - if (links) { - erts_link_tree_foreach_delete(&links, - erts_proc_exit_handle_link, - (void *) &pectxt); - ASSERT(!links); - } + ASSERT(p->scheduler_data); + ASSERT(p->scheduler_data->current_process == p); + ASSERT(p->scheduler_data->free_process == NULL); - if (monitors) { - erts_monitor_tree_foreach_delete(&monitors, - erts_proc_exit_handle_monitor, - (void *) &pectxt); - ASSERT(!monitors); - } + p->scheduler_data->current_process = NULL; + p->scheduler_data->free_process = p; - if (lt_monitors) { - erts_monitor_list_foreach_delete(<_monitors, - erts_proc_exit_handle_monitor, - (void *) &pectxt); - ASSERT(!lt_monitors); - } + if (state & (ERTS_PSFLG_DIRTY_RUNNING + | ERTS_PSFLG_DIRTY_RUNNING_SYS)) { + p->flags |= F_DELAYED_DEL_PROC; + delay_del_proc = 1; + /* + * The dirty scheduler decrease refc + * when done with the process... + */ + } - /* - * erts_proc_sig_handle_exit() implements yielding. - * However, this function cannot handle it yet... loop - * until done... - */ - while (!0) { - int reds = CONTEXT_REDS; - if (erts_proc_sig_handle_exit(p, &reds)) - break; + erts_schedule_thr_prgr_later_cleanup_op( + (void (*)(void*))erts_proc_dec_refc, + (void *) &p->common, + &p->common.u.release, + sizeof(Process)); + + break; + } } + if (trap_state != &static_state) { + erts_free(ERTS_ALC_T_CONT_EXIT_TRAP, trap_state); + p->u.terminate = NULL; + } + ERTS_CHK_HAVE_ONLY_MAIN_PROC_LOCK(p); + if (IS_TRACED_FL(p, F_TRACE_SCHED_EXIT)) + trace_sched(p, curr_locks, am_out_exited); + erts_flush_trace_messages(p, ERTS_PROC_LOCK_MAIN); ERTS_TRACER_CLEAR(&ERTS_TRACER(p)); @@ -12467,15 +12964,30 @@ erts_continue_exit_process(Process *p) yield: -#ifdef DEBUG ASSERT(yield_allowed); -#endif ERTS_LC_ASSERT(curr_locks == erts_proc_lc_my_proc_locks(p)); ERTS_LC_ASSERT(ERTS_PROC_LOCK_MAIN & curr_locks); + ASSERT(erts_proc_read_refc(p) > 0); + + if (trap_state == &static_state) { + trap_state = erts_alloc(ERTS_ALC_T_CONT_EXIT_TRAP, sizeof(*trap_state)); + sys_memcpy(trap_state, &static_state, sizeof(*trap_state)); + p->u.terminate = trap_state; + } + + ASSERT(p->scheduler_data); + ASSERT(p->scheduler_data->current_process == p); + ASSERT(p->scheduler_data->free_process == NULL); + + if (trap_state->phase >= ERTS_CONTINUE_EXIT_FREE) { + p->scheduler_data->current_process = NULL; + p->scheduler_data->free_process = p; + } p->i = (BeamInstr *) beam_continue_exit; + /* Why is this lock take??? */ if (!(curr_locks & ERTS_PROC_LOCK_STATUS)) { erts_proc_lock(p, ERTS_PROC_LOCK_STATUS); curr_locks |= ERTS_PROC_LOCK_STATUS; |