diff options
Diffstat (limited to 'erts/emulator')
| -rw-r--r-- | erts/emulator/beam/erl_async.c | 4 | ||||
| -rw-r--r-- | erts/emulator/beam/erl_db_hash.c | 63 | ||||
| -rw-r--r-- | erts/emulator/beam/erl_init.c | 32 | ||||
| -rw-r--r-- | erts/emulator/beam/erl_process.c | 329 | ||||
| -rw-r--r-- | erts/emulator/beam/erl_process.h | 4 | ||||
| -rw-r--r-- | erts/emulator/beam/io.c | 6 | ||||
| -rw-r--r-- | erts/emulator/drivers/common/inet_drv.c | 23 | ||||
| -rw-r--r-- | erts/emulator/valgrind/suppress.patched.3.6.0 | 18 | ||||
| -rw-r--r-- | erts/emulator/valgrind/suppress.standard | 18 |
9 files changed, 417 insertions, 80 deletions
diff --git a/erts/emulator/beam/erl_async.c b/erts/emulator/beam/erl_async.c index f321ed21aa..cb975d64b0 100644 --- a/erts/emulator/beam/erl_async.c +++ b/erts/emulator/beam/erl_async.c @@ -253,7 +253,9 @@ erts_get_async_ready_queue(Uint sched_id) static ERTS_INLINE void async_add(ErtsAsync *a, ErtsAsyncQ* q) { +#ifdef USE_VM_PROBES int len; +#endif if (is_internal_port(a->port)) { #if ERTS_USE_ASYNC_READY_Q @@ -291,7 +293,9 @@ static ERTS_INLINE ErtsAsync *async_get(ErtsThrQ_t *q, int saved_fin_deq = 0; ErtsThrQFinDeQ_t fin_deq; #endif +#ifdef USE_VM_PROBES int len; +#endif while (1) { ErtsAsync *a = (ErtsAsync *) erts_thr_q_dequeue(q); diff --git a/erts/emulator/beam/erl_db_hash.c b/erts/emulator/beam/erl_db_hash.c index c726be5fb4..2fea4671e1 100644 --- a/erts/emulator/beam/erl_db_hash.c +++ b/erts/emulator/beam/erl_db_hash.c @@ -105,7 +105,20 @@ #define NSEG_2 256 /* Size of second segment table */ #define NSEG_INC 128 /* Number of segments to grow after that */ -#define SEGTAB(tb) ((struct segment**)erts_smp_atomic_read_ddrb(&(tb)->segtab)) +#ifdef ERTS_SMP +# define DB_USING_FINE_LOCKING(TB) (((TB))->common.type & DB_FINE_LOCKED) +#else +# define DB_USING_FINE_LOCKING(TB) 0 +#endif + +#ifdef ETHR_ORDERED_READ_DEPEND +#define SEGTAB(tb) ((struct segment**) erts_smp_atomic_read_nob(&(tb)->segtab)) +#else +#define SEGTAB(tb) \ + (DB_USING_FINE_LOCKING(tb) \ + ? ((struct segment**) erts_smp_atomic_read_ddrb(&(tb)->segtab)) \ + : ((struct segment**) erts_smp_atomic_read_nob(&(tb)->segtab))) +#endif #define NACTIVE(tb) ((int)erts_smp_atomic_read_nob(&(tb)->nactive)) #define NITEMS(tb) ((int)erts_smp_atomic_read_nob(&(tb)->common.nitems)) @@ -122,7 +135,9 @@ */ static ERTS_INLINE Uint hash_to_ix(DbTableHash* tb, HashValue hval) { - Uint mask = erts_smp_atomic_read_acqb(&tb->szm); + Uint mask = (DB_USING_FINE_LOCKING(tb) + ? erts_smp_atomic_read_acqb(&tb->szm) + : erts_smp_atomic_read_nob(&tb->szm)); Uint ix = hval & mask; if (ix >= erts_smp_atomic_read_nob(&tb->nactive)) { ix &= mask>>1; @@ -319,7 +334,10 @@ struct ext_segment { static ERTS_INLINE void SET_SEGTAB(DbTableHash* tb, struct segment** segtab) { - erts_smp_atomic_set_wb(&tb->segtab, (erts_aint_t) segtab); + if (DB_USING_FINE_LOCKING(tb)) + erts_smp_atomic_set_wb(&tb->segtab, (erts_aint_t) segtab); + else + erts_smp_atomic_set_nob(&tb->segtab, (erts_aint_t) segtab); #ifdef VALGRIND tb->top_ptr_to_segment_with_active_segtab = EXTSEG(segtab); #endif @@ -2501,6 +2519,28 @@ static Eterm build_term_list(Process* p, HashDbTerm* ptr1, HashDbTerm* ptr2, return list; } +static ERTS_INLINE int +begin_resizing(DbTableHash* tb) +{ + if (DB_USING_FINE_LOCKING(tb)) + return !erts_smp_atomic_xchg_acqb(&tb->is_resizing, 1); + else { + if (erts_smp_atomic_read_nob(&tb->is_resizing)) + return 0; + erts_smp_atomic_set_nob(&tb->is_resizing, 1); + return 1; + } +} + +static ERTS_INLINE void +done_resizing(DbTableHash* tb) +{ + if (DB_USING_FINE_LOCKING(tb)) + erts_smp_atomic_set_relb(&tb->is_resizing, 0); + else + erts_smp_atomic_set_nob(&tb->is_resizing, 0); +} + /* Grow table with one new bucket. ** Allocate new segment if needed. */ @@ -2513,9 +2553,8 @@ static void grow(DbTableHash* tb, int nactive) int from_ix; int szm; - if (erts_smp_atomic_xchg_acqb(&tb->is_resizing, 1)) { + if (!begin_resizing(tb)) return; /* already in progress */ - } if (NACTIVE(tb) != nactive) { goto abort; /* already done (race) */ } @@ -2547,9 +2586,12 @@ static void grow(DbTableHash* tb, int nactive) } erts_smp_atomic_inc_nob(&tb->nactive); if (from_ix == 0) { - erts_smp_atomic_set_relb(&tb->szm, szm); + if (DB_USING_FINE_LOCKING(tb)) + erts_smp_atomic_set_relb(&tb->szm, szm); + else + erts_smp_atomic_set_nob(&tb->szm, szm); } - erts_smp_atomic_set_relb(&tb->is_resizing, 0); + done_resizing(tb); /* Finally, let's split the bucket. We try to do it in a smart way to keep link order and avoid unnecessary updates of next-pointers */ @@ -2581,7 +2623,7 @@ static void grow(DbTableHash* tb, int nactive) return; abort: - erts_smp_atomic_set_relb(&tb->is_resizing, 0); + done_resizing(tb); } @@ -2590,9 +2632,8 @@ abort: */ static void shrink(DbTableHash* tb, int nactive) { - if (erts_smp_atomic_xchg_acqb(&tb->is_resizing, 1)) { + if (!begin_resizing(tb)) return; /* already in progress */ - } if (NACTIVE(tb) == nactive) { erts_smp_rwmtx_t* lck; int src_ix = nactive - 1; @@ -2639,7 +2680,7 @@ static void shrink(DbTableHash* tb, int nactive) } /*else already done */ - erts_smp_atomic_set_relb(&tb->is_resizing, 0); + done_resizing(tb); } diff --git a/erts/emulator/beam/erl_init.c b/erts/emulator/beam/erl_init.c index ca4385dd3a..7bf8d14708 100644 --- a/erts/emulator/beam/erl_init.c +++ b/erts/emulator/beam/erl_init.c @@ -511,10 +511,14 @@ void erts_usage(void) erts_fprintf(stderr, "-rg amount set reader groups limit\n"); erts_fprintf(stderr, "-sbt type set scheduler bind type, valid types are:\n"); erts_fprintf(stderr, " u|ns|ts|ps|s|nnts|nnps|tnnps|db\n"); + erts_fprintf(stderr, "-sbwt val set scheduler busy wait threshold, valid values are:\n"); + erts_fprintf(stderr, " none|very_short|short|medium|long|very_long.\n"); erts_fprintf(stderr, "-scl bool enable/disable compaction of scheduler load,\n"); erts_fprintf(stderr, " see the erl(1) documentation for more info.\n"); erts_fprintf(stderr, "-sct cput set cpu topology,\n"); erts_fprintf(stderr, " see the erl(1) documentation for more info.\n"); + erts_fprintf(stderr, "-sws val set scheduler wakeup strategy, valid values are:\n"); + erts_fprintf(stderr, " default|legacy|proposal.\n"); erts_fprintf(stderr, "-swt val set scheduler wakeup threshold, valid values are:\n"); erts_fprintf(stderr, " very_low|low|medium|high|very_high.\n"); erts_fprintf(stderr, "-sss size suggested stack size in kilo words for scheduler threads,\n"); @@ -790,6 +794,10 @@ early_init(int *argc, char **argv) /* } } +#ifndef USE_THREADS + erts_async_max_threads = 0; +#endif + #ifdef ERTS_SMP no_schedulers = schdlrs; no_schedulers_online = schdlrs_onln; @@ -1198,6 +1206,16 @@ erl_start(int argc, char **argv) erts_usage(); } } + else if (has_prefix("bwt", sub_param)) { + arg = get_arg(sub_param+3, argv[i+1], &i); + if (erts_sched_set_busy_wait_threshold(arg) != 0) { + erts_fprintf(stderr, "bad scheduler busy wait threshold: %s\n", + arg); + erts_usage(); + } + VERBOSE(DEBUG_SYSTEM, + ("scheduler wakup threshold: %s\n", arg)); + } else if (has_prefix("cl", sub_param)) { arg = get_arg(sub_param+2, argv[i+1], &i); if (sys_strcmp("true", arg) == 0) @@ -1258,13 +1276,23 @@ erl_start(int argc, char **argv) erts_use_sender_punish = 0; else if (sys_strcmp("wt", sub_param) == 0) { arg = get_arg(sub_param+2, argv[i+1], &i); - if (erts_sched_set_wakeup_limit(arg) != 0) { + if (erts_sched_set_wakeup_other_thresold(arg) != 0) { erts_fprintf(stderr, "scheduler wakeup threshold: %s\n", arg); erts_usage(); } VERBOSE(DEBUG_SYSTEM, - ("scheduler wakup threshold: %s\n", arg)); + ("scheduler wakeup threshold: %s\n", arg)); + } + else if (sys_strcmp("ws", sub_param) == 0) { + arg = get_arg(sub_param+2, argv[i+1], &i); + if (erts_sched_set_wakeup_other_type(arg) != 0) { + erts_fprintf(stderr, "scheduler wakeup strategy: %s\n", + arg); + erts_usage(); + } + VERBOSE(DEBUG_SYSTEM, + ("scheduler wakeup threshold: %s\n", arg)); } else if (has_prefix("ss", sub_param)) { /* suggested stack size (Kilo Words) for scheduler threads */ diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c index 95d408f79d..09ca536188 100644 --- a/erts/emulator/beam/erl_process.c +++ b/erts/emulator/beam/erl_process.c @@ -52,21 +52,22 @@ #define ERTS_SCHED_SPIN_UNTIL_YIELD 100 -#define ERTS_SCHED_SYS_SLEEP_SPINCOUNT 10 +#define ERTS_SCHED_SYS_SLEEP_SPINCOUNT_VERY_LONG 40 +#define ERTS_SCHED_AUX_WORK_SLEEP_SPINCOUNT_FACT_VERY_LONG 1000 +#define ERTS_SCHED_SYS_SLEEP_SPINCOUNT_LONG 20 +#define ERTS_SCHED_AUX_WORK_SLEEP_SPINCOUNT_FACT_LONG 1000 +#define ERTS_SCHED_SYS_SLEEP_SPINCOUNT_MEDIUM 10 +#define ERTS_SCHED_AUX_WORK_SLEEP_SPINCOUNT_FACT_MEDIUM 1000 +#define ERTS_SCHED_SYS_SLEEP_SPINCOUNT_SHORT 10 +#define ERTS_SCHED_AUX_WORK_SLEEP_SPINCOUNT_FACT_SHORT 0 +#define ERTS_SCHED_SYS_SLEEP_SPINCOUNT_VERY_SHORT 5 +#define ERTS_SCHED_AUX_WORK_SLEEP_SPINCOUNT_FACT_VERY_SHORT 0 +#define ERTS_SCHED_SYS_SLEEP_SPINCOUNT_NONE 0 +#define ERTS_SCHED_AUX_WORK_SLEEP_SPINCOUNT_FACT_NONE 0 + #define ERTS_SCHED_TSE_SLEEP_SPINCOUNT_FACT 1000 -#define ERTS_SCHED_TSE_SLEEP_SPINCOUNT \ - (ERTS_SCHED_SYS_SLEEP_SPINCOUNT*ERTS_SCHED_TSE_SLEEP_SPINCOUNT_FACT) #define ERTS_SCHED_SUSPEND_SLEEP_SPINCOUNT 0 -#define ERTS_WAKEUP_OTHER_LIMIT_VERY_HIGH (200*CONTEXT_REDS) -#define ERTS_WAKEUP_OTHER_LIMIT_HIGH (50*CONTEXT_REDS) -#define ERTS_WAKEUP_OTHER_LIMIT_MEDIUM (10*CONTEXT_REDS) -#define ERTS_WAKEUP_OTHER_LIMIT_LOW (CONTEXT_REDS) -#define ERTS_WAKEUP_OTHER_LIMIT_VERY_LOW (CONTEXT_REDS/10) - -#define ERTS_WAKEUP_OTHER_DEC 10 -#define ERTS_WAKEUP_OTHER_FIXED_INC (CONTEXT_REDS/10) - #if 0 || defined(DEBUG) #define ERTS_FAKE_SCHED_BIND_PRINT_SORTED_CPU_DATA #endif @@ -123,14 +124,18 @@ Uint erts_no_schedulers; Uint erts_max_processes = ERTS_DEFAULT_MAX_PROCESSES; Uint erts_process_tab_index_mask; -static int wakeup_other_limit; - int erts_sched_thread_suggested_stack_size = -1; #ifdef ERTS_ENABLE_LOCK_CHECK ErtsLcPSDLocks erts_psd_required_locks[ERTS_PSD_SIZE]; #endif +static struct { + int aux_work; + int tse; + int sys_schedule; +} sched_busy_wait; + #ifdef ERTS_SMP int erts_disable_proc_not_running_opt; @@ -2046,7 +2051,7 @@ scheduler_wait(int *fcalls, ErtsSchedulerData *esdp, ErtsRunQueue *rq) erts_smp_runq_unlock(rq); - spincount = ERTS_SCHED_TSE_SLEEP_SPINCOUNT; + spincount = sched_busy_wait.tse; tse_wait: @@ -2097,7 +2102,7 @@ scheduler_wait(int *fcalls, ErtsSchedulerData *esdp, ErtsRunQueue *rq) } flgs = sched_prep_cont_spin_wait(ssi); - spincount = ERTS_SCHED_TSE_SLEEP_SPINCOUNT; + spincount = sched_busy_wait.aux_work; if (!(flgs & ERTS_SSI_FLG_WAITING)) { ASSERT(!(flgs & ERTS_SSI_FLG_SLEEPING)); @@ -2134,7 +2139,9 @@ scheduler_wait(int *fcalls, ErtsSchedulerData *esdp, ErtsRunQueue *rq) ASSERT(working); sched_wall_time_change(esdp, working = 0); - spincount = ERTS_SCHED_SYS_SLEEP_SPINCOUNT; + spincount = sched_busy_wait.sys_schedule; + if (spincount == 0) + goto sys_aux_work; while (spincount-- > 0) { @@ -3560,31 +3567,280 @@ erts_debug_nbalance(void) #endif } +/* Wakeup other schedulers */ + +typedef enum { + ERTS_SCHED_WAKEUP_OTHER_THRESHOLD_VERY_HIGH, + ERTS_SCHED_WAKEUP_OTHER_THRESHOLD_HIGH, + ERTS_SCHED_WAKEUP_OTHER_THRESHOLD_MEDIUM, + ERTS_SCHED_WAKEUP_OTHER_THRESHOLD_LOW, + ERTS_SCHED_WAKEUP_OTHER_THRESHOLD_VERY_LOW +} ErtsSchedWakeupOtherThreshold; + +typedef enum { + ERTS_SCHED_WAKEUP_OTHER_TYPE_PROPOSAL, + ERTS_SCHED_WAKEUP_OTHER_TYPE_LEGACY +} ErtsSchedWakeupOtherType; + +/* First proposal */ + +#define ERTS_WAKEUP_OTHER_LIMIT_VERY_HIGH (200*CONTEXT_REDS) +#define ERTS_WAKEUP_OTHER_LIMIT_HIGH (50*CONTEXT_REDS) +#define ERTS_WAKEUP_OTHER_LIMIT_MEDIUM (10*CONTEXT_REDS) +#define ERTS_WAKEUP_OTHER_LIMIT_LOW (CONTEXT_REDS) +#define ERTS_WAKEUP_OTHER_LIMIT_VERY_LOW (CONTEXT_REDS/10) + +#define ERTS_WAKEUP_OTHER_DEC_SHIFT_VERY_HIGH 3 +#define ERTS_WAKEUP_OTHER_DEC_SHIFT_HIGH 1 +#define ERTS_WAKEUP_OTHER_DEC_SHIFT_MEDIUM 0 +#define ERTS_WAKEUP_OTHER_DEC_SHIFT_LOW -2 +#define ERTS_WAKEUP_OTHER_DEC_SHIFT_VERY_LOW -5 + +#define ERTS_WAKEUP_OTHER_DEC_SHIFT 2 +#define ERTS_WAKEUP_OTHER_FIXED_INC (CONTEXT_REDS/10) + +/* To be legacy */ + +#define ERTS_WAKEUP_OTHER_LIMIT_VERY_HIGH_LEGACY (200*CONTEXT_REDS) +#define ERTS_WAKEUP_OTHER_LIMIT_HIGH_LEGACY (50*CONTEXT_REDS) +#define ERTS_WAKEUP_OTHER_LIMIT_MEDIUM_LEGACY (10*CONTEXT_REDS) +#define ERTS_WAKEUP_OTHER_LIMIT_LOW_LEGACY (CONTEXT_REDS) +#define ERTS_WAKEUP_OTHER_LIMIT_VERY_LOW_LEGACY (CONTEXT_REDS/10) + +#define ERTS_WAKEUP_OTHER_DEC_LEGACY 10 +#define ERTS_WAKEUP_OTHER_FIXED_INC_LEGACY (CONTEXT_REDS/10) + +#ifdef ERTS_SMP + +static struct { + ErtsSchedWakeupOtherThreshold threshold; + ErtsSchedWakeupOtherType type; + int limit; + int dec_shift; + int dec_mask; + void (*check)(ErtsRunQueue *rq); +} wakeup_other; + +static void +wakeup_other_check(ErtsRunQueue *rq) +{ + int wo_reds = rq->wakeup_other_reds; + if (wo_reds) { + int left_len = rq->len - 1; + if (left_len < 1) { + int wo_reduce = wo_reds << wakeup_other.dec_shift; + wo_reduce &= wakeup_other.dec_mask; + rq->wakeup_other -= wo_reduce; + if (rq->wakeup_other < 0) + rq->wakeup_other = 0; + } + else { + rq->wakeup_other += (left_len*wo_reds + + ERTS_WAKEUP_OTHER_FIXED_INC); + if (rq->wakeup_other > wakeup_other.limit) { + int empty_rqs = + erts_smp_atomic32_read_acqb(&no_empty_run_queues); + if (empty_rqs != 0) + wake_scheduler_on_empty_runq(rq); + rq->wakeup_other = 0; + } + } + rq->wakeup_other_reds = 0; + } +} + +static void +wakeup_other_set_limit(void) +{ + switch (wakeup_other.threshold) { + case ERTS_SCHED_WAKEUP_OTHER_THRESHOLD_VERY_HIGH: + wakeup_other.limit = ERTS_WAKEUP_OTHER_LIMIT_VERY_HIGH; + wakeup_other.dec_shift = ERTS_WAKEUP_OTHER_DEC_SHIFT_VERY_HIGH; + break; + case ERTS_SCHED_WAKEUP_OTHER_THRESHOLD_HIGH: + wakeup_other.limit = ERTS_WAKEUP_OTHER_LIMIT_HIGH; + wakeup_other.dec_shift = ERTS_WAKEUP_OTHER_DEC_SHIFT_HIGH; + break; + case ERTS_SCHED_WAKEUP_OTHER_THRESHOLD_MEDIUM: + wakeup_other.limit = ERTS_WAKEUP_OTHER_LIMIT_MEDIUM; + wakeup_other.dec_shift = ERTS_WAKEUP_OTHER_DEC_SHIFT_MEDIUM; + break; + case ERTS_SCHED_WAKEUP_OTHER_THRESHOLD_LOW: + wakeup_other.limit = ERTS_WAKEUP_OTHER_LIMIT_LOW; + wakeup_other.dec_shift = ERTS_WAKEUP_OTHER_DEC_SHIFT_LOW; + break; + case ERTS_SCHED_WAKEUP_OTHER_THRESHOLD_VERY_LOW: + wakeup_other.limit = ERTS_WAKEUP_OTHER_LIMIT_VERY_LOW; + wakeup_other.dec_shift = ERTS_WAKEUP_OTHER_DEC_SHIFT_VERY_LOW; + break; + } + if (wakeup_other.dec_shift < 0) + wakeup_other.dec_mask = (1 << (sizeof(wakeup_other.dec_mask)*8 + + wakeup_other.dec_shift)) - 1; + else { + wakeup_other.dec_mask = 0; + wakeup_other.dec_mask = ~wakeup_other.dec_mask; + } +} + +static void +wakeup_other_check_legacy(ErtsRunQueue *rq) +{ + int wo_reds = rq->wakeup_other_reds; + if (wo_reds) { + if (rq->len < 2) { + rq->wakeup_other -= ERTS_WAKEUP_OTHER_DEC_LEGACY*wo_reds; + if (rq->wakeup_other < 0) + rq->wakeup_other = 0; + } + else if (rq->wakeup_other < wakeup_other.limit) + rq->wakeup_other += rq->len*wo_reds + ERTS_WAKEUP_OTHER_FIXED_INC_LEGACY; + else { + if (erts_smp_atomic32_read_acqb(&no_empty_run_queues) != 0) { + wake_scheduler_on_empty_runq(rq); + rq->wakeup_other = 0; + } + rq->wakeup_other = 0; + } + } + rq->wakeup_other_reds = 0; +} + +static void +wakeup_other_set_limit_legacy(void) +{ + switch (wakeup_other.threshold) { + case ERTS_SCHED_WAKEUP_OTHER_THRESHOLD_VERY_HIGH: + wakeup_other.limit = ERTS_WAKEUP_OTHER_LIMIT_VERY_HIGH_LEGACY; + break; + case ERTS_SCHED_WAKEUP_OTHER_THRESHOLD_HIGH: + wakeup_other.limit = ERTS_WAKEUP_OTHER_LIMIT_HIGH_LEGACY; + break; + case ERTS_SCHED_WAKEUP_OTHER_THRESHOLD_MEDIUM: + wakeup_other.limit = ERTS_WAKEUP_OTHER_LIMIT_MEDIUM_LEGACY; + break; + case ERTS_SCHED_WAKEUP_OTHER_THRESHOLD_LOW: + wakeup_other.limit = ERTS_WAKEUP_OTHER_LIMIT_LOW_LEGACY; + break; + case ERTS_SCHED_WAKEUP_OTHER_THRESHOLD_VERY_LOW: + wakeup_other.limit = ERTS_WAKEUP_OTHER_LIMIT_VERY_LOW_LEGACY; + break; + } +} + +static void +set_wakeup_other_data(void) +{ + switch (wakeup_other.type) { + case ERTS_SCHED_WAKEUP_OTHER_TYPE_PROPOSAL: + wakeup_other.check = wakeup_other_check; + wakeup_other_set_limit(); + break; + case ERTS_SCHED_WAKEUP_OTHER_TYPE_LEGACY: + wakeup_other.check = wakeup_other_check_legacy; + wakeup_other_set_limit_legacy(); + break; + } +} + +#endif + void erts_early_init_scheduling(int no_schedulers) { aux_work_timeout_early_init(no_schedulers); - wakeup_other_limit = ERTS_WAKEUP_OTHER_LIMIT_MEDIUM; +#ifdef ERTS_SMP + wakeup_other.threshold = ERTS_SCHED_WAKEUP_OTHER_THRESHOLD_MEDIUM; + wakeup_other.type = ERTS_SCHED_WAKEUP_OTHER_TYPE_LEGACY; +#endif + 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); } int -erts_sched_set_wakeup_limit(char *str) +erts_sched_set_wakeup_other_thresold(char *str) { + ErtsSchedWakeupOtherThreshold threshold; if (sys_strcmp(str, "very_high") == 0) - wakeup_other_limit = ERTS_WAKEUP_OTHER_LIMIT_VERY_HIGH; + threshold = ERTS_SCHED_WAKEUP_OTHER_THRESHOLD_VERY_HIGH; else if (sys_strcmp(str, "high") == 0) - wakeup_other_limit = ERTS_WAKEUP_OTHER_LIMIT_HIGH; + threshold = ERTS_SCHED_WAKEUP_OTHER_THRESHOLD_HIGH; else if (sys_strcmp(str, "medium") == 0) - wakeup_other_limit = ERTS_WAKEUP_OTHER_LIMIT_MEDIUM; + threshold = ERTS_SCHED_WAKEUP_OTHER_THRESHOLD_MEDIUM; else if (sys_strcmp(str, "low") == 0) - wakeup_other_limit = ERTS_WAKEUP_OTHER_LIMIT_LOW; + threshold = ERTS_SCHED_WAKEUP_OTHER_THRESHOLD_LOW; else if (sys_strcmp(str, "very_low") == 0) - wakeup_other_limit = ERTS_WAKEUP_OTHER_LIMIT_VERY_LOW; + threshold = ERTS_SCHED_WAKEUP_OTHER_THRESHOLD_VERY_LOW; + else + return EINVAL; +#ifdef ERTS_SMP + wakeup_other.threshold = threshold; + set_wakeup_other_data(); +#endif + return 0; +} + +int +erts_sched_set_wakeup_other_type(char *str) +{ + ErtsSchedWakeupOtherType type; + if (sys_strcmp(str, "proposal") == 0) + type = ERTS_SCHED_WAKEUP_OTHER_TYPE_PROPOSAL; + else if (sys_strcmp(str, "default") == 0) + type = ERTS_SCHED_WAKEUP_OTHER_TYPE_LEGACY; + else if (sys_strcmp(str, "legacy") == 0) + type = ERTS_SCHED_WAKEUP_OTHER_TYPE_LEGACY; else return EINVAL; +#ifdef ERTS_SMP + wakeup_other.type = type; +#endif return 0; } +int +erts_sched_set_busy_wait_threshold(char *str) +{ + int sys_sched; + int aux_work_fact; + + if (sys_strcmp(str, "very_long") == 0) { + sys_sched = ERTS_SCHED_SYS_SLEEP_SPINCOUNT_VERY_LONG; + aux_work_fact = ERTS_SCHED_AUX_WORK_SLEEP_SPINCOUNT_FACT_VERY_LONG; + } + else if (sys_strcmp(str, "long") == 0) { + sys_sched = ERTS_SCHED_SYS_SLEEP_SPINCOUNT_LONG; + aux_work_fact = ERTS_SCHED_AUX_WORK_SLEEP_SPINCOUNT_FACT_LONG; + } + else if (sys_strcmp(str, "medium") == 0) { + sys_sched = ERTS_SCHED_SYS_SLEEP_SPINCOUNT_MEDIUM; + aux_work_fact = ERTS_SCHED_AUX_WORK_SLEEP_SPINCOUNT_FACT_MEDIUM; + } + else if (sys_strcmp(str, "short") == 0) { + sys_sched = ERTS_SCHED_SYS_SLEEP_SPINCOUNT_SHORT; + aux_work_fact = ERTS_SCHED_AUX_WORK_SLEEP_SPINCOUNT_FACT_SHORT; + } + else if (sys_strcmp(str, "very_short") == 0) { + sys_sched = ERTS_SCHED_SYS_SLEEP_SPINCOUNT_VERY_SHORT; + aux_work_fact = ERTS_SCHED_AUX_WORK_SLEEP_SPINCOUNT_FACT_VERY_SHORT; + } + else if (sys_strcmp(str, "none") == 0) { + sys_sched = ERTS_SCHED_SYS_SLEEP_SPINCOUNT_NONE; + aux_work_fact = ERTS_SCHED_AUX_WORK_SLEEP_SPINCOUNT_FACT_NONE; + } + else { + return EINVAL; + } + + sched_busy_wait.sys_schedule = sys_sched; + sched_busy_wait.tse = sys_sched*ERTS_SCHED_TSE_SLEEP_SPINCOUNT_FACT; + sched_busy_wait.aux_work = sys_sched*aux_work_fact; + + return 0; +} static void init_aux_work_data(ErtsAuxWorkData *awdp, ErtsSchedulerData *esdp) { @@ -3613,6 +3869,10 @@ erts_init_scheduling(int no_schedulers, int no_schedulers_online) init_misc_op_list_alloc(); +#ifdef ERTS_SMP + set_wakeup_other_data(); +#endif + ASSERT(no_schedulers_online <= no_schedulers); ASSERT(no_schedulers_online >= 1); ASSERT(no_schedulers >= 1); @@ -6502,26 +6762,7 @@ Process *schedule(Process *p, int calls) exec_misc_ops(rq); #ifdef ERTS_SMP - { - int wo_reds = rq->wakeup_other_reds; - if (wo_reds) { - if (rq->len < 2) { - rq->wakeup_other -= ERTS_WAKEUP_OTHER_DEC*wo_reds; - if (rq->wakeup_other < 0) - rq->wakeup_other = 0; - } - else if (rq->wakeup_other < wakeup_other_limit) - rq->wakeup_other += rq->len*wo_reds + ERTS_WAKEUP_OTHER_FIXED_INC; - else { - if (erts_smp_atomic32_read_acqb(&no_empty_run_queues) != 0) { - wake_scheduler_on_empty_runq(rq); - rq->wakeup_other = 0; - } - rq->wakeup_other = 0; - } - } - rq->wakeup_other_reds = 0; - } + wakeup_other.check(rq); #endif /* diff --git a/erts/emulator/beam/erl_process.h b/erts/emulator/beam/erl_process.h index ca194ab90c..28aaedf2e2 100644 --- a/erts/emulator/beam/erl_process.h +++ b/erts/emulator/beam/erl_process.h @@ -1096,7 +1096,9 @@ ErtsProcList *erts_proclist_create(Process *); void erts_proclist_destroy(ErtsProcList *); int erts_proclist_same(ErtsProcList *, Process *); -int erts_sched_set_wakeup_limit(char *str); +int erts_sched_set_wakeup_other_thresold(char *str); +int erts_sched_set_wakeup_other_type(char *str); +int erts_sched_set_busy_wait_threshold(char *str); #if defined(ERTS_SMP) && defined(ERTS_ENABLE_LOCK_CHECK) int erts_dbg_check_halloc_lock(Process *p); diff --git a/erts/emulator/beam/io.c b/erts/emulator/beam/io.c index eb8db73bae..204bff299e 100644 --- a/erts/emulator/beam/io.c +++ b/erts/emulator/beam/io.c @@ -3260,6 +3260,8 @@ driver_deliver_term(ErlDrvPort port, Uint size = ptr[1]; Uint offset = ptr[2]; + erts_smp_atomic_add_nob(&erts_bytes_in, (erts_aint_t) size); + if (size <= ERL_ONHEAP_BIN_LIMIT) { ErlHeapBin* hbp = (ErlHeapBin *) hp; hp += heap_bin_size(size); @@ -3291,6 +3293,9 @@ driver_deliver_term(ErlDrvPort port, case ERL_DRV_BUF2BINARY: { /* char*, size */ byte *bufp = (byte *) ptr[0]; Uint size = (Uint) ptr[1]; + + erts_smp_atomic_add_nob(&erts_bytes_in, (erts_aint_t) size); + if (size <= ERL_ONHEAP_BIN_LIMIT) { ErlHeapBin* hbp = (ErlHeapBin *) hp; hp += heap_bin_size(size); @@ -3327,6 +3332,7 @@ driver_deliver_term(ErlDrvPort port, } case ERL_DRV_STRING: /* char*, length */ + erts_smp_atomic_add_nob(&erts_bytes_in, (erts_aint_t) ptr[1]); mess = buf_to_intlist(&hp, (char*)ptr[0], ptr[1], NIL); ptr += 2; break; diff --git a/erts/emulator/drivers/common/inet_drv.c b/erts/emulator/drivers/common/inet_drv.c index bf376f0494..76a9b55179 100644 --- a/erts/emulator/drivers/common/inet_drv.c +++ b/erts/emulator/drivers/common/inet_drv.c @@ -3124,6 +3124,7 @@ static int tcp_message(inet_descriptor* desc, const char* buf, int len) int i = 0; DEBUGF(("tcp_message(%ld): len = %d\r\n", (long)desc->port, len)); + /* XXX fprintf(stderr,"tcp_message send.\r\n"); */ i = LOAD_ATOM(spec, i, am_tcp); i = LOAD_PORT(spec, i, desc->dport); @@ -5426,6 +5427,7 @@ static int inet_set_opts(inet_descriptor* desc, char* ptr, int len) if (IS_SCTP(desc)) return sctp_set_opts(desc, ptr, len); #endif + /* XXX { int i; for(i=0;i<len;++i) fprintf(stderr,"0x%02X, ", (unsigned) ptr[i]); fprintf(stderr,"\r\n");} */ while(len >= 5) { opt = *ptr++; @@ -5755,10 +5757,16 @@ skip_os_setopt: if (desc->active != old_active) sock_select(desc, (FD_READ|FD_CLOSE), (desc->active>0)); + /* XXX: UDP sockets could also trigger immediate read here NIY */ if ((desc->stype==SOCK_STREAM) && desc->active) { if (!old_active || (desc->htype != old_htype)) { /* passive => active change OR header type change in active mode */ - return 1; + /* Return > 1 if only active changed to INET_ONCE -> direct read if + header type is unchanged. */ + /* XXX fprintf(stderr,"desc->htype == %d, old_htype == %d, + desc->active == %d, old_active == %d\r\n",(int)desc->htype, + (int) old_htype, (int) desc->active, (int) old_active );*/ + return 1+(desc->htype == old_htype && desc->active == INET_ONCE); } return 0; } @@ -7592,17 +7600,27 @@ static ErlDrvSSizeT inet_ctl(inet_descriptor* desc, int cmd, char* buf, case INET_REQ_SETOPTS: { /* set options */ DEBUGF(("inet_ctl(%ld): SETOPTS\r\n", (long)desc->port)); + /* XXX fprintf(stderr,"inet_ctl(%ld): SETOPTS (len = %d)\r\n", (long)desc->port,(int) len); */ switch(inet_set_opts(desc, buf, len)) { case -1: return ctl_error(EINVAL, rbuf, rsize); case 0: return ctl_reply(INET_REP_OK, NULL, 0, rbuf, rsize); - default: /* active/passive change!! */ + case 1: /* * Let's hope that the descriptor really is a tcp_descriptor here. */ + /* fprintf(stderr,"Triggered tcp_deliver by setopt.\r\n"); */ tcp_deliver((tcp_descriptor *) desc, 0); return ctl_reply(INET_REP_OK, NULL, 0, rbuf, rsize); + default: + /* fprintf(stderr,"Triggered tcp_recv by setopt.\r\n"); */ + /* + * Same as above, but active changed to once w/o header type + * change, so try a read instead of just deliver. + */ + tcp_recv((tcp_descriptor *) desc, 0); + return ctl_reply(INET_REP_OK, NULL, 0, rbuf, rsize); } } @@ -9196,6 +9214,7 @@ static int tcp_inet_input(tcp_descriptor* desc, HANDLE event) #endif ASSERT(!INETP(desc)->is_ignored); DEBUGF(("tcp_inet_input(%ld) {s=%d\r\n", port, desc->inet.s)); + /* XXX fprintf(stderr,"tcp_inet_input(%ld) {s=%d}\r\n",(long) desc->inet.port, desc->inet.s); */ if (desc->inet.state == INET_STATE_ACCEPTING) { SOCKET s; unsigned int len; diff --git a/erts/emulator/valgrind/suppress.patched.3.6.0 b/erts/emulator/valgrind/suppress.patched.3.6.0 index 37960af17a..62ba032520 100644 --- a/erts/emulator/valgrind/suppress.patched.3.6.0 +++ b/erts/emulator/valgrind/suppress.patched.3.6.0 @@ -57,18 +57,16 @@ fun:putenv fun:erts_sys_putenv fun:os_putenv_2 - fun:process_main + ... } { -Leak in libc putenv -Memcheck:Leak -fun:malloc -fun:erts_sys_alloc -... -fun:erts_alloc -fun:erts_sys_putenv -fun:os_putenv_2 -fun:process_main + Leak in libc putenv + Memcheck:Leak + ... + fun:erts_alloc + fun:erts_sys_putenv + fun:os_putenv_2 + ... } { erronous warning diff --git a/erts/emulator/valgrind/suppress.standard b/erts/emulator/valgrind/suppress.standard index 54677c36be..5a129bfd10 100644 --- a/erts/emulator/valgrind/suppress.standard +++ b/erts/emulator/valgrind/suppress.standard @@ -54,18 +54,16 @@ fun:putenv fun:erts_sys_putenv fun:os_putenv_2 - fun:process_main + ... } { -Leak in libc putenv -Memcheck:Leak -fun:malloc -fun:erts_sys_alloc -... -fun:erts_alloc -fun:erts_sys_putenv -fun:os_putenv_2 -fun:process_main + Leak in libc putenv + Memcheck:Leak + ... + fun:erts_alloc + fun:erts_sys_putenv + fun:os_putenv_2 + ... } { erronous warning |