From 1c86a620d74f4f9383c4956dafd3e2486300dc0d Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Bj=C3=B6rn-Egil=20Dahlberg?= Date: Wed, 17 Jun 2015 17:33:29 +0200 Subject: erts: Remove HALFWORD_HEAP definition --- erts/emulator/beam/erl_process.c | 27 +++++---------------------- 1 file changed, 5 insertions(+), 22 deletions(-) (limited to 'erts/emulator/beam/erl_process.c') diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c index 7b3d12ce09..a21c1e5582 100644 --- a/erts/emulator/beam/erl_process.c +++ b/erts/emulator/beam/erl_process.c @@ -119,7 +119,7 @@ #define RUNQ_SET_RQ(X, RQ) erts_smp_atomic_set_nob((X), (erts_aint_t) (RQ)) #ifdef DEBUG -# if defined(ARCH_64) && !HALFWORD_HEAP +# if defined(ARCH_64) # define ERTS_DBG_SET_INVALID_RUNQP(RQP, N) \ (RUNQ_SET_RQ((RQP), (0xdeadbeefdead0003LL | ((N) << 4))) # define ERTS_DBG_VERIFY_VALID_RUNQP(RQP) \ @@ -971,25 +971,10 @@ typedef struct { erts_smp_atomic32_t refc; } ErtsSchedWallTimeReq; -#if !HALFWORD_HEAP ERTS_SCHED_PREF_QUICK_ALLOC_IMPL(swtreq, - ErtsSchedWallTimeReq, - 5, - ERTS_ALC_T_SCHED_WTIME_REQ) -#else -static ERTS_INLINE ErtsSchedWallTimeReq * -swtreq_alloc(void) -{ - return erts_alloc(ERTS_ALC_T_SCHED_WTIME_REQ, - sizeof(ErtsSchedWallTimeReq)); -} - -static ERTS_INLINE void -swtreq_free(ErtsSchedWallTimeReq *ptr) -{ - erts_free(ERTS_ALC_T_SCHED_WTIME_REQ, ptr); -} -#endif + ErtsSchedWallTimeReq, + 5, + ERTS_ALC_T_SCHED_WTIME_REQ) static void reply_sched_wall_time(void *vswtrp) @@ -5797,9 +5782,7 @@ erts_init_scheduling(int no_schedulers, int no_schedulers_online #endif init_misc_aux_work(); -#if !HALFWORD_HEAP init_swtreq_alloc(); -#endif erts_atomic32_init_nob(&debug_wait_completed_count, 0); /* debug only */ debug_wait_completed_flags = 0; @@ -8286,7 +8269,7 @@ add_pend_suspend(Process *suspendee, sizeof(ErtsPendingSuspend)); psp->next = NULL; #ifdef DEBUG -#if defined(ARCH_64) && !HALFWORD_HEAP +#if defined(ARCH_64) psp->end = (ErtsPendingSuspend *) 0xdeaddeaddeaddead; #else psp->end = (ErtsPendingSuspend *) 0xdeaddead; -- cgit v1.2.3 From 17bcc73e511eee06ca64d51edb401f8340fe9abc Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Bj=C3=B6rn-Egil=20Dahlberg?= Date: Thu, 18 Jun 2015 14:59:28 +0200 Subject: erts: Remove halfword pointer compression * Removed COMPRESS_POINTER and EXPAND_POINTER --- erts/emulator/beam/erl_process.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'erts/emulator/beam/erl_process.c') diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c index a21c1e5582..135f09c04d 100644 --- a/erts/emulator/beam/erl_process.c +++ b/erts/emulator/beam/erl_process.c @@ -12501,7 +12501,7 @@ stack_element_dump(int to, void *to_arg, Eterm* sp, int yreg) } if (is_CP(x)) { - erts_print(to, to_arg, "Return addr %p (", (Eterm *) EXPAND_POINTER(x)); + erts_print(to, to_arg, "Return addr %p (", (Eterm *) x); print_function_from_pc(to, to_arg, cp_val(x)); erts_print(to, to_arg, ")\n"); yreg = 0; -- cgit v1.2.3 From 3bda47f2f26dd417fbd65d5f46b2ab8411a2a41f Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Bj=C3=B6rn-Egil=20Dahlberg?= Date: Wed, 1 Jul 2015 15:09:58 +0200 Subject: erts: Remove halfword !HEAP_ON_C_STACK --- erts/emulator/beam/erl_process.c | 3 --- 1 file changed, 3 deletions(-) (limited to 'erts/emulator/beam/erl_process.c') diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c index 135f09c04d..338be6d8e1 100644 --- a/erts/emulator/beam/erl_process.c +++ b/erts/emulator/beam/erl_process.c @@ -5495,9 +5495,6 @@ init_scheduler_data(ErtsSchedulerData* esdp, int num, 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 #ifdef ERTS_DIRTY_SCHEDULERS if (ERTS_RUNQ_IX_IS_DIRTY(runq->ix)) { esdp->no = 0; -- cgit v1.2.3 From c431a065ba515d27830f01c852f70940efb3003b Mon Sep 17 00:00:00 2001 From: Lukas Larsson Date: Thu, 2 Jul 2015 11:13:32 +0200 Subject: ose: Remove all code related to the OSE port The OSE port is no longer supported and this commit removed it and any changes related to it. The things that were general improvements have been left in the code. --- erts/emulator/beam/erl_process.c | 68 ++++------------------------------------ 1 file changed, 6 insertions(+), 62 deletions(-) (limited to 'erts/emulator/beam/erl_process.c') diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c index 135f09c04d..5b13f74dcb 100644 --- a/erts/emulator/beam/erl_process.c +++ b/erts/emulator/beam/erl_process.c @@ -58,11 +58,7 @@ #define ERTS_PROC_MIN_CONTEXT_SWITCH_REDS_COST (CONTEXT_REDS/10) -#ifndef ERTS_SCHED_MIN_SPIN #define ERTS_SCHED_SPIN_UNTIL_YIELD 100 -#else -#define ERTS_SCHED_SPIN_UNTIL_YIELD 1 -#endif #define ERTS_SCHED_SYS_SLEEP_SPINCOUNT_VERY_LONG 40 #define ERTS_SCHED_AUX_WORK_SLEEP_SPINCOUNT_FACT_VERY_LONG 1000 @@ -152,12 +148,7 @@ extern BeamInstr beam_apply[]; extern BeamInstr beam_exit[]; extern BeamInstr beam_continue_exit[]; -#ifdef __OSE__ -/* Eager check I/O not supported on OSE yet. */ -int erts_eager_check_io = 0; -#else int erts_eager_check_io = 1; -#endif int erts_sched_compact_load; int erts_sched_balance_util = 0; Uint erts_no_schedulers; @@ -2521,19 +2512,10 @@ try_set_sys_scheduling(void) #endif static ERTS_INLINE int -prepare_for_sys_schedule(ErtsSchedulerData *esdp, int non_blocking) +prepare_for_sys_schedule(int non_blocking) { if (non_blocking && erts_eager_check_io) { #ifdef ERTS_SMP -#ifdef ERTS_SCHED_ONLY_POLL_SCHED_1 - if (esdp->no != 1) { - /* If we are not scheduler 1 and ERTS_SCHED_ONLY_POLL_SCHED_1 is used - then we make sure to wake scheduler 1 */ - ErtsRunQueue *rq = ERTS_RUNQ_IX(0); - wake_scheduler(rq); - return 0; - } -#endif return try_set_sys_scheduling(); #else return 1; @@ -2543,16 +2525,6 @@ prepare_for_sys_schedule(ErtsSchedulerData *esdp, int non_blocking) #ifdef ERTS_SMP while (!erts_port_task_have_outstanding_io_tasks() && try_set_sys_scheduling()) { -#ifdef ERTS_SCHED_ONLY_POLL_SCHED_1 - if (esdp->no != 1) { - /* If we are not scheduler 1 and ERTS_SCHED_ONLY_POLL_SCHED_1 is used - then we make sure to wake scheduler 1 */ - ErtsRunQueue *rq = ERTS_RUNQ_IX(0); - clear_sys_scheduling(); - wake_scheduler(rq); - return 0; - } -#endif if (!erts_port_task_have_outstanding_io_tasks()) return 1; clear_sys_scheduling(); @@ -2876,8 +2848,6 @@ aux_thread(void *unused) erts_thr_progress_active(NULL, thr_prgr_active = 0); erts_thr_progress_prepare_wait(NULL); - ERTS_SCHED_FAIR_YIELD(); - flgs = sched_spin_wait(ssi, 0); if (flgs & ERTS_SSI_FLG_SLEEPING) { @@ -2945,7 +2915,7 @@ scheduler_wait(int *fcalls, ErtsSchedulerData *esdp, ErtsRunQueue *rq) * be waiting in erl_sys_schedule() */ - if (ERTS_SCHEDULER_IS_DIRTY(esdp) || !prepare_for_sys_schedule(esdp, 0)) { + if (ERTS_SCHEDULER_IS_DIRTY(esdp) || !prepare_for_sys_schedule(0)) { sched_waiting(esdp->no, rq); @@ -3010,8 +2980,6 @@ scheduler_wait(int *fcalls, ErtsSchedulerData *esdp, ErtsRunQueue *rq) erts_thr_progress_prepare_wait(esdp); } - ERTS_SCHED_FAIR_YIELD(); - flgs = sched_spin_wait(ssi, spincount); if (flgs & ERTS_SSI_FLG_SLEEPING) { ASSERT(flgs & ERTS_SSI_FLG_WAITING); @@ -3082,13 +3050,8 @@ scheduler_wait(int *fcalls, ErtsSchedulerData *esdp, ErtsRunQueue *rq) #ifdef ERTS_DIRTY_SCHEDULERS ASSERT(!ERTS_SCHEDULER_IS_DIRTY(esdp)); #endif - -#ifdef ERTS_SCHED_ONLY_POLL_SCHED_1 - ASSERT(esdp->no == 1); -#endif sched_waiting_sys(esdp->no, rq); - erts_smp_runq_unlock(rq); ASSERT(working); @@ -3158,7 +3121,7 @@ scheduler_wait(int *fcalls, ErtsSchedulerData *esdp, ErtsRunQueue *rq) * Got to check that we still got I/O tasks; otherwise * we have to continue checking for I/O... */ - if (!prepare_for_sys_schedule(esdp, 0)) { + if (!prepare_for_sys_schedule(0)) { spincount *= ERTS_SCHED_TSE_SLEEP_SPINCOUNT_FACT; goto tse_wait; } @@ -3180,7 +3143,7 @@ scheduler_wait(int *fcalls, ErtsSchedulerData *esdp, ErtsRunQueue *rq) * Got to check that we still got I/O tasks; otherwise * we have to wait in erl_sys_schedule() after all... */ - if (!prepare_for_sys_schedule(esdp, 0)) { + if (!prepare_for_sys_schedule(0)) { /* * Not allowed to wait in erl_sys_schedule; * do tse wait instead... @@ -5301,17 +5264,11 @@ erts_early_init_scheduling(int no_schedulers) wakeup_other.threshold = ERTS_SCHED_WAKEUP_OTHER_THRESHOLD_MEDIUM; wakeup_other.type = ERTS_SCHED_WAKEUP_OTHER_TYPE_DEFAULT; #endif -#ifndef ERTS_SCHED_MIN_SPIN sched_busy_wait.sys_schedule = ERTS_SCHED_SYS_SLEEP_SPINCOUNT_MEDIUM; sched_busy_wait.tse = (ERTS_SCHED_SYS_SLEEP_SPINCOUNT_MEDIUM * ERTS_SCHED_TSE_SLEEP_SPINCOUNT_FACT); sched_busy_wait.aux_work = (ERTS_SCHED_SYS_SLEEP_SPINCOUNT_MEDIUM * ERTS_SCHED_AUX_WORK_SLEEP_SPINCOUNT_FACT_MEDIUM); -#else - sched_busy_wait.sys_schedule = ERTS_SCHED_SYS_SLEEP_SPINCOUNT_NONE; - sched_busy_wait.tse = ERTS_SCHED_SYS_SLEEP_SPINCOUNT_NONE; - sched_busy_wait.aux_work = ERTS_SCHED_SYS_SLEEP_SPINCOUNT_NONE; -#endif } int @@ -8187,18 +8144,12 @@ erts_start_schedulers(void) erts_snprintf(opts.name, 16, "%lu_scheduler", actual + 1); -#ifdef __OSE__ - /* This should be done in the bind strategy */ - opts.coreNo = (actual+1) % ose_num_cpus(); -#endif - res = ethr_thr_create(&esdp->tid, sched_thread_func, (void*)esdp, &opts); if (res != 0) { break; } } - erts_no_schedulers = actual; #ifdef ERTS_DIRTY_SCHEDULERS @@ -8227,10 +8178,6 @@ erts_start_schedulers(void) erts_snprintf(opts.name, 16, "aux"); -#ifdef __OSE__ - opts.coreNo = 0; -#endif /* __OSE__ */ - res = ethr_thr_create(&aux_tid, aux_thread, NULL, &opts); if (res != 0) erl_exit(1, "Failed to create aux thread\n"); @@ -8250,7 +8197,6 @@ erts_start_schedulers(void) actual, actual == 1 ? " was" : "s were"); erts_send_error_to_logger_nogl(dsbufp); } - } #endif /* ERTS_SMP */ @@ -9391,12 +9337,10 @@ Process *schedule(Process *p, int calls) int leader_update = ERTS_SCHEDULER_IS_DIRTY(esdp) ? 0 : erts_thr_progress_update(esdp); aux_work = erts_atomic32_read_acqb(&esdp->ssi->aux_work); - if (aux_work | leader_update | ERTS_SCHED_FAIR) { + if (aux_work | leader_update) { erts_smp_runq_unlock(rq); if (leader_update) erts_thr_progress_leader_update(esdp); - else if (ERTS_SCHED_FAIR) - ERTS_SCHED_FAIR_YIELD(); if (aux_work) handle_aux_work(&esdp->aux_work_data, aux_work, 0); erts_smp_runq_lock(rq); @@ -9472,7 +9416,7 @@ Process *schedule(Process *p, int calls) } else if (!ERTS_SCHEDULER_IS_DIRTY(esdp) && (fcalls > input_reductions && - prepare_for_sys_schedule(esdp, !0))) { + prepare_for_sys_schedule(!0))) { ErtsMonotonicTime current_time; /* * Schedule system-level activities. -- cgit v1.2.3 From 76777477d8722b785afa8da0166b4577fb7e047c Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Bj=C3=B6rn-Egil=20Dahlberg?= Date: Fri, 9 Oct 2015 17:18:31 +0200 Subject: erts: Remove vheap mature from process control block Binary vheap mature is not necessary for binary gc. --- erts/emulator/beam/erl_process.c | 2 -- 1 file changed, 2 deletions(-) (limited to 'erts/emulator/beam/erl_process.c') diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c index 879b523c38..15a6d5d651 100644 --- a/erts/emulator/beam/erl_process.c +++ b/erts/emulator/beam/erl_process.c @@ -10834,7 +10834,6 @@ erl_create_process(Process* parent, /* Parent of process (default group leader). p->bin_vheap_sz = p->min_vheap_size; p->bin_old_vheap_sz = p->min_vheap_size; p->bin_old_vheap = 0; - p->bin_vheap_mature = 0; p->sys_task_qs = NULL; @@ -11054,7 +11053,6 @@ void erts_init_empty_process(Process *p) p->bin_old_vheap_sz = BIN_VH_MIN_SIZE; p->bin_old_vheap = 0; p->sys_task_qs = NULL; - p->bin_vheap_mature = 0; ERTS_PTMR_INIT(p); p->next = NULL; p->off_heap.first = NULL; -- cgit v1.2.3 From 3ac08f9b668613a4292436979eacc61863c2ab94 Mon Sep 17 00:00:00 2001 From: Rickard Green Date: Wed, 16 Sep 2015 15:02:50 +0200 Subject: Fragmented young heap generation and off_heap_message_queue option * The youngest generation of the heap can now consist of multiple blocks. Heap fragments and message fragments are added to the youngest generation when needed without triggering a GC. After a GC the youngest generation is contained in one single block. * The off_heap_message_queue process flag has been added. When enabled all message data in the queue is kept off heap. When a message is selected from the queue, the message fragment (or heap fragment) containing the actual message is attached to the youngest generation. Messages stored off heap is not part of GC. --- erts/emulator/beam/erl_process.c | 346 ++++++++++++++++++++------------------- 1 file changed, 180 insertions(+), 166 deletions(-) (limited to 'erts/emulator/beam/erl_process.c') diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c index bad9da90ea..e490ffea5a 100644 --- a/erts/emulator/beam/erl_process.c +++ b/erts/emulator/beam/erl_process.c @@ -148,6 +148,7 @@ extern BeamInstr beam_apply[]; extern BeamInstr beam_exit[]; extern BeamInstr beam_continue_exit[]; +int erts_default_spo_flags = 0; int erts_eager_check_io = 1; int erts_sched_compact_load; int erts_sched_balance_util = 0; @@ -351,7 +352,8 @@ struct erts_system_profile_flags_t erts_system_profile_flags; typedef enum { ERTS_PSTT_GC, /* Garbage Collect */ - ERTS_PSTT_CPC /* Check Process Code */ + ERTS_PSTT_CPC, /* Check Process Code */ + ERTS_PSTT_COHMQ /* Change off heap message queue */ } ErtsProcSysTaskType; #define ERTS_MAX_PROC_SYS_TASK_ARGS 2 @@ -982,7 +984,7 @@ reply_sched_wall_time(void *vswtrp) Eterm **hpp; Uint sz, *szp; ErlOffHeap *ohp = NULL; - ErlHeapFragment *bp = NULL; + ErtsMessage *mp = NULL; ASSERT(esdp); #ifdef ERTS_DIRTY_SCHEDULERS @@ -1038,12 +1040,12 @@ reply_sched_wall_time(void *vswtrp) if (hpp) break; - hp = erts_alloc_message_heap(sz, &bp, &ohp, rp, &rp_locks); + mp = erts_alloc_message_heap(rp, &rp_locks, sz, &hp, &ohp); szp = NULL; hpp = &hp; } - erts_queue_message(rp, &rp_locks, bp, msg, NIL); + erts_queue_message(rp, &rp_locks, mp, msg, NIL); if (swtrp->req_sched == esdp->no) rp_locks &= ~ERTS_PROC_LOCK_MAIN; @@ -6294,22 +6296,99 @@ erts_schedule_process(Process *p, erts_aint32_t state, ErtsProcLocks locks) schedule_process(p, state, locks); } -static void -schedule_process_sys_task(Process *p, erts_aint32_t state, Process *proxy) +static int +schedule_process_sys_task(Process *p, erts_aint32_t prio, ErtsProcSysTask *st) { - /* - * Expects status lock to be locked when called, and - * returns with status lock unlocked... - */ - erts_aint32_t a = state, n, enq_prio = -1; + int res; + int locked; + ErtsProcSysTaskQs *stqs, *free_stqs; + erts_aint32_t state, a, n, enq_prio; int enqueue; /* < 0 -> use proxy */ - unsigned int prof_runnable_procs = erts_system_profile_flags.runnable_procs; + unsigned int prof_runnable_procs; + + res = 1; /* prepare for success */ + st->next = st->prev = st; /* Prep for empty prio queue */ + state = erts_smp_atomic32_read_nob(&p->state); + prof_runnable_procs = erts_system_profile_flags.runnable_procs; + locked = 0; + free_stqs = NULL; + if (state & ERTS_PSFLG_ACTIVE_SYS) + stqs = NULL; + else { + alloc_qs: + stqs = proc_sys_task_queues_alloc(); + stqs->qmask = 1 << prio; + stqs->ncount = 0; + stqs->q[PRIORITY_MAX] = NULL; + stqs->q[PRIORITY_HIGH] = NULL; + stqs->q[PRIORITY_NORMAL] = NULL; + stqs->q[PRIORITY_LOW] = NULL; + stqs->q[prio] = st; + } + + if (!locked) { + locked = 1; + erts_smp_proc_lock(p, ERTS_PROC_LOCK_STATUS); + + state = erts_smp_atomic32_read_nob(&p->state); + if (state & ERTS_PSFLG_EXITING) { + free_stqs = stqs; + res = 0; + goto cleanup; + } + } + + if (!p->sys_task_qs) { + if (stqs) + p->sys_task_qs = stqs; + else + goto alloc_qs; + } + else { + free_stqs = stqs; + stqs = p->sys_task_qs; + if (!stqs->q[prio]) { + stqs->q[prio] = st; + stqs->qmask |= 1 << prio; + } + else { + st->next = stqs->q[prio]; + st->prev = stqs->q[prio]->prev; + st->next->prev = st; + st->prev->next = st; + ASSERT(stqs->qmask & (1 << prio)); + } + } + + if (ERTS_PSFLGS_GET_ACT_PRIO(state) > prio) { + erts_aint32_t n, a, e; + /* Need to elevate actual prio */ + + a = state; + do { + if (ERTS_PSFLGS_GET_ACT_PRIO(a) <= prio) { + n = a; + break; + } + n = e = a; + n &= ~ERTS_PSFLGS_ACT_PRIO_MASK; + n |= (prio << ERTS_PSFLGS_ACT_PRIO_OFFSET); + a = erts_smp_atomic32_cmpxchg_nob(&p->state, n, e); + } while (a != e); + state = n; + } + + + a = state; + enq_prio = -1; /* Status lock prevents out of order "runnable proc" trace msgs */ ERTS_SMP_LC_ASSERT(ERTS_PROC_LOCK_STATUS & erts_proc_lc_my_proc_locks(p)); - if (!prof_runnable_procs) + if (!prof_runnable_procs) { erts_smp_proc_unlock(p, ERTS_PROC_LOCK_STATUS); + locked = 0; + } ASSERT(!(state & ERTS_PSFLG_PROXY)); @@ -6317,8 +6396,10 @@ schedule_process_sys_task(Process *p, erts_aint32_t state, Process *proxy) erts_aint32_t e; n = e = a; - if (a & ERTS_PSFLG_FREE) + if (a & ERTS_PSFLG_FREE) { + res = 0; goto cleanup; /* We don't want to schedule free processes... */ + } enqueue = ERTS_ENQUEUE_NOT; n |= ERTS_PSFLG_ACTIVE_SYS; @@ -6342,29 +6423,24 @@ schedule_process_sys_task(Process *p, erts_aint32_t state, Process *proxy) } erts_smp_proc_unlock(p, ERTS_PROC_LOCK_STATUS); - prof_runnable_procs = 0; + locked = 0; } - if (enqueue != ERTS_ENQUEUE_NOT) { - Process *sched_p; - if (enqueue > 0) - sched_p = p; - else { - sched_p = make_proxy_proc(proxy, p, enq_prio); - proxy = NULL; - } - add2runq(sched_p, n, enq_prio); - } + if (enqueue != ERTS_ENQUEUE_NOT) + add2runq(enqueue > 0 ? p : make_proxy_proc(NULL, p, enq_prio), + n, enq_prio); cleanup: - if (prof_runnable_procs) + if (locked) erts_smp_proc_unlock(p, ERTS_PROC_LOCK_STATUS); - if (proxy) - free_proxy_proc(proxy); + if (free_stqs) + proc_sys_task_queues_free(free_stqs); ERTS_SMP_LC_ASSERT(!(ERTS_PROC_LOCK_STATUS & erts_proc_lc_my_proc_locks(p))); + + return res; } static ERTS_INLINE int @@ -9696,13 +9772,13 @@ Process *schedule(Process *p, int calls) } } - if (!(state & ERTS_PSFLG_EXITING) - && ((FLAGS(p) & F_FORCE_GC) - || (MSO(p).overhead > BIN_VHEAP_SZ(p)))) { - reds -= erts_garbage_collect(p, 0, p->arg_reg, p->arity); - if (reds <= 0) { - p->fcalls = reds; - goto sched_out_proc; + if (ERTS_IS_GC_DESIRED(p)) { + if (!(state & ERTS_PSFLG_EXITING) && !(p->flags & F_DISABLE_GC)) { + reds -= erts_garbage_collect(p, 0, p->arg_reg, p->arity); + if (reds <= 0) { + p->fcalls = reds; + goto sched_out_proc; + } } } @@ -9742,7 +9818,7 @@ notify_sys_task_executed(Process *c_p, ErtsProcSysTask *st, Eterm st_result) if (rp) { ErtsProcLocks rp_locks; ErlOffHeap *ohp; - ErlHeapFragment* bp; + ErtsMessage *mp; Eterm *hp, msg, req_id, result; Uint st_result_sz, hsz; #ifdef DEBUG @@ -9754,11 +9830,7 @@ notify_sys_task_executed(Process *c_p, ErtsProcSysTask *st, Eterm st_result) st_result_sz = is_immed(st_result) ? 0 : size_object(st_result); hsz = st->req_id_sz + st_result_sz + 4 /* 3-tuple */; - hp = erts_alloc_message_heap(hsz, - &bp, - &ohp, - rp, - &rp_locks); + mp = erts_alloc_message_heap(rp, &rp_locks, hsz, &hp, &ohp); #ifdef DEBUG hp_start = hp; @@ -9783,7 +9855,7 @@ notify_sys_task_executed(Process *c_p, ErtsProcSysTask *st, Eterm st_result) ASSERT(hp_start + hsz == hp); #endif - erts_queue_message(rp, &rp_locks, bp, msg, NIL); + erts_queue_message(rp, &rp_locks, mp, msg, NIL); if (c_p == rp) rp_locks &= ~ERTS_PROC_LOCK_MAIN; @@ -10005,6 +10077,10 @@ execute_sys_tasks(Process *c_p, erts_aint32_t *statep, int in_reds) st = NULL; } break; + case ERTS_PSTT_COHMQ: + reds += erts_complete_off_heap_message_queue_change(c_p); + st_res = am_true; + break; default: ERTS_INTERNAL_ERROR("Invalid process sys task type"); st_res = am_false; @@ -10047,6 +10123,9 @@ cleanup_sys_tasks(Process *c_p, erts_aint32_t in_state, int in_reds) case ERTS_PSTT_CPC: st_res = am_false; break; + case ERTS_PSTT_COHMQ: + st_res = am_false; + break; default: ERTS_INTERNAL_ERROR("Invalid process sys task type"); st_res = am_false; @@ -10065,10 +10144,8 @@ BIF_RETTYPE erts_internal_request_system_task_3(BIF_ALIST_3) { Process *rp = erts_proc_lookup(BIF_ARG_1); - ErtsProcSysTaskQs *stqs, *free_stqs = NULL; ErtsProcSysTask *st = NULL; - erts_aint32_t prio, rp_state; - int rp_locked; + erts_aint32_t prio; Eterm noproc_res, req_type; if (!rp && !is_internal_pid(BIF_ARG_1)) { @@ -10125,7 +10202,6 @@ erts_internal_request_system_task_3(BIF_ALIST_3) } st = erts_alloc(ERTS_ALC_T_PROC_SYS_TSK, ERTS_PROC_SYS_TASK_SIZE(tot_sz)); - st->next = st->prev = st; /* Prep for empty prio queue */ ERTS_INIT_OFF_HEAP(&st->off_heap); hp = &st->heap[0]; @@ -10169,95 +10245,11 @@ erts_internal_request_system_task_3(BIF_ALIST_3) goto badarg; } - rp_state = erts_smp_atomic32_read_nob(&rp->state); - - rp_locked = 0; - - free_stqs = NULL; - if (rp_state & ERTS_PSFLG_ACTIVE_SYS) - stqs = NULL; - else { - alloc_qs: - stqs = proc_sys_task_queues_alloc(); - stqs->qmask = 1 << prio; - stqs->ncount = 0; - stqs->q[PRIORITY_MAX] = NULL; - stqs->q[PRIORITY_HIGH] = NULL; - stqs->q[PRIORITY_NORMAL] = NULL; - stqs->q[PRIORITY_LOW] = NULL; - stqs->q[prio] = st; - } - - if (!rp_locked) { - rp_locked = 1; - erts_smp_proc_lock(rp, ERTS_PROC_LOCK_STATUS); - - rp_state = erts_smp_atomic32_read_nob(&rp->state); - if (rp_state & ERTS_PSFLG_EXITING) { - erts_smp_proc_unlock(rp, ERTS_PROC_LOCK_STATUS); - rp = NULL; - free_stqs = stqs; - goto noproc; - } + if (!schedule_process_sys_task(rp, prio, st)) { + noproc: + notify_sys_task_executed(BIF_P, st, noproc_res); } - if (!rp->sys_task_qs) { - if (stqs) - rp->sys_task_qs = stqs; - else - goto alloc_qs; - } - else { - if (stqs) - free_stqs = stqs; - stqs = rp->sys_task_qs; - if (!stqs->q[prio]) { - stqs->q[prio] = st; - stqs->qmask |= 1 << prio; - } - else { - st->next = stqs->q[prio]; - st->prev = stqs->q[prio]->prev; - st->next->prev = st; - st->prev->next = st; - ASSERT(stqs->qmask & (1 << prio)); - } - } - - if (ERTS_PSFLGS_GET_ACT_PRIO(rp_state) > prio) { - erts_aint32_t n, a, e; - /* Need to elevate actual prio */ - - a = rp_state; - do { - if (ERTS_PSFLGS_GET_ACT_PRIO(a) <= prio) { - n = a; - break; - } - n = e = a; - n &= ~ERTS_PSFLGS_ACT_PRIO_MASK; - n |= (prio << ERTS_PSFLGS_ACT_PRIO_OFFSET); - a = erts_smp_atomic32_cmpxchg_nob(&rp->state, n, e); - } while (a != e); - rp_state = n; - } - - /* - * schedule_process_sys_task() unlocks status - * lock on process. - */ - schedule_process_sys_task(rp, rp_state, NULL); - - if (free_stqs) - proc_sys_task_queues_free(free_stqs); - - BIF_RET(am_ok); - -noproc: - - notify_sys_task_executed(BIF_P, st, noproc_res); - if (free_stqs) - proc_sys_task_queues_free(free_stqs); BIF_RET(am_ok); badarg: @@ -10266,11 +10258,35 @@ badarg: erts_cleanup_offheap(&st->off_heap); erts_free(ERTS_ALC_T_PROC_SYS_TSK, st); } - if (free_stqs) - proc_sys_task_queues_free(free_stqs); BIF_ERROR(BIF_P, BADARG); } +void +erts_schedule_complete_off_heap_message_queue_change(Eterm pid) +{ + Process *rp = erts_proc_lookup(pid); + if (rp) { + ErtsProcSysTask *st; + erts_aint32_t state; + int i; + + st = erts_alloc(ERTS_ALC_T_PROC_SYS_TSK, + ERTS_PROC_SYS_TASK_SIZE(0)); + st->type = ERTS_PSTT_COHMQ; + st->requester = NIL; + st->reply_tag = NIL; + st->req_id = NIL; + st->req_id_sz = 0; + for (i = 0; i < ERTS_MAX_PROC_SYS_TASK_ARGS; i++) + st->arg[i] = NIL; + ERTS_INIT_OFF_HEAP(&st->off_heap); + state = erts_smp_atomic32_read_nob(&rp->state); + + if (!schedule_process_sys_task(rp, ERTS_PSFLGS_GET_USR_PRIO(state), st)) + erts_free(ERTS_ALC_T_PROC_SYS_TSK, st); + } +} + static void save_gc_task(Process *c_p, ErtsProcSysTask *st, int prio) { @@ -10716,6 +10732,7 @@ erl_create_process(Process* parent, /* Parent of process (default group leader). Eterm args, /* Arguments for function (must be well-formed list). */ ErlSpawnOpts* so) /* Options for spawn. */ { + Uint flags = erts_default_process_flags; ErtsRunQueue *rq = NULL; Process *p; Sint arity; /* Number of arguments. */ @@ -10753,6 +10770,11 @@ erl_create_process(Process* parent, /* Parent of process (default group leader). state |= (((prio & ERTS_PSFLGS_PRIO_MASK) << ERTS_PSFLGS_ACT_PRIO_OFFSET) | ((prio & ERTS_PSFLGS_PRIO_MASK) << ERTS_PSFLGS_USR_PRIO_OFFSET)); + if (so->flags & SPO_OFF_HEAP_MSGQ) { + state |= ERTS_PSFLG_OFF_HEAP_MSGQ; + flags |= F_OFF_HEAP_MSGQ; + } + if (!rq) rq = erts_get_runq_proc(parent); @@ -10775,7 +10797,7 @@ erl_create_process(Process* parent, /* Parent of process (default group leader). BM_SWAP_TIMER(size,system); heap_need = arg_size; - p->flags = erts_default_process_flags; + p->flags = flags; p->static_flags = 0; if (so->flags & SPO_SYSTEM_PROC) @@ -10824,6 +10846,8 @@ erl_create_process(Process* parent, /* Parent of process (default group leader). p->stop = p->hend = p->heap + sz; p->htop = p->heap; p->heap_sz = sz; + p->abandoned_heap = NULL; + p->live_hf_end = ERTS_INVALID_HFRAG_PTR; p->catches = 0; p->bin_vheap_sz = p->min_vheap_size; @@ -10894,6 +10918,7 @@ erl_create_process(Process* parent, /* Parent of process (default group leader). p->accessor_bif_timers = NULL; #endif p->mbuf = NULL; + p->msg_frag = NULL; p->mbuf_sz = 0; p->psd = NULL; p->dictionary = NULL; @@ -11029,6 +11054,8 @@ void erts_init_empty_process(Process *p) p->stop = NULL; p->hend = NULL; p->heap = NULL; + p->abandoned_heap = NULL; + p->live_hf_end = ERTS_INVALID_HFRAG_PTR; p->gen_gcs = 0; p->max_gen_gcs = 0; p->min_heap_size = 0; @@ -11061,6 +11088,7 @@ void erts_init_empty_process(Process *p) p->old_htop = NULL; p->old_heap = NULL; p->mbuf = NULL; + p->msg_frag = NULL; p->mbuf_sz = 0; p->psd = NULL; ERTS_P_MONITORS(p) = NULL; @@ -11149,6 +11177,8 @@ erts_debug_verify_clean_empty_process(Process* p) ASSERT(p->htop == NULL); ASSERT(p->stop == NULL); ASSERT(p->hend == NULL); + ASSERT(p->abandoned_heap == NULL); + ASSERT(p->live_hf_end == ERTS_INVALID_HFRAG_PTR); ASSERT(p->heap == NULL); ASSERT(p->common.id == ERTS_INVALID_PID); ASSERT(ERTS_TRACER_PROC(p) == NIL); @@ -11226,8 +11256,6 @@ erts_cleanup_empty_process(Process* p) static void delete_process(Process* p) { - ErlMessage* mp; - VERBOSE(DEBUG_PROCESSES, ("Removing process: %T\n",p->common.id)); /* Cleanup psd */ @@ -11283,24 +11311,8 @@ delete_process(Process* p) erts_erase_dicts(p); /* free all pending messages */ - mp = p->msg.first; - while(mp != NULL) { - ErlMessage* next_mp = mp->next; - if (mp->data.attached) { - if (is_value(mp->m[0])) - free_message_buffer(mp->data.heap_frag); - else { - if (is_not_nil(mp->m[1])) { - ErlHeapFragment *heap_frag; - heap_frag = (ErlHeapFragment *) mp->data.dist_ext->ext_endp; - erts_cleanup_offheap(&heap_frag->off_heap); - } - erts_free_dist_ext_copy(mp->data.dist_ext); - } - } - free_message(mp); - mp = next_mp; - } + erts_cleanup_messages(p->msg.first); + p->msg.first = NULL; ASSERT(!p->nodes_monitors); ASSERT(!p->suspend_monitors); @@ -11488,6 +11500,9 @@ static ERTS_INLINE void send_exit_message(Process *to, ErtsProcLocks *to_locksp, Eterm exit_term, Uint term_size, Eterm token) { + ErtsMessage *mp; + ErlOffHeap *ohp; + if (token == NIL #ifdef USE_VM_PROBES || token == am_have_dt_utag @@ -11495,14 +11510,12 @@ send_exit_message(Process *to, ErtsProcLocks *to_locksp, ) { Eterm* hp; Eterm mess; - ErlHeapFragment* bp; - ErlOffHeap *ohp; - hp = erts_alloc_message_heap(term_size, &bp, &ohp, to, to_locksp); + mp = erts_alloc_message_heap(to, to_locksp, + term_size, &hp, &ohp); mess = copy_struct(exit_term, term_size, &hp, ohp); - erts_queue_message(to, to_locksp, bp, mess, NIL); + erts_queue_message(to, to_locksp, mp, mess, NIL); } else { - ErlHeapFragment* bp; Eterm* hp; Eterm mess; Eterm temp_token; @@ -11510,13 +11523,14 @@ send_exit_message(Process *to, ErtsProcLocks *to_locksp, ASSERT(is_tuple(token)); sz_token = size_object(token); - bp = new_message_buffer(term_size+sz_token); - hp = bp->mem; - mess = copy_struct(exit_term, term_size, &hp, &bp->off_heap); + + mp = erts_alloc_message_heap(to, to_locksp, + term_size+sz_token, &hp, &ohp); + mess = copy_struct(exit_term, term_size, &hp, ohp); /* the trace token must in this case be updated by the caller */ seq_trace_output(token, mess, SEQ_TRACE_SEND, to->common.id, NULL); - temp_token = copy_struct(token, sz_token, &hp, &bp->off_heap); - erts_queue_message(to, to_locksp, bp, mess, temp_token); + temp_token = copy_struct(token, sz_token, &hp, ohp); + erts_queue_message(to, to_locksp, mp, mess, temp_token); } } -- cgit v1.2.3 From 9c6f45b901ee701553afe34c0b33b7d931d73fd9 Mon Sep 17 00:00:00 2001 From: Rickard Green Date: Wed, 11 Nov 2015 11:39:32 +0100 Subject: Bump reductions on GC --- erts/emulator/beam/erl_process.c | 10 +++++----- 1 file changed, 5 insertions(+), 5 deletions(-) (limited to 'erts/emulator/beam/erl_process.c') diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c index e490ffea5a..5e7ed0d151 100644 --- a/erts/emulator/beam/erl_process.c +++ b/erts/emulator/beam/erl_process.c @@ -9774,7 +9774,7 @@ Process *schedule(Process *p, int calls) if (ERTS_IS_GC_DESIRED(p)) { if (!(state & ERTS_PSFLG_EXITING) && !(p->flags & F_DISABLE_GC)) { - reds -= erts_garbage_collect(p, 0, p->arg_reg, p->arity); + reds -= erts_garbage_collect_nobump(p, 0, p->arg_reg, p->arity); if (reds <= 0) { p->fcalls = reds; goto sched_out_proc; @@ -10057,10 +10057,10 @@ execute_sys_tasks(Process *c_p, erts_aint32_t *statep, int in_reds) else { if (!garbage_collected) { FLAGS(c_p) |= F_NEED_FULLSWEEP; - reds += erts_garbage_collect(c_p, - 0, - c_p->arg_reg, - c_p->arity); + reds += erts_garbage_collect_nobump(c_p, + 0, + c_p->arg_reg, + c_p->arity); garbage_collected = 1; } st_res = am_true; -- cgit v1.2.3 From 244e9d5855d1b1f160d667b5cf369defee72829d Mon Sep 17 00:00:00 2001 From: "Nikolaos S. Papaspyrou" Date: Sat, 9 Jun 2012 01:37:29 +0300 Subject: Enable shcopy for spawning processes --- erts/emulator/beam/erl_process.c | 16 ++++++++++++++++ 1 file changed, 16 insertions(+) (limited to 'erts/emulator/beam/erl_process.c') diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c index 9acce8acb6..f2da5289d3 100644 --- a/erts/emulator/beam/erl_process.c +++ b/erts/emulator/beam/erl_process.c @@ -10747,6 +10747,11 @@ erl_create_process(Process* parent, /* Parent of process (default group leader). Eterm res = THE_NON_VALUE; erts_aint32_t state = 0; erts_aint32_t prio = (erts_aint32_t) PRIORITY_NORMAL; +#ifdef SHCOPY_SPAWN + unsigned shflags = 0; /* could be taken from so->flags, if necessary */ + shcopy_info info; + INITIALIZE_INFO(info); +#endif #ifdef ERTS_SMP erts_smp_proc_lock(parent, ERTS_PROC_LOCKS_ALL_MINOR); @@ -10798,7 +10803,11 @@ erl_create_process(Process* parent, /* Parent of process (default group leader). BM_COUNT(processes_spawned); BM_SWAP_TIMER(system,size); +#ifdef SHCOPY_SPAWN + arg_size = copy_shared_calculate(args, &info, shflags); +#else arg_size = size_object(args); +#endif BM_SWAP_TIMER(size,system); heap_need = arg_size; @@ -10876,7 +10885,12 @@ erl_create_process(Process* parent, /* Parent of process (default group leader). BM_MESSAGE(args,p,parent); BM_START_TIMER(system); BM_SWAP_TIMER(system,copy); +#ifdef SHCOPY_SPAWN + p->arg_reg[2] = copy_shared_perform(args, arg_size, &info, &p->htop, &p->off_heap, shflags); + DESTROY_INFO(info); +#else p->arg_reg[2] = copy_struct(args, arg_size, &p->htop, &p->off_heap); +#endif BM_MESSAGE_COPIED(arg_size); BM_SWAP_TIMER(copy,system); p->arity = 3; @@ -11260,6 +11274,8 @@ static void delete_process(Process* p) { VERBOSE(DEBUG_PROCESSES, ("Removing process: %T\n",p->common.id)); + VERBOSE_DEBUG("[pid=%T] delete process: %p %p %p %p\n", p->common.id, + HEAP_START(p), HEAP_END(p), OLD_HEAP(p), OLD_HEND(p)); /* Cleanup psd */ -- cgit v1.2.3 From 277e8e77384ed6628009243e63d62f0555d10c69 Mon Sep 17 00:00:00 2001 From: "Nikolaos S. Papaspyrou" Date: Mon, 11 Jun 2012 15:17:01 +0300 Subject: Add -debug +vc flag for debuging SHCOPY This is very verbose, you have been warned. It should work with the copy-spy.py script, which may be a bit outdated. --- erts/emulator/beam/erl_process.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) (limited to 'erts/emulator/beam/erl_process.c') diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c index f2da5289d3..96d17306a5 100644 --- a/erts/emulator/beam/erl_process.c +++ b/erts/emulator/beam/erl_process.c @@ -11274,8 +11274,8 @@ static void delete_process(Process* p) { VERBOSE(DEBUG_PROCESSES, ("Removing process: %T\n",p->common.id)); - VERBOSE_DEBUG("[pid=%T] delete process: %p %p %p %p\n", p->common.id, - HEAP_START(p), HEAP_END(p), OLD_HEAP(p), OLD_HEND(p)); + VERBOSE(DEBUG_SHCOPY, ("[pid=%T] delete process: %p %p %p %p\n", p->common.id, + HEAP_START(p), HEAP_END(p), OLD_HEAP(p), OLD_HEND(p))); /* Cleanup psd */ -- cgit v1.2.3 From 748c73f1687b2375d4c607487f40036ba990c4c7 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Bj=C3=B6rn-Egil=20Dahlberg?= Date: Thu, 17 Sep 2015 16:17:10 +0200 Subject: Refactor copy sharing --- erts/emulator/beam/erl_process.c | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) (limited to 'erts/emulator/beam/erl_process.c') diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c index 96d17306a5..a691a3c773 100644 --- a/erts/emulator/beam/erl_process.c +++ b/erts/emulator/beam/erl_process.c @@ -10749,8 +10749,8 @@ erl_create_process(Process* parent, /* Parent of process (default group leader). erts_aint32_t prio = (erts_aint32_t) PRIORITY_NORMAL; #ifdef SHCOPY_SPAWN unsigned shflags = 0; /* could be taken from so->flags, if necessary */ - shcopy_info info; - INITIALIZE_INFO(info); + erts_shcopy_t info; + INITIALIZE_SHCOPY(info); #endif #ifdef ERTS_SMP @@ -10887,7 +10887,7 @@ erl_create_process(Process* parent, /* Parent of process (default group leader). BM_SWAP_TIMER(system,copy); #ifdef SHCOPY_SPAWN p->arg_reg[2] = copy_shared_perform(args, arg_size, &info, &p->htop, &p->off_heap, shflags); - DESTROY_INFO(info); + DESTROY_SHCOPY(info); #else p->arg_reg[2] = copy_struct(args, arg_size, &p->htop, &p->off_heap); #endif -- cgit v1.2.3 From 5d764f988ab09326d24e39a172083b09ab364c6c Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Bj=C3=B6rn-Egil=20Dahlberg?= Date: Wed, 18 Nov 2015 17:58:14 +0100 Subject: Refactor sharing preserved copy flags The TMPBUF option is no longer needed due to is_literal test and NONE was only used for initial debugging. So we remove the entire option. --- erts/emulator/beam/erl_process.c | 5 ++--- 1 file changed, 2 insertions(+), 3 deletions(-) (limited to 'erts/emulator/beam/erl_process.c') diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c index a691a3c773..23616f36bf 100644 --- a/erts/emulator/beam/erl_process.c +++ b/erts/emulator/beam/erl_process.c @@ -10748,7 +10748,6 @@ erl_create_process(Process* parent, /* Parent of process (default group leader). erts_aint32_t state = 0; erts_aint32_t prio = (erts_aint32_t) PRIORITY_NORMAL; #ifdef SHCOPY_SPAWN - unsigned shflags = 0; /* could be taken from so->flags, if necessary */ erts_shcopy_t info; INITIALIZE_SHCOPY(info); #endif @@ -10804,7 +10803,7 @@ erl_create_process(Process* parent, /* Parent of process (default group leader). BM_SWAP_TIMER(system,size); #ifdef SHCOPY_SPAWN - arg_size = copy_shared_calculate(args, &info, shflags); + arg_size = copy_shared_calculate(args, &info); #else arg_size = size_object(args); #endif @@ -10886,7 +10885,7 @@ erl_create_process(Process* parent, /* Parent of process (default group leader). BM_START_TIMER(system); BM_SWAP_TIMER(system,copy); #ifdef SHCOPY_SPAWN - p->arg_reg[2] = copy_shared_perform(args, arg_size, &info, &p->htop, &p->off_heap, shflags); + p->arg_reg[2] = copy_shared_perform(args, arg_size, &info, &p->htop, &p->off_heap); DESTROY_SHCOPY(info); #else p->arg_reg[2] = copy_struct(args, arg_size, &p->htop, &p->off_heap); -- cgit v1.2.3 From 02f038355d16e9b6474837727878f58e4ca669c8 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Bj=C3=B6rn-Egil=20Dahlberg?= Date: Wed, 18 Nov 2015 20:12:38 +0100 Subject: Use sharing preserving copy error messages and exceptions --- erts/emulator/beam/erl_process.c | 33 +++++++++++++++++++++++---------- 1 file changed, 23 insertions(+), 10 deletions(-) (limited to 'erts/emulator/beam/erl_process.c') diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c index 23616f36bf..f347d57a9d 100644 --- a/erts/emulator/beam/erl_process.c +++ b/erts/emulator/beam/erl_process.c @@ -11520,31 +11520,44 @@ send_exit_message(Process *to, ErtsProcLocks *to_locksp, { ErtsMessage *mp; ErlOffHeap *ohp; + Eterm* hp; + Eterm mess; +#ifdef SHCOPY_SEND + erts_shcopy_t info; +#endif if (token == NIL #ifdef USE_VM_PROBES || token == am_have_dt_utag #endif ) { - Eterm* hp; - Eterm mess; - - mp = erts_alloc_message_heap(to, to_locksp, - term_size, &hp, &ohp); +#ifdef SHCOPY_SEND + INITIALIZE_SHCOPY(info); + term_size = copy_shared_calculate(exit_term, &info); + mp = erts_alloc_message_heap(to, to_locksp, term_size, &hp, &ohp); + mess = copy_shared_perform(exit_term, term_size, &info, &hp, ohp); + DESTROY_SHCOPY(info); +#else + mp = erts_alloc_message_heap(to, to_locksp, term_size, &hp, &ohp); mess = copy_struct(exit_term, term_size, &hp, ohp); +#endif erts_queue_message(to, to_locksp, mp, mess, NIL); } else { - Eterm* hp; - Eterm mess; Eterm temp_token; Uint sz_token; ASSERT(is_tuple(token)); sz_token = size_object(token); - - mp = erts_alloc_message_heap(to, to_locksp, - term_size+sz_token, &hp, &ohp); +#ifdef SHCOPY_SEND + INITIALIZE_SHCOPY(info); + term_size = copy_shared_calculate(exit_term, &info); + mp = erts_alloc_message_heap(to, to_locksp, term_size, &hp, &ohp); + mess = copy_shared_perform(exit_term, term_size, &info, &hp, ohp); + DESTROY_SHCOPY(info); +#else + mp = erts_alloc_message_heap(to, to_locksp, term_size+sz_token, &hp, &ohp); mess = copy_struct(exit_term, term_size, &hp, ohp); +#endif /* the trace token must in this case be updated by the caller */ seq_trace_output(token, mess, SEQ_TRACE_SEND, to->common.id, NULL); temp_token = copy_struct(token, sz_token, &hp, ohp); -- cgit v1.2.3 From 6ff15f23c68db356bbad6ab5f939c191b58d453d Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Bj=C3=B6rn-Egil=20Dahlberg?= Date: Thu, 19 Nov 2015 19:36:08 +0100 Subject: Refactor have seq_trace token test --- erts/emulator/beam/erl_process.c | 12 ++---------- 1 file changed, 2 insertions(+), 10 deletions(-) (limited to 'erts/emulator/beam/erl_process.c') diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c index f347d57a9d..5907dd4567 100644 --- a/erts/emulator/beam/erl_process.c +++ b/erts/emulator/beam/erl_process.c @@ -11526,11 +11526,7 @@ send_exit_message(Process *to, ErtsProcLocks *to_locksp, erts_shcopy_t info; #endif - if (token == NIL -#ifdef USE_VM_PROBES - || token == am_have_dt_utag -#endif - ) { + if (!have_seqtrace(token)) { #ifdef SHCOPY_SEND INITIALIZE_SHCOPY(info); term_size = copy_shared_calculate(exit_term, &info); @@ -11670,11 +11666,7 @@ send_exit_signal(Process *c_p, /* current process if and only if ((state & ERTS_PSFLG_TRAP_EXIT) && (reason != am_kill || (flags & ERTS_XSIG_FLG_IGN_KILL))) { - if (is_not_nil(token) -#ifdef USE_VM_PROBES - && token != am_have_dt_utag -#endif - && token_update) + if (have_seqtrace(token)) seq_trace_update_send(token_update); if (is_value(exit_tuple)) send_exit_message(rp, rp_locks, exit_tuple, exit_tuple_sz, token); -- cgit v1.2.3 From 5cf556df2e178ce78d0c4197a73654d593f58f92 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Bj=C3=B6rn-Egil=20Dahlberg?= Date: Tue, 24 Nov 2015 16:05:04 +0100 Subject: Fix seq_trace refactoring bug --- erts/emulator/beam/erl_process.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'erts/emulator/beam/erl_process.c') diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c index 5907dd4567..a8b7211793 100644 --- a/erts/emulator/beam/erl_process.c +++ b/erts/emulator/beam/erl_process.c @@ -11666,7 +11666,7 @@ send_exit_signal(Process *c_p, /* current process if and only if ((state & ERTS_PSFLG_TRAP_EXIT) && (reason != am_kill || (flags & ERTS_XSIG_FLG_IGN_KILL))) { - if (have_seqtrace(token)) + if (have_seqtrace(token) && token_update) seq_trace_update_send(token_update); if (is_value(exit_tuple)) send_exit_message(rp, rp_locks, exit_tuple, exit_tuple_sz, token); -- cgit v1.2.3 From bc08768e5bad4baa2eaec4c6798ead9ded0a8889 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Bj=C3=B6rn-Egil=20Dahlberg?= Date: Tue, 24 Nov 2015 18:35:11 +0100 Subject: Fix seq_trace token copy size --- erts/emulator/beam/erl_process.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'erts/emulator/beam/erl_process.c') diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c index a8b7211793..a814462668 100644 --- a/erts/emulator/beam/erl_process.c +++ b/erts/emulator/beam/erl_process.c @@ -11547,7 +11547,7 @@ send_exit_message(Process *to, ErtsProcLocks *to_locksp, #ifdef SHCOPY_SEND INITIALIZE_SHCOPY(info); term_size = copy_shared_calculate(exit_term, &info); - mp = erts_alloc_message_heap(to, to_locksp, term_size, &hp, &ohp); + mp = erts_alloc_message_heap(to, to_locksp, term_size+sz_token, &hp, &ohp); mess = copy_shared_perform(exit_term, term_size, &info, &hp, ohp); DESTROY_SHCOPY(info); #else -- cgit v1.2.3 From ac529b0326496e52f3289464f9410001bc3bde6d Mon Sep 17 00:00:00 2001 From: Rickard Green Date: Mon, 16 Nov 2015 11:21:49 +0100 Subject: Fix memory leaks --- erts/emulator/beam/erl_process.c | 17 +++++++++++------ 1 file changed, 11 insertions(+), 6 deletions(-) (limited to 'erts/emulator/beam/erl_process.c') diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c index 9acce8acb6..ebb4d323e6 100644 --- a/erts/emulator/beam/erl_process.c +++ b/erts/emulator/beam/erl_process.c @@ -11259,6 +11259,7 @@ erts_cleanup_empty_process(Process* p) static void delete_process(Process* p) { + Eterm *heap; VERBOSE(DEBUG_PROCESSES, ("Removing process: %T\n",p->common.id)); /* Cleanup psd */ @@ -11283,16 +11284,17 @@ delete_process(Process* p) * Release heaps. Clobber contents in DEBUG build. */ - -#ifdef DEBUG - sys_memset(p->heap, DEBUG_BAD_BYTE, p->heap_sz*sizeof(Eterm)); -#endif - #ifdef HIPE hipe_delete_process(&p->hipe); #endif - ERTS_HEAP_FREE(ERTS_ALC_T_HEAP, (void*) p->heap, p->heap_sz*sizeof(Eterm)); + heap = p->abandoned_heap ? p->abandoned_heap : p->heap; + +#ifdef DEBUG + sys_memset(heap, DEBUG_BAD_BYTE, p->heap_sz*sizeof(Eterm)); +#endif + + ERTS_HEAP_FREE(ERTS_ALC_T_HEAP, (void*) heap, p->heap_sz*sizeof(Eterm)); if (p->old_heap != NULL) { #ifdef DEBUG @@ -11311,6 +11313,9 @@ delete_process(Process* p) free_message_buffer(p->mbuf); } + if (p->msg_frag) + erts_cleanup_messages(p->msg_frag); + erts_erase_dicts(p); /* free all pending messages */ -- cgit v1.2.3 From ccbb5ba4d0198f3ba21b83ae1aca3fe52a7f2c80 Mon Sep 17 00:00:00 2001 From: Rickard Green Date: Tue, 17 Nov 2015 15:21:56 +0100 Subject: Remove unused variable --- erts/emulator/beam/erl_process.c | 4 +--- 1 file changed, 1 insertion(+), 3 deletions(-) (limited to 'erts/emulator/beam/erl_process.c') diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c index ebb4d323e6..ebe9361b8d 100644 --- a/erts/emulator/beam/erl_process.c +++ b/erts/emulator/beam/erl_process.c @@ -335,7 +335,6 @@ static ErtsAlignedSchedulerSleepInfo *aligned_dirty_io_sched_sleep_info; static Uint last_reductions; static Uint last_exact_reductions; -Uint erts_default_process_flags; Eterm erts_system_monitor; Eterm erts_system_monitor_long_gc; Uint erts_system_monitor_long_schedule; @@ -682,7 +681,6 @@ erts_init_process(int ncpu, int proc_tab_size, int legacy_proc_tab) last_reductions = 0; last_exact_reductions = 0; - erts_default_process_flags = 0; } void @@ -10737,7 +10735,7 @@ erl_create_process(Process* parent, /* Parent of process (default group leader). Eterm args, /* Arguments for function (must be well-formed list). */ ErlSpawnOpts* so) /* Options for spawn. */ { - Uint flags = erts_default_process_flags; + Uint flags = 0; ErtsRunQueue *rq = NULL; Process *p; Sint arity; /* Number of arguments. */ -- cgit v1.2.3 From a5dd6499d53ed596e2f8aa17ee35ff87cd32fe60 Mon Sep 17 00:00:00 2001 From: Rickard Green Date: Thu, 19 Nov 2015 17:07:06 +0100 Subject: Distinguish between GC disabled by BIFs and other disabled GC Processes remember heap fragments that are known to be fully live due to creation in a just called BIF that yields in the live_hf_end field. This field must not be used if we have not disabled GC in a BIF. F_DELAY_GC has been introduced in order to distinguish between to two different scenarios. - F_DISABLE_GC should *only* be used by BIFs. This when the BIF needs to yield while preventig a GC. - F_DELAY_GC should only be used when GC is temporarily disabled while the process is scheduled. A process must not be scheduled out while F_DELAY_GC is set. --- erts/emulator/beam/erl_process.c | 2 ++ 1 file changed, 2 insertions(+) (limited to 'erts/emulator/beam/erl_process.c') diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c index ebe9361b8d..d17355207b 100644 --- a/erts/emulator/beam/erl_process.c +++ b/erts/emulator/beam/erl_process.c @@ -9262,6 +9262,8 @@ Process *schedule(Process *p, int calls) } else { sched_out_proc: + ASSERT(!(p->flags & F_DELAY_GC)); + #ifdef ERTS_SMP ERTS_SMP_CHK_HAVE_ONLY_MAIN_PROC_LOCK(p); esdp = p->scheduler_data; -- cgit v1.2.3 From 19c4689eea86f26c5af9b8f712c227ce4f62310b Mon Sep 17 00:00:00 2001 From: Rickard Green Date: Tue, 24 Nov 2015 15:57:55 +0100 Subject: Replace off_heap_message_queue option with message_queue_data option The message_queue_data option can have the values - off_heap - on_heap - mixed --- erts/emulator/beam/erl_process.c | 4 ++++ 1 file changed, 4 insertions(+) (limited to 'erts/emulator/beam/erl_process.c') diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c index d17355207b..7a31aa3e33 100644 --- a/erts/emulator/beam/erl_process.c +++ b/erts/emulator/beam/erl_process.c @@ -10779,6 +10779,10 @@ erl_create_process(Process* parent, /* Parent of process (default group leader). state |= ERTS_PSFLG_OFF_HEAP_MSGQ; flags |= F_OFF_HEAP_MSGQ; } + else if (so->flags & SPO_ON_HEAP_MSGQ) { + state |= ERTS_PSFLG_ON_HEAP_MSGQ; + flags |= F_ON_HEAP_MSGQ; + } if (!rq) rq = erts_get_runq_proc(parent); -- cgit v1.2.3 From 14c7fefd51be035a44bfe42127fb4b9df92d760b Mon Sep 17 00:00:00 2001 From: Lukas Larsson Date: Mon, 6 Jul 2015 17:13:52 +0200 Subject: erts: Create forker process for spawn driver Instead of forking from the beam process, we create a separate process in which all forks are done. This has several advantages: 1) performance: * don't have to close all fd's in the world * fork only has to copy stuff from a small process * work is done in a completely seperate process * a 3x performance increase has been measured, can be made even greater (10x) if we cache the environment in child setup 2) stability * the exec is done in another process than beam, which means that if the file that we exec to is on an nfs that is not available right now we will not block a scheduler until the nfs returns. 3) simplicity * don't have to deal with SIGCHLD in the erts Unfortunately, this solution also implies some badness. 1) There will always be a seperate process running together with beam on unix. This could be confusing and undesirable. 2) We have to transfer the entire environment to child_setup for each command. OTP-13088 --- erts/emulator/beam/erl_process.c | 37 ------------------------------------- 1 file changed, 37 deletions(-) (limited to 'erts/emulator/beam/erl_process.c') diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c index 15a6d5d651..03cb3dc254 100644 --- a/erts/emulator/beam/erl_process.c +++ b/erts/emulator/beam/erl_process.c @@ -497,9 +497,6 @@ dbg_chk_aux_work_val(erts_aint32_t value) #if HAVE_ERTS_MSEG valid |= ERTS_SSI_AUX_WORK_MSEG_CACHE_CHECK; #endif -#ifdef ERTS_SMP_SCHEDULERS_NEED_TO_CHECK_CHILDREN - valid |= ERTS_SSI_AUX_WORK_CHECK_CHILDREN; -#endif #ifdef ERTS_SSI_AUX_WORK_REAP_PORTS valid |= ERTS_SSI_AUX_WORK_REAP_PORTS; #endif @@ -586,8 +583,6 @@ erts_pre_init_process(void) = "MISC_THR_PRGR"; erts_aux_work_flag_descr[ERTS_SSI_AUX_WORK_MISC_IX] = "MISC"; - erts_aux_work_flag_descr[ERTS_SSI_AUX_WORK_CHECK_CHILDREN_IX] - = "CHECK_CHILDREN"; erts_aux_work_flag_descr[ERTS_SSI_AUX_WORK_SET_TMO_IX] = "SET_TMO"; erts_aux_work_flag_descr[ERTS_SSI_AUX_WORK_MSEG_CACHE_CHECK_IX] @@ -2100,34 +2095,6 @@ erts_debug_wait_completed(Process *c_p, int flags) } -#ifdef ERTS_SMP_SCHEDULERS_NEED_TO_CHECK_CHILDREN -void -erts_smp_notify_check_children_needed(void) -{ - int i; - for (i = 0; i < erts_no_schedulers; i++) - set_aux_work_flags_wakeup_nob(ERTS_SCHED_SLEEP_INFO_IX(i), - ERTS_SSI_AUX_WORK_CHECK_CHILDREN); -#ifdef ERTS_DIRTY_SCHEDULERS - for (i = 0; i < erts_no_dirty_cpu_schedulers; i++) - set_aux_work_flags_wakeup_nob(ERTS_DIRTY_CPU_SCHED_SLEEP_INFO_IX(i), - ERTS_SSI_AUX_WORK_CHECK_CHILDREN); - for (i = 0; i < erts_no_dirty_io_schedulers; i++) - set_aux_work_flags_wakeup_nob(ERTS_DIRTY_IO_SCHED_SLEEP_INFO_IX(i), - ERTS_SSI_AUX_WORK_CHECK_CHILDREN); -#endif -} - -static ERTS_INLINE erts_aint32_t -handle_check_children(ErtsAuxWorkData *awdp, erts_aint32_t aux_work, int waiting) -{ - unset_aux_work_flags(awdp->ssi, ERTS_SSI_AUX_WORK_CHECK_CHILDREN); - erts_check_children(); - return aux_work & ~ERTS_SSI_AUX_WORK_CHECK_CHILDREN; -} - -#endif - static void notify_reap_ports_relb(void) { @@ -2281,10 +2248,6 @@ handle_aux_work(ErtsAuxWorkData *awdp, erts_aint32_t orig_aux_work, int waiting) HANDLE_AUX_WORK(ERTS_SSI_AUX_WORK_MISC, handle_misc_aux_work); -#ifdef ERTS_SMP_SCHEDULERS_NEED_TO_CHECK_CHILDREN - HANDLE_AUX_WORK(ERTS_SSI_AUX_WORK_CHECK_CHILDREN, - handle_check_children); -#endif HANDLE_AUX_WORK(ERTS_SSI_AUX_WORK_SET_TMO, handle_setup_aux_work_timer); -- cgit v1.2.3 From 47a7d8b9d701b81355a02f4bad8d16a327bc1588 Mon Sep 17 00:00:00 2001 From: Steve Vinoski Date: Mon, 21 Dec 2015 23:07:10 -0500 Subject: Do not allow aux work on dirty schedulers The nature of aux work is such that dirty schedulers should not attempt to perform it. Modify the code to ensure that dirty schedulers avoid aux work. Also fix an incorrect assumption about the size of a Uint in the ErtsDirtySchedId type. --- erts/emulator/beam/erl_process.c | 75 +++++++++++++++++++--------------------- 1 file changed, 36 insertions(+), 39 deletions(-) (limited to 'erts/emulator/beam/erl_process.c') diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c index d583118e7b..eae05f1651 100644 --- a/erts/emulator/beam/erl_process.c +++ b/erts/emulator/beam/erl_process.c @@ -2239,6 +2239,7 @@ handle_aux_work(ErtsAuxWorkData *awdp, erts_aint32_t orig_aux_work, int waiting) erts_aint32_t aux_work = orig_aux_work; erts_aint32_t ignore = 0; + ASSERT(!ERTS_SCHEDULER_IS_DIRTY(awdp->esdp)); #ifdef ERTS_SMP haw_thr_prgr_current_reset(awdp); #endif @@ -2972,14 +2973,13 @@ scheduler_wait(int *fcalls, ErtsSchedulerData *esdp, ErtsRunQueue *rq) ErtsMonotonicTime current_time; aux_work = erts_atomic32_read_acqb(&ssi->aux_work); - if (aux_work) { - if (!ERTS_SCHEDULER_IS_DIRTY(esdp) && !thr_prgr_active) { + if (aux_work && !ERTS_SCHEDULER_IS_DIRTY(esdp)) { + if (!thr_prgr_active) { erts_thr_progress_active(esdp, thr_prgr_active = 1); sched_wall_time_change(esdp, 1); } aux_work = handle_aux_work(&esdp->aux_work_data, aux_work, 1); - if (aux_work && !ERTS_SCHEDULER_IS_DIRTY(esdp) - && erts_thr_progress_update(esdp)) + if (aux_work && erts_thr_progress_update(esdp)) erts_thr_progress_leader_update(esdp); } @@ -3131,19 +3131,16 @@ scheduler_wait(int *fcalls, ErtsSchedulerData *esdp, ErtsRunQueue *rq) #endif aux_work = erts_atomic32_read_acqb(&ssi->aux_work); - if (aux_work) { - if (!ERTS_SCHEDULER_IS_DIRTY(esdp)) { - if (!working) - sched_wall_time_change(esdp, working = 1); + if (aux_work && !ERTS_SCHEDULER_IS_DIRTY(esdp)) { + if (!working) + sched_wall_time_change(esdp, working = 1); #ifdef ERTS_SMP - if (!thr_prgr_active) - erts_thr_progress_active(esdp, thr_prgr_active = 1); + if (!thr_prgr_active) + erts_thr_progress_active(esdp, thr_prgr_active = 1); #endif - } aux_work = handle_aux_work(&esdp->aux_work_data, aux_work, 1); #ifdef ERTS_SMP - if (!ERTS_SCHEDULER_IS_DIRTY(esdp) && aux_work && - erts_thr_progress_update(esdp)) + if (aux_work && erts_thr_progress_update(esdp)) erts_thr_progress_leader_update(esdp); #endif } @@ -6798,18 +6795,19 @@ suspend_scheduler(ErtsSchedulerData *esdp) & ERTS_RUNQ_FLGS_QMASK); aux_work = erts_atomic32_read_acqb(&ssi->aux_work); if (aux_work|qmask) { - if (!ERTS_SCHEDULER_IS_DIRTY(esdp) && !thr_prgr_active) { - erts_thr_progress_active(esdp, thr_prgr_active = 1); - sched_wall_time_change(esdp, 1); + if (!ERTS_SCHEDULER_IS_DIRTY(esdp)) { + if (!thr_prgr_active) { + erts_thr_progress_active(esdp, thr_prgr_active = 1); + sched_wall_time_change(esdp, 1); + } + if (aux_work) + aux_work = handle_aux_work(&esdp->aux_work_data, + aux_work, + 1); + + if (aux_work && erts_thr_progress_update(esdp)) + erts_thr_progress_leader_update(esdp); } - if (aux_work) - aux_work = handle_aux_work(&esdp->aux_work_data, - aux_work, - 1); - - if (!ERTS_SCHEDULER_IS_DIRTY(esdp) && - (aux_work && erts_thr_progress_update(esdp))) - erts_thr_progress_leader_update(esdp); if (qmask) { #ifdef ERTS_DIRTY_SCHEDULERS if (ERTS_SCHEDULER_IS_DIRTY(esdp)) { @@ -7026,17 +7024,18 @@ suspend_scheduler(ErtsSchedulerData *esdp) & ERTS_RUNQ_FLGS_QMASK); aux_work = erts_atomic32_read_acqb(&ssi->aux_work); if (aux_work|qmask) { - if (!ERTS_SCHEDULER_IS_DIRTY(esdp) && !thr_prgr_active) { - erts_thr_progress_active(esdp, thr_prgr_active = 1); - sched_wall_time_change(esdp, 1); + if (!ERTS_SCHEDULER_IS_DIRTY(esdp)) { + if (!thr_prgr_active) { + erts_thr_progress_active(esdp, thr_prgr_active = 1); + sched_wall_time_change(esdp, 1); + } + if (aux_work) + aux_work = handle_aux_work(&esdp->aux_work_data, + aux_work, + 1); + if (aux_work && erts_thr_progress_update(esdp)) + erts_thr_progress_leader_update(esdp); } - if (aux_work) - aux_work = handle_aux_work(&esdp->aux_work_data, - aux_work, - 1); - if (!ERTS_SCHEDULER_IS_DIRTY(esdp) && aux_work && - erts_thr_progress_update(esdp)) - erts_thr_progress_leader_update(esdp); if (qmask) { erts_smp_runq_lock(esdp->run_queue); evacuate_run_queue(esdp->run_queue, &sbp); @@ -9407,10 +9406,9 @@ Process *schedule(Process *p, int calls) suspend_scheduler(esdp); #endif - { + if (!ERTS_SCHEDULER_IS_DIRTY(esdp)) { erts_aint32_t aux_work; - int leader_update = ERTS_SCHEDULER_IS_DIRTY(esdp) ? 0 - : erts_thr_progress_update(esdp); + int leader_update = erts_thr_progress_update(esdp); aux_work = erts_atomic32_read_acqb(&esdp->ssi->aux_work); if (aux_work | leader_update | ERTS_SCHED_FAIR) { erts_smp_runq_unlock(rq); @@ -9423,8 +9421,7 @@ Process *schedule(Process *p, int calls) erts_smp_runq_lock(rq); } - ERTS_SMP_LC_ASSERT(ERTS_SCHEDULER_IS_DIRTY(esdp) - || !erts_thr_progress_is_blocking()); + ERTS_SMP_LC_ASSERT(!erts_thr_progress_is_blocking()); } ERTS_SMP_LC_ASSERT(erts_smp_lc_runq_is_locked(rq)); -- cgit v1.2.3 From 37f58ad6bff2bf2bac4f3f20c2684e8cee66af03 Mon Sep 17 00:00:00 2001 From: Rickard Green Date: Tue, 15 Dec 2015 09:40:30 +0100 Subject: Light weight statistics of run queue lengths - statistics(total_run_queue_lengths) - statistics(run_queue_lengths) - statistics(total_active_tasks) - statistics(active_tasks) Conflicts: erts/emulator/beam/erl_process.c --- erts/emulator/beam/erl_process.c | 65 +++++++++++++++++++++++++++------------- 1 file changed, 44 insertions(+), 21 deletions(-) (limited to 'erts/emulator/beam/erl_process.c') diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c index d583118e7b..b47aae0f74 100644 --- a/erts/emulator/beam/erl_process.c +++ b/erts/emulator/beam/erl_process.c @@ -3149,7 +3149,7 @@ scheduler_wait(int *fcalls, ErtsSchedulerData *esdp, ErtsRunQueue *rq) } #ifndef ERTS_SMP - if (rq->len != 0 || rq->misc.start) + if (erts_smp_atomic32_read_dirty(&rq->len) != 0 || rq->misc.start) goto sys_woken; #else flgs = erts_smp_atomic32_read_acqb(&ssi->flags); @@ -3248,7 +3248,7 @@ scheduler_wait(int *fcalls, ErtsSchedulerData *esdp, ErtsRunQueue *rq) } #ifndef ERTS_SMP - if (rq->len == 0 && !rq->misc.start) + if (erts_smp_atomic32_read_dirty(&rq->len) == 0 && !rq->misc.start) goto sys_aux_work; sys_woken: #else @@ -4965,7 +4965,7 @@ 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 = rq->len; + rq->max_len = erts_smp_atomic32_read_dirty(&rq->len); for (pix = 0; pix < ERTS_NO_PRIO_LEVELS; pix++) { ErtsRunQueueInfo *rqi; rqi = (pix == ERTS_PORT_PRIO_LEVEL @@ -5123,7 +5123,7 @@ wakeup_other_check(ErtsRunQueue *rq, Uint32 flags) { int wo_reds = rq->wakeup_other_reds; if (wo_reds) { - int left_len = rq->len - 1; + int left_len = erts_smp_atomic32_read_dirty(&rq->len) - 1; if (left_len < 1) { int wo_reduce = wo_reds << wakeup_other.dec_shift; wo_reduce &= wakeup_other.dec_mask; @@ -5196,7 +5196,7 @@ wakeup_other_check_legacy(ErtsRunQueue *rq, Uint32 flags) { int wo_reds = rq->wakeup_other_reds; if (wo_reds) { - erts_aint32_t len = rq->len; + erts_aint32_t len = erts_smp_atomic32_read_dirty(&rq->len); if (len < 2) { rq->wakeup_other -= ERTS_WAKEUP_OTHER_DEC_LEGACY*wo_reds; if (rq->wakeup_other < 0) @@ -5292,7 +5292,7 @@ runq_supervisor(void *unused) ErtsRunQueue *rq = ERTS_RUNQ_IX(ix); if (ERTS_RUNQ_FLGS_GET(rq) & ERTS_RUNQ_FLG_NONEMPTY) { erts_smp_runq_lock(rq); - if (rq->len != 0) + if (erts_smp_atomic32_read_dirty(&rq->len) != 0) wake_scheduler_on_empty_runq(rq); /* forced wakeup... */ erts_smp_runq_unlock(rq); } @@ -5642,7 +5642,7 @@ erts_init_scheduling(int no_schedulers, int no_schedulers_online } rq->out_of_work_count = 0; rq->max_len = 0; - rq->len = 0; + erts_smp_atomic32_set_nob(&rq->len, 0); rq->wakeup_other = 0; rq->wakeup_other_reds = 0; rq->halt_in_progress = 0; @@ -7939,6 +7939,9 @@ sched_thread_func(void *vesdp) erts_sched_init_time_sup(esdp); + (void) ERTS_RUNQ_FLGS_SET_NOB(esdp->run_queue, + ERTS_RUNQ_FLG_EXEC); + #ifdef ERTS_SMP tse = erts_tse_fetch(); erts_tse_prepare_timed(tse); @@ -8947,24 +8950,39 @@ resume_process_1(BIF_ALIST_1) } Uint -erts_run_queues_len(Uint *qlen) +erts_run_queues_len(Uint *qlen, int atomic_queues_read, int incl_active_sched) { int i = 0; Uint len = 0; - ERTS_ATOMIC_FOREACH_RUNQ(rq, - { - Sint pqlen = 0; - int pix; - for (pix = 0; pix < ERTS_NO_PROC_PRIO_LEVELS; pix++) - pqlen += RUNQ_READ_LEN(&rq->procs.prio_info[pix].len); + if (atomic_queues_read) + ERTS_ATOMIC_FOREACH_RUNQ(rq, + { + Sint rq_len = (Sint) erts_smp_atomic32_read_dirty(&rq->len); + ASSERT(rq_len >= 0); + if (incl_active_sched + && (ERTS_RUNQ_FLGS_GET_NOB(rq) & ERTS_RUNQ_FLG_EXEC)) { + rq_len++; + } + if (qlen) + qlen[i++] = rq_len; + len += (Uint) rq_len; + } + ); + else { + for (i = 0; i < erts_no_run_queues; i++) { + ErtsRunQueue *rq = ERTS_RUNQ_IX(i); + Sint rq_len = (Sint) erts_smp_atomic32_read_nob(&rq->len); + ASSERT(rq_len >= 0); + if (incl_active_sched + && (ERTS_RUNQ_FLGS_GET_NOB(rq) & ERTS_RUNQ_FLG_EXEC)) { + rq_len++; + } + if (qlen) + qlen[i] = rq_len; + len += (Uint) rq_len; + } - if (pqlen < 0) - pqlen = 0; - if (qlen) - qlen[i++] = pqlen; - len += pqlen; } - ); return len; } @@ -9391,8 +9409,10 @@ Process *schedule(Process *p, int calls) if (flags & (ERTS_RUNQ_FLG_CHK_CPU_BIND|ERTS_RUNQ_FLG_SUSPENDED)) { if (flags & ERTS_RUNQ_FLG_SUSPENDED) { + (void) ERTS_RUNQ_FLGS_UNSET_NOB(rq, ERTS_RUNQ_FLG_EXEC); suspend_scheduler(esdp); - flags = ERTS_RUNQ_FLGS_GET_NOB(rq); + flags = ERTS_RUNQ_FLGS_SET_NOB(rq, ERTS_RUNQ_FLG_EXEC); + flags |= ERTS_RUNQ_FLG_EXEC; } if (flags & ERTS_RUNQ_FLG_CHK_CPU_BIND) { flags = ERTS_RUNQ_FLGS_UNSET(rq, ERTS_RUNQ_FLG_CHK_CPU_BIND); @@ -9483,7 +9503,10 @@ Process *schedule(Process *p, int calls) } #endif + (void) ERTS_RUNQ_FLGS_UNSET(rq, ERTS_RUNQ_FLG_EXEC); scheduler_wait(&fcalls, esdp, rq); + flags = ERTS_RUNQ_FLGS_SET_NOB(rq, ERTS_RUNQ_FLG_EXEC); + flags |= ERTS_RUNQ_FLG_EXEC; #ifdef ERTS_SMP non_empty_runq(rq); -- cgit v1.2.3 From 4c763443365591e170308a1c5f11a4586734ca4e Mon Sep 17 00:00:00 2001 From: Sverker Eriksson Date: Tue, 12 Jan 2016 16:57:59 +0100 Subject: erts: Optimize erlang:check_process_code by ignoring literals. erts_internal:check_process_code will be called again anyway (with option {copy_literals, true}) before the module is actually purged. No need to check literals twice. --- erts/emulator/beam/erl_process.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) (limited to 'erts/emulator/beam/erl_process.c') diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c index 9495436ba6..d36d866b2f 100644 --- a/erts/emulator/beam/erl_process.c +++ b/erts/emulator/beam/erl_process.c @@ -10055,7 +10055,7 @@ execute_sys_tasks(Process *c_p, erts_aint32_t *statep, int in_reds) case ERTS_PSTT_CPC: st_res = erts_check_process_code(c_p, st->arg[0], - st->arg[1] == am_true, + unsigned_val(st->arg[1]), &reds); if (is_non_value(st_res)) { /* Needed gc, but gc was disabled */ @@ -10219,7 +10219,7 @@ erts_internal_request_system_task_3(BIF_ALIST_3) case am_check_process_code: if (is_not_atom(st->arg[0])) goto badarg; - if (st->arg[1] != am_true && st->arg[1] != am_false) + if (is_not_small(st->arg[1]) || (unsigned_val(st->arg[1]) & ~ERTS_CPC_ALL)) goto badarg; noproc_res = am_false; st->type = ERTS_PSTT_CPC; -- cgit v1.2.3 From 52738d8bee488c262cf514ce4df49e87dfa47bc3 Mon Sep 17 00:00:00 2001 From: Sverker Eriksson Date: Mon, 18 Jan 2016 18:47:30 +0100 Subject: erts: Fix race between receive timeout and exit signal Must re-read 'state' after seizing proc locks as other thread may have set EXITING. --- erts/emulator/beam/erl_process.c | 2 ++ 1 file changed, 2 insertions(+) (limited to 'erts/emulator/beam/erl_process.c') diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c index 7b3d12ce09..d767e1bb5a 100644 --- a/erts/emulator/beam/erl_process.c +++ b/erts/emulator/beam/erl_process.c @@ -9674,6 +9674,8 @@ Process *schedule(Process *p, int calls) erts_smp_proc_lock(p, ERTS_PROC_LOCK_MAIN|ERTS_PROC_LOCK_STATUS); + state = erts_smp_atomic32_read_nob(&p->state); + if (erts_sched_stat.enabled) { int prio; UWord old = ERTS_PROC_SCHED_ID(p, -- cgit v1.2.3 From 693db9ef9c5a67b36215f21c32f91a986fc5630d Mon Sep 17 00:00:00 2001 From: Steve Vinoski Date: Mon, 18 Jan 2016 14:04:24 -0500 Subject: Fix dirty scheduler check in handle_aux_work --- erts/emulator/beam/erl_process.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'erts/emulator/beam/erl_process.c') diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c index 96d9a2f8b4..dd8bc9a698 100644 --- a/erts/emulator/beam/erl_process.c +++ b/erts/emulator/beam/erl_process.c @@ -2239,7 +2239,7 @@ handle_aux_work(ErtsAuxWorkData *awdp, erts_aint32_t orig_aux_work, int waiting) erts_aint32_t aux_work = orig_aux_work; erts_aint32_t ignore = 0; - ASSERT(!ERTS_SCHEDULER_IS_DIRTY(awdp->esdp)); + ASSERT(!awdp->esdp || !ERTS_SCHEDULER_IS_DIRTY(awdp->esdp)); #ifdef ERTS_SMP haw_thr_prgr_current_reset(awdp); #endif -- cgit v1.2.3 From 858c6f7fa44f7b2dc363b359198d6522dd60e914 Mon Sep 17 00:00:00 2001 From: Rickard Green Date: Tue, 5 Jan 2016 16:55:04 +0100 Subject: Introduce time warp safe trace timestamp formats New timestamp options for trace, sequential trace, and system profile: - monotonic_timestamp - strict_monotonic_timestamp --- erts/emulator/beam/erl_process.c | 1 + 1 file changed, 1 insertion(+) (limited to 'erts/emulator/beam/erl_process.c') diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c index b47aae0f74..c1d2e8db8e 100644 --- a/erts/emulator/beam/erl_process.c +++ b/erts/emulator/beam/erl_process.c @@ -353,6 +353,7 @@ struct erts_system_monitor_flags_t erts_system_monitor_flags; /* system performance monitor */ Eterm erts_system_profile; struct erts_system_profile_flags_t erts_system_profile_flags; +int erts_system_profile_ts_type = ERTS_TRACE_FLG_NOW_TIMESTAMP; #if ERTS_MAX_PROCESSES > 0x7fffffff #error "Need to store process_count in another type" -- cgit v1.2.3 From d76ee58c07f32dfc0652844ec2b513af2105ffa1 Mon Sep 17 00:00:00 2001 From: Lukas Larsson Date: Wed, 22 Oct 2014 17:15:31 +0200 Subject: erts: Add ERTS_WRITE_UNLIKELY ERTS_WRITE_UNLIKELY can be used to place global variables in a specific section where only data that is very rarely modified sits. This is used to improve cache locality. --- erts/emulator/beam/erl_process.c | 24 ++++++++++++------------ 1 file changed, 12 insertions(+), 12 deletions(-) (limited to 'erts/emulator/beam/erl_process.c') diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c index 7d194f9840..3902632ef8 100644 --- a/erts/emulator/beam/erl_process.c +++ b/erts/emulator/beam/erl_process.c @@ -148,14 +148,14 @@ extern BeamInstr beam_apply[]; extern BeamInstr beam_exit[]; extern BeamInstr beam_continue_exit[]; -int erts_default_spo_flags = 0; -int erts_eager_check_io = 1; -int erts_sched_compact_load; -int erts_sched_balance_util = 0; -Uint erts_no_schedulers; +int ERTS_WRITE_UNLIKELY(erts_default_spo_flags) = 0; +int ERTS_WRITE_UNLIKELY(erts_eager_check_io) = 1; +int ERTS_WRITE_UNLIKELY(erts_sched_compact_load); +int ERTS_WRITE_UNLIKELY(erts_sched_balance_util) = 0; +Uint ERTS_WRITE_UNLIKELY(erts_no_schedulers); #ifdef ERTS_DIRTY_SCHEDULERS -Uint erts_no_dirty_cpu_schedulers; -Uint erts_no_dirty_io_schedulers; +Uint ERTS_WRITE_UNLIKELY(erts_no_dirty_cpu_schedulers); +Uint ERTS_WRITE_UNLIKELY(erts_no_dirty_io_schedulers); #endif static char *erts_aux_work_flag_descr[ERTS_SSI_AUX_WORK_NO_FLAGS] = {0}; @@ -295,7 +295,7 @@ do { \ erts_sched_stat_t erts_sched_stat; #ifdef USE_THREADS -static erts_tsd_key_t sched_data_key; +static erts_tsd_key_t ERTS_WRITE_UNLIKELY(sched_data_key); #endif static erts_smp_atomic32_t function_calls; @@ -335,10 +335,10 @@ static ErtsAlignedSchedulerSleepInfo *aligned_dirty_io_sched_sleep_info; static Uint last_reductions; static Uint last_exact_reductions; -Eterm erts_system_monitor; -Eterm erts_system_monitor_long_gc; -Uint erts_system_monitor_long_schedule; -Eterm erts_system_monitor_large_heap; +Eterm ERTS_WRITE_UNLIKELY(erts_system_monitor); +Eterm ERTS_WRITE_UNLIKELY(erts_system_monitor_long_gc); +Uint ERTS_WRITE_UNLIKELY(erts_system_monitor_long_schedule); +Eterm ERTS_WRITE_UNLIKELY(erts_system_monitor_large_heap); struct erts_system_monitor_flags_t erts_system_monitor_flags; /* system performance monitor */ -- cgit v1.2.3 From 664ed2a6fd2b324bb6b56db3d3eca853cfda8f61 Mon Sep 17 00:00:00 2001 From: Lukas Larsson Date: Fri, 12 Sep 2014 16:38:00 +0200 Subject: erts: Add microstate accounting Microstate accounting is a way to track which state the different threads within ERTS are in. The main usage area is to pin point performance bottlenecks by checking which states the threads are in and then from there figuring out why and where to optimize. Since checking whether microstate accounting is on or off is relatively expensive if done in a short loop only a few of the states are enabled by default and more states can be enabled through configure. I've done some benchmarking and the overhead with it turned off is not noticible and with it on it is a fraction of a percent. If you enable the extra states, depending on the benchmark, the ovehead when turned off is about 1% and when turned on somewhere inbetween 5-15%. OTP-12345 --- erts/emulator/beam/erl_process.c | 48 ++++++++++++++++++++++++++++++++++++++-- 1 file changed, 46 insertions(+), 2 deletions(-) (limited to 'erts/emulator/beam/erl_process.c') diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c index 3902632ef8..8e1ebe795c 100644 --- a/erts/emulator/beam/erl_process.c +++ b/erts/emulator/beam/erl_process.c @@ -941,6 +941,12 @@ sched_wall_time_change(ErtsSchedulerData *esdp, int working) } } } + if (!working) { + ERTS_MSACC_SET_STATE_M_X(ERTS_MSACC_STATE_BUSY_WAIT); + } else { + ERTS_MSACC_SET_STATE_M_X(ERTS_MSACC_STATE_OTHER); + } + } typedef struct { @@ -1696,15 +1702,17 @@ handle_delayed_dealloc(ErtsAuxWorkData *awdp, erts_aint32_t aux_work, int waitin int need_thr_progress = 0; ErtsThrPrgrVal wakeup = ERTS_THR_PRGR_INVALID; int more_work = 0; - + ERTS_MSACC_PUSH_STATE_M_X(); #ifdef ERTS_DIRTY_SCHEDULERS ASSERT(!awdp->esdp || !ERTS_SCHEDULER_IS_DIRTY(awdp->esdp)); #endif unset_aux_work_flags(ssi, ERTS_SSI_AUX_WORK_DD); + ERTS_MSACC_SET_STATE_CACHED_M_X(ERTS_MSACC_STATE_ALLOC); erts_alloc_scheduler_handle_delayed_dealloc((void *) awdp->esdp, &need_thr_progress, &wakeup, &more_work); + ERTS_MSACC_POP_STATE_M_X(); if (more_work) { if (set_aux_work_flags(ssi, ERTS_SSI_AUX_WORK_DD) & ERTS_SSI_AUX_WORK_DD_THR_PRGR) { @@ -2176,12 +2184,15 @@ handle_aux_work(ErtsAuxWorkData *awdp, erts_aint32_t orig_aux_work, int waiting) ERTS_DBG_CHK_AUX_WORK_VAL(aux_work); \ if (!(aux_work & ~ignore)) { \ ERTS_DBG_CHK_AUX_WORK_VAL(aux_work); \ + ERTS_MSACC_UPDATE_CACHE(); \ + ERTS_MSACC_POP_STATE_M(); \ return aux_work; \ } \ } erts_aint32_t aux_work = orig_aux_work; erts_aint32_t ignore = 0; + ERTS_MSACC_PUSH_AND_SET_STATE_M(ERTS_MSACC_STATE_AUX); ASSERT(!awdp->esdp || !ERTS_SCHEDULER_IS_DIRTY(awdp->esdp)); #ifdef ERTS_SMP @@ -2272,6 +2283,8 @@ handle_aux_work(ErtsAuxWorkData *awdp, erts_aint32_t orig_aux_work, int waiting) haw_thr_prgr_current_check_progress(awdp); #endif + ERTS_MSACC_UPDATE_CACHE(); + ERTS_MSACC_POP_STATE_M(); return aux_work; #undef HANDLE_AUX_WORK @@ -2779,6 +2792,8 @@ aux_thread(void *unused) ssi->event = erts_tse_fetch(); + erts_msacc_init_thread("aux", 1, 1); + callbacks.arg = (void *) ssi; callbacks.wakeup = thr_prgr_wakeup; callbacks.prepare_wait = thr_prgr_prep_wait; @@ -2789,6 +2804,7 @@ aux_thread(void *unused) init_aux_work_data(awdp, NULL, NULL); awdp->ssi = ssi; + sched_prep_spin_wait(ssi); while (1) { @@ -2842,6 +2858,9 @@ scheduler_wait(int *fcalls, ErtsSchedulerData *esdp, ErtsRunQueue *rq) #ifdef ERTS_SMP int thr_prgr_active = 1; erts_aint32_t flgs; +#endif + ERTS_MSACC_PUSH_STATE_M(); +#ifdef ERTS_SMP ERTS_SMP_LC_ASSERT(erts_smp_lc_runq_is_locked(rq)); @@ -2898,6 +2917,7 @@ scheduler_wait(int *fcalls, ErtsSchedulerData *esdp, ErtsRunQueue *rq) 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); } @@ -2959,7 +2979,9 @@ scheduler_wait(int *fcalls, ErtsSchedulerData *esdp, ErtsRunQueue *rq) - 1) + 1; } else timeout = -1; + ERTS_MSACC_SET_STATE_CACHED_M(ERTS_MSACC_STATE_SLEEP); res = erts_tse_twait(ssi->event, timeout); + ERTS_MSACC_POP_STATE_M(); current_time = ERTS_SCHEDULER_IS_DIRTY(esdp) ? 0 : erts_get_monotonic_time(esdp); } while (res == EINTR); @@ -3028,9 +3050,13 @@ scheduler_wait(int *fcalls, ErtsSchedulerData *esdp, ErtsRunQueue *rq) if (working) sched_wall_time_change(esdp, working = 0); + ERTS_MSACC_SET_STATE_CACHED_M(ERTS_MSACC_STATE_CHECK_IO); + ASSERT(!erts_port_task_have_outstanding_io_tasks()); erl_sys_schedule(1); /* Might give us something to do */ + ERTS_MSACC_POP_STATE_M(); + if (!ERTS_SCHEDULER_IS_DIRTY(esdp)) { current_time = erts_get_monotonic_time(esdp); if (current_time >= erts_next_timeout_time(esdp->next_tmo_ref)) @@ -3051,6 +3077,7 @@ scheduler_wait(int *fcalls, ErtsSchedulerData *esdp, ErtsRunQueue *rq) erts_thr_progress_active(esdp, thr_prgr_active = 1); #endif aux_work = handle_aux_work(&esdp->aux_work_data, aux_work, 1); + ERTS_MSACC_UPDATE_CACHE(); #ifdef ERTS_SMP if (aux_work && erts_thr_progress_update(esdp)) erts_thr_progress_leader_update(esdp); @@ -3148,8 +3175,12 @@ scheduler_wait(int *fcalls, ErtsSchedulerData *esdp, ErtsRunQueue *rq) ASSERT(!erts_port_task_have_outstanding_io_tasks()); + ERTS_MSACC_SET_STATE_CACHED_M(ERTS_MSACC_STATE_CHECK_IO); + erl_sys_schedule(0); + ERTS_MSACC_POP_STATE_M(); + if (!ERTS_SCHEDULER_IS_DIRTY(esdp)) { ErtsMonotonicTime current_time = erts_get_monotonic_time(esdp); if (current_time >= erts_next_timeout_time(esdp->next_tmo_ref)) @@ -7926,6 +7957,8 @@ sched_thread_func(void *vesdp) callbacks.wait = thr_prgr_wait; callbacks.finalize_wait = thr_prgr_fin_wait; + erts_msacc_init_thread("scheduler", no, 1); + erts_thr_progress_register_managed_thread(esdp, &callbacks, 0); erts_alloc_register_scheduler(vesdp); #endif @@ -9204,6 +9237,8 @@ Process *schedule(Process *p, int calls) Uint32 flags; erts_aint32_t state = 0; /* Supress warning... */ + ERTS_MSACC_DECLARE_CACHE(); + #ifdef USE_VM_PROBES if (p != NULL && DTRACE_ENABLED(process_unscheduled)) { DTRACE_CHARBUF(process_buf, DTRACE_TERM_BUF_SIZE); @@ -9308,6 +9343,8 @@ Process *schedule(Process *p, int calls) erts_smp_proc_unlock(p, ERTS_PROC_LOCK_MAIN|ERTS_PROC_LOCK_STATUS); + ERTS_MSACC_SET_STATE_CACHED_M(ERTS_MSACC_STATE_OTHER); + if (state & ERTS_PSFLG_FREE) { #ifdef ERTS_SMP ASSERT(esdp->free_process == p); @@ -9468,7 +9505,7 @@ Process *schedule(Process *p, int calls) scheduler_wait(&fcalls, esdp, rq); flags = ERTS_RUNQ_FLGS_SET_NOB(rq, ERTS_RUNQ_FLG_EXEC); flags |= ERTS_RUNQ_FLG_EXEC; - + ERTS_MSACC_UPDATE_CACHE(); #ifdef ERTS_SMP non_empty_runq(rq); #endif @@ -9483,6 +9520,8 @@ Process *schedule(Process *p, int calls) * Schedule system-level activities. */ + ERTS_MSACC_PUSH_STATE_CACHED_M(); + erts_smp_atomic32_set_relb(&function_calls, 0); fcalls = 0; @@ -9490,7 +9529,9 @@ Process *schedule(Process *p, int calls) erts_sys_schedule_interrupt(0); #endif erts_smp_runq_unlock(rq); + ERTS_MSACC_SET_STATE_CACHED_M(ERTS_MSACC_STATE_CHECK_IO); erl_sys_schedule(1); + ERTS_MSACC_POP_STATE_M(); current_time = erts_get_monotonic_time(esdp); if (current_time >= erts_next_timeout_time(esdp->next_tmo_ref)) @@ -9568,9 +9609,12 @@ Process *schedule(Process *p, int calls) case 0: /* No process at all */ default: ASSERT(qmask == 0); + ERTS_MSACC_SET_STATE_CACHED_M(ERTS_MSACC_STATE_OTHER); goto check_activities_to_run; } + ERTS_MSACC_SET_STATE_CACHED_M(ERTS_MSACC_STATE_EMULATOR); + BM_START_TIMER(system); /* -- cgit v1.2.3 From a7f48d4972be0c7984d5cb0e08e73260c0fdbe1b Mon Sep 17 00:00:00 2001 From: Rickard Green Date: Wed, 27 Jan 2016 14:45:32 +0100 Subject: Ensure that work is done on the correct type of schedulers Only the actual call to the dirty nif is allowed to execute on dirty schedulers. The dirty nif is not allowed to execute on normal schedulers if dirty schedulers are available. Arrival of exit signals and system tasks, while a process was scheduled for execution on a dirty scheduler, could mess up the process internal state. Preparation for dirty system task has been made, but is currently unused. --- erts/emulator/beam/erl_process.c | 650 +++++++++++++++++++++++++-------------- 1 file changed, 412 insertions(+), 238 deletions(-) (limited to 'erts/emulator/beam/erl_process.c') diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c index d583118e7b..1765315ed3 100644 --- a/erts/emulator/beam/erl_process.c +++ b/erts/emulator/beam/erl_process.c @@ -2969,7 +2969,7 @@ scheduler_wait(int *fcalls, ErtsSchedulerData *esdp, ErtsRunQueue *rq) sched_wall_time_change(esdp, thr_prgr_active); while (1) { - ErtsMonotonicTime current_time; + ErtsMonotonicTime current_time = 0; aux_work = erts_atomic32_read_acqb(&ssi->aux_work); if (aux_work) { @@ -3635,6 +3635,13 @@ check_requeue_process(ErtsRunQueue *rq, int prio_q) return 0; } +static ERTS_INLINE void +free_proxy_proc(Process *proxy) +{ + ASSERT(erts_smp_atomic32_read_nob(&proxy->state) & ERTS_PSFLG_PROXY); + erts_free(ERTS_ALC_T_PROC, proxy); +} + #ifdef ERTS_SMP static ErtsRunQueue * @@ -3949,9 +3956,6 @@ evacuate_run_queue(ErtsRunQueue *rq, erts_aint32_t state; Process *proc; int notify = 0; -#ifdef ERTS_DIRTY_SCHEDULERS - int requeue; -#endif to_rq = NULL; #ifdef ERTS_DIRTY_SCHEDULERS @@ -3966,49 +3970,94 @@ evacuate_run_queue(ErtsRunQueue *rq, proc = dequeue_process(rq, prio_q, &state); while (proc) { -#ifdef ERTS_DIRTY_SCHEDULERS - requeue = 1; + Process *real_proc; + int prio; + erts_aint32_t max_qbit, qbit, real_state; + + prio = ERTS_PSFLGS_GET_PRQ_PRIO(state); + qbit = ((erts_aint32_t) 1) << prio; + + if (!(state & ERTS_PSFLG_PROXY)) { + real_proc = proc; + real_state = state; + } + else { + real_proc = erts_proc_lookup_raw(proc->common.id); + if (!real_proc) { + free_proxy_proc(proc); + goto handle_next_proc; + } + real_state = erts_smp_atomic32_read_acqb(&real_proc->state); + } + + max_qbit = (state >> ERTS_PSFLGS_IN_PRQ_MASK_OFFSET); + max_qbit &= ERTS_PSFLGS_QMASK; + max_qbit |= 1 << ERTS_PSFLGS_QMASK_BITS; + max_qbit &= -max_qbit; + + if (qbit > max_qbit) { + /* Process already queued with higher prio; drop it... */ + if (real_proc != proc) + free_proxy_proc(proc); + else { + erts_aint32_t clr_bits; +#ifdef DEBUG + erts_aint32_t old; #endif - if (ERTS_PSFLG_BOUND & state) { - /* Bound processes get stuck here... */ - proc->next = NULL; - if (sbpp->last) - sbpp->last->next = proc; - else - sbpp->first = proc; - sbpp->last = proc; -#ifdef ERTS_DIRTY_SCHEDULERS - requeue = 0; + + clr_bits = ERTS_PSFLG_IN_RUNQ; + clr_bits |= qbit << ERTS_PSFLGS_IN_PRQ_MASK_OFFSET; + +#ifdef DEBUG + old = +#else + (void) #endif + erts_smp_atomic32_read_band_mb(&proc->state, + ~clr_bits); + ASSERT((old & clr_bits) == clr_bits); + + } + + goto handle_next_proc; } + #ifdef ERTS_DIRTY_SCHEDULERS - else if (state & ERTS_PSFLG_DIRTY_CPU_PROC_IN_Q) { + + if (ERTS_RUNQ_IX_IS_DIRTY(rq->ix)) { + erts_aint32_t dqbit = qbit; #ifdef DEBUG - erts_aint32_t old = -#else - (void) + erts_aint32_t old_dqbit; #endif - erts_smp_atomic32_read_band_nob(&proc->state, - ~(ERTS_PSFLG_DIRTY_CPU_PROC - | ERTS_PSFLG_DIRTY_CPU_PROC_IN_Q)); - /* assert that no other dirty flags are set */ - ASSERT(!(old & (ERTS_PSFLG_DIRTY_IO_PROC|ERTS_PSFLG_DIRTY_IO_PROC_IN_Q))); - } else if (state & ERTS_PSFLG_DIRTY_IO_PROC_IN_Q) { + + if (rq == ERTS_DIRTY_CPU_RUNQ) + dqbit <<= ERTS_PDSFLGS_IN_CPU_PRQ_MASK_OFFSET; + else { + ASSERT(rq == ERTS_DIRTY_IO_RUNQ); + dqbit <<= ERTS_PDSFLGS_IN_IO_PRQ_MASK_OFFSET; + } + #ifdef DEBUG - erts_aint32_t old = + old_dqbit = (int) #else - (void) + (void) #endif - erts_smp_atomic32_read_band_nob(&proc->state, - ~(ERTS_PSFLG_DIRTY_IO_PROC - | ERTS_PSFLG_DIRTY_IO_PROC_IN_Q)); - /* assert that no other dirty flags are set */ - ASSERT(!(old & (ERTS_PSFLG_DIRTY_CPU_PROC|ERTS_PSFLG_DIRTY_CPU_PROC_IN_Q))); + erts_smp_atomic32_read_band_mb(&real_proc->dirty_state, + ~dqbit); + ASSERT(old_dqbit & dqbit); } - if (requeue) { -#else - else { #endif + + if (ERTS_PSFLG_BOUND & real_state) { + /* Bound processes get stuck here... */ + proc->next = NULL; + if (sbpp->last) + sbpp->last->next = proc; + else + sbpp->first = proc; + sbpp->last = proc; + } + else { int prio = (int) ERTS_PSFLGS_GET_PRQ_PRIO(state); erts_smp_runq_unlock(rq); @@ -4031,6 +4080,8 @@ evacuate_run_queue(ErtsRunQueue *rq, erts_smp_runq_lock(rq); } + + handle_next_proc: proc = dequeue_process(rq, prio_q, &state); } if (notify) @@ -5974,19 +6025,96 @@ make_proxy_proc(Process *prev_proxy, Process *proc, erts_aint32_t prio) return proxy; } -static ERTS_INLINE void -free_proxy_proc(Process *proxy) -{ - ASSERT(erts_smp_atomic32_read_nob(&proxy->state) & ERTS_PSFLG_PROXY); - erts_free(ERTS_ALC_T_PROC, proxy); -} - #define ERTS_ENQUEUE_NOT 0 #define ERTS_ENQUEUE_NORMAL_QUEUE 1 -#ifdef ERTS_DIRTY_SCHEDULERS #define ERTS_ENQUEUE_DIRTY_CPU_QUEUE 2 #define ERTS_ENQUEUE_DIRTY_IO_QUEUE 3 + +#ifdef ERTS_DIRTY_SCHEDULERS + +static int +check_dirty_enqueue_in_prio_queue(Process *c_p, + erts_aint32_t *newp, + erts_aint32_t actual, + erts_aint32_t aprio, + erts_aint32_t qbit) +{ + int queue; + erts_aint32_t dact, max_qbit; + + /* Termination should be done on an ordinary scheduler */ + if (actual & ERTS_PSFLG_EXITING) { + *newp &= ~ERTS_PSFLGS_DIRTY_WORK; + return ERTS_ENQUEUE_NORMAL_QUEUE; + } + + /* + * If we have system tasks, we enqueue on ordinary run-queue + * and take care of those system tasks first. + */ + if (actual & ERTS_PSFLG_ACTIVE_SYS) + return ERTS_ENQUEUE_NORMAL_QUEUE; + + dact = erts_smp_atomic32_read_mb(&c_p->dirty_state); + if (actual & (ERTS_PSFLG_DIRTY_ACTIVE_SYS + | ERTS_PSFLG_DIRTY_CPU_PROC)) { + max_qbit = ((dact >> ERTS_PDSFLGS_IN_CPU_PRQ_MASK_OFFSET) + & ERTS_PDSFLGS_QMASK); + queue = ERTS_ENQUEUE_DIRTY_CPU_QUEUE; + } + else { + ASSERT(actual & ERTS_PSFLG_DIRTY_IO_PROC); + max_qbit = ((dact >> ERTS_PDSFLGS_IN_IO_PRQ_MASK_OFFSET) + & ERTS_PDSFLGS_QMASK); + queue = ERTS_ENQUEUE_DIRTY_IO_QUEUE; + } + + max_qbit |= 1 << ERTS_PSFLGS_QMASK_BITS; + max_qbit &= -max_qbit; + + if (qbit >= max_qbit) + return ERTS_ENQUEUE_NOT; /* Already queued in higher or equal prio */ + if ((actual & (ERTS_PSFLG_IN_RUNQ|ERTS_PSFLGS_USR_PRIO_MASK)) + != (aprio << ERTS_PSFLGS_USR_PRIO_OFFSET)) { + /* + * Process struct already enqueued, or actual prio not + * equal to user prio, i.e., enqueue using proxy. + */ + return -1*queue; + } + + *newp |= ERTS_PSFLG_IN_RUNQ; + return queue; +} + +static ERTS_INLINE int +fin_dirty_enq_s_change(Process *p, + int pstruct_reserved, + erts_aint32_t enq_prio, + int qmask_offset) +{ + erts_aint32_t qbit = 1 << enq_prio; + qbit <<= qmask_offset; + + if (qbit & erts_smp_atomic32_read_bor_mb(&p->dirty_state, qbit)) { + /* Already enqueue by someone else... */ + if (pstruct_reserved) { + /* We reserved process struct for enqueue; clear it... */ +#ifdef DEBUG + erts_aint32_t old = +#else + (void) #endif + erts_smp_atomic32_read_band_nob(&p->state, ~ERTS_PSFLG_IN_RUNQ); + ASSERT(old & ERTS_PSFLG_IN_RUNQ); + } + return 0; + } + + return !0; +} + +#endif /* ERTS_DIRTY_SCHEDULERS */ static ERTS_INLINE int check_enqueue_in_prio_queue(Process *c_p, @@ -6002,61 +6130,14 @@ check_enqueue_in_prio_queue(Process *c_p, *prq_prio_p = aprio; #ifdef ERTS_DIRTY_SCHEDULERS - if (actual & (ERTS_PSFLG_DIRTY_CPU_PROC|ERTS_PSFLG_DIRTY_IO_PROC)) { - /* - * If we have system tasks of a priority higher - * or equal to the user priority, we enqueue - * on ordinary run-queue and take care of - * those system tasks first. - */ - if (actual & ERTS_PSFLG_ACTIVE_SYS) { - erts_aint32_t uprio, stprio, qmask; - uprio = (actual >> ERTS_PSFLGS_USR_PRIO_OFFSET) & ERTS_PSFLGS_PRIO_MASK; - if (aprio < uprio) - goto enqueue_normal_runq; /* system tasks with higher prio */ - erts_smp_proc_lock(c_p, ERTS_PROC_LOCK_STATUS); - qmask = c_p->sys_task_qs->qmask; - erts_smp_proc_unlock(c_p, ERTS_PROC_LOCK_STATUS); - switch (qmask & -qmask) { - case MAX_BIT: - stprio = PRIORITY_MAX; - break; - case HIGH_BIT: - stprio = PRIORITY_HIGH; - break; - case NORMAL_BIT: - stprio = PRIORITY_NORMAL; - break; - case LOW_BIT: - stprio = PRIORITY_LOW; - break; - default: - stprio = PRIORITY_LOW+1; - break; - } - if (stprio <= uprio) - goto enqueue_normal_runq; /* system tasks with higher prio */ - } - - /* Enqueue in dirty run queue if not already enqueued */ - if (actual & (ERTS_PSFLG_DIRTY_CPU_PROC_IN_Q|ERTS_PSFLG_DIRTY_IO_PROC_IN_Q)) - return ERTS_ENQUEUE_NOT; /* already in queue */ - if (actual & ERTS_PSFLG_DIRTY_CPU_PROC) { - *newp |= ERTS_PSFLG_DIRTY_CPU_PROC_IN_Q; - if (actual & ERTS_PSFLG_IN_RUNQ) - return -ERTS_ENQUEUE_DIRTY_CPU_QUEUE; /* use proxy */ - *newp |= ERTS_PSFLG_IN_RUNQ; - return ERTS_ENQUEUE_DIRTY_CPU_QUEUE; - } - *newp |= ERTS_PSFLG_DIRTY_IO_PROC_IN_Q; - if (actual & ERTS_PSFLG_IN_RUNQ) - return -ERTS_ENQUEUE_DIRTY_IO_QUEUE; /* use proxy */ - *newp |= ERTS_PSFLG_IN_RUNQ; - return ERTS_ENQUEUE_DIRTY_IO_QUEUE; + if (actual & ERTS_PSFLGS_DIRTY_WORK) { + int res = check_dirty_enqueue_in_prio_queue(c_p, newp, actual, + aprio, qbit); + if (res != ERTS_ENQUEUE_NORMAL_QUEUE) + return res; } - - enqueue_normal_runq: #endif + max_qbit = (actual >> ERTS_PSFLGS_IN_PRQ_MASK_OFFSET) & ERTS_PSFLGS_QMASK; max_qbit |= 1 << ERTS_PSFLGS_QMASK_BITS; max_qbit &= -max_qbit; @@ -6088,6 +6169,65 @@ check_enqueue_in_prio_queue(Process *c_p, return ERTS_ENQUEUE_NORMAL_QUEUE; } +static ERTS_INLINE ErtsRunQueue * +select_enqueue_run_queue(int enqueue, int enq_prio, Process *p, erts_aint32_t state) +{ + + switch (enqueue) { + + case ERTS_ENQUEUE_NOT: + + return NULL; + +#ifdef ERTS_DIRTY_SCHEDULERS + + case ERTS_ENQUEUE_DIRTY_CPU_QUEUE: + case -ERTS_ENQUEUE_DIRTY_CPU_QUEUE: + + if (fin_dirty_enq_s_change(p, enqueue > 0, enq_prio, + ERTS_PDSFLGS_IN_CPU_PRQ_MASK_OFFSET)) + return ERTS_DIRTY_CPU_RUNQ; + + return NULL; + + + case ERTS_ENQUEUE_DIRTY_IO_QUEUE: + case -ERTS_ENQUEUE_DIRTY_IO_QUEUE: + + if (fin_dirty_enq_s_change(p, enqueue > 0, enq_prio, + ERTS_PDSFLGS_IN_IO_PRQ_MASK_OFFSET)) + return ERTS_DIRTY_IO_RUNQ; + + return NULL; + +#endif + + default: { + ErtsRunQueue* runq; + + ASSERT(enqueue == ERTS_ENQUEUE_NORMAL_QUEUE + || enqueue == -ERTS_ENQUEUE_NORMAL_QUEUE); + + runq = erts_get_runq_proc(p); + +#ifdef ERTS_SMP + if (!(ERTS_PSFLG_BOUND & state)) { + ErtsRunQueue *new_runq = erts_check_emigration_need(runq, enq_prio); + if (new_runq) { + RUNQ_SET_RQ(&p->run_queue, new_runq); + runq = new_runq; + } + } +#endif + + ASSERT(runq); + + return runq; + } + } +} + + /* * schedule_out_process() return with c_rq locked. */ @@ -6096,11 +6236,7 @@ schedule_out_process(ErtsRunQueue *c_rq, erts_aint32_t state, Process *p, Proces { erts_aint32_t a, e, n, enq_prio = -1; int enqueue; /* < 0 -> use proxy */ - Process* sched_p; ErtsRunQueue* runq; -#ifdef ERTS_SMP - int check_emigration_need; -#endif a = state; @@ -6112,7 +6248,7 @@ schedule_out_process(ErtsRunQueue *c_rq, erts_aint32_t state, Process *p, Proces enqueue = ERTS_ENQUEUE_NOT; n &= ~(ERTS_PSFLG_RUNNING|ERTS_PSFLG_RUNNING_SYS); - if (a & ERTS_PSFLG_ACTIVE_SYS + if (a & (ERTS_PSFLG_ACTIVE_SYS|ERTS_PSFLG_DIRTY_ACTIVE_SYS) || (a & (ERTS_PSFLG_ACTIVE|ERTS_PSFLG_SUSPENDED)) == ERTS_PSFLG_ACTIVE) { enqueue = check_enqueue_in_prio_queue(p, &enq_prio, &n, a); } @@ -6121,16 +6257,17 @@ schedule_out_process(ErtsRunQueue *c_rq, erts_aint32_t state, Process *p, Proces break; } - switch (enqueue) { - case ERTS_ENQUEUE_NOT: + runq = select_enqueue_run_queue(enqueue, enq_prio, p, n); + + if (!runq) { + if (erts_system_profile_flags.runnable_procs) { /* Status lock prevents out of order "runnable proc" trace msgs */ ERTS_SMP_LC_ASSERT(ERTS_PROC_LOCK_STATUS & erts_proc_lc_my_proc_locks(p)); - if (!(a & ERTS_PSFLG_ACTIVE_SYS) - && (!(a & ERTS_PSFLG_ACTIVE) - || (a & ERTS_PSFLG_SUSPENDED))) { + if (!(a & (ERTS_PSFLG_ACTIVE_SYS|ERTS_PSFLG_DIRTY_ACTIVE_SYS)) + && (!(a & ERTS_PSFLG_ACTIVE) || (a & ERTS_PSFLG_SUSPENDED))) { /* Process inactive */ profile_runnable_proc(p, am_inactive); } @@ -6143,98 +6280,76 @@ schedule_out_process(ErtsRunQueue *c_rq, erts_aint32_t state, Process *p, Proces return 0; -#ifdef ERTS_DIRTY_SCHEDULERS -#ifdef ERTS_SMP - case ERTS_ENQUEUE_DIRTY_CPU_QUEUE: - case -ERTS_ENQUEUE_DIRTY_CPU_QUEUE: - runq = ERTS_DIRTY_CPU_RUNQ; - ASSERT(ERTS_SCHEDULER_IS_DIRTY_CPU(runq->scheduler)); -#ifdef ERTS_SMP - check_emigration_need = 0; -#endif - break; + } + else { + Process* sched_p; - case ERTS_ENQUEUE_DIRTY_IO_QUEUE: - case -ERTS_ENQUEUE_DIRTY_IO_QUEUE: - runq = ERTS_DIRTY_IO_RUNQ; - ASSERT(ERTS_SCHEDULER_IS_DIRTY_IO(runq->scheduler)); -#ifdef ERTS_SMP - check_emigration_need = 0; -#endif - break; -#endif -#endif + ASSERT(!(n & ERTS_PSFLG_SUSPENDED) || (n & (ERTS_PSFLG_ACTIVE_SYS + | ERTS_PSFLG_DIRTY_ACTIVE_SYS))); - default: - ASSERT(enqueue == ERTS_ENQUEUE_NORMAL_QUEUE - || enqueue == -ERTS_ENQUEUE_NORMAL_QUEUE); + if (enqueue < 0) + sched_p = make_proxy_proc(proxy, p, enq_prio); + else { + sched_p = p; + if (proxy) + free_proxy_proc(proxy); + } - runq = erts_get_runq_proc(p); -#ifdef ERTS_SMP - check_emigration_need = !(ERTS_PSFLG_BOUND & n); -#endif - break; - } + ASSERT(runq); - ASSERT(!(n & ERTS_PSFLG_SUSPENDED) || (n & ERTS_PSFLG_ACTIVE_SYS)); + erts_smp_runq_lock(runq); - if (enqueue < 0) - sched_p = make_proxy_proc(proxy, p, enq_prio); - else { - sched_p = p; - if (proxy) - free_proxy_proc(proxy); - } + /* Enqueue the process */ + enqueue_process(runq, (int) enq_prio, sched_p); -#ifdef ERTS_SMP - if (check_emigration_need) { - ErtsRunQueue *new_runq = erts_check_emigration_need(runq, enq_prio); - if (new_runq) { - RUNQ_SET_RQ(&sched_p->run_queue, new_runq); - runq = new_runq; - } - } -#endif + if (runq == c_rq) + return 1; - ASSERT(runq); + erts_smp_runq_unlock(runq); - erts_smp_runq_lock(runq); + smp_notify_inc_runq(runq); - /* Enqueue the process */ - enqueue_process(runq, (int) enq_prio, sched_p); + erts_smp_runq_lock(c_rq); - if (runq == c_rq) return 1; - erts_smp_runq_unlock(runq); - smp_notify_inc_runq(runq); - erts_smp_runq_lock(c_rq); - return 1; + } + } static ERTS_INLINE void -add2runq(Process *p, erts_aint32_t state, erts_aint32_t prio) +add2runq(int enqueue, erts_aint32_t prio, + Process *proc, erts_aint32_t state, + Process **proxy) { - ErtsRunQueue *runq = erts_get_runq_proc(p); + ErtsRunQueue *runq; -#ifdef ERTS_SMP - if (!(ERTS_PSFLG_BOUND & state)) { - ErtsRunQueue *new_runq = erts_check_emigration_need(runq, (int) prio); - if (new_runq) { - RUNQ_SET_RQ(&p->run_queue, new_runq); - runq = new_runq; - } - } -#endif - ASSERT(runq); + runq = select_enqueue_run_queue(enqueue, prio, proc, state); - erts_smp_runq_lock(runq); + if (runq) { + Process *sched_p; - /* Enqueue the process */ - enqueue_process(runq, (int) prio, p); + if (enqueue > 0) + sched_p = proc; + else { + Process *pxy; - erts_smp_runq_unlock(runq); - smp_notify_inc_runq(runq); + if (!proxy) + pxy = NULL; + else { + pxy = *proxy; + *proxy = NULL; + } + sched_p = make_proxy_proc(pxy, proc, prio); + } + + erts_smp_runq_lock(runq); + + /* Enqueue the process */ + enqueue_process(runq, (int) prio, sched_p); + erts_smp_runq_unlock(runq); + smp_notify_inc_runq(runq); + } } static ERTS_INLINE int @@ -6340,10 +6455,7 @@ schedule_process(Process *p, erts_aint32_t in_state, ErtsProcLocks locks) &state, &enq_prio, locks); - if (enqueue != ERTS_ENQUEUE_NOT) - add2runq(enqueue > 0 ? p : make_proxy_proc(NULL, p, enq_prio), - state, - enq_prio); + add2runq(enqueue, enq_prio, p, state, NULL); } void @@ -6403,16 +6515,7 @@ schedule_process_sys_task(Process *p, erts_aint32_t state, Process *proxy) prof_runnable_procs = 0; } - if (enqueue != ERTS_ENQUEUE_NOT) { - Process *sched_p; - if (enqueue > 0) - sched_p = p; - else { - sched_p = make_proxy_proc(proxy, p, enq_prio); - proxy = NULL; - } - add2runq(sched_p, n, enq_prio); - } + add2runq(enqueue, enq_prio, p, n, &proxy); cleanup: @@ -6508,10 +6611,7 @@ resume_process(Process *p, ErtsProcLocks locks) &state, &enq_prio, locks); - if (enqueue) - add2runq(enqueue > 0 ? p : make_proxy_proc(NULL, p, enq_prio), - state, - enq_prio); + add2runq(enqueue, enq_prio, p, state, NULL); } int @@ -9361,15 +9461,16 @@ Process *schedule(Process *p, int calls) #ifdef ERTS_SMP ErtsMigrationPaths *mps; ErtsMigrationPath *mp; - ErtsProcList *pnd_xtrs = rq->procs.pending_exiters; - if (erts_proclist_fetch(&pnd_xtrs, NULL)) { + + if (!ERTS_SCHEDULER_IS_DIRTY(esdp)) { + ErtsProcList *pnd_xtrs = rq->procs.pending_exiters; + if (erts_proclist_fetch(&pnd_xtrs, NULL)) { rq->procs.pending_exiters = NULL; erts_smp_runq_unlock(rq); handle_pending_exiters(pnd_xtrs); erts_smp_runq_lock(rq); } - if (!ERTS_SCHEDULER_IS_DIRTY(esdp)) { if (rq->check_balance_reds <= 0) check_balance(rq); @@ -9564,11 +9665,12 @@ Process *schedule(Process *p, int calls) pick_next_process: { erts_aint32_t psflg_band_mask; int prio_q; - int qmask; + int qmask, qbit; flags = ERTS_RUNQ_FLGS_GET_NOB(rq); qmask = (int) (flags & ERTS_RUNQ_FLGS_PROCS_QMASK); - switch (qmask & -qmask) { + qbit = qmask & -qmask; + switch (qbit) { case MAX_BIT: prio_q = PRIORITY_MAX; break; @@ -9596,20 +9698,11 @@ Process *schedule(Process *p, int calls) ASSERT(p); /* Wrong qmask in rq->flags? */ - psflg_band_mask = ~(((erts_aint32_t) 1) << (ERTS_PSFLGS_GET_PRQ_PRIO(state) - + ERTS_PSFLGS_IN_PRQ_MASK_OFFSET)); - -#ifdef ERTS_DIRTY_SCHEDULERS - ASSERT((state & (ERTS_PSFLG_DIRTY_CPU_PROC|ERTS_PSFLG_DIRTY_IO_PROC)) != - (ERTS_PSFLG_DIRTY_CPU_PROC|ERTS_PSFLG_DIRTY_IO_PROC)); - if (state & (ERTS_PSFLG_DIRTY_CPU_PROC|ERTS_PSFLG_DIRTY_IO_PROC)) { - ASSERT((ERTS_SCHEDULER_IS_DIRTY_CPU(esdp) && (state & ERTS_PSFLG_DIRTY_CPU_PROC)) || - (ERTS_SCHEDULER_IS_DIRTY_IO(esdp) && (state & ERTS_PSFLG_DIRTY_IO_PROC))); - if (!ERTS_SCHEDULER_IS_DIRTY(esdp) && !(state & ERTS_PSFLG_ACTIVE_SYS)) - goto pick_next_process; - state &= ~(ERTS_PSFLG_DIRTY_CPU_PROC_IN_Q|ERTS_PSFLG_DIRTY_IO_PROC_IN_Q); - } -#endif + if (ERTS_SCHEDULER_IS_DIRTY(esdp)) + psflg_band_mask = ~((erts_aint32_t) 0); + else + psflg_band_mask = ~(((erts_aint32_t) 1) << (ERTS_PSFLGS_GET_PRQ_PRIO(state) + + ERTS_PSFLGS_IN_PRQ_MASK_OFFSET)); if (!(state & ERTS_PSFLG_PROXY)) psflg_band_mask &= ~ERTS_PSFLG_IN_RUNQ; @@ -9632,9 +9725,11 @@ Process *schedule(Process *p, int calls) | ERTS_PSFLG_RUNNING_SYS))) { tmp = state & (ERTS_PSFLG_SUSPENDED | ERTS_PSFLG_PENDING_EXIT - | ERTS_PSFLG_ACTIVE_SYS); + | ERTS_PSFLG_ACTIVE_SYS + | ERTS_PSFLG_DIRTY_ACTIVE_SYS); if (tmp != ERTS_PSFLG_SUSPENDED) { - if (state & ERTS_PSFLG_ACTIVE_SYS) + if (state & (ERTS_PSFLG_ACTIVE_SYS + | ERTS_PSFLG_DIRTY_ACTIVE_SYS)) new |= ERTS_PSFLG_RUNNING_SYS; else new |= ERTS_PSFLG_RUNNING; @@ -9647,7 +9742,8 @@ Process *schedule(Process *p, int calls) | ERTS_PSFLG_FREE)) || ((state & (ERTS_PSFLG_SUSPENDED | ERTS_PSFLG_PENDING_EXIT - | ERTS_PSFLG_ACTIVE_SYS)) + | ERTS_PSFLG_ACTIVE_SYS + | ERTS_PSFLG_DIRTY_ACTIVE_SYS)) == ERTS_PSFLG_SUSPENDED)) { if (state & ERTS_PSFLG_FREE) erts_proc_dec_refc(p); @@ -9666,10 +9762,42 @@ Process *schedule(Process *p, int calls) esdp->current_process = p; + + reds = context_reds; + +#ifdef ERTS_SMP + + erts_smp_runq_unlock(rq); + +#ifdef ERTS_DIRTY_SCHEDULERS + if (ERTS_SCHEDULER_IS_DIRTY(esdp)) { +#ifdef DEBUG + int old_dqbit; +#endif + int dqbit = qbit; + + if (rq == ERTS_DIRTY_CPU_RUNQ) + dqbit <<= ERTS_PDSFLGS_IN_CPU_PRQ_MASK_OFFSET; + else { + ASSERT(rq == ERTS_DIRTY_IO_RUNQ); + dqbit <<= ERTS_PDSFLGS_IN_IO_PRQ_MASK_OFFSET; + } + +#ifdef DEBUG + old_dqbit = (int) +#else + (void) +#endif + erts_smp_atomic32_read_band_mb(&p->dirty_state, ~dqbit); + ASSERT(old_dqbit & dqbit); + } +#endif /* ERTS_DIRTY_SCHEDULERS */ + +#endif /* ERTS_SMP */ + } #ifdef ERTS_SMP - erts_smp_runq_unlock(rq); if (flags & ERTS_RUNQ_FLG_PROTECTED) (void) ERTS_RUNQ_FLGS_UNSET(rq, ERTS_RUNQ_FLG_PROTECTED); @@ -9702,15 +9830,56 @@ Process *schedule(Process *p, int calls) erts_smp_spin_unlock(&erts_sched_stat.lock); } - if (ERTS_PROC_PENDING_EXIT(p)) { + ASSERT(!p->scheduler_data); + p->scheduler_data = esdp; + + state = erts_smp_atomic32_read_nob(&p->state); + +#ifdef ERTS_DIRTY_SCHEDULERS + if (!ERTS_SCHEDULER_IS_DIRTY(esdp)) { + if (!!(state & ERTS_PSFLGS_DIRTY_WORK) + & !(state & ERTS_PSFLG_ACTIVE_SYS)) { + /* Migrate to dirty scheduler... */ + sunlock_sched_out_proc: + erts_smp_proc_unlock(p, ERTS_PROC_LOCK_STATUS); + p->fcalls = reds; + goto sched_out_proc; + + } + } + else { + if (state & (ERTS_PSFLG_ACTIVE_SYS + | ERTS_PSFLG_PENDING_EXIT + | ERTS_PSFLG_EXITING)) { + /* Migrate to normal scheduler... */ + goto sunlock_sched_out_proc; + } + if ((state & ERTS_PSFLG_DIRTY_ACTIVE_SYS) + && rq == ERTS_DIRTY_IO_RUNQ) { + /* Migrate to dirty cpu scheduler... */ + goto sunlock_sched_out_proc; + } + + ASSERT((state & ERTS_PSFLG_DIRTY_ACTIVE_SYS) + || *p->i == (BeamInstr) em_call_nif); + + ASSERT(rq == ERTS_DIRTY_CPU_RUNQ + ? (state & (ERTS_PSFLG_DIRTY_CPU_PROC + | ERTS_PSFLG_DIRTY_ACTIVE_SYS)) + : (rq == ERTS_DIRTY_IO_RUNQ + && (state & ERTS_PSFLG_DIRTY_IO_PROC))); + } +#endif + + if (state & ERTS_PSFLG_PENDING_EXIT) { erts_handle_pending_exit(p, ERTS_PROC_LOCK_MAIN|ERTS_PROC_LOCK_STATUS); state = erts_smp_atomic32_read_nob(&p->state); } - ASSERT(!p->scheduler_data); - p->scheduler_data = esdp; -#endif - reds = context_reds; + +#endif /* ERTS_SMP */ + + p->fcalls = reds; if (IS_TRACED(p)) { if (state & ERTS_PSFLG_EXITING) { @@ -9739,7 +9908,8 @@ Process *schedule(Process *p, int calls) reds -= execute_sys_tasks(p, &state, reds); if (reds <= 0 #ifdef ERTS_DIRTY_SCHEDULERS - || (state & (ERTS_PSFLG_DIRTY_CPU_PROC|ERTS_PSFLG_DIRTY_IO_PROC)) + || ERTS_SCHEDULER_IS_DIRTY(esdp) + || (state & ERTS_PSFLGS_DIRTY_WORK) #endif ) { p->fcalls = reds; @@ -9787,7 +9957,6 @@ Process *schedule(Process *p, int calls) proxy_p = NULL; } - p->fcalls = reds; ERTS_SMP_CHK_HAVE_ONLY_MAIN_PROC_LOCK(p); /* Never run a suspended process */ @@ -10732,6 +10901,9 @@ static void early_init_process_struct(void *varg, Eterm data) Process *proc = arg->proc; proc->common.id = make_internal_pid(data); +#ifdef ERTS_DIRTY_SCHEDULERS + erts_smp_atomic32_init_nob(&proc->dirty_state, 0); +#endif erts_smp_atomic32_init_relb(&proc->state, arg->state); #ifdef ERTS_SMP @@ -11194,6 +11366,9 @@ void erts_init_empty_process(Process *p) p->last_old_htop = NULL; #endif +#ifdef ERTS_DIRTY_SCHEDULERS + erts_smp_atomic32_init_nob(&p->dirty_state, 0); +#endif erts_smp_atomic32_init_nob(&p->state, (erts_aint32_t) PRIORITY_NORMAL); #ifdef ERTS_SMP @@ -11395,7 +11570,9 @@ set_proc_exiting(Process *p, ERTS_SMP_LC_ASSERT(erts_proc_lc_my_proc_locks(p) == ERTS_PROC_LOCKS_ALL); enqueue = change_proc_schedule_state(p, - ERTS_PSFLG_SUSPENDED|ERTS_PSFLG_PENDING_EXIT, + (ERTS_PSFLG_SUSPENDED + | ERTS_PSFLG_PENDING_EXIT + | ERTS_PSFLGS_DIRTY_WORK), ERTS_PSFLG_EXITING|ERTS_PSFLG_ACTIVE, &state, &enq_prio, @@ -11429,10 +11606,7 @@ set_proc_exiting(Process *p, } #endif - if (enqueue) - add2runq(enqueue > 0 ? p : make_proxy_proc(NULL, p, enq_prio), - state, - enq_prio); + add2runq(enqueue, enq_prio, p, state, NULL); } static ERTS_INLINE erts_aint32_t @@ -11535,6 +11709,11 @@ save_pending_exiter(Process *p) else rq = esdp->run_queue; +#ifdef ERTS_DIRTY_SCHEDULERS + if (ERTS_RUNQ_IX_IS_DIRTY(rq->ix)) + rq = ERTS_RUNQ_IX(0); /* Handle on ordinary scheduler */ +#endif + plp = proclist_create(p); erts_smp_runq_lock(rq); @@ -11544,13 +11723,8 @@ save_pending_exiter(Process *p) non_empty_runq(rq); erts_smp_runq_unlock(rq); -#ifdef ERTS_DIRTY_SCHEDULERS - if (ERTS_RUNQ_IX_IS_DIRTY(rq->ix)) - wake_dirty_schedulers(rq, 0); - else -#endif - wake_scheduler(rq); + wake_scheduler(rq); } #endif -- cgit v1.2.3 From a5e66520d92d283d14f1352dcfd0ce889c27a0c5 Mon Sep 17 00:00:00 2001 From: Rickard Green Date: Tue, 23 Feb 2016 18:31:52 +0100 Subject: Do not wait for main lock when looking up process not running --- erts/emulator/beam/erl_process.c | 102 +++++++++++++++++++++++++++------------ 1 file changed, 71 insertions(+), 31 deletions(-) (limited to 'erts/emulator/beam/erl_process.c') diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c index b24f26138c..adea745471 100644 --- a/erts/emulator/beam/erl_process.c +++ b/erts/emulator/beam/erl_process.c @@ -6462,9 +6462,6 @@ suspend_process(Process *c_p, Process *p) if (suspended) { - ASSERT(!(ERTS_PSFLG_RUNNING & state) - || p == erts_get_current_process()); - if (suspended > 0 && erts_system_profile_flags.runnable_procs) { /* 'state' is before our change... */ @@ -8411,9 +8408,8 @@ pid2proc_not_running(Process *c_p, ErtsProcLocks c_p_locks, resume_process(rp, rp_locks); } else { - rp = erts_pid2proc(c_p, c_p_locks|ERTS_PROC_LOCK_STATUS, - pid, pid_locks|ERTS_PROC_LOCK_STATUS); + pid, ERTS_PROC_LOCK_STATUS); if (!rp) { c_p->flags &= ~F_P2PNR_RESCHED; @@ -8422,40 +8418,84 @@ pid2proc_not_running(Process *c_p, ErtsProcLocks c_p_locks, ASSERT(!(c_p->flags & F_P2PNR_RESCHED)); - if (suspend) { - if (suspend_process(c_p, rp)) - goto done; - } - else { - if (!((ERTS_PSFLG_RUNNING|ERTS_PSFLG_RUNNING_SYS) - & erts_smp_atomic32_read_acqb(&rp->state))) - goto done; + /* + * Suspend the other process in order to prevent + * it from being selected for normal execution. + * This will however not prevent it from being + * selected for execution of a system task. If + * it is selected for execution of a system task + * we might be blocked for quite a while if the + * try-lock below fails. That is, there is room + * for improvement here... + */ - } + if (!suspend_process(c_p, rp)) { + /* Other process running */ - /* Other process running */ + ASSERT(ERTS_PSFLG_RUNNING + & erts_smp_atomic32_read_nob(&rp->state)); - /* - * If we got pending suspenders and suspend ourselves waiting - * to suspend another process we might deadlock. - * In this case we have to yield, be suspended by - * someone else and then do it all over again. - */ - if (!c_p->pending_suspenders) { - /* Mark rp pending for suspend by c_p */ - add_pend_suspend(rp, c_p->common.id, handle_pend_sync_suspend); - ASSERT(is_nil(c_p->suspendee)); + running: - /* Suspend c_p; when rp is suspended c_p will be resumed. */ - suspend_process(c_p, c_p); - c_p->flags |= F_P2PNR_RESCHED; + /* + * If we got pending suspenders and suspend ourselves waiting + * to suspend another process we might deadlock. + * In this case we have to yield, be suspended by + * someone else and then do it all over again. + */ + if (!c_p->pending_suspenders) { + /* Mark rp pending for suspend by c_p */ + add_pend_suspend(rp, c_p->common.id, handle_pend_sync_suspend); + ASSERT(is_nil(c_p->suspendee)); + + /* Suspend c_p; when rp is suspended c_p will be resumed. */ + suspend_process(c_p, c_p); + c_p->flags |= F_P2PNR_RESCHED; + } + /* Yield (caller is assumed to yield immediately in bif). */ + erts_smp_proc_unlock(rp, ERTS_PROC_LOCK_STATUS); + rp = ERTS_PROC_LOCK_BUSY; + } + else { + ErtsProcLocks need_locks = pid_locks & ~ERTS_PROC_LOCK_STATUS; + if (need_locks && erts_smp_proc_trylock(rp, need_locks) == EBUSY) { + if (ERTS_PSFLG_RUNNING_SYS + & erts_smp_atomic32_read_nob(&rp->state)) { + /* Executing system task... */ + resume_process(rp, ERTS_PROC_LOCK_STATUS); + goto running; + } + erts_smp_proc_unlock(rp, ERTS_PROC_LOCK_STATUS); + /* + * If we are unlucky, the process just got selected for + * execution of a system task. In this case we may be + * blocked here for quite a while... Execution of system + * tasks are fortunately quite rare events. We try to + * avoid this by checking if it is in a state executing + * system tasks (above), but it will not prevent all + * scenarios for a long block here... + */ + rp = erts_pid2proc(c_p, c_p_locks|ERTS_PROC_LOCK_STATUS, + pid, pid_locks|ERTS_PROC_LOCK_STATUS); + if (!rp) + goto done; + } + + /* + * The previous suspend has prevented the process + * from being selected for normal execution regardless + * of locks held or not held on it... + */ + ASSERT(!(ERTS_PSFLG_RUNNING + & erts_smp_atomic32_read_nob(&rp->state))); + + if (!suspend) + resume_process(rp, pid_locks|ERTS_PROC_LOCK_STATUS); } - /* Yield (caller is assumed to yield immediately in bif). */ - erts_smp_proc_unlock(rp, pid_locks|ERTS_PROC_LOCK_STATUS); - rp = ERTS_PROC_LOCK_BUSY; } done: + if (rp && rp != ERTS_PROC_LOCK_BUSY && !(pid_locks & ERTS_PROC_LOCK_STATUS)) erts_smp_proc_unlock(rp, ERTS_PROC_LOCK_STATUS); if (unlock_c_p_status) -- cgit v1.2.3