aboutsummaryrefslogtreecommitdiffstats
path: root/erts/emulator/beam/erl_process.c
diff options
context:
space:
mode:
authorRickard Green <[email protected]>2017-01-17 12:39:55 +0100
committerRickard Green <[email protected]>2017-01-17 12:39:55 +0100
commit8e77d78810c2b77a81e261234b351d30e374d629 (patch)
treedb5c8dc37898b5c8e8bcd9d91b1c1e35ca3ff4dc /erts/emulator/beam/erl_process.c
parent894cceaa2191bbb59474c1a1bfc657493c5ea79e (diff)
parent0d20adbfbb0918c8d6bfc629e9e4ff89d2ef1a64 (diff)
downloadotp-8e77d78810c2b77a81e261234b351d30e374d629.tar.gz
otp-8e77d78810c2b77a81e261234b351d30e374d629.tar.bz2
otp-8e77d78810c2b77a81e261234b351d30e374d629.zip
Merge branch 'rickard/ds-features-20'
OTP-14152 * rickard/ds-features-20: Remove experimental disclaimer on dirty schedulers Scheduler wall time support for dirty schedulers Always return info from system_info(dirty_[cpu|io]_schedulers*) Improve etp-commands Switch between scheduler types when multi-scheduling is blocked Perform potentially long GC on dirty schedulers if available Return and exception trace for nif-export scheduled BIFs Optimize handling of BIF errors Support for dirty BIFs Reduction counting on non-tail return
Diffstat (limited to 'erts/emulator/beam/erl_process.c')
-rw-r--r--erts/emulator/beam/erl_process.c1751
1 files changed, 1255 insertions, 496 deletions
diff --git a/erts/emulator/beam/erl_process.c b/erts/emulator/beam/erl_process.c
index 8ce5e28773..f80ebdf31b 100644
--- a/erts/emulator/beam/erl_process.c
+++ b/erts/emulator/beam/erl_process.c
@@ -48,6 +48,7 @@
#include "erl_bif_unique.h"
#define ERTS_WANT_TIMER_WHEEL_API
#include "erl_time.h"
+#include "erl_nfunc_sched.h"
#define ERTS_CHECK_TIME_REDS CONTEXT_REDS
#define ERTS_DELAYED_WAKEUP_INFINITY (~(Uint64) 0)
@@ -106,9 +107,31 @@
#define LOW_BIT (1 << PRIORITY_LOW)
#define PORT_BIT (1 << ERTS_PORT_PRIO_LEVEL)
-#define ERTS_EMPTY_RUNQ(RQ) \
- ((ERTS_RUNQ_FLGS_GET_NOB((RQ)) & ERTS_RUNQ_FLGS_QMASK) == 0 \
- && (RQ)->misc.start == NULL)
+#define ERTS_IS_RUNQ_EMPTY_FLGS(FLGS) \
+ (!((FLGS) & (ERTS_RUNQ_FLGS_QMASK|ERTS_RUNQ_FLG_MISC_OP)))
+
+#define ERTS_IS_RUNQ_EMPTY_PORTS_FLGS(FLGS) \
+ (!((FLGS) & (PORT_BIT|ERTS_RUNQ_FLG_MISC_OP)))
+
+#define ERTS_EMPTY_RUNQ(RQ) \
+ ERTS_IS_RUNQ_EMPTY_FLGS(ERTS_RUNQ_FLGS_GET_NOB((RQ)))
+
+#define ERTS_EMPTY_RUNQ_PORTS(RQ) \
+ ERTS_IS_RUNQ_EMPTY_FLGS(ERTS_RUNQ_FLGS_GET_NOB((RQ)))
+
+static ERTS_INLINE int
+runq_got_work_to_execute_flags(Uint32 flags)
+{
+ if (flags & ERTS_RUNQ_FLG_HALTING)
+ return !ERTS_IS_RUNQ_EMPTY_PORTS_FLGS(flags);
+ return !ERTS_IS_RUNQ_EMPTY_FLGS(flags);
+}
+
+static ERTS_INLINE int
+runq_got_work_to_execute(ErtsRunQueue *rq)
+{
+ return runq_got_work_to_execute_flags(ERTS_RUNQ_FLGS_GET_NOB(rq));
+}
#undef RUNQ_READ_RQ
#undef RUNQ_SET_RQ
@@ -140,9 +163,6 @@ do { \
# define ERTS_DBG_VERIFY_VALID_RUNQP(RQP)
#endif
-#define ERTS_EMPTY_RUNQ_PORTS(RQ) \
- (RUNQ_READ_LEN(&(RQ)->ports.info.len) == 0 && (RQ)->misc.start == NULL)
-
const Process erts_invalid_process = {{ERTS_INVALID_PID}};
extern BeamInstr beam_apply[];
@@ -193,170 +213,145 @@ static ErtsAuxWorkData *aux_thread_aux_work_data;
#define ERTS_SCHDLR_SSPND_CHNG_ONLN (((erts_aint32_t) 1) << 2)
#define ERTS_SCHDLR_SSPND_CHNG_DCPU_ONLN (((erts_aint32_t) 1) << 3)
-typedef enum {
- ERTS_SCHED_NORMAL,
- ERTS_SCHED_DIRTY_CPU,
- ERTS_SCHED_DIRTY_IO
-} ErtsSchedType;
-
typedef struct {
int ongoing;
ErtsProcList *blckrs;
ErtsProcList *chngq;
} ErtsMultiSchedulingBlock;
+typedef struct {
+ Uint32 normal;
+#ifdef ERTS_DIRTY_SCHEDULERS
+ Uint32 dirty_cpu;
+ Uint32 dirty_io;
+#endif
+} ErtsSchedTypeCounters;
+
static struct {
erts_smp_mtx_t mtx;
- Uint32 online;
- Uint32 curr_online;
- Uint32 active;
+ ErtsSchedTypeCounters online;
+ ErtsSchedTypeCounters curr_online;
+ ErtsSchedTypeCounters active;
erts_smp_atomic32_t changing;
ErtsProcList *chngq;
Eterm changer;
ErtsMultiSchedulingBlock nmsb; /* Normal multi Scheduling Block */
ErtsMultiSchedulingBlock msb; /* Multi Scheduling Block */
+#ifdef ERTS_DIRTY_SCHEDULERS
+ ErtsSchedType last_msb_dirty_type;
+#endif
} schdlr_sspnd;
-#define ERTS_SCHDLR_SSPND_S_BITS 10
-#define ERTS_SCHDLR_SSPND_DCS_BITS 11
-#define ERTS_SCHDLR_SSPND_DIS_BITS 11
-
-#define ERTS_SCHDLR_SSPND_S_MASK ((1 << ERTS_SCHDLR_SSPND_S_BITS)-1)
-#define ERTS_SCHDLR_SSPND_DCS_MASK ((1 << ERTS_SCHDLR_SSPND_DCS_BITS)-1)
-#define ERTS_SCHDLR_SSPND_DIS_MASK ((1 << ERTS_SCHDLR_SSPND_DIS_BITS)-1)
-
-#define ERTS_SCHDLR_SSPND_S_SHIFT 0
-#define ERTS_SCHDLR_SSPND_DCS_SHIFT (ERTS_SCHDLR_SSPND_S_SHIFT \
- + ERTS_SCHDLR_SSPND_S_BITS)
-#define ERTS_SCHDLR_SSPND_DIS_SHIFT (ERTS_SCHDLR_SSPND_DCS_SHIFT \
- + ERTS_SCHDLR_SSPND_DCS_BITS)
-
-#if (ERTS_SCHDLR_SSPND_S_BITS \
- + ERTS_SCHDLR_SSPND_DCS_BITS \
- + ERTS_SCHDLR_SSPND_DIS_BITS) > 32
-# error Wont fit in Uint32
-#endif
+static void init_scheduler_suspend(void);
-#if (ERTS_MAX_NO_OF_SCHEDULERS-1) > ERTS_SCHDLR_SSPND_S_MASK
-# error Max no schedulers wont fit in its bit-field
-#endif
-#if ERTS_MAX_NO_OF_DIRTY_CPU_SCHEDULERS > ERTS_SCHDLR_SSPND_DCS_MASK
-# error Max no dirty cpu schedulers wont fit in its bit-field
-#endif
-#if ERTS_MAX_NO_OF_DIRTY_IO_SCHEDULERS > ERTS_SCHDLR_SSPND_DIS_MASK
-# error Max no dirty io schedulers wont fit in its bit-field
+static ERTS_INLINE Uint32
+schdlr_sspnd_eq_nscheds(ErtsSchedTypeCounters *val1p, ErtsSchedTypeCounters *val2p)
+{
+ int res = val1p->normal == val2p->normal;
+#ifdef ERTS_DIRTY_SCHEDULERS
+ res &= val1p->dirty_cpu == val2p->dirty_cpu;
+ res &= val1p->dirty_io == val2p->dirty_io;
#endif
-
-#define ERTS_SCHDLR_SSPND_MAKE_NSCHEDS_VAL(S, DCS, DIS) \
- ((((Uint32) (((S) & ERTS_SCHDLR_SSPND_S_MASK))-1) \
- << ERTS_SCHDLR_SSPND_S_SHIFT) \
- | ((((Uint32) ((DCS) & ERTS_SCHDLR_SSPND_DCS_MASK)) \
- << ERTS_SCHDLR_SSPND_DCS_SHIFT)) \
- | ((((Uint32) ((DIS) & ERTS_SCHDLR_SSPND_DIS_MASK)) \
- << ERTS_SCHDLR_SSPND_DIS_SHIFT)))
-
-static void init_scheduler_suspend(void);
+ return res;
+}
static ERTS_INLINE Uint32
-schdlr_sspnd_get_nscheds(Uint32 *valp, ErtsSchedType type)
+schdlr_sspnd_get_nscheds(ErtsSchedTypeCounters *valp,
+ ErtsSchedType type)
{
- Uint32 res = (Uint32) (*valp);
switch (type) {
case ERTS_SCHED_NORMAL:
- res >>= ERTS_SCHDLR_SSPND_S_SHIFT;
- res &= (Uint32) ERTS_SCHDLR_SSPND_S_MASK;
- res++;
- break;
+ return valp->normal;
+#ifdef ERTS_DIRTY_SCHEDULERS
case ERTS_SCHED_DIRTY_CPU:
- res >>= ERTS_SCHDLR_SSPND_DCS_SHIFT;
- res &= (Uint32) ERTS_SCHDLR_SSPND_DCS_MASK;
- break;
+ return valp->dirty_cpu;
case ERTS_SCHED_DIRTY_IO:
- res >>= ERTS_SCHDLR_SSPND_DIS_SHIFT;
- res &= (Uint32) ERTS_SCHDLR_SSPND_DIS_MASK;
- break;
+ return valp->dirty_io;
+#else
+ case ERTS_SCHED_DIRTY_CPU:
+ case ERTS_SCHED_DIRTY_IO:
+ return 0;
+#endif
default:
ERTS_INTERNAL_ERROR("Invalid scheduler type");
return 0;
}
+}
+static ERTS_INLINE Uint32
+schdlr_sspnd_get_nscheds_tot(ErtsSchedTypeCounters *valp)
+{
+ Uint32 res = valp->normal;
+#ifdef ERTS_DIRTY_SCHEDULERS
+ res += valp->dirty_cpu;
+ res += valp->dirty_io;
+#endif
return res;
}
static ERTS_INLINE void
-schdlr_sspnd_dec_nscheds(Uint32 *valp, ErtsSchedType type)
+schdlr_sspnd_dec_nscheds(ErtsSchedTypeCounters *valp,
+ ErtsSchedType type)
{
ASSERT(schdlr_sspnd_get_nscheds(valp, type) > 0);
switch (type) {
case ERTS_SCHED_NORMAL:
- *valp -= ((Uint32) 1) << ERTS_SCHDLR_SSPND_S_SHIFT;
+ valp->normal--;
break;
+#ifdef ERTS_DIRTY_SCHEDULERS
case ERTS_SCHED_DIRTY_CPU:
- *valp -= ((Uint32) 1) << ERTS_SCHDLR_SSPND_DCS_SHIFT;
+ valp->dirty_cpu--;
break;
case ERTS_SCHED_DIRTY_IO:
- *valp -= ((Uint32) 1) << ERTS_SCHDLR_SSPND_DIS_SHIFT;
+ valp->dirty_io--;
break;
+#endif
default:
ERTS_INTERNAL_ERROR("Invalid scheduler type");
}
}
static ERTS_INLINE void
-schdlr_sspnd_inc_nscheds(Uint32 *valp, ErtsSchedType type)
+schdlr_sspnd_inc_nscheds(ErtsSchedTypeCounters *valp,
+ ErtsSchedType type)
{
switch (type) {
case ERTS_SCHED_NORMAL:
- ASSERT(schdlr_sspnd_get_nscheds(valp, type)
- < ERTS_MAX_NO_OF_SCHEDULERS-1);
- *valp += ((Uint32) 1) << ERTS_SCHDLR_SSPND_S_SHIFT;
+ valp->normal++;
break;
+#ifdef ERTS_DIRTY_SCHEDULERS
case ERTS_SCHED_DIRTY_CPU:
- ASSERT(schdlr_sspnd_get_nscheds(valp, type)
- < ERTS_MAX_NO_OF_DIRTY_CPU_SCHEDULERS);
- *valp += ((Uint32) 1) << ERTS_SCHDLR_SSPND_DCS_SHIFT;
+ valp->dirty_cpu++;
break;
case ERTS_SCHED_DIRTY_IO:
- ASSERT(schdlr_sspnd_get_nscheds(valp, type)
- < ERTS_MAX_NO_OF_DIRTY_IO_SCHEDULERS);
- *valp += ((Uint32) 1) << ERTS_SCHDLR_SSPND_DIS_SHIFT;
+ valp->dirty_io++;
break;
+#endif
default:
ERTS_INTERNAL_ERROR("Invalid scheduler type");
}
}
static ERTS_INLINE void
-schdlr_sspnd_set_nscheds(Uint32 *valp, ErtsSchedType type, Uint32 no)
+schdlr_sspnd_set_nscheds(ErtsSchedTypeCounters *valp,
+ ErtsSchedType type, Uint32 no)
{
- Uint32 val = *valp;
-
switch (type) {
case ERTS_SCHED_NORMAL:
- ASSERT(no > 0);
- val &= ~(((Uint32) ERTS_SCHDLR_SSPND_S_MASK)
- << ERTS_SCHDLR_SSPND_S_SHIFT);
- val |= (((no-1) & ((Uint32) ERTS_SCHDLR_SSPND_S_MASK))
- << ERTS_SCHDLR_SSPND_S_SHIFT);
+ valp->normal = no;
break;
+#ifdef ERTS_DIRTY_SCHEDULERS
case ERTS_SCHED_DIRTY_CPU:
- val &= ~(((Uint32) ERTS_SCHDLR_SSPND_DCS_MASK)
- << ERTS_SCHDLR_SSPND_DCS_SHIFT);
- val |= ((no & ((Uint32) ERTS_SCHDLR_SSPND_DCS_MASK))
- << ERTS_SCHDLR_SSPND_DCS_SHIFT);
+ valp->dirty_cpu = no;
break;
case ERTS_SCHED_DIRTY_IO:
- val &= ~(((Uint32) ERTS_SCHDLR_SSPND_DIS_MASK)
- << ERTS_SCHDLR_SSPND_DIS_SHIFT);
- val |= ((no & ((Uint32) ERTS_SCHDLR_SSPND_DIS_MASK))
- << ERTS_SCHDLR_SSPND_DIS_SHIFT);
+ valp->dirty_io = no;
break;
+#endif
default:
ERTS_INTERNAL_ERROR("Invalid scheduler type");
}
-
- *valp = val;
}
static struct {
@@ -809,15 +804,28 @@ erts_late_init_process(void)
}
+#define ERTS_SCHED_WTIME_IDLE ~((Uint64) 0)
+
static void
-init_sched_wall_time(ErtsSchedWallTime *swtp)
+init_sched_wall_time(ErtsSchedulerData *esdp, Uint64 time_stamp)
{
- swtp->need = erts_sched_balance_util;
- swtp->enabled = 0;
- swtp->start = 0;
- swtp->working.total = 0;
- swtp->working.start = 0;
- swtp->working.currently = 0;
+#ifdef ERTS_DIRTY_SCHEDULERS
+ if (esdp->type != ERTS_SCHED_NORMAL) {
+ erts_atomic32_init_nob(&esdp->sched_wall_time.u.mod, 0);
+ esdp->sched_wall_time.enabled = 1;
+ esdp->sched_wall_time.start = time_stamp;
+ esdp->sched_wall_time.working.total = 0;
+ esdp->sched_wall_time.working.start = ERTS_SCHED_WTIME_IDLE;
+ }
+ else
+#endif
+ {
+ esdp->sched_wall_time.u.need = erts_sched_balance_util;
+ esdp->sched_wall_time.enabled = 0;
+ esdp->sched_wall_time.start = 0;
+ esdp->sched_wall_time.working.total = 0;
+ esdp->sched_wall_time.working.start = 0;
+ }
}
static ERTS_INLINE Uint64
@@ -1008,31 +1016,145 @@ init_runq_sched_util(ErtsRunQueueSchedUtil *rqsu, int enabled)
#endif /* ERTS_HAVE_SCHED_UTIL_BALANCING_SUPPORT */
-static ERTS_INLINE void
+#ifdef ERTS_DIRTY_SCHEDULERS
+
+typedef struct {
+ Uint64 working;
+ Uint64 total;
+} ErtsDirtySchedWallTime;
+
+static void
+read_dirty_sched_wall_time(ErtsSchedulerData *esdp, ErtsDirtySchedWallTime *info)
+{
+ erts_aint32_t mod1;
+ Uint64 working, start, ts;
+
+ mod1 = erts_atomic32_read_nob(&esdp->sched_wall_time.u.mod);
+
+ while (1) {
+ erts_aint32_t mod2;
+
+ /* Spin until values are not written... */
+ while (1) {
+ if ((mod1 & 1) == 0)
+ break;
+ ERTS_SPIN_BODY;
+ mod1 = erts_atomic32_read_nob(&esdp->sched_wall_time.u.mod);
+ }
+
+ ERTS_THR_READ_MEMORY_BARRIER;
+
+ working = esdp->sched_wall_time.working.total;
+ start = esdp->sched_wall_time.working.start;
+
+ ERTS_THR_READ_MEMORY_BARRIER;
+
+ mod2 = erts_atomic32_read_nob(&esdp->sched_wall_time.u.mod);
+ if (mod1 == mod2)
+ break;
+ mod1 = mod2;
+ }
+
+ ts = sched_wall_time_ts();
+ ts -= esdp->sched_wall_time.start;
+
+ info->total = ts;
+
+ if (start == ERTS_SCHED_WTIME_IDLE || ts < start)
+ info->working = working;
+ else
+ info->working = working + (ts - start);
+
+ if (info->working > info->total)
+ info->working = info->total;
+}
+
+#endif
+
+#ifdef ERTS_SMP
+
+static void
+dirty_sched_wall_time_change(ErtsSchedulerData *esdp, int working)
+{
+ erts_aint32_t mod;
+ Uint64 ts = sched_wall_time_ts();
+
+ ts -= esdp->sched_wall_time.start;
+
+ /*
+ * This thread is the only thread writing in
+ * this sched_wall_time struct. We set 'mod' to
+ * an odd value while writing...
+ */
+ mod = erts_atomic32_read_dirty(&esdp->sched_wall_time.u.mod);
+ ASSERT((mod & 1) == 0);
+ mod++;
+
+ erts_atomic32_set_nob(&esdp->sched_wall_time.u.mod, mod);
+ ERTS_THR_WRITE_MEMORY_BARRIER;
+
+ if (working) {
+ ASSERT(esdp->sched_wall_time.working.start
+ == ERTS_SCHED_WTIME_IDLE);
+
+ esdp->sched_wall_time.working.start = ts;
+
+ }
+ else {
+ Uint64 total;
+
+ ASSERT(esdp->sched_wall_time.working.start
+ != ERTS_SCHED_WTIME_IDLE);
+
+ total = esdp->sched_wall_time.working.total;
+ total += ts - esdp->sched_wall_time.working.start;
+
+ esdp->sched_wall_time.working.total = total;
+ esdp->sched_wall_time.working.start = ERTS_SCHED_WTIME_IDLE;
+
+
+ }
+
+ ERTS_THR_WRITE_MEMORY_BARRIER;
+ mod++;
+ erts_atomic32_set_nob(&esdp->sched_wall_time.u.mod, mod);
+
+#if 0
+ if (!working) {
+ ERTS_MSACC_SET_STATE_M_X(ERTS_MSACC_STATE_BUSY_WAIT);
+ } else {
+ ERTS_MSACC_SET_STATE_M_X(ERTS_MSACC_STATE_OTHER);
+ }
+#endif
+}
+
+#endif /* ERTS_SMP */
+
+static void
sched_wall_time_change(ErtsSchedulerData *esdp, int working)
{
- if (esdp->sched_wall_time.need) {
+ if (esdp->sched_wall_time.u.need) {
Uint64 ts = sched_wall_time_ts();
#if ERTS_HAVE_SCHED_UTIL_BALANCING_SUPPORT
- update_avg_sched_util(esdp, ts, working);
+ update_avg_sched_util(esdp, ts, working);
#endif
if (esdp->sched_wall_time.enabled) {
if (working) {
-#ifdef DEBUG
- ASSERT(!esdp->sched_wall_time.working.currently);
- esdp->sched_wall_time.working.currently = 1;
-#endif
+ ASSERT(esdp->sched_wall_time.working.start
+ == ERTS_SCHED_WTIME_IDLE);
ts -= esdp->sched_wall_time.start;
esdp->sched_wall_time.working.start = ts;
}
else {
-#ifdef DEBUG
- ASSERT(esdp->sched_wall_time.working.currently);
- esdp->sched_wall_time.working.currently = 0;
-#endif
+ ASSERT(esdp->sched_wall_time.working.start
+ != ERTS_SCHED_WTIME_IDLE);
ts -= esdp->sched_wall_time.start;
ts -= esdp->sched_wall_time.working.start;
esdp->sched_wall_time.working.total += ts;
+#ifdef DEBUG
+ esdp->sched_wall_time.working.start
+ = ERTS_SCHED_WTIME_IDLE;
+#endif
}
}
}
@@ -1052,6 +1174,10 @@ typedef struct {
Eterm ref_heap[REF_THING_SIZE];
Uint req_sched;
erts_smp_atomic32_t refc;
+#ifdef ERTS_DIRTY_SCHEDULERS
+ int want_dirty_cpu;
+ int want_dirty_io;
+#endif
} ErtsSchedWallTimeReq;
typedef struct {
@@ -1073,6 +1199,7 @@ ERTS_SCHED_PREF_QUICK_ALLOC_IMPL(screq,
5,
ERTS_ALC_T_SYS_CHECK_REQ)
+
static void
reply_sched_wall_time(void *vswtrp)
{
@@ -1096,23 +1223,23 @@ reply_sched_wall_time(void *vswtrp)
#endif
if (swtrp->set) {
if (!swtrp->enable && esdp->sched_wall_time.enabled) {
- esdp->sched_wall_time.need = erts_sched_balance_util;
+ esdp->sched_wall_time.u.need = erts_sched_balance_util;
esdp->sched_wall_time.enabled = 0;
}
else if (swtrp->enable && !esdp->sched_wall_time.enabled) {
Uint64 ts = sched_wall_time_ts();
- esdp->sched_wall_time.need = 1;
+ esdp->sched_wall_time.u.need = 1;
esdp->sched_wall_time.enabled = 1;
esdp->sched_wall_time.start = ts;
esdp->sched_wall_time.working.total = 0;
esdp->sched_wall_time.working.start = 0;
- esdp->sched_wall_time.working.currently = 1;
}
}
if (esdp->sched_wall_time.enabled) {
Uint64 ts = sched_wall_time_ts();
- ASSERT(esdp->sched_wall_time.working.currently);
+ ASSERT(esdp->sched_wall_time.working.start
+ != ERTS_SCHED_WTIME_IDLE);
ts -= esdp->sched_wall_time.start;
total = ts;
ts -= esdp->sched_wall_time.working.start;
@@ -1123,30 +1250,117 @@ reply_sched_wall_time(void *vswtrp)
hpp = NULL;
szp = &sz;
- while (1) {
- if (hpp)
- ref_copy = STORE_NC(hpp, ohp, swtrp->ref);
- else
- *szp += REF_THING_SIZE;
+#ifdef ERTS_DIRTY_SCHEDULERS
+ if (esdp->sched_wall_time.enabled
+ && swtrp->req_sched == esdp->no
+ && (swtrp->want_dirty_cpu || swtrp->want_dirty_io)) {
+ /* Reply with info about this scheduler and all dirty schedulers... */
+ ErtsDirtySchedWallTime *dswt;
+ int ix, no_dirty_scheds, want_dcpu, want_dio, soffset;
+
+ want_dcpu = swtrp->want_dirty_cpu;
+ want_dio = swtrp->want_dirty_io;
+
+ no_dirty_scheds = 0;
+ if (want_dcpu)
+ no_dirty_scheds += erts_no_dirty_cpu_schedulers;
+ if (want_dio)
+ no_dirty_scheds += erts_no_dirty_io_schedulers;
+
+ ASSERT(no_dirty_scheds);
+
+ dswt = erts_alloc(ERTS_ALC_T_TMP,
+ sizeof(ErtsDirtySchedWallTime)
+ * no_dirty_scheds);
+
+ for (ix = 0; ix < no_dirty_scheds; ix++) {
+ ErtsSchedulerData *esdp;
+ if (want_dcpu && ix < erts_no_dirty_cpu_schedulers)
+ esdp = &erts_aligned_dirty_cpu_scheduler_data[ix].esd;
+ else {
+ int dio_ix = ix - erts_no_dirty_cpu_schedulers;
+ esdp = &erts_aligned_dirty_io_scheduler_data[dio_ix].esd;
+ }
+ read_dirty_sched_wall_time(esdp, &dswt[ix]);
+ }
- if (swtrp->set)
- msg = ref_copy;
- else {
- msg = (!esdp->sched_wall_time.enabled
- ? am_notsup
- : erts_bld_tuple(hpp, szp, 3,
- make_small(esdp->no),
- erts_bld_uint64(hpp, szp, working),
- erts_bld_uint64(hpp, szp, total)));
+ soffset = erts_no_schedulers + 1;
- msg = erts_bld_tuple(hpp, szp, 2, ref_copy, msg);
- }
- if (hpp)
- break;
+ if (!want_dcpu) {
+ ASSERT(want_dio);
+ soffset += erts_no_dirty_cpu_schedulers;
+ }
+
+ while (1) {
+ if (hpp)
+ ref_copy = STORE_NC(hpp, ohp, swtrp->ref);
+ else
+ *szp += REF_THING_SIZE;
+
+ ASSERT(!swtrp->set);
+
+ /* info about dirty schedulers... */
+ msg = NIL;
+ for (ix = no_dirty_scheds-1; ix >= 0; ix--) {
+ msg = erts_bld_cons(hpp, szp,
+ erts_bld_tuple(hpp, szp, 3,
+ make_small(ix+soffset),
+ erts_bld_uint64(hpp, szp,
+ dswt[ix].working),
+ erts_bld_uint64(hpp, szp,
+ dswt[ix].total)),
+ msg);
+ }
+ /* info about this scheduler... */
+ msg = erts_bld_cons(hpp, szp,
+ erts_bld_tuple(hpp, szp, 3,
+ make_small(esdp->no),
+ erts_bld_uint64(hpp, szp, working),
+ erts_bld_uint64(hpp, szp, total)),
+ msg);
+
+ msg = erts_bld_tuple(hpp, szp, 2, ref_copy, msg);
+
+ if (hpp)
+ break;
+
+ mp = erts_alloc_message_heap(rp, &rp_locks, sz, &hp, &ohp);
+ szp = NULL;
+ hpp = &hp;
+ }
+
+ erts_free(ERTS_ALC_T_TMP, dswt);
+ }
+ else
+#endif
+ {
+ /* Reply with info about this scheduler only... */
+
+ while (1) {
+ if (hpp)
+ ref_copy = STORE_NC(hpp, ohp, swtrp->ref);
+ else
+ *szp += REF_THING_SIZE;
+
+ if (swtrp->set)
+ msg = ref_copy;
+ else {
+ msg = (!esdp->sched_wall_time.enabled
+ ? am_undefined
+ : erts_bld_tuple(hpp, szp, 3,
+ make_small(esdp->no),
+ erts_bld_uint64(hpp, szp, working),
+ erts_bld_uint64(hpp, szp, total)));
+
+ msg = erts_bld_tuple(hpp, szp, 2, ref_copy, msg);
+ }
+ if (hpp)
+ break;
- mp = erts_alloc_message_heap(rp, &rp_locks, sz, &hp, &ohp);
- szp = NULL;
- hpp = &hp;
+ mp = erts_alloc_message_heap(rp, &rp_locks, sz, &hp, &ohp);
+ szp = NULL;
+ hpp = &hp;
+ }
}
erts_queue_message(rp, rp_locks, mp, msg, am_system);
@@ -1164,7 +1378,8 @@ reply_sched_wall_time(void *vswtrp)
}
Eterm
-erts_sched_wall_time_request(Process *c_p, int set, int enable)
+erts_sched_wall_time_request(Process *c_p, int set, int enable,
+ int want_dirty_cpu, int want_dirty_io)
{
ErtsSchedulerData *esdp = erts_proc_sched_data(c_p);
Eterm ref;
@@ -1186,6 +1401,10 @@ erts_sched_wall_time_request(Process *c_p, int set, int enable)
swtrp->proc = c_p;
swtrp->ref = STORE_NC(&hp, NULL, ref);
swtrp->req_sched = esdp->no;
+#ifdef ERTS_DIRTY_SCHEDULERS
+ swtrp->want_dirty_cpu = want_dirty_cpu;
+ swtrp->want_dirty_io = want_dirty_io;
+#endif
erts_smp_atomic32_init_nob(&swtrp->refc,
(erts_aint32_t) erts_no_schedulers);
@@ -2298,7 +2517,8 @@ static ERTS_INLINE erts_aint32_t
handle_reap_ports(ErtsAuxWorkData *awdp, erts_aint32_t aux_work, int waiting)
{
unset_aux_work_flags(awdp->ssi, ERTS_SSI_AUX_WORK_REAP_PORTS);
- awdp->esdp->run_queue->halt_in_progress = 1;
+ ERTS_RUNQ_FLGS_SET(awdp->esdp->run_queue, ERTS_RUNQ_FLG_HALTING);
+
if (erts_smp_atomic32_dec_read_acqb(&erts_halt_progress) == 0) {
int i, max = erts_ptab_max(&erts_port);
erts_smp_atomic32_set_nob(&erts_halt_progress, 1);
@@ -2838,11 +3058,12 @@ static erts_aint32_t
sched_prep_spin_wait(ErtsSchedulerSleepInfo *ssi)
{
erts_aint32_t oflgs;
- erts_aint32_t nflgs = (ERTS_SSI_FLG_SLEEPING
- | ERTS_SSI_FLG_WAITING);
+ erts_aint32_t nflgs;
erts_aint32_t xflgs = 0;
do {
+ nflgs = (xflgs & ERTS_SSI_FLG_MSB_EXEC);
+ nflgs |= ERTS_SSI_FLG_SLEEPING|ERTS_SSI_FLG_WAITING;
oflgs = erts_smp_atomic32_cmpxchg_acqb(&ssi->flags, nflgs, xflgs);
if (oflgs == xflgs)
return nflgs;
@@ -2864,7 +3085,7 @@ sched_prep_cont_spin_wait(ErtsSchedulerSleepInfo *ssi)
if (oflgs == xflgs)
return nflgs;
xflgs = oflgs;
- nflgs |= oflgs & ERTS_SSI_FLG_SUSPENDED;
+ nflgs |= oflgs & (ERTS_SSI_FLG_SUSPENDED|ERTS_SSI_FLG_MSB_EXEC);
} while (oflgs & ERTS_SSI_FLG_WAITING);
return oflgs;
}
@@ -2914,7 +3135,7 @@ sched_set_sleeptype(ErtsSchedulerSleepInfo *ssi, erts_aint32_t sleep_type)
return oflgs;
}
xflgs = oflgs;
- nflgs |= oflgs & ERTS_SSI_FLG_SUSPENDED;
+ nflgs |= oflgs & (ERTS_SSI_FLG_SUSPENDED|ERTS_SSI_FLG_MSB_EXEC);
}
}
@@ -3049,6 +3270,8 @@ aux_thread(void *unused)
return NULL;
}
+static void suspend_scheduler(ErtsSchedulerData *esdp);
+
#endif /* ERTS_SMP */
static void
@@ -3107,8 +3330,10 @@ scheduler_wait(int *fcalls, ErtsSchedulerData *esdp, ErtsRunQueue *rq)
tse_wait:
- if (!ERTS_SCHEDULER_IS_DIRTY(esdp) && thr_prgr_active != working)
- sched_wall_time_change(esdp, thr_prgr_active);
+ if (ERTS_SCHEDULER_IS_DIRTY(esdp))
+ dirty_sched_wall_time_change(esdp, working = 0);
+ else if (thr_prgr_active != working)
+ sched_wall_time_change(esdp, working = thr_prgr_active);
while (1) {
ErtsMonotonicTime current_time = 0;
@@ -3214,13 +3439,17 @@ scheduler_wait(int *fcalls, ErtsSchedulerData *esdp, ErtsRunQueue *rq)
}
- if (flgs & ~ERTS_SSI_FLG_SUSPENDED)
- erts_smp_atomic32_read_band_nob(&ssi->flags, ERTS_SSI_FLG_SUSPENDED);
+ if (flgs & ~(ERTS_SSI_FLG_SUSPENDED|ERTS_SSI_FLG_MSB_EXEC))
+ erts_smp_atomic32_read_band_nob(&ssi->flags,
+ (ERTS_SSI_FLG_SUSPENDED
+ | ERTS_SSI_FLG_MSB_EXEC));
- if (!ERTS_SCHEDULER_IS_DIRTY(esdp) && !thr_prgr_active) {
+ if (ERTS_SCHEDULER_IS_DIRTY(esdp))
+ dirty_sched_wall_time_change(esdp, working = 1);
+ else if (!thr_prgr_active) {
erts_thr_progress_active(esdp, thr_prgr_active = 1);
sched_wall_time_change(esdp, 1);
- }
+ }
erts_smp_runq_lock(rq);
sched_active(esdp->no, rq);
@@ -3414,8 +3643,10 @@ scheduler_wait(int *fcalls, ErtsSchedulerData *esdp, ErtsRunQueue *rq)
erts_smp_runq_lock(rq);
}
clear_sys_scheduling();
- if (flgs & ~ERTS_SSI_FLG_SUSPENDED)
- erts_smp_atomic32_read_band_nob(&ssi->flags, ERTS_SSI_FLG_SUSPENDED);
+ if (flgs & ~(ERTS_SSI_FLG_SUSPENDED|ERTS_SSI_FLG_MSB_EXEC))
+ erts_smp_atomic32_read_band_nob(&ssi->flags,
+ (ERTS_SSI_FLG_SUSPENDED
+ | ERTS_SSI_FLG_MSB_EXEC));
#endif
if (!working)
sched_wall_time_change(esdp, working = 1);
@@ -3438,17 +3669,54 @@ ssi_flags_set_wake(ErtsSchedulerSleepInfo *ssi)
oflgs = erts_smp_atomic32_cmpxchg_relb(&ssi->flags, nflgs, xflgs);
if (oflgs == xflgs)
return oflgs;
- nflgs = oflgs & ERTS_SSI_FLG_SUSPENDED;
+ nflgs = oflgs & (ERTS_SSI_FLG_SUSPENDED|ERTS_SSI_FLG_MSB_EXEC);
xflgs = oflgs;
}
}
+static ERTS_INLINE void
+ssi_wake(ErtsSchedulerSleepInfo *ssi)
+{
+ erts_sched_finish_poke(ssi, ssi_flags_set_wake(ssi));
+}
+
+#ifdef ERTS_DIRTY_SCHEDULERS
+
static void
-wake_scheduler(ErtsRunQueue *rq)
+dcpu_sched_ix_suspend_wake(Uint ix)
{
- ErtsSchedulerSleepInfo *ssi;
- erts_aint32_t flgs;
+ ErtsSchedulerSleepInfo* ssi = ERTS_DIRTY_CPU_SCHED_SLEEP_INFO_IX(ix);
+ erts_smp_atomic32_read_bor_nob(&ssi->flags, ERTS_SSI_FLG_SUSPENDED);
+ ssi_wake(ssi);
+}
+
+static void
+dio_sched_ix_suspend_wake(Uint ix)
+{
+ ErtsSchedulerSleepInfo* ssi = ERTS_DIRTY_IO_SCHED_SLEEP_INFO_IX(ix);
+ erts_smp_atomic32_read_bor_nob(&ssi->flags, ERTS_SSI_FLG_SUSPENDED);
+ ssi_wake(ssi);
+}
+static void
+dcpu_sched_ix_wake(Uint ix)
+{
+ ssi_wake(ERTS_DIRTY_CPU_SCHED_SLEEP_INFO_IX(ix));
+}
+
+#if 0
+static void
+dio_sched_ix_wake(Uint ix)
+{
+ ssi_wake(ERTS_DIRTY_IO_SCHED_SLEEP_INFO_IX(ix));
+}
+#endif
+
+#endif
+
+static void
+wake_scheduler(ErtsRunQueue *rq)
+{
/*
* The unlocked run queue is not strictly necessary
* from a thread safety or deadlock prevention
@@ -3460,10 +3728,7 @@ wake_scheduler(ErtsRunQueue *rq)
ERTS_SMP_LC_ASSERT(!erts_smp_lc_runq_is_locked(rq)
|| ERTS_RUNQ_IX_IS_DIRTY(rq->ix));
- ssi = rq->scheduler->ssi;
-
- flgs = ssi_flags_set_wake(ssi);
- erts_sched_finish_poke(ssi, flgs);
+ ssi_wake(rq->scheduler->ssi);
}
#ifdef ERTS_DIRTY_SCHEDULERS
@@ -3510,6 +3775,13 @@ wake_dirty_schedulers(ErtsRunQueue *rq, int one)
} while (ssi);
}
}
+
+static void
+wake_dirty_scheduler(ErtsRunQueue *rq)
+{
+ wake_dirty_schedulers(rq, 1);
+}
+
#endif
#define ERTS_NO_USED_RUNQS_SHIFT 16
@@ -3650,7 +3922,7 @@ smp_notify_inc_runq(ErtsRunQueue *runq)
if (runq) {
#ifdef ERTS_DIRTY_SCHEDULERS
if (ERTS_RUNQ_IX_IS_DIRTY(runq->ix))
- wake_dirty_schedulers(runq, 1);
+ wake_dirty_scheduler(runq);
else
#endif
wake_scheduler(runq);
@@ -3954,8 +4226,7 @@ suspend_run_queue(ErtsRunQueue *rq)
wake_scheduler(rq);
}
-static void scheduler_ix_resume_wake(Uint ix);
-static void scheduler_ssi_resume_wake(ErtsSchedulerSleepInfo *ssi);
+static void nrml_sched_ix_resume_wake(Uint ix);
static ERTS_INLINE void
resume_run_queue(ErtsRunQueue *rq)
@@ -3963,16 +4234,19 @@ resume_run_queue(ErtsRunQueue *rq)
int pix;
Uint32 oflgs;
+ ASSERT(!ERTS_RUNQ_IX_IS_DIRTY(rq->ix));
+
erts_smp_runq_lock(rq);
oflgs = ERTS_RUNQ_FLGS_READ_BSET(rq,
(ERTS_RUNQ_FLG_OUT_OF_WORK
| ERTS_RUNQ_FLG_HALFTIME_OUT_OF_WORK
- | ERTS_RUNQ_FLG_SUSPENDED),
+ | ERTS_RUNQ_FLG_SUSPENDED
+ | ERTS_RUNQ_FLG_MSB_EXEC),
(ERTS_RUNQ_FLG_OUT_OF_WORK
| ERTS_RUNQ_FLG_HALFTIME_OUT_OF_WORK));
- if (oflgs & ERTS_RUNQ_FLG_SUSPENDED) {
+ if (oflgs & (ERTS_RUNQ_FLG_SUSPENDED|ERTS_RUNQ_FLG_MSB_EXEC)) {
erts_aint32_t len;
rq->check_balance_reds = ERTS_RUNQ_CALL_CHECK_BALANCE_REDS;
@@ -3991,10 +4265,7 @@ resume_run_queue(ErtsRunQueue *rq)
erts_smp_runq_unlock(rq);
-#ifdef ERTS_DIRTY_SCHEDULERS
- if (!ERTS_RUNQ_IX_IS_DIRTY(rq->ix))
-#endif
- scheduler_ix_resume_wake(rq->ix);
+ nrml_sched_ix_resume_wake(rq->ix);
}
typedef struct {
@@ -4052,28 +4323,22 @@ evacuate_run_queue(ErtsRunQueue *rq,
int prio_q;
ErtsRunQueue *to_rq;
ErtsMigrationPaths *mps;
- ErtsMigrationPath *mp = NULL;
+ ErtsMigrationPath *mp;
ERTS_SMP_LC_ASSERT(erts_smp_lc_runq_is_locked(rq));
(void) ERTS_RUNQ_FLGS_UNSET(rq, ERTS_RUNQ_FLG_PROTECTED);
-#ifdef ERTS_DIRTY_SCHEDULERS
- if (!ERTS_RUNQ_IX_IS_DIRTY(rq->ix))
-#endif
- {
- mps = erts_get_migration_paths_managed();
- mp = &mps->mpath[rq->ix];
- }
+ ASSERT(!ERTS_RUNQ_IX_IS_DIRTY(rq->ix));
+
+ mps = erts_get_migration_paths_managed();
+ mp = &mps->mpath[rq->ix];
/* Evacuate scheduled misc ops */
if (rq->misc.start) {
ErtsMiscOpList *start, *end;
-#ifdef ERTS_DIRTY_SCHEDULERS
- ASSERT(!ERTS_RUNQ_IX_IS_DIRTY(rq->ix));
-#endif
to_rq = mp->misc_evac_runq;
if (!to_rq)
return;
@@ -4082,6 +4347,7 @@ evacuate_run_queue(ErtsRunQueue *rq,
end = rq->misc.end;
rq->misc.start = NULL;
rq->misc.end = NULL;
+ ERTS_RUNQ_FLGS_UNSET_NOB(rq, ERTS_RUNQ_FLG_MISC_OP);
erts_smp_runq_unlock(rq);
erts_smp_runq_lock(to_rq);
@@ -4102,9 +4368,6 @@ evacuate_run_queue(ErtsRunQueue *rq,
if (rq->ports.start) {
Port *prt;
-#ifdef ERTS_DIRTY_SCHEDULERS
- ASSERT(!ERTS_RUNQ_IX_IS_DIRTY(rq->ix));
-#endif
to_rq = mp->prio[ERTS_PORT_PRIO_LEVEL].runq;
if (!to_rq)
return;
@@ -4142,15 +4405,10 @@ evacuate_run_queue(ErtsRunQueue *rq,
int notify = 0;
to_rq = NULL;
-#ifdef ERTS_DIRTY_SCHEDULERS
- if (!ERTS_RUNQ_IX_IS_DIRTY(rq->ix))
-#endif
- {
- if (!mp->prio[prio_q].runq)
- return;
- if (prio_q == PRIORITY_NORMAL && !mp->prio[PRIORITY_LOW].runq)
- return;
- }
+ if (!mp->prio[prio_q].runq)
+ return;
+ if (prio_q == PRIORITY_NORMAL && !mp->prio[PRIORITY_LOW].runq)
+ return;
proc = dequeue_process(rq, prio_q, &state);
while (proc) {
@@ -4206,11 +4464,6 @@ evacuate_run_queue(ErtsRunQueue *rq,
goto handle_next_proc;
}
-#ifdef ERTS_DIRTY_SCHEDULERS
- if (ERTS_RUNQ_IX_IS_DIRTY(rq->ix))
- clear_proc_dirty_queue_bit(real_proc, rq, qbit);
-#endif
-
if (ERTS_PSFLG_BOUND & real_state) {
/* Bound processes get stuck here... */
proc->next = NULL;
@@ -4224,16 +4477,7 @@ evacuate_run_queue(ErtsRunQueue *rq,
int prio = (int) ERTS_PSFLGS_GET_PRQ_PRIO(state);
erts_smp_runq_unlock(rq);
-#ifdef ERTS_DIRTY_SCHEDULERS
- if (ERTS_RUNQ_IX_IS_DIRTY(rq->ix))
- /*
- * dirty run queues evacuate only to run
- * queue 0 during multi-scheduling blocking
- */
- to_rq = ERTS_RUNQ_IX(0);
- else
-#endif
- to_rq = mp->prio[prio].runq;
+ to_rq = mp->prio[prio].runq;
RUNQ_SET_RQ(&proc->run_queue, to_rq);
erts_smp_runq_lock(to_rq);
@@ -4268,7 +4512,7 @@ try_steal_task_from_victim(ErtsRunQueue *rq, int *rq_lockedp, ErtsRunQueue *vrq,
erts_smp_runq_lock(vrq);
- if (rq->halt_in_progress)
+ if (ERTS_RUNQ_FLGS_GET_NOB(rq) & ERTS_RUNQ_FLG_HALTING)
goto no_procs;
/*
@@ -4367,8 +4611,7 @@ check_possible_steal_victim(ErtsRunQueue *rq, int *rq_lockedp, int vix)
{
ErtsRunQueue *vrq = ERTS_RUNQ_IX(vix);
Uint32 flags = ERTS_RUNQ_FLGS_GET(vrq);
- if ((flags & (ERTS_RUNQ_FLG_NONEMPTY
- | ERTS_RUNQ_FLG_PROTECTED)) == ERTS_RUNQ_FLG_NONEMPTY)
+ if (runq_got_work_to_execute_flags(flags) & (!(flags & ERTS_RUNQ_FLG_PROTECTED)))
return try_steal_task_from_victim(rq, rq_lockedp, vrq, flags);
else
return 0;
@@ -4436,11 +4679,9 @@ try_steal_task(ErtsRunQueue *rq)
if (!rq_locked)
erts_smp_runq_lock(rq);
- if (!res)
- res = rq->halt_in_progress ?
- !ERTS_EMPTY_RUNQ_PORTS(rq) : !ERTS_EMPTY_RUNQ(rq);
-
- return res;
+ if (res)
+ return res;
+ return runq_got_work_to_execute(rq);
}
/* Run queue balancing */
@@ -5352,7 +5593,7 @@ wakeup_other_check(ErtsRunQueue *rq, Uint32 flags)
#ifdef ERTS_DIRTY_SCHEDULERS
if (ERTS_RUNQ_IX_IS_DIRTY(rq->ix)) {
if (rq->waiting) {
- wake_dirty_schedulers(rq, 1);
+ wake_dirty_scheduler(rq);
}
} else
#endif
@@ -5699,7 +5940,8 @@ init_scheduler_data(ErtsSchedulerData* esdp, int num,
ErtsSchedulerSleepInfo* ssi,
ErtsRunQueue* runq,
char** daww_ptr, size_t daww_sz,
- Process *shadow_proc)
+ Process *shadow_proc,
+ Uint64 time_stamp)
{
esdp->timer_wheel = NULL;
#ifdef ERTS_SMP
@@ -5715,13 +5957,29 @@ init_scheduler_data(ErtsSchedulerData* esdp, int num,
erts_alloc_permanent_cache_aligned(ERTS_ALC_T_BEAM_REGISTER,
MAX_REG * sizeof(FloatDef));
#ifdef ERTS_DIRTY_SCHEDULERS
+ esdp->run_queue = runq;
if (ERTS_RUNQ_IX_IS_DIRTY(runq->ix)) {
esdp->no = 0;
+ if (runq == ERTS_DIRTY_CPU_RUNQ)
+ esdp->type = ERTS_SCHED_DIRTY_CPU;
+ else {
+ ASSERT(runq == ERTS_DIRTY_IO_RUNQ);
+ esdp->type = ERTS_SCHED_DIRTY_IO;
+ }
ERTS_DIRTY_SCHEDULER_NO(esdp) = (Uint) num;
+ if (num == 1) {
+ /*
+ * Multi-scheduling block functionality depends
+ * on finding dirty scheduler number 1 here...
+ */
+ runq->scheduler = esdp;
+ }
}
else {
+ esdp->type = ERTS_SCHED_NORMAL;
esdp->no = (Uint) num;
ERTS_DIRTY_SCHEDULER_NO(esdp) = 0;
+ runq->scheduler = esdp;
}
esdp->dirty_shadow_process = shadow_proc;
if (shadow_proc) {
@@ -5733,6 +5991,8 @@ init_scheduler_data(ErtsSchedulerData* esdp, int num,
shadow_proc->static_flags = ERTS_STC_FLG_SHADOW_PROC;
}
#else
+ runq->scheduler = esdp;
+ esdp->run_queue = runq;
esdp->no = (Uint) num;
#endif
@@ -5745,9 +6005,6 @@ init_scheduler_data(ErtsSchedulerData* esdp, int num,
erts_init_atom_cache_map(&esdp->atom_cache_map);
- esdp->run_queue = runq;
- esdp->run_queue->scheduler = esdp;
-
esdp->last_monotonic_time = 0;
esdp->check_time_reds = 0;
@@ -5766,7 +6023,7 @@ init_scheduler_data(ErtsSchedulerData* esdp, int num,
esdp->reductions = 0;
- init_sched_wall_time(&esdp->sched_wall_time);
+ init_sched_wall_time(esdp, time_stamp);
erts_port_task_handle_init(&esdp->nosuspend_port_task_handle);
}
@@ -5863,7 +6120,6 @@ erts_init_scheduling(int no_schedulers, int no_schedulers_online
erts_smp_atomic32_set_nob(&rq->len, 0);
rq->wakeup_other = 0;
rq->wakeup_other_reds = 0;
- rq->halt_in_progress = 0;
rq->procs.pending_exiters = NULL;
rq->procs.context_switches = 0;
@@ -5981,17 +6237,18 @@ erts_init_scheduling(int no_schedulers, int no_schedulers_online
erts_aligned_scheduler_data =
erts_alloc_permanent_cache_aligned(ERTS_ALC_T_SCHDLR_DATA,
- n*sizeof(ErtsAlignedSchedulerData));
+ n*sizeof(ErtsAlignedSchedulerData));
for (ix = 0; ix < n; ix++) {
ErtsSchedulerData *esdp = ERTS_SCHEDULER_IX(ix);
init_scheduler_data(esdp, ix+1, ERTS_SCHED_SLEEP_INFO_IX(ix),
ERTS_RUNQ_IX(ix), &daww_ptr, daww_sz,
- NULL);
+ NULL, 0);
}
#ifdef ERTS_DIRTY_SCHEDULERS
{
+ Uint64 ts = sched_wall_time_ts();
int dirty_scheds = no_dirty_cpu_schedulers + no_dirty_io_schedulers;
int adspix = 0;
ErtsAlignedDirtyShadowProcess *adsp =
@@ -6011,13 +6268,13 @@ erts_init_scheduling(int no_schedulers, int no_schedulers_online
ErtsSchedulerData *esdp = ERTS_DIRTY_CPU_SCHEDULER_IX(ix);
init_scheduler_data(esdp, ix+1, ERTS_DIRTY_CPU_SCHED_SLEEP_INFO_IX(ix),
ERTS_DIRTY_CPU_RUNQ, NULL, 0,
- &adsp[adspix++].dsp);
+ &adsp[adspix++].dsp, ts);
}
for (ix = 0; ix < no_dirty_io_schedulers; ix++) {
ErtsSchedulerData *esdp = ERTS_DIRTY_IO_SCHEDULER_IX(ix);
init_scheduler_data(esdp, ix+1, ERTS_DIRTY_IO_SCHED_SLEEP_INFO_IX(ix),
ERTS_DIRTY_IO_RUNQ, NULL, 0,
- &adsp[adspix++].dsp);
+ &adsp[adspix++].dsp, ts);
}
}
#endif
@@ -6444,10 +6701,30 @@ schedule_out_process(ErtsRunQueue *c_rq, erts_aint32_t state, Process *p,
int enqueue; /* < 0 -> use proxy */
ErtsRunQueue* runq;
- if (is_normal_sched)
- running_flgs = ERTS_PSFLG_RUNNING|ERTS_PSFLG_RUNNING_SYS;
- else
+ if (!is_normal_sched)
running_flgs = ERTS_PSFLG_DIRTY_RUNNING|ERTS_PSFLG_DIRTY_RUNNING_SYS;
+ else {
+ running_flgs = ERTS_PSFLG_RUNNING|ERTS_PSFLG_RUNNING_SYS;
+#ifdef ERTS_DIRTY_SCHEDULERS
+ if (state & ERTS_PSFLG_DIRTY_ACTIVE_SYS
+ && (p->flags & (F_DELAY_GC|F_DISABLE_GC))) {
+ /*
+ * Delay dirty GC; will be enabled automatically
+ * again by next GC...
+ */
+
+ /*
+ * No normal execution until dirty CLA or hibernat has
+ * been handled...
+ */
+ ASSERT(!(p->flags & (F_DIRTY_CLA | F_DIRTY_GC_HIBERNATE)));
+
+ state = erts_smp_atomic32_read_band_nob(&p->state,
+ ~ERTS_PSFLG_DIRTY_ACTIVE_SYS);
+ state &= ~ERTS_PSFLG_DIRTY_ACTIVE_SYS;
+ }
+#endif
+ }
a = state;
@@ -6959,15 +7236,8 @@ resume_process(Process *p, ErtsProcLocks locks)
#ifdef ERTS_SMP
-static void
-scheduler_ix_resume_wake(Uint ix)
-{
- ErtsSchedulerSleepInfo *ssi = ERTS_SCHED_SLEEP_INFO_IX(ix);
- scheduler_ssi_resume_wake(ssi);
-}
-
-static void
-scheduler_ssi_resume_wake(ErtsSchedulerSleepInfo *ssi)
+static ERTS_INLINE void
+sched_resume_wake__(ErtsSchedulerSleepInfo *ssi)
{
erts_aint32_t xflgs = (ERTS_SSI_FLG_SLEEPING
| ERTS_SSI_FLG_TSE_SLEEPING
@@ -6981,9 +7251,31 @@ scheduler_ssi_resume_wake(ErtsSchedulerSleepInfo *ssi)
break;
}
xflgs = oflgs;
- } while (oflgs & ERTS_SSI_FLG_SUSPENDED);
+ } while (oflgs & (ERTS_SSI_FLG_MSB_EXEC|ERTS_SSI_FLG_SUSPENDED));
}
+static void
+nrml_sched_ix_resume_wake(Uint ix)
+{
+ sched_resume_wake__(ERTS_SCHED_SLEEP_INFO_IX(ix));
+}
+
+#ifdef ERTS_DIRTY_SCHEDULERS
+
+static void
+dcpu_sched_ix_resume_wake(Uint ix)
+{
+ sched_resume_wake__(ERTS_DIRTY_CPU_SCHED_SLEEP_INFO_IX(ix));
+}
+
+static void
+dio_sched_ix_resume_wake(Uint ix)
+{
+ sched_resume_wake__(ERTS_DIRTY_IO_SCHED_SLEEP_INFO_IX(ix));
+}
+
+#endif
+
static erts_aint32_t
sched_prep_spin_suspended(ErtsSchedulerSleepInfo *ssi, erts_aint32_t xpct)
{
@@ -7063,9 +7355,18 @@ static void
init_scheduler_suspend(void)
{
erts_smp_mtx_init(&schdlr_sspnd.mtx, "schdlr_sspnd");
- schdlr_sspnd.online = ERTS_SCHDLR_SSPND_MAKE_NSCHEDS_VAL(1, 0, 0);
- schdlr_sspnd.curr_online = ERTS_SCHDLR_SSPND_MAKE_NSCHEDS_VAL(1, 0, 0);
- schdlr_sspnd.active = ERTS_SCHDLR_SSPND_MAKE_NSCHEDS_VAL(1, 0, 0);
+ schdlr_sspnd.online.normal = 1;
+ schdlr_sspnd.curr_online.normal = 1;
+ schdlr_sspnd.active.normal = 1;
+#ifdef ERTS_DIRTY_SCHEDULERS
+ schdlr_sspnd.online.dirty_cpu = 0;
+ schdlr_sspnd.curr_online.dirty_cpu = 0;
+ schdlr_sspnd.active.dirty_cpu = 0;
+ schdlr_sspnd.online.dirty_io = 0;
+ schdlr_sspnd.curr_online.dirty_io = 0;
+ schdlr_sspnd.active.dirty_io = 0;
+ schdlr_sspnd.last_msb_dirty_type = ERTS_SCHED_DIRTY_IO;
+#endif
erts_smp_atomic32_init_nob(&schdlr_sspnd.changing, 0);
schdlr_sspnd.chngq = NULL;
schdlr_sspnd.changer = am_false;
@@ -7126,6 +7427,252 @@ schdlr_sspnd_resume_procs(ErtsSchedType sched_type,
}
}
+#ifdef ERTS_DIRTY_SCHEDULERS
+
+static ERTS_INLINE int
+have_dirty_work(void)
+{
+ return !(ERTS_EMPTY_RUNQ(ERTS_DIRTY_CPU_RUNQ)
+ | ERTS_EMPTY_RUNQ(ERTS_DIRTY_IO_RUNQ));
+}
+
+#define ERTS_MSB_NONE_PRIO_BIT PORT_BIT
+
+static ERTS_INLINE Uint32
+msb_runq_prio_bit(Uint32 flgs)
+{
+ int pbit;
+
+ pbit = (int) (flgs & ERTS_RUNQ_FLGS_PROCS_QMASK);
+ if (flgs & PORT_BIT) {
+ /* rate ports as proc prio high */
+ pbit |= HIGH_BIT;
+ }
+ if (flgs & ERTS_RUNQ_FLG_MISC_OP) {
+ /* rate misc ops as proc prio normal */
+ pbit |= NORMAL_BIT;
+ }
+ if (flgs & LOW_BIT) {
+ /* rate low prio as normal (avoid starvation) */
+ pbit |= NORMAL_BIT;
+ }
+ if (!pbit)
+ pbit = (int) ERTS_MSB_NONE_PRIO_BIT;
+ else
+ pbit &= -pbit; /* least significant bit set... */
+ ASSERT(pbit);
+
+ /* High prio low value; low prio high value... */
+ return (Uint32) pbit;
+}
+
+static ERTS_INLINE void
+msb_runq_prio_bits(Uint32 *nrmlp, Uint32 *dcpup, Uint32 *diop)
+{
+ Uint32 flgs = ERTS_RUNQ_FLGS_GET(ERTS_RUNQ_IX(0));
+ if (flgs & ERTS_RUNQ_FLG_HALTING) {
+ /*
+ * Emulator is halting; only execute port jobs
+ * on normal scheduler. Ensure that we switch
+ * to the normal scheduler.
+ */
+ *nrmlp = HIGH_BIT;
+ *dcpup = ERTS_MSB_NONE_PRIO_BIT;
+ *diop = ERTS_MSB_NONE_PRIO_BIT;
+ }
+ else {
+ *nrmlp = msb_runq_prio_bit(flgs);
+
+ flgs = ERTS_RUNQ_FLGS_GET(ERTS_DIRTY_CPU_RUNQ);
+ *dcpup = msb_runq_prio_bit(flgs);
+
+ flgs = ERTS_RUNQ_FLGS_GET(ERTS_DIRTY_IO_RUNQ);
+ *diop = msb_runq_prio_bit(flgs);
+ }
+}
+
+static int
+msb_scheduler_type_switch(ErtsSchedType sched_type,
+ ErtsSchedulerData *esdp,
+ long no)
+{
+ Uint32 nrml_prio, dcpu_prio, dio_prio;
+ ErtsSchedType exec_type;
+ ErtsRunQueue *exec_rq;
+#ifdef DEBUG
+ erts_aint32_t dbg_val;
+#endif
+
+ ASSERT(schdlr_sspnd.msb.ongoing);
+
+ /*
+ * This function determines how to switch
+ * between scheduler types when multi-scheduling
+ * is blocked.
+ *
+ * If no dirty work exist, we always select
+ * execution of normal scheduler. If nothing
+ * executes, normal scheduler 1 should be waiting
+ * in sys_schedule(), otherwise we cannot react
+ * on I/O events.
+ *
+ * We unconditionally switch back to normal
+ * scheduler after executing dirty in order to
+ * make sure we check for I/O...
+ */
+
+ msb_runq_prio_bits(&nrml_prio, &dcpu_prio, &dio_prio);
+
+ exec_type = ERTS_SCHED_NORMAL;
+ if (sched_type == ERTS_SCHED_NORMAL) {
+
+ /*
+ * Check priorities of work in the
+ * different run-queues and determine
+ * run-queue with highest prio job...
+ */
+
+ if ((dcpu_prio == ERTS_MSB_NONE_PRIO_BIT)
+ & (dio_prio == ERTS_MSB_NONE_PRIO_BIT)) {
+ /*
+ * No dirty work exist; continue on normal
+ * scheduler...
+ */
+ return 0;
+ }
+
+ if (dcpu_prio < nrml_prio) {
+ exec_type = ERTS_SCHED_DIRTY_CPU;
+ if (dio_prio < dcpu_prio)
+ exec_type = ERTS_SCHED_DIRTY_IO;
+ }
+ else {
+ if (dio_prio < nrml_prio)
+ exec_type = ERTS_SCHED_DIRTY_IO;
+ }
+
+ /*
+ * Make sure to alternate between dirty types
+ * inbetween normal execution if highest
+ * priorities are equal.
+ */
+
+ if (exec_type == ERTS_SCHED_NORMAL) {
+ if (dcpu_prio == nrml_prio)
+ exec_type = ERTS_SCHED_DIRTY_CPU;
+ else if (dio_prio == nrml_prio)
+ exec_type = ERTS_SCHED_DIRTY_IO;
+ else {
+ /*
+ * Normal work has higher prio than
+ * dirty work; continue on normal
+ * scheduler...
+ */
+ return 0;
+ }
+ }
+
+ ASSERT(exec_type != ERTS_SCHED_NORMAL);
+ if (dio_prio == dcpu_prio) {
+ /* Alter between dirty types... */
+ if (schdlr_sspnd.last_msb_dirty_type == ERTS_SCHED_DIRTY_IO)
+ exec_type = ERTS_SCHED_DIRTY_CPU;
+ else
+ exec_type = ERTS_SCHED_DIRTY_IO;
+ }
+ }
+
+ ASSERT(sched_type != exec_type);
+
+ if (exec_type != ERTS_SCHED_NORMAL)
+ schdlr_sspnd.last_msb_dirty_type = exec_type;
+ else {
+ erts_aint32_t calls;
+ /*
+ * Going back to normal scheduler after
+ * dirty execution; make sure it will check
+ * for I/O...
+ */
+ if (ERTS_USE_MODIFIED_TIMING())
+ calls = ERTS_MODIFIED_TIMING_INPUT_REDS + 1;
+ else
+ calls = INPUT_REDUCTIONS + 1;
+ erts_smp_atomic32_set_nob(&function_calls, calls);
+
+ if ((nrml_prio == ERTS_MSB_NONE_PRIO_BIT)
+ & ((dcpu_prio != ERTS_MSB_NONE_PRIO_BIT)
+ | (dio_prio != ERTS_MSB_NONE_PRIO_BIT))) {
+ /*
+ * We have dirty work, but an empty
+ * normal run-queue.
+ *
+ * Since the normal run-queue is
+ * empty, the normal scheduler will
+ * go to sleep when selected for
+ * execution. We have dirty work to
+ * do, so we only want it to check
+ * I/O, and then come back here and
+ * switch to dirty execution.
+ *
+ * To prevent the scheduler from going
+ * to sleep we trick it into believing
+ * it has work to do...
+ */
+ ERTS_RUNQ_FLGS_SET_NOB(ERTS_RUNQ_IX(0),
+ ERTS_RUNQ_FLG_MISC_OP);
+ }
+ }
+
+ /*
+ * Suspend this scheduler and wake up scheduler
+ * number one of another type...
+ */
+#ifdef DEBUG
+ dbg_val =
+#else
+ (void)
+#endif
+ erts_smp_atomic32_read_bset_mb(&esdp->ssi->flags,
+ (ERTS_SSI_FLG_SUSPENDED
+ | ERTS_SSI_FLG_MSB_EXEC),
+ ERTS_SSI_FLG_SUSPENDED);
+ ASSERT(dbg_val & ERTS_SSI_FLG_MSB_EXEC);
+
+ switch (exec_type) {
+ case ERTS_SCHED_NORMAL:
+ exec_rq = ERTS_RUNQ_IX(0);
+ break;
+ case ERTS_SCHED_DIRTY_CPU:
+ exec_rq = ERTS_DIRTY_CPU_RUNQ;
+ break;
+ case ERTS_SCHED_DIRTY_IO:
+ exec_rq = ERTS_DIRTY_IO_RUNQ;
+ break;
+ default:
+ ERTS_INTERNAL_ERROR("Invalid scheduler type");
+ exec_rq = NULL;
+ break;
+ }
+
+#ifdef DEBUG
+ dbg_val =
+#else
+ (void)
+#endif
+ erts_smp_atomic32_read_bset_mb(&exec_rq->scheduler->ssi->flags,
+ (ERTS_SSI_FLG_SUSPENDED
+ | ERTS_SSI_FLG_MSB_EXEC),
+ ERTS_SSI_FLG_MSB_EXEC);
+ ASSERT(dbg_val & ERTS_SSI_FLG_SUSPENDED);
+
+ wake_scheduler(exec_rq);
+
+ return 1; /* suspend this scheduler... */
+
+}
+
+#endif
+
static void
suspend_scheduler(ErtsSchedulerData *esdp)
{
@@ -7151,32 +7698,51 @@ suspend_scheduler(ErtsSchedulerData *esdp)
* Regardless of why a scheduler is suspended, it ends up here.
*/
-#ifdef ERTS_DIRTY_SCHEDULERS
- if (ERTS_SCHEDULER_IS_DIRTY(esdp)) {
+
+#if !defined(ERTS_DIRTY_SCHEDULERS)
+
+ sched_type = ERTS_SCHED_NORMAL;
+ online_flag = ERTS_SCHDLR_SSPND_CHNG_ONLN;
+ no = esdp->no;
+ ASSERT(no != 1);
+
+#else
+
+ sched_type = esdp->type;
+ switch (sched_type) {
+ case ERTS_SCHED_NORMAL:
+ online_flag = ERTS_SCHDLR_SSPND_CHNG_ONLN;
+ no = esdp->no;
+ break;
+ case ERTS_SCHED_DIRTY_CPU:
+ online_flag = ERTS_SCHDLR_SSPND_CHNG_DCPU_ONLN;
+ no = ERTS_DIRTY_SCHEDULER_NO(esdp);
+ break;
+ case ERTS_SCHED_DIRTY_IO:
+ online_flag = 0;
no = ERTS_DIRTY_SCHEDULER_NO(esdp);
- if (ERTS_RUNQ_IS_DIRTY_CPU_RUNQ(esdp->run_queue)) {
- online_flag = ERTS_SCHDLR_SSPND_CHNG_DCPU_ONLN;
- sched_type = ERTS_SCHED_DIRTY_CPU;
- }
- else {
- online_flag = 0;
- sched_type = ERTS_SCHED_DIRTY_IO;
- }
+ break;
+ default:
+ ERTS_INTERNAL_ERROR("Invalid scheduler type");
+ return;
}
- else
-#endif
- {
- online_flag = ERTS_SCHDLR_SSPND_CHNG_ONLN;
- no = esdp->no;
- sched_type = ERTS_SCHED_NORMAL;
+
+ if (erts_smp_atomic32_read_nob(&ssi->flags) & ERTS_SSI_FLG_MSB_EXEC) {
+ ASSERT(no == 1);
+ if (!msb_scheduler_type_switch(sched_type, esdp, no))
+ return;
+ /* Suspend and let scheduler 1 of another type execute... */
}
- ASSERT(sched_type != ERTS_SCHED_NORMAL || no != 1);
+#endif
- if (sched_type != ERTS_SCHED_NORMAL)
+ if (sched_type != ERTS_SCHED_NORMAL) {
erts_smp_runq_unlock(esdp->run_queue);
+ dirty_sched_wall_time_change(esdp, 0);
+ }
else {
- evacuate_run_queue(esdp->run_queue, &sbp);
+ if (no != 1)
+ evacuate_run_queue(esdp->run_queue, &sbp);
erts_smp_runq_unlock(esdp->run_queue);
@@ -7184,10 +7750,8 @@ suspend_scheduler(ErtsSchedulerData *esdp)
if (erts_system_profile_flags.scheduler)
profile_scheduler(make_small(esdp->no), am_inactive);
-
- sched_wall_time_change(esdp, 0);
-
}
+
erts_smp_mtx_lock(&schdlr_sspnd.mtx);
flgs = sched_prep_spin_suspended(ssi, ERTS_SSI_FLG_SUSPENDED);
@@ -7195,9 +7759,6 @@ suspend_scheduler(ErtsSchedulerData *esdp)
schdlr_sspnd_dec_nscheds(&schdlr_sspnd.active, sched_type);
- ASSERT(schdlr_sspnd_get_nscheds(&schdlr_sspnd.active,
- ERTS_SCHED_NORMAL) >= 1);
-
changing = erts_smp_atomic32_read_nob(&schdlr_sspnd.changing);
while (1) {
@@ -7219,9 +7780,13 @@ suspend_scheduler(ErtsSchedulerData *esdp)
ERTS_SCHED_NORMAL) == 1) {
clr_flg = ERTS_SCHDLR_SSPND_CHNG_NMSB;
}
- else if (schdlr_sspnd.active
- == ERTS_SCHDLR_SSPND_MAKE_NSCHEDS_VAL(1, 0, 0)) {
- clr_flg = ERTS_SCHDLR_SSPND_CHNG_MSB;
+ else if (schdlr_sspnd_get_nscheds(&schdlr_sspnd.active,
+ ERTS_SCHED_NORMAL) == 1
+ && schdlr_sspnd_get_nscheds(&schdlr_sspnd.active,
+ ERTS_SCHED_DIRTY_CPU) == 0
+ && schdlr_sspnd_get_nscheds(&schdlr_sspnd.active,
+ ERTS_SCHED_DIRTY_IO) == 0) {
+ clr_flg = ERTS_SCHDLR_SSPND_CHNG_MSB;
}
if (clr_flg) {
@@ -7293,10 +7858,7 @@ suspend_scheduler(ErtsSchedulerData *esdp)
}
}
- if (curr_online
- && (sched_type == ERTS_SCHED_NORMAL
- ? !(schdlr_sspnd.msb.ongoing|schdlr_sspnd.nmsb.ongoing)
- : !schdlr_sspnd.msb.ongoing)) {
+ if (curr_online) {
flgs = erts_smp_atomic32_read_acqb(&ssi->flags);
if (!(flgs & ERTS_SSI_FLG_SUSPENDED))
break;
@@ -7312,13 +7874,11 @@ suspend_scheduler(ErtsSchedulerData *esdp)
if (sched_type != ERTS_SCHED_NORMAL)
aux_work = 0;
else {
- erts_aint32_t qmask;
- qmask = (ERTS_RUNQ_FLGS_GET(esdp->run_queue)
- & ERTS_RUNQ_FLGS_QMASK);
+ int evacuate = no == 1 ? 0 : !ERTS_EMPTY_RUNQ(esdp->run_queue);
aux_work = erts_atomic32_read_acqb(&ssi->aux_work);
- if (aux_work|qmask) {
+ if (aux_work|evacuate) {
if (!thr_prgr_active) {
erts_thr_progress_active(esdp, thr_prgr_active = 1);
sched_wall_time_change(esdp, 1);
@@ -7330,7 +7890,7 @@ suspend_scheduler(ErtsSchedulerData *esdp)
if (aux_work && erts_thr_progress_update(esdp))
erts_thr_progress_leader_update(esdp);
- if (qmask) {
+ if (evacuate) {
erts_smp_runq_lock(esdp->run_queue);
evacuate_run_queue(esdp->run_queue, &sbp);
erts_smp_runq_unlock(esdp->run_queue);
@@ -7445,7 +8005,8 @@ suspend_scheduler(ErtsSchedulerData *esdp)
if (changing) {
if ((changing & ERTS_SCHDLR_SSPND_CHNG_MSB)
&& !schdlr_sspnd.msb.ongoing
- && schdlr_sspnd.online == schdlr_sspnd.active) {
+ && schdlr_sspnd_eq_nscheds(&schdlr_sspnd.online,
+ &schdlr_sspnd.active)) {
erts_smp_atomic32_read_band_nob(&schdlr_sspnd.changing,
~ERTS_SCHDLR_SSPND_CHNG_MSB);
}
@@ -7460,30 +8021,27 @@ suspend_scheduler(ErtsSchedulerData *esdp)
}
}
ASSERT(no <= schdlr_sspnd_get_nscheds(&schdlr_sspnd.online, sched_type));
- ASSERT((sched_type == ERTS_SCHED_NORMAL
- ? !(schdlr_sspnd.msb.ongoing|schdlr_sspnd.nmsb.ongoing)
- : !schdlr_sspnd.msb.ongoing));
}
erts_smp_mtx_unlock(&schdlr_sspnd.mtx);
- ASSERT(!resume.msb.chngrs);
schdlr_sspnd_resume_procs(sched_type, &resume);
ASSERT(curr_online);
- if (!ERTS_SCHEDULER_IS_DIRTY(esdp)) {
- if (erts_system_profile_flags.scheduler)
- profile_scheduler(make_small(esdp->no), am_active);
+ if (sched_type != ERTS_SCHED_NORMAL)
+ dirty_sched_wall_time_change(esdp, 1);
+ else {
+ (void) erts_get_monotonic_time(esdp);
+ if (erts_system_profile_flags.scheduler)
+ profile_scheduler(make_small(esdp->no), am_active);
- if (!thr_prgr_active) {
- erts_thr_progress_active(esdp, thr_prgr_active = 1);
- sched_wall_time_change(esdp, 1);
- }
+ if (!thr_prgr_active) {
+ erts_thr_progress_active(esdp, thr_prgr_active = 1);
+ sched_wall_time_change(esdp, 1);
+ }
}
- if (sched_type == ERTS_SCHED_NORMAL)
- (void) erts_get_monotonic_time(esdp);
erts_smp_runq_lock(esdp->run_queue);
non_empty_runq(esdp->run_queue);
@@ -7753,10 +8311,8 @@ erts_set_schedulers_online(Process *p,
erts_sched_poke(ssi);
}
} else {
- for (ix = dirty_online; ix < dirty_no; ix++) {
- ssi = ERTS_DIRTY_CPU_SCHED_SLEEP_INFO_IX(ix);
- scheduler_ssi_resume_wake(ssi);
- }
+ for (ix = dirty_online; ix < dirty_no; ix++)
+ dcpu_sched_ix_resume_wake(ix);
}
}
if (!dirty_only)
@@ -7784,19 +8340,19 @@ erts_set_schedulers_online(Process *p,
else /* if decrease */ {
#ifdef ERTS_DIRTY_SCHEDULERS
if (change_dirty) {
- ErtsSchedulerSleepInfo* ssi;
if (schdlr_sspnd.msb.ongoing) {
- for (ix = dirty_no; ix < dirty_online; ix++) {
- ssi = ERTS_DIRTY_CPU_SCHED_SLEEP_INFO_IX(ix);
- erts_sched_poke(ssi);
- }
- } else {
- for (ix = dirty_no; ix < dirty_online; ix++) {
- ssi = ERTS_DIRTY_CPU_SCHED_SLEEP_INFO_IX(ix);
- erts_smp_atomic32_read_bor_nob(&ssi->flags,
- ERTS_SSI_FLG_SUSPENDED);
- }
- wake_dirty_schedulers(ERTS_DIRTY_CPU_RUNQ, 0);
+ for (ix = dirty_no; ix < dirty_online; ix++)
+ erts_sched_poke(ERTS_DIRTY_CPU_SCHED_SLEEP_INFO_IX(ix));
+ }
+ else {
+ for (ix = dirty_no; ix < dirty_online; ix++)
+ dcpu_sched_ix_suspend_wake(ix);
+ /*
+ * Newly suspended scheduler may have just been
+ * about to handle a task. Make sure someone takes
+ * care of such a task...
+ */
+ dcpu_sched_ix_wake(0);
}
}
if (!dirty_only)
@@ -7859,9 +8415,6 @@ erts_block_multi_scheduling(Process *p, ErtsProcLocks plocks, int on, int normal
{
int resume_proc, ix, res, have_unlocked_plocks = 0;
ErtsProcList *plp;
-#ifdef ERTS_DIRTY_SCHEDULERS
- ErtsSchedulerSleepInfo* ssi;
-#endif
ErtsMultiSchedulingBlock *msbp;
erts_aint32_t chng_flg;
int have_blckd_flg;
@@ -7908,8 +8461,9 @@ erts_block_multi_scheduling(Process *p, ErtsProcLocks plocks, int on, int normal
erts_proclist_store_last(&msbp->blckrs, plp);
p->flags |= have_blckd_flg;
ASSERT(normal
- ? 1 == schdlr_sspnd_get_nscheds(&schdlr_sspnd.active, ERTS_SCHED_NORMAL)
- : schdlr_sspnd.active == ERTS_SCHDLR_SSPND_MAKE_NSCHEDS_VAL(1, 0, 0));
+ ? 1 == schdlr_sspnd_get_nscheds(&schdlr_sspnd.active,
+ ERTS_SCHED_NORMAL)
+ : schdlr_sspnd_get_nscheds_tot(&schdlr_sspnd.active) == 1);
ASSERT(erts_proc_sched_data(p)->no == 1);
if (schdlr_sspnd.msb.ongoing)
res = ERTS_SCHDLR_SSPND_DONE_MSCHED_BLOCKED;
@@ -7927,59 +8481,41 @@ erts_block_multi_scheduling(Process *p, ErtsProcLocks plocks, int on, int normal
}
ASSERT(!msbp->ongoing);
msbp->ongoing = 1;
- if (schdlr_sspnd.active == ERTS_SCHDLR_SSPND_MAKE_NSCHEDS_VAL(1, 0, 0)
- || (normal && schdlr_sspnd_get_nscheds(&schdlr_sspnd.active,
- ERTS_SCHED_NORMAL) == 1)) {
- ASSERT(erts_proc_sched_data(p)->no == 1);
- plp = proclist_create(p);
- erts_proclist_store_last(&msbp->blckrs, plp);
- if (schdlr_sspnd.msb.ongoing)
- res = ERTS_SCHDLR_SSPND_DONE_MSCHED_BLOCKED;
- else
- res = ERTS_SCHDLR_SSPND_DONE_NMSCHED_BLOCKED;
- }
- else {
- erts_smp_atomic32_read_bor_nob(&schdlr_sspnd.changing,
- chng_flg);
- change_no_used_runqs(1);
- for (ix = 1; ix < erts_no_run_queues; ix++)
- suspend_run_queue(ERTS_RUNQ_IX(ix));
- for (ix = 1; ix < online; ix++) {
- ErtsRunQueue *rq = ERTS_RUNQ_IX(ix);
- wake_scheduler(rq);
- }
+ erts_smp_atomic32_read_bor_nob(&schdlr_sspnd.changing,
+ chng_flg);
+ change_no_used_runqs(1);
+ for (ix = 1; ix < erts_no_run_queues; ix++)
+ suspend_run_queue(ERTS_RUNQ_IX(ix));
-#ifdef ERTS_DIRTY_SCHEDULERS
- if (!normal) {
- for (ix = 0; ix < erts_no_dirty_cpu_schedulers; ix++) {
- ssi = ERTS_DIRTY_CPU_SCHED_SLEEP_INFO_IX(ix);
- erts_smp_atomic32_read_bor_nob(&ssi->flags,
- ERTS_SSI_FLG_SUSPENDED);
- }
- wake_dirty_schedulers(ERTS_DIRTY_CPU_RUNQ, 0);
+ for (ix = 1; ix < online; ix++) {
+ ErtsRunQueue *rq = ERTS_RUNQ_IX(ix);
+ wake_scheduler(rq);
+ }
- for (ix = 0; ix < erts_no_dirty_io_schedulers; ix++) {
- ssi = ERTS_DIRTY_IO_SCHED_SLEEP_INFO_IX(ix);
- erts_smp_atomic32_read_bor_nob(&ssi->flags,
- ERTS_SSI_FLG_SUSPENDED);
- }
- wake_dirty_schedulers(ERTS_DIRTY_IO_RUNQ, 0);
- }
+#ifdef ERTS_DIRTY_SCHEDULERS
+ if (!normal) {
+ ERTS_RUNQ_FLGS_SET_NOB(ERTS_RUNQ_IX(0), ERTS_RUNQ_FLG_MSB_EXEC);
+ erts_smp_atomic32_read_bor_nob(&ERTS_RUNQ_IX(0)->scheduler->ssi->flags,
+ ERTS_SSI_FLG_MSB_EXEC);
+ for (ix = 0; ix < erts_no_dirty_cpu_schedulers; ix++)
+ dcpu_sched_ix_suspend_wake(ix);
+ for (ix = 0; ix < erts_no_dirty_io_schedulers; ix++)
+ dio_sched_ix_suspend_wake(ix);
+ }
#endif
- wait_until_msb:
+ wait_until_msb:
- ASSERT(chng_flg & erts_smp_atomic32_read_nob(&schdlr_sspnd.changing));
+ ASSERT(chng_flg & erts_smp_atomic32_read_nob(&schdlr_sspnd.changing));
- plp = proclist_create(p);
- erts_proclist_store_last(&msbp->chngq, plp);
- resume_proc = 0;
- if (schdlr_sspnd.msb.ongoing)
- res = ERTS_SCHDLR_SSPND_YIELD_DONE_MSCHED_BLOCKED;
- else
- res = ERTS_SCHDLR_SSPND_YIELD_DONE_NMSCHED_BLOCKED;
- }
+ plp = proclist_create(p);
+ erts_proclist_store_last(&msbp->chngq, plp);
+ resume_proc = 0;
+ if (schdlr_sspnd.msb.ongoing)
+ res = ERTS_SCHDLR_SSPND_YIELD_DONE_MSCHED_BLOCKED;
+ else
+ res = ERTS_SCHDLR_SSPND_YIELD_DONE_NMSCHED_BLOCKED;
ASSERT(erts_proc_sched_data(p));
}
}
@@ -8011,27 +8547,19 @@ erts_block_multi_scheduling(Process *p, ErtsProcLocks plocks, int on, int normal
}
}
if (!msbp->blckrs && !msbp->chngq) {
- int online = (int) schdlr_sspnd_get_nscheds(&schdlr_sspnd.online,
- ERTS_SCHED_NORMAL);
+ int online;
erts_smp_atomic32_read_bor_nob(&schdlr_sspnd.changing,
chng_flg);
p->flags &= ~have_blckd_flg;
msbp->ongoing = 0;
- if (online == 1) {
- /* No normal schedulers to resume */
- ASSERT(schdlr_sspnd_get_nscheds(&schdlr_sspnd.active,
- ERTS_SCHED_NORMAL) == 1);
-#ifndef ERTS_DIRTY_SCHEDULERS
- erts_smp_atomic32_read_band_nob(&schdlr_sspnd.changing,
- ~chng_flg);
-#endif
- }
- else if (!(schdlr_sspnd.msb.ongoing|schdlr_sspnd.nmsb.ongoing)) {
- if (plocks) {
+ if (!(schdlr_sspnd.msb.ongoing|schdlr_sspnd.nmsb.ongoing)) {
+ if (plocks) {
have_unlocked_plocks = 1;
erts_smp_proc_unlock(p, plocks);
}
+ online = (int) schdlr_sspnd_get_nscheds(&schdlr_sspnd.online,
+ ERTS_SCHED_NORMAL);
change_no_used_runqs(online);
/* Resume all online run queues */
@@ -8042,19 +8570,15 @@ erts_block_multi_scheduling(Process *p, ErtsProcLocks plocks, int on, int normal
suspend_run_queue(ERTS_RUNQ_IX(ix));
}
#ifdef ERTS_DIRTY_SCHEDULERS
- if (!normal) {
- ASSERT(!schdlr_sspnd.msb.ongoing);
+ if (!schdlr_sspnd.msb.ongoing) {
+ /* Get rid of msb-exec flag in run-queue of scheduler 1 */
+ resume_run_queue(ERTS_RUNQ_IX(0));
online = (int) schdlr_sspnd_get_nscheds(&schdlr_sspnd.online,
ERTS_SCHED_DIRTY_CPU);
- for (ix = 0; ix < online; ix++) {
- ssi = ERTS_DIRTY_CPU_SCHED_SLEEP_INFO_IX(ix);
- scheduler_ssi_resume_wake(ssi);
- }
-
- for (ix = 0; ix < erts_no_dirty_io_schedulers; ix++) {
- ssi = ERTS_DIRTY_IO_SCHED_SLEEP_INFO_IX(ix);
- scheduler_ssi_resume_wake(ssi);
- }
+ for (ix = 0; ix < online; ix++)
+ dcpu_sched_ix_resume_wake(ix);
+ for (ix = 0; ix < erts_no_dirty_io_schedulers; ix++)
+ dio_sched_ix_resume_wake(ix);
}
#endif
}
@@ -8228,6 +8752,8 @@ sched_dirty_cpu_thread_func(void *vesdp)
callbacks.wait = NULL;
callbacks.finalize_wait = NULL;
+ dirty_sched_wall_time_change(esdp, 1);
+
esdp->thr_id += erts_no_schedulers;
erts_msacc_init_thread("dirty_cpu_scheduler", no, 0);
@@ -8275,6 +8801,8 @@ sched_dirty_io_thread_func(void *vesdp)
callbacks.wait = NULL;
callbacks.finalize_wait = NULL;
+ dirty_sched_wall_time_change(esdp, 1);
+
esdp->thr_id += erts_no_schedulers + erts_no_dirty_cpu_schedulers;
erts_msacc_init_thread("dirty_io_scheduler", no, 0);
@@ -8647,8 +9175,7 @@ pid2proc_not_running(Process *c_p, ErtsProcLocks c_p_locks,
erts_aint32_t state;
state = erts_smp_atomic32_read_nob(&rp->state);
ASSERT((state & ERTS_PSFLG_PENDING_EXIT)
- || !(state & (ERTS_PSFLG_RUNNING
- | ERTS_PSFLG_DIRTY_RUNNING_SYS)));
+ || !(state & ERTS_PSFLG_RUNNING));
}
#endif
@@ -9720,7 +10247,7 @@ Process *erts_schedule(ErtsSchedulerData *esdp, Process *p, int calls)
if (!is_normal_sched) {
if (erts_smp_atomic32_read_acqb(&esdp->ssi->flags)
- & ERTS_SSI_FLG_SUSPENDED) {
+ & (ERTS_SSI_FLG_SUSPENDED|ERTS_SSI_FLG_MSB_EXEC)) {
suspend_scheduler(esdp);
}
}
@@ -9730,8 +10257,10 @@ Process *erts_schedule(ErtsSchedulerData *esdp, Process *p, int calls)
ASSERT(is_normal_sched);
- if (flags & (ERTS_RUNQ_FLG_CHK_CPU_BIND|ERTS_RUNQ_FLG_SUSPENDED)) {
- if (flags & ERTS_RUNQ_FLG_SUSPENDED) {
+ if (flags & (ERTS_RUNQ_FLG_CHK_CPU_BIND
+ | ERTS_RUNQ_FLG_SUSPENDED
+ | ERTS_RUNQ_FLG_MSB_EXEC)) {
+ if (flags & (ERTS_RUNQ_FLG_SUSPENDED|ERTS_RUNQ_FLG_MSB_EXEC)) {
(void) ERTS_RUNQ_FLGS_UNSET_NOB(rq, ERTS_RUNQ_FLG_EXEC);
suspend_scheduler(esdp);
flags = ERTS_RUNQ_FLGS_SET_NOB(rq, ERTS_RUNQ_FLG_EXEC);
@@ -9770,13 +10299,12 @@ Process *erts_schedule(ErtsSchedulerData *esdp, Process *p, int calls)
flags = ERTS_RUNQ_FLGS_GET_NOB(rq);
- if (!is_normal_sched && rq->halt_in_progress) {
+ if (!is_normal_sched & !!(flags & ERTS_RUNQ_FLG_HALTING)) {
/* Wait for emulator to terminate... */
while (1)
erts_milli_sleep(1000*1000);
}
- else if ((!(flags & ERTS_RUNQ_FLGS_QMASK) && !rq->misc.start)
- || (rq->halt_in_progress && ERTS_EMPTY_RUNQ_PORTS(rq))) {
+ else if (!runq_got_work_to_execute_flags(flags)) {
/* Prepare for scheduler wait */
#ifdef ERTS_SMP
ERTS_SMP_LC_ASSERT(erts_smp_lc_runq_is_locked(rq));
@@ -9790,21 +10318,44 @@ Process *erts_schedule(ErtsSchedulerData *esdp, Process *p, int calls)
if (flags & ERTS_RUNQ_FLG_INACTIVE)
empty_runq(rq);
else {
- if (is_normal_sched && try_steal_task(rq))
- goto continue_check_activities_to_run;
-
- empty_runq(rq);
-
- /*
- * Check for ERTS_RUNQ_FLG_SUSPENDED has to be done
- * after trying to steal a task.
- */
- flags = ERTS_RUNQ_FLGS_GET_NOB(rq);
- if (flags & ERTS_RUNQ_FLG_SUSPENDED) {
- non_empty_runq(rq);
- flags |= ERTS_RUNQ_FLG_NONEMPTY;
- goto continue_check_activities_to_run_known_flags;
+ ASSERT(!runq_got_work_to_execute(rq));
+ if (!is_normal_sched) {
+ /* Dirty scheduler */
+ if (erts_smp_atomic32_read_acqb(&esdp->ssi->flags)
+ & (ERTS_SSI_FLG_SUSPENDED|ERTS_SSI_FLG_MSB_EXEC)) {
+ /* Go suspend... */
+ goto continue_check_activities_to_run_known_flags;
+ }
+ }
+ else {
+ /* Normal scheduler */
+ if (try_steal_task(rq))
+ goto continue_check_activities_to_run;
+ /*
+ * Check for suspend has to be done after trying
+ * to steal a task...
+ */
+ flags = ERTS_RUNQ_FLGS_GET_NOB(rq);
+ if ((flags & ERTS_RUNQ_FLG_SUSPENDED)
+#ifdef ERTS_DIRTY_SCHEDULERS
+ /* If multi scheduling block and we have
+ * dirty work, suspend and let dirty
+ * scheduler handle work... */
+ || ((((flags & (ERTS_RUNQ_FLG_HALTING
+ | ERTS_RUNQ_FLG_MSB_EXEC))
+ == ERTS_RUNQ_FLG_MSB_EXEC))
+ && have_dirty_work())
+#endif
+ ) {
+ non_empty_runq(rq);
+ flags |= ERTS_RUNQ_FLG_NONEMPTY;
+ /*
+ * Go suspend...
+ */
+ goto continue_check_activities_to_run_known_flags;
+ }
}
+ empty_runq(rq);
}
#endif
@@ -9856,7 +10407,7 @@ Process *erts_schedule(ErtsSchedulerData *esdp, Process *p, int calls)
#endif
}
- if (rq->misc.start)
+ if (flags & ERTS_RUNQ_FLG_MISC_OP)
exec_misc_ops(rq);
#ifdef ERTS_SMP
@@ -9867,13 +10418,15 @@ Process *erts_schedule(ErtsSchedulerData *esdp, Process *p, int calls)
* Find a new port to run.
*/
- if (RUNQ_READ_LEN(&rq->ports.info.len)) {
+ flags = ERTS_RUNQ_FLGS_GET_NOB(rq);
+
+ if (flags & PORT_BIT) {
int have_outstanding_io;
have_outstanding_io = erts_port_task_execute(rq, &esdp->current_port);
if ((!erts_eager_check_io
&& have_outstanding_io
&& fcalls > 2*input_reductions)
- || rq->halt_in_progress) {
+ || (flags & ERTS_RUNQ_FLG_HALTING)) {
/*
* If we have performed more than 2*INPUT_REDUCTIONS since
* last call to erl_sys_schedule() and we still haven't
@@ -9986,6 +10539,7 @@ Process *erts_schedule(ErtsSchedulerData *esdp, Process *p, int calls)
| ERTS_PSFLG_FREE
| ERTS_PSFLG_RUNNING_SYS
+ | ERTS_PSFLG_DIRTY_RUNNING_SYS
| ERTS_PSFLG_EXITING))
== ERTS_PSFLG_EXITING)
& (!!is_normal_sched))
@@ -9997,7 +10551,14 @@ Process *erts_schedule(ErtsSchedulerData *esdp, Process *p, int calls)
| ERTS_PSFLG_PENDING_EXIT
| ERTS_PSFLG_ACTIVE_SYS
| ERTS_PSFLG_DIRTY_ACTIVE_SYS))
- != ERTS_PSFLG_SUSPENDED));
+ != ERTS_PSFLG_SUSPENDED)
+#ifdef ERTS_DIRTY_SCHEDULERS
+ & (!(state & (ERTS_PSFLG_EXITING
+ | ERTS_PSFLG_PENDING_EXIT))
+ | (!!is_normal_sched))
+#endif
+ );
+
if (run_process) {
if (state & (ERTS_PSFLG_ACTIVE_SYS
| ERTS_PSFLG_DIRTY_ACTIVE_SYS))
@@ -10102,9 +10663,6 @@ Process *erts_schedule(ErtsSchedulerData *esdp, Process *p, int calls)
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))
@@ -10141,65 +10699,70 @@ Process *erts_schedule(ErtsSchedulerData *esdp, Process *p, int calls)
}
}
- if (state & (ERTS_PSFLG_RUNNING_SYS
- | ERTS_PSFLG_DIRTY_RUNNING_SYS)) {
- /*
- * GC is normally never delayed when a process
- * is scheduled out, but might be when executing
- * hand written beam assembly in
- * prim_eval:'receive'. If GC is delayed we are
- * not allowed to execute system tasks.
- */
- if (!(p->flags & F_DELAY_GC)) {
- int cost = execute_sys_tasks(p, &state, reds);
- calls += cost;
- reds -= cost;
- if (reds <= 0
+ if (is_normal_sched) {
+
+ if (state & ERTS_PSFLG_RUNNING_SYS) {
+ /*
+ * GC is normally never delayed when a process
+ * is scheduled out, but might be when executing
+ * hand written beam assembly in
+ * prim_eval:'receive'. If GC is delayed we are
+ * not allowed to execute system tasks.
+ */
+ if (!(p->flags & F_DELAY_GC)) {
+ int cost = execute_sys_tasks(p, &state, reds);
+ calls += cost;
+ reds -= cost;
+ if (reds <= 0)
+ goto sched_out_proc;
#ifdef ERTS_DIRTY_SCHEDULERS
- || !is_normal_sched
- || (state & ERTS_PSFLGS_DIRTY_WORK)
+ if (state & ERTS_PSFLGS_DIRTY_WORK)
+ goto sched_out_proc;
#endif
- ) {
- goto sched_out_proc;
- }
- }
+ }
- ASSERT(state & psflg_running_sys);
- ASSERT(!(state & psflg_running));
+ ASSERT(state & psflg_running_sys);
+ ASSERT(!(state & psflg_running));
- while (1) {
- erts_aint32_t n, e;
+ while (1) {
+ erts_aint32_t n, e;
- if (((state & (ERTS_PSFLG_SUSPENDED
- | ERTS_PSFLG_ACTIVE)) != ERTS_PSFLG_ACTIVE)
- && !(state & ERTS_PSFLG_EXITING)) {
- goto sched_out_proc;
- }
+ if (((state & (ERTS_PSFLG_SUSPENDED
+ | ERTS_PSFLG_ACTIVE)) != ERTS_PSFLG_ACTIVE)
+ && !(state & ERTS_PSFLG_EXITING)) {
+ goto sched_out_proc;
+ }
- n = e = state;
- n &= ~psflg_running_sys;
- n |= psflg_running;
+ n = e = state;
+ n &= ~psflg_running_sys;
+ n |= psflg_running;
- state = erts_smp_atomic32_cmpxchg_mb(&p->state, n, e);
- if (state == e) {
- state = n;
- break;
- }
+ state = erts_smp_atomic32_cmpxchg_mb(&p->state, n, e);
+ if (state == e) {
+ state = n;
+ break;
+ }
- ASSERT(state & psflg_running_sys);
- ASSERT(!(state & psflg_running));
- }
- }
+ ASSERT(state & psflg_running_sys);
+ ASSERT(!(state & psflg_running));
+ }
+ }
- if (ERTS_IS_GC_DESIRED(p) && !ERTS_SCHEDULER_IS_DIRTY_IO(esdp)) {
- if (!(state & ERTS_PSFLG_EXITING) && !(p->flags & (F_DELAY_GC|F_DISABLE_GC))) {
- int cost = scheduler_gc_proc(p, reds);
- calls += cost;
- reds -= cost;
- if (reds <= 0)
- goto sched_out_proc;
- }
- }
+ if (ERTS_IS_GC_DESIRED(p)) {
+ if (!(state & ERTS_PSFLG_EXITING)
+ && !(p->flags & (F_DELAY_GC|F_DISABLE_GC))) {
+ int cost = scheduler_gc_proc(p, reds);
+ calls += cost;
+ reds -= cost;
+ if (reds <= 0)
+ goto sched_out_proc;
+#ifdef ERTS_DIRTY_SCHEDULERS
+ if (p->flags & (F_DIRTY_MAJOR_GC|F_DIRTY_MINOR_GC))
+ goto sched_out_proc;
+#endif
+ }
+ }
+ }
if (proxy_p) {
free_proxy_proc(proxy_p);
@@ -10211,7 +10774,13 @@ Process *erts_schedule(ErtsSchedulerData *esdp, Process *p, int calls)
ERTS_SMP_CHK_HAVE_ONLY_MAIN_PROC_LOCK(p);
/* Never run a suspended process */
- ASSERT(!(ERTS_PSFLG_SUSPENDED & erts_smp_atomic32_read_nob(&p->state)));
+#ifdef DEBUG
+ {
+ erts_aint32_t dstate = erts_smp_atomic32_read_nob(&p->state);
+ ASSERT(!(ERTS_PSFLG_SUSPENDED & dstate)
+ || (ERTS_PSFLG_DIRTY_RUNNING_SYS & dstate));
+ }
+#endif
ASSERT(erts_proc_read_refc(p) > 0);
@@ -10232,9 +10801,18 @@ Process *erts_schedule(ErtsSchedulerData *esdp, Process *p, int calls)
}
static int
-notify_sys_task_executed(Process *c_p, ErtsProcSysTask *st, Eterm st_result)
+notify_sys_task_executed(Process *c_p, ErtsProcSysTask *st,
+ Eterm st_result, int normal_sched)
{
- Process *rp = erts_proc_lookup(st->requester);
+ Process *rp;
+#ifdef ERTS_DIRTY_SCHEDULERS
+ if (!normal_sched)
+ rp = erts_pid2proc_opt(c_p, ERTS_PROC_LOCK_MAIN,
+ st->requester, 0,
+ ERTS_P2P_FLG_INC_REFC);
+ else
+#endif
+ rp = erts_proc_lookup(st->requester);
if (rp) {
ErtsProcLocks rp_locks;
ErlOffHeap *ohp;
@@ -10282,6 +10860,11 @@ notify_sys_task_executed(Process *c_p, ErtsProcSysTask *st, Eterm st_result)
if (rp_locks)
erts_smp_proc_unlock(rp, rp_locks);
+
+#ifdef ERTS_DIRTY_SCHEDULERS
+ if (!normal_sched)
+ erts_proc_dec_refc(rp);
+#endif
}
erts_cleanup_offheap(&st->off_heap);
@@ -10437,18 +11020,23 @@ done:
}
static void save_gc_task(Process *c_p, ErtsProcSysTask *st, int prio);
+#ifdef ERTS_DIRTY_SCHEDULERS
+static void save_dirty_task(Process *c_p, ErtsProcSysTask *st);
+#endif
static int
execute_sys_tasks(Process *c_p, erts_aint32_t *statep, int in_reds)
{
- int garbage_collected = 0;
+ int minor_gc = 0, major_gc = 0;
erts_aint32_t state = *statep;
int reds = in_reds;
int qmask = 0;
+ ASSERT(!ERTS_SCHEDULER_IS_DIRTY(erts_proc_sched_data(c_p)));
ERTS_SMP_LC_ASSERT(erts_proc_lc_my_proc_locks(c_p) == ERTS_PROC_LOCK_MAIN);
do {
+ ErtsProcSysTaskType type;
ErtsProcSysTask *st;
int st_prio;
Eterm st_res;
@@ -10466,7 +11054,9 @@ execute_sys_tasks(Process *c_p, erts_aint32_t *statep, int in_reds)
if (!st)
break;
- switch (st->type) {
+ type = st->type;
+
+ switch (type) {
case ERTS_PSTT_GC_MAJOR:
case ERTS_PSTT_GC_MINOR:
if (c_p->flags & F_DISABLE_GC) {
@@ -10475,12 +11065,23 @@ execute_sys_tasks(Process *c_p, erts_aint32_t *statep, int in_reds)
reds--;
}
else {
- if (!garbage_collected) {
- if (st->type == ERTS_PSTT_GC_MAJOR) {
+ if (!minor_gc
+ || (!major_gc && type == ERTS_PSTT_GC_MAJOR)) {
+ if (type == ERTS_PSTT_GC_MAJOR) {
FLAGS(c_p) |= F_NEED_FULLSWEEP;
}
reds -= scheduler_gc_proc(c_p, reds);
- garbage_collected = 1;
+#ifdef ERTS_DIRTY_SCHEDULERS
+ if (c_p->flags & (F_DIRTY_MAJOR_GC|F_DIRTY_MINOR_GC)) {
+ save_dirty_task(c_p, st);
+ st = NULL;
+ break;
+ }
+#endif
+ if (type == ERTS_PSTT_GC_MAJOR)
+ minor_gc = major_gc = 1;
+ else
+ minor_gc = 1;
}
st_res = am_true;
}
@@ -10507,20 +11108,31 @@ execute_sys_tasks(Process *c_p, erts_aint32_t *statep, int in_reds)
case ERTS_PSTT_CLA: {
int fcalls;
int cla_reds = 0;
+ int do_gc;
+
if (!ERTS_PROC_GET_SAVED_CALLS_BUF(c_p))
fcalls = reds;
else
fcalls = reds - CONTEXT_REDS;
- st_res = erts_proc_copy_literal_area(c_p,
- &cla_reds,
- fcalls,
- st->arg[0] == am_true);
+ do_gc = st->arg[0] == am_true;
+ st_res = erts_proc_copy_literal_area(c_p, &cla_reds,
+ fcalls, do_gc);
reds -= cla_reds;
if (is_non_value(st_res)) {
+#ifdef ERTS_DIRTY_SCHEDULERS
+ if (c_p->flags & F_DIRTY_CLA) {
+ save_dirty_task(c_p, st);
+ st = NULL;
+ break;
+ }
+#endif
/* Needed gc, but gc was disabled */
save_gc_task(c_p, st, st_prio);
st = NULL;
+ break;
}
+ if (do_gc) /* We did a major gc */
+ minor_gc = major_gc = 1;
break;
}
case ERTS_PSTT_COHMQ:
@@ -10539,7 +11151,7 @@ execute_sys_tasks(Process *c_p, erts_aint32_t *statep, int in_reds)
}
if (st)
- reds += notify_sys_task_executed(c_p, st, st_res);
+ reds += notify_sys_task_executed(c_p, st, st_res, 1);
state = erts_smp_atomic32_read_acqb(&c_p->state);
} while (qmask && reds > 0);
@@ -10567,9 +11179,18 @@ cleanup_sys_tasks(Process *c_p, erts_aint32_t in_state, int in_reds)
Eterm st_res;
int st_prio;
- st = fetch_sys_task(c_p, state, &qmask, &st_prio);
- if (!st)
- break;
+#ifdef ERTS_DIRTY_SCHEDULERS
+ if (c_p->dirty_sys_tasks) {
+ st = c_p->dirty_sys_tasks;
+ c_p->dirty_sys_tasks = st->next;
+ }
+ else
+#endif
+ {
+ st = fetch_sys_task(c_p, state, &qmask, &st_prio);
+ if (!st)
+ break;
+ }
switch (st->type) {
case ERTS_PSTT_GC_MAJOR:
@@ -10593,7 +11214,7 @@ cleanup_sys_tasks(Process *c_p, erts_aint32_t in_state, int in_reds)
break;
}
- reds += notify_sys_task_executed(c_p, st, st_res);
+ reds += notify_sys_task_executed(c_p, st, st_res, 1);
state = erts_smp_atomic32_read_acqb(&c_p->state);
} while (qmask && reds < max_reds);
@@ -10603,6 +11224,92 @@ cleanup_sys_tasks(Process *c_p, erts_aint32_t in_state, int in_reds)
#ifdef ERTS_DIRTY_SCHEDULERS
+void
+erts_execute_dirty_system_task(Process *c_p)
+{
+ Eterm cla_res = THE_NON_VALUE;
+ ErtsProcSysTask *stasks;
+
+ /*
+ * If multiple operations, perform them in the following
+ * order (in order to avoid unnecessary GC):
+ * 1. Copy Literal Area (implies major GC).
+ * 2. GC Hibernate (implies major GC if not woken).
+ * 3. Major GC (implies minor GC).
+ * 4. Minor GC.
+ *
+ * System task requests are handled after the actual
+ * operations have been performed...
+ */
+
+ ASSERT(!(c_p->flags & (F_DELAY_GC|F_DISABLE_GC)));
+
+ if (c_p->flags & F_DIRTY_CLA) {
+ int cla_reds = 0;
+ cla_res = erts_proc_copy_literal_area(c_p, &cla_reds, c_p->fcalls, 1);
+ ASSERT(is_value(cla_res));
+ }
+
+ if (c_p->flags & F_DIRTY_GC_HIBERNATE) {
+ erts_smp_proc_lock(c_p, ERTS_PROC_LOCK_MSGQ|ERTS_PROC_LOCK_STATUS);
+ ERTS_SMP_MSGQ_MV_INQ2PRIVQ(c_p);
+ if (c_p->msg.len)
+ c_p->flags &= ~F_DIRTY_GC_HIBERNATE; /* operation aborted... */
+ else {
+ erts_smp_proc_unlock(c_p, ERTS_PROC_LOCK_MSGQ|ERTS_PROC_LOCK_STATUS);
+ c_p->fvalue = NIL;
+ erts_garbage_collect_hibernate(c_p);
+ ERTS_VERIFY_UNUSED_TEMP_ALLOC(c_p);
+ erts_smp_proc_lock(c_p, ERTS_PROC_LOCK_MSGQ|ERTS_PROC_LOCK_STATUS);
+ ASSERT(!ERTS_PROC_IS_EXITING(c_p));
+ }
+ erts_smp_proc_unlock(c_p, ERTS_PROC_LOCK_MSGQ|ERTS_PROC_LOCK_STATUS);
+ }
+
+ if (c_p->flags & (F_DIRTY_MAJOR_GC|F_DIRTY_MINOR_GC)) {
+ if (c_p->flags & F_DIRTY_MAJOR_GC)
+ c_p->flags |= F_NEED_FULLSWEEP;
+ (void) erts_garbage_collect_nobump(c_p, 0, c_p->arg_reg,
+ c_p->arity, c_p->fcalls);
+ }
+
+ ASSERT(!(c_p->flags & (F_DIRTY_CLA
+ | F_DIRTY_GC_HIBERNATE
+ | F_DIRTY_MAJOR_GC
+ | F_DIRTY_MINOR_GC)));
+
+ stasks = c_p->dirty_sys_tasks;
+ c_p->dirty_sys_tasks = NULL;
+
+ while (stasks) {
+ Eterm st_res;
+ ErtsProcSysTask *st = stasks;
+ stasks = st->next;
+
+ switch (st->type) {
+ case ERTS_PSTT_CLA:
+ ASSERT(is_value(st_res));
+ st_res = cla_res;
+ break;
+ case ERTS_PSTT_GC_MAJOR:
+ st_res = am_true;
+ break;
+ case ERTS_PSTT_GC_MINOR:
+ st_res = am_true;
+ break;
+
+ default:
+ ERTS_INTERNAL_ERROR("Not supported dirty system task");
+ break;
+ }
+
+ (void) notify_sys_task_executed(c_p, st, st_res, 0);
+
+ }
+
+ erts_smp_atomic32_read_band_relb(&c_p->state, ~ERTS_PSFLG_DIRTY_ACTIVE_SYS);
+}
+
static BIF_RETTYPE
dispatch_system_task(Process *c_p, erts_aint_t fail_state,
ErtsProcSysTask *st, Eterm target,
@@ -10805,7 +11512,7 @@ request_system_task(Process *c_p, Eterm requester, Eterm target,
ERTS_INTERNAL_ERROR("Unknown failure schedule_process_sys_task()");
failure = am_internal_error;
}
- notify_sys_task_executed(c_p, st, failure);
+ notify_sys_task_executed(c_p, st, failure, 1);
}
ERTS_BIF_PREP_RET(ret, am_ok);
@@ -11020,6 +11727,15 @@ save_gc_task(Process *c_p, ErtsProcSysTask *st, int prio)
}
}
+#ifdef ERTS_DIRTY_SCHEDULERS
+static void
+save_dirty_task(Process *c_p, ErtsProcSysTask *st)
+{
+ st->next = c_p->dirty_sys_tasks;
+ c_p->dirty_sys_tasks = st;
+}
+#endif
+
int
erts_set_gc_state(Process *c_p, int enable)
{
@@ -11242,6 +11958,8 @@ erts_schedule_misc_op(void (*func)(void *), void *arg)
non_empty_runq(rq);
#endif
+ ERTS_RUNQ_FLGS_SET_NOB(rq, ERTS_RUNQ_FLG_MISC_OP);
+
erts_smp_runq_unlock(rq);
smp_notify_inc_runq(rq);
@@ -11272,6 +11990,9 @@ exec_misc_ops(ErtsRunQueue *rq)
rq->misc.end = NULL;
}
+ if (!rq->misc.start)
+ ERTS_RUNQ_FLGS_UNSET_NOB(rq, ERTS_RUNQ_FLG_MISC_OP);
+
erts_smp_runq_unlock(rq);
while (molp) {
@@ -11356,6 +12077,7 @@ static void early_init_process_struct(void *varg, Eterm data)
proc->common.id = make_internal_pid(data);
#ifdef ERTS_DIRTY_SCHEDULERS
erts_smp_atomic32_init_nob(&proc->dirty_state, 0);
+ proc->dirty_sys_tasks = NULL;
#endif
erts_smp_atomic32_init_relb(&proc->state, arg->state);
@@ -11864,6 +12586,7 @@ void erts_init_empty_process(Process *p)
#ifdef ERTS_DIRTY_SCHEDULERS
erts_smp_atomic32_init_nob(&p->dirty_state, 0);
+ p->dirty_sys_tasks = NULL;
#endif
erts_smp_atomic32_init_nob(&p->state, (erts_aint32_t) PRIORITY_NORMAL);
@@ -11975,7 +12698,6 @@ delete_process(Process* p)
ErtsPSD *psd;
struct saved_calls *scb;
process_breakpoint_time_t *pbt;
- void *nif_export;
VERBOSE(DEBUG_PROCESSES, ("Removing process: %T\n",p->common.id));
VERBOSE(DEBUG_SHCOPY, ("[pid=%T] delete process: %p %p %p %p\n", p->common.id,
@@ -11992,9 +12714,7 @@ delete_process(Process* p)
if (pbt)
erts_free(ERTS_ALC_T_BPD, (void *) pbt);
- nif_export = ERTS_PROC_SET_NIF_TRAP_EXPORT(p, NULL);
- if (nif_export)
- erts_destroy_nif_export(nif_export);
+ erts_destroy_nif_export(p);
/* Cleanup psd */
@@ -12438,7 +13158,9 @@ send_exit_signal(Process *c_p, /* current process if and only
}
set_proc_exiting(c_p, state, rsn, NULL);
}
- else if (!(state & (ERTS_PSFLG_RUNNING|ERTS_PSFLG_RUNNING_SYS))) {
+ else if (!(state & (ERTS_PSFLG_RUNNING
+ | ERTS_PSFLG_RUNNING_SYS
+ | ERTS_PSFLG_DIRTY_RUNNING_SYS))) {
/* Process not running ... */
ErtsProcLocks need_locks = ~(*rp_locks) & ERTS_PROC_LOCKS_ALL;
ErlHeapFragment *bp = NULL;
@@ -12465,8 +13187,7 @@ send_exit_signal(Process *c_p, /* current process if and only
ErlOffHeap *ohp;
Uint rsn_sz = size_object(rsn);
#ifdef ERTS_DIRTY_SCHEDULERS
- if (state & (ERTS_PSFLG_DIRTY_RUNNING
- | ERTS_PSFLG_DIRTY_RUNNING_SYS)) {
+ if (state & ERTS_PSFLG_DIRTY_RUNNING) {
bp = new_message_buffer(rsn_sz);
ohp = &bp->off_heap;
hp = &bp->mem[0];
@@ -12483,7 +13204,7 @@ send_exit_signal(Process *c_p, /* current process if and only
set_proc_exiting(rp, state, rsn_cpy, bp);
}
else { /* Process running... */
-
+
/*
* The pending exit will be discovered when the process
* is scheduled out if not discovered earlier.
@@ -12505,8 +13226,35 @@ send_exit_signal(Process *c_p, /* current process if and only
&bp->off_heap);
rp->pending_exit.bp = bp;
}
- erts_smp_atomic32_read_bor_relb(&rp->state,
- ERTS_PSFLG_PENDING_EXIT);
+
+ /*
+ * If no dirty work has been scheduled, pending exit will
+ * be discovered when the process is scheduled. If dirty work
+ * has been scheduled, we may need to add it to a normal run
+ * queue...
+ */
+#ifndef ERTS_DIRTY_SCHEDULERS
+ (void) erts_smp_atomic32_read_bor_relb(&rp->state,
+ ERTS_PSFLG_PENDING_EXIT);
+#else
+ {
+ erts_aint32_t a = erts_smp_atomic32_read_nob(&rp->state);
+ while (1) {
+ erts_aint32_t n, e;
+ int dwork;
+ n = e = a;
+ n |= ERTS_PSFLG_PENDING_EXIT;
+ dwork = !!(n & ERTS_PSFLGS_DIRTY_WORK);
+ n &= ~ERTS_PSFLGS_DIRTY_WORK;
+ a = erts_smp_atomic32_cmpxchg_mb(&rp->state, n, e);
+ if (a == e) {
+ if (dwork)
+ erts_schedule_process(rp, n, *rp_locks);
+ break;
+ }
+ }
+ }
+#endif
}
}
/* else:
@@ -13356,6 +14104,8 @@ erts_print_scheduler_info(fmtfn_t to, void *to_arg, ErtsSchedulerData *esdp) {
erts_print(to, to_arg, "WAITING"); break;
case ERTS_SSI_FLG_SUSPENDED:
erts_print(to, to_arg, "SUSPENDED"); break;
+ case ERTS_SSI_FLG_MSB_EXEC:
+ erts_print(to, to_arg, "MSB_EXEC"); break;
default:
erts_print(to, to_arg, "UNKNOWN(%d)", flg); break;
}
@@ -13465,6 +14215,12 @@ erts_print_scheduler_info(fmtfn_t to, void *to_arg, ErtsSchedulerData *esdp) {
erts_print(to, to_arg, "NONEMPTY"); break;
case ERTS_RUNQ_FLG_PROTECTED:
erts_print(to, to_arg, "PROTECTED"); break;
+ case ERTS_RUNQ_FLG_EXEC:
+ erts_print(to, to_arg, "EXEC"); break;
+ case ERTS_RUNQ_FLG_MSB_EXEC:
+ erts_print(to, to_arg, "MSB_EXEC"); break;
+ case ERTS_RUNQ_FLG_MISC_OP:
+ erts_print(to, to_arg, "MISC_OP"); break;
default:
erts_print(to, to_arg, "UNKNOWN(%d)", flg); break;
}
@@ -13514,11 +14270,11 @@ erts_print_scheduler_info(fmtfn_t to, void *to_arg, ErtsSchedulerData *esdp) {
* A nice system halt closing all open port goes as follows:
* 1) This function schedules the aux work ERTS_SSI_AUX_WORK_REAP_PORTS
* on all schedulers, then schedules itself out.
- * 2) All shedulers detect this and set the flag halt_in_progress
+ * 2) All shedulers detect this and set the flag ERTS_RUNQ_FLG_HALTING
* on their run queue. The last scheduler sets all non-closed ports
* ERTS_PORT_SFLG_HALT. Global atomic erts_halt_progress is used
* as refcount to determine which is last.
- * 3) While the run ques has flag halt_in_progress no processes
+ * 3) While the run queues has flag ERTS_RUNQ_FLG_HALTING no processes
* will be scheduled, only ports.
* 4) When the last port closes that scheduler calls erlang:halt/1.
* The same global atomic is used as refcount.
@@ -13533,8 +14289,8 @@ void erts_halt(int code)
erts_no_schedulers,
-1)) {
#ifdef ERTS_DIRTY_SCHEDULERS
- ERTS_DIRTY_CPU_RUNQ->halt_in_progress = 1;
- ERTS_DIRTY_IO_RUNQ->halt_in_progress = 1;
+ ERTS_RUNQ_FLGS_SET(ERTS_DIRTY_CPU_RUNQ, ERTS_RUNQ_FLG_HALTING);
+ ERTS_RUNQ_FLGS_SET(ERTS_DIRTY_IO_RUNQ, ERTS_RUNQ_FLG_HALTING);
#endif
erts_halt_code = code;
notify_reap_ports_relb();
@@ -13548,6 +14304,9 @@ erts_dbg_check_halloc_lock(Process *p)
ErtsSchedulerData *esdp;
if (ERTS_PROC_LOCK_MAIN & erts_proc_lc_my_proc_locks(p))
return 1;
+ if ((p->static_flags & ERTS_STC_FLG_SHADOW_PROC)
+ && ERTS_SCHEDULER_IS_DIRTY(erts_get_scheduler_data()))
+ return 1;
if (p->common.id == ERTS_INVALID_PID)
return 1;
esdp = erts_proc_sched_data(p);