1 // Copyright 2009 The Go Authors. All rights reserved.
2 // Use of this source code is governed by a BSD-style
3 // license that can be found in the LICENSE file.
17 #ifdef USING_SPLIT_STACK
19 /* FIXME: These are not declared anywhere. */
21 extern void __splitstack_getcontext(void *context
[10]);
23 extern void __splitstack_setcontext(void *context
[10]);
25 extern void *__splitstack_makecontext(size_t, void *context
[10], size_t *);
27 extern void * __splitstack_resetcontext(void *context
[10], size_t *);
29 extern void *__splitstack_find(void *, void *, size_t *, void **, void **,
32 extern void __splitstack_block_signals (int *, int *);
34 extern void __splitstack_block_signals_context (void *context
[10], int *,
39 #if defined(USING_SPLIT_STACK) && defined(LINKER_SUPPORTS_SPLIT_STACK)
40 # ifdef PTHREAD_STACK_MIN
41 # define StackMin PTHREAD_STACK_MIN
43 # define StackMin 8192
46 # define StackMin 2 * 1024 * 1024
49 static void schedule(G
*);
51 typedef struct Sched Sched
;
54 G runtime_g0
; // idle goroutine for m0
63 #ifndef SETCONTEXT_CLOBBERS_TLS
71 fixcontext(ucontext_t
*c
__attribute__ ((unused
)))
77 # if defined(__x86_64__) && defined(__sun__)
79 // x86_64 Solaris 10 and 11 have a bug: setcontext switches the %fs
80 // register to that of the thread which called getcontext. The effect
81 // is that the address of all __thread variables changes. This bug
82 // also affects pthread_self() and pthread_getspecific. We work
83 // around it by clobbering the context field directly to keep %fs the
86 static __thread greg_t fs
;
94 fs
= c
.uc_mcontext
.gregs
[REG_FSBASE
];
98 fixcontext(ucontext_t
* c
)
100 c
->uc_mcontext
.gregs
[REG_FSBASE
] = fs
;
105 # error unknown case for SETCONTEXT_CLOBBERS_TLS
111 // We can not always refer to the TLS variables directly. The
112 // compiler will call tls_get_addr to get the address of the variable,
113 // and it may hold it in a register across a call to schedule. When
114 // we get back from the call we may be running in a different thread,
115 // in which case the register now points to the TLS variable for a
116 // different thread. We use non-inlinable functions to avoid this
119 G
* runtime_g(void) __attribute__ ((noinline
, no_split_stack
));
127 M
* runtime_m(void) __attribute__ ((noinline
, no_split_stack
));
135 int32 runtime_gcwaiting
;
139 // The go scheduler's job is to match ready-to-run goroutines (`g's)
140 // with waiting-for-work schedulers (`m's). If there are ready g's
141 // and no waiting m's, ready() will start a new m running in a new
142 // OS thread, so that all ready g's can run simultaneously, up to a limit.
143 // For now, m's never go away.
145 // By default, Go keeps only one kernel thread (m) running user code
146 // at a single time; other threads may be blocked in the operating system.
147 // Setting the environment variable $GOMAXPROCS or calling
148 // runtime.GOMAXPROCS() will change the number of user threads
149 // allowed to execute simultaneously. $GOMAXPROCS is thus an
150 // approximation of the maximum number of cores to use.
152 // Even a program that can run without deadlock in a single process
153 // might use more m's if given the chance. For example, the prime
154 // sieve will use as many m's as there are primes (up to runtime_sched.mmax),
155 // allowing different stages of the pipeline to execute in parallel.
156 // We could revisit this choice, only kicking off new m's for blocking
157 // system calls, but that would limit the amount of parallel computation
158 // that go would try to do.
160 // In general, one could imagine all sorts of refinements to the
161 // scheduler, but the goal now is just to get something working on
167 G
*gfree
; // available g's (status == Gdead)
170 G
*ghead
; // g's waiting to run
172 int32 gwait
; // number of g's waiting to run
173 int32 gcount
; // number of g's that are alive
174 int32 grunning
; // number of g's running on cpu or in syscall
176 M
*mhead
; // m's waiting for work
177 int32 mwait
; // number of m's waiting for work
178 int32 mcount
; // number of m's that have been created
180 volatile uint32 atomic
; // atomic scheduling word (see below)
182 int32 profilehz
; // cpu profiling rate
184 bool init
; // running initialization
185 bool lockmain
; // init called runtime.LockOSThread
187 Note stopped
; // one g can set waitstop and wait here for m's to stop
190 // The atomic word in sched is an atomic uint32 that
191 // holds these fields.
193 // [15 bits] mcpu number of m's executing on cpu
194 // [15 bits] mcpumax max number of m's allowed on cpu
195 // [1 bit] waitstop some g is waiting on stopped
196 // [1 bit] gwaiting gwait != 0
198 // These fields are the information needed by entersyscall
199 // and exitsyscall to decide whether to coordinate with the
200 // scheduler. Packing them into a single machine word lets
201 // them use a fast path with a single atomic read/write and
202 // no lock/unlock. This greatly reduces contention in
203 // syscall- or cgo-heavy multithreaded programs.
205 // Except for entersyscall and exitsyscall, the manipulations
206 // to these fields only happen while holding the schedlock,
207 // so the routines holding schedlock only need to worry about
208 // what entersyscall and exitsyscall do, not the other routines
209 // (which also use the schedlock).
211 // In particular, entersyscall and exitsyscall only read mcpumax,
212 // waitstop, and gwaiting. They never write them. Thus, writes to those
213 // fields can be done (holding schedlock) without fear of write conflicts.
214 // There may still be logic conflicts: for example, the set of waitstop must
215 // be conditioned on mcpu >= mcpumax or else the wait may be a
216 // spurious sleep. The Promela model in proc.p verifies these accesses.
219 mcpuMask
= (1<<mcpuWidth
) - 1,
221 mcpumaxShift
= mcpuShift
+ mcpuWidth
,
222 waitstopShift
= mcpumaxShift
+ mcpuWidth
,
223 gwaitingShift
= waitstopShift
+1,
225 // The max value of GOMAXPROCS is constrained
226 // by the max value we can store in the bit fields
227 // of the atomic word. Reserve a few high values
228 // so that we can detect accidental decrement
230 maxgomaxprocs
= mcpuMask
- 10,
233 #define atomic_mcpu(v) (((v)>>mcpuShift)&mcpuMask)
234 #define atomic_mcpumax(v) (((v)>>mcpumaxShift)&mcpuMask)
235 #define atomic_waitstop(v) (((v)>>waitstopShift)&1)
236 #define atomic_gwaiting(v) (((v)>>gwaitingShift)&1)
239 int32 runtime_gomaxprocs
;
240 bool runtime_singleproc
;
242 static bool canaddmcpu(void);
244 // An m that is waiting for notewakeup(&m->havenextg). This may
245 // only be accessed while the scheduler lock is held. This is used to
246 // minimize the number of times we call notewakeup while the scheduler
247 // lock is held, since the m will normally move quickly to lock the
248 // scheduler itself, producing lock contention.
251 // Scheduling helpers. Sched must be locked.
252 static void gput(G
*); // put/get on ghead/gtail
253 static G
* gget(void);
254 static void mput(M
*); // put/get on mhead
256 static void gfput(G
*); // put/get on gfree
257 static G
* gfget(void);
258 static void matchmg(void); // match m's to g's
259 static void readylocked(G
*); // ready, but sched is locked
260 static void mnextg(M
*, G
*);
261 static void mcommoninit(M
*);
269 v
= runtime_sched
.atomic
;
271 w
&= ~(mcpuMask
<<mcpumaxShift
);
272 w
|= n
<<mcpumaxShift
;
273 if(runtime_cas(&runtime_sched
.atomic
, v
, w
))
278 // First function run by a new goroutine. This replaces gogocall.
284 fn
= (void (*)(void*))(g
->entry
);
289 // Switch context to a different goroutine. This is like longjmp.
290 static void runtime_gogo(G
*) __attribute__ ((noinline
));
292 runtime_gogo(G
* newg
)
294 #ifdef USING_SPLIT_STACK
295 __splitstack_setcontext(&newg
->stack_context
[0]);
298 newg
->fromgogo
= true;
299 fixcontext(&newg
->context
);
300 setcontext(&newg
->context
);
301 runtime_throw("gogo setcontext returned");
304 // Save context and call fn passing g as a parameter. This is like
305 // setjmp. Because getcontext always returns 0, unlike setjmp, we use
306 // g->fromgogo as a code. It will be true if we got here via
307 // setcontext. g == nil the first time this is called in a new m.
308 static void runtime_mcall(void (*)(G
*)) __attribute__ ((noinline
));
310 runtime_mcall(void (*pfn
)(G
*))
314 #ifndef USING_SPLIT_STACK
318 // Ensure that all registers are on the stack for the garbage
320 __builtin_unwind_init();
325 runtime_throw("runtime: mcall called on m->g0 stack");
329 #ifdef USING_SPLIT_STACK
330 __splitstack_getcontext(&g
->stack_context
[0]);
334 gp
->fromgogo
= false;
335 getcontext(&gp
->context
);
337 // When we return from getcontext, we may be running
338 // in a new thread. That means that m and g may have
339 // changed. They are global variables so we will
340 // reload them, but the addresses of m and g may be
341 // cached in our local stack frame, and those
342 // addresses may be wrong. Call functions to reload
343 // the values for this thread.
347 if (gp
== nil
|| !gp
->fromgogo
) {
348 #ifdef USING_SPLIT_STACK
349 __splitstack_setcontext(&mp
->g0
->stack_context
[0]);
351 mp
->g0
->entry
= (byte
*)pfn
;
354 // It's OK to set g directly here because this case
355 // can not occur if we got here via a setcontext to
356 // the getcontext call just above.
359 fixcontext(&mp
->g0
->context
);
360 setcontext(&mp
->g0
->context
);
361 runtime_throw("runtime: mcall function returned");
365 // Keep trace of scavenger's goroutine for deadlock detection.
368 // The bootstrap sequence is:
372 // make & queue new G
373 // call runtime_mstart
375 // The new G calls runtime_main.
377 runtime_schedinit(void)
391 runtime_mallocinit();
398 // Allocate internal symbol table representation now,
399 // so that we don't need to call malloc when we crash.
400 // runtime_findfunc(0);
402 runtime_gomaxprocs
= 1;
403 p
= runtime_getenv("GOMAXPROCS");
404 if(p
!= nil
&& (n
= runtime_atoi(p
)) != 0) {
405 if(n
> maxgomaxprocs
)
407 runtime_gomaxprocs
= n
;
409 // wait for the main goroutine to start before taking
410 // GOMAXPROCS into account.
412 runtime_singleproc
= runtime_gomaxprocs
== 1;
414 canaddmcpu(); // mcpu++ to account for bootstrap m
415 m
->helpgc
= 1; // flag to tell schedule() to mcpu--
416 runtime_sched
.grunning
++;
418 // Can not enable GC until all roots are registered.
419 // mstats.enablegc = 1;
423 extern void main_init(void) __asm__ ("__go_init_main");
424 extern void main_main(void) __asm__ ("main.main");
426 // The main goroutine.
430 // Lock the main goroutine onto this, the main OS thread,
431 // during initialization. Most programs won't care, but a few
432 // do require certain calls to be made by the main thread.
433 // Those can arrange for main.main to run in the main thread
434 // by calling runtime.LockOSThread during initialization
435 // to preserve the lock.
436 runtime_LockOSThread();
437 // From now on, newgoroutines may use non-main threads.
438 setmcpumax(runtime_gomaxprocs
);
439 runtime_sched
.init
= true;
440 scvg
= __go_go(runtime_MHeap_Scavenger
, nil
);
442 runtime_sched
.init
= false;
443 if(!runtime_sched
.lockmain
)
444 runtime_UnlockOSThread();
446 // For gccgo we have to wait until after main is initialized
447 // to enable GC, because initializing main registers the GC
451 // The deadlock detection has false negatives.
452 // Let scvg start up, to eliminate the false negative
453 // for the trivial program func main() { select{} }.
462 // Lock the scheduler.
466 runtime_lock(&runtime_sched
);
469 // Unlock the scheduler.
477 runtime_unlock(&runtime_sched
);
479 runtime_notewakeup(&m
->havenextg
);
485 g
->status
= Gmoribund
;
490 runtime_goroutineheader(G
*g
)
509 status
= g
->waitreason
;
520 runtime_printf("goroutine %d [%s]:\n", g
->goid
, status
);
524 runtime_tracebackothers(G
*me
)
528 for(g
= runtime_allg
; g
!= nil
; g
= g
->alllink
) {
529 if(g
== me
|| g
->status
== Gdead
)
531 runtime_printf("\n");
532 runtime_goroutineheader(g
);
533 // runtime_traceback(g->sched.pc, g->sched.sp, 0, g);
537 // Mark this g as m's idle goroutine.
538 // This functionality might be used in environments where programs
539 // are limited to a single thread, to simulate a select-driven
540 // network server. It is not exposed via the standard runtime API.
542 runtime_idlegoroutine(void)
545 runtime_throw("g is already an idle goroutine");
552 m
->id
= runtime_sched
.mcount
++;
553 m
->fastrand
= 0x49f6428aUL
+ m
->id
+ runtime_cputicks();
556 m
->mcache
= runtime_allocmcache();
558 runtime_callers(1, m
->createstack
, nelem(m
->createstack
));
560 // Add to runtime_allm so garbage collector doesn't free m
561 // when it is just in a register or thread-local storage.
562 m
->alllink
= runtime_allm
;
563 // runtime_NumCgoCall() iterates over allm w/o schedlock,
564 // so we need to publish it safely.
565 runtime_atomicstorep(&runtime_allm
, m
);
568 // Try to increment mcpu. Report whether succeeded.
575 v
= runtime_sched
.atomic
;
576 if(atomic_mcpu(v
) >= atomic_mcpumax(v
))
578 if(runtime_cas(&runtime_sched
.atomic
, v
, v
+(1<<mcpuShift
)))
583 // Put on `g' queue. Sched must be locked.
589 // If g is wired, hand it off directly.
590 if((m
= g
->lockedm
) != nil
&& canaddmcpu()) {
595 // If g is the idle goroutine for an m, hand it off.
596 if(g
->idlem
!= nil
) {
597 if(g
->idlem
->idleg
!= nil
) {
598 runtime_printf("m%d idle out of sync: g%d g%d\n",
600 g
->idlem
->idleg
->goid
, g
->goid
);
601 runtime_throw("runtime: double idle");
608 if(runtime_sched
.ghead
== nil
)
609 runtime_sched
.ghead
= g
;
611 runtime_sched
.gtail
->schedlink
= g
;
612 runtime_sched
.gtail
= g
;
615 // if it transitions to nonzero, set atomic gwaiting bit.
616 if(runtime_sched
.gwait
++ == 0)
617 runtime_xadd(&runtime_sched
.atomic
, 1<<gwaitingShift
);
620 // Report whether gget would return something.
624 return runtime_sched
.ghead
!= nil
|| m
->idleg
!= nil
;
627 // Get from `g' queue. Sched must be locked.
633 g
= runtime_sched
.ghead
;
635 runtime_sched
.ghead
= g
->schedlink
;
636 if(runtime_sched
.ghead
== nil
)
637 runtime_sched
.gtail
= nil
;
639 // if it transitions to zero, clear atomic gwaiting bit.
640 if(--runtime_sched
.gwait
== 0)
641 runtime_xadd(&runtime_sched
.atomic
, -1<<gwaitingShift
);
642 } else if(m
->idleg
!= nil
) {
649 // Put on `m' list. Sched must be locked.
653 m
->schedlink
= runtime_sched
.mhead
;
654 runtime_sched
.mhead
= m
;
655 runtime_sched
.mwait
++;
658 // Get an `m' to run `g'. Sched must be locked.
664 // if g has its own m, use it.
665 if(g
&& (m
= g
->lockedm
) != nil
)
668 // otherwise use general m pool.
669 if((m
= runtime_sched
.mhead
) != nil
){
670 runtime_sched
.mhead
= m
->schedlink
;
671 runtime_sched
.mwait
--;
676 // Mark g ready to run.
685 // Mark g ready to run. Sched is already locked.
686 // G might be running already and about to stop.
687 // The sched lock protects g->status from changing underfoot.
692 // Running on another machine.
693 // Ready it when it stops.
699 if(g
->status
== Grunnable
|| g
->status
== Grunning
) {
700 runtime_printf("goroutine %d has status %d\n", g
->goid
, g
->status
);
701 runtime_throw("bad g->status in ready");
703 g
->status
= Grunnable
;
709 // Same as readylocked but a different symbol so that
710 // debuggers can set a breakpoint here and catch all
713 newprocreadylocked(G
*g
)
718 // Pass g to m for running.
719 // Caller has already incremented mcpu.
723 runtime_sched
.grunning
++;
728 runtime_notewakeup(&mwakeup
->havenextg
);
733 // Get the next goroutine that m should run.
734 // Sched must be locked on entry, is unlocked on exit.
735 // Makes sure that at most $GOMAXPROCS g's are
736 // running on cpus (not in system calls) at any given time.
744 if(atomic_mcpu(runtime_sched
.atomic
) >= maxgomaxprocs
)
745 runtime_throw("negative mcpu");
747 // If there is a g waiting as m->nextg, the mcpu++
748 // happened before it was passed to mnextg.
749 if(m
->nextg
!= nil
) {
756 if(m
->lockedg
!= nil
) {
757 // We can only run one g, and it's not available.
758 // Make sure some other cpu is running to handle
759 // the ordinary run queue.
760 if(runtime_sched
.gwait
!= 0) {
762 // m->lockedg might have been on the queue.
763 if(m
->nextg
!= nil
) {
771 // Look for work on global queue.
772 while(haveg() && canaddmcpu()) {
775 runtime_throw("gget inconsistency");
778 mnextg(gp
->lockedm
, gp
);
781 runtime_sched
.grunning
++;
786 // The while loop ended either because the g queue is empty
787 // or because we have maxed out our m procs running go
788 // code (mcpu >= mcpumax). We need to check that
789 // concurrent actions by entersyscall/exitsyscall cannot
790 // invalidate the decision to end the loop.
792 // We hold the sched lock, so no one else is manipulating the
793 // g queue or changing mcpumax. Entersyscall can decrement
794 // mcpu, but if does so when there is something on the g queue,
795 // the gwait bit will be set, so entersyscall will take the slow path
796 // and use the sched lock. So it cannot invalidate our decision.
798 // Wait on global m queue.
802 // Look for deadlock situation.
803 // There is a race with the scavenger that causes false negatives:
804 // if the scavenger is just starting, then we have
805 // scvg != nil && grunning == 0 && gwait == 0
806 // and we do not detect a deadlock. It is possible that we should
807 // add that case to the if statement here, but it is too close to Go 1
808 // to make such a subtle change. Instead, we work around the
809 // false negative in trivial programs by calling runtime.gosched
810 // from the main goroutine just before main.main.
811 // See runtime_main above.
813 // On a related note, it is also possible that the scvg == nil case is
814 // wrong and should include gwait, but that does not happen in
815 // standard Go programs, which all start the scavenger.
817 if((scvg
== nil
&& runtime_sched
.grunning
== 0) ||
818 (scvg
!= nil
&& runtime_sched
.grunning
== 1 && runtime_sched
.gwait
== 0 &&
819 (scvg
->status
== Grunning
|| scvg
->status
== Gsyscall
))) {
820 runtime_throw("all goroutines are asleep - deadlock!");
825 runtime_noteclear(&m
->havenextg
);
827 // Stoptheworld is waiting for all but its cpu to go to stop.
828 // Entersyscall might have decremented mcpu too, but if so
829 // it will see the waitstop and take the slow path.
830 // Exitsyscall never increments mcpu beyond mcpumax.
831 v
= runtime_atomicload(&runtime_sched
.atomic
);
832 if(atomic_waitstop(v
) && atomic_mcpu(v
) <= atomic_mcpumax(v
)) {
833 // set waitstop = 0 (known to be 1)
834 runtime_xadd(&runtime_sched
.atomic
, -1<<waitstopShift
);
835 runtime_notewakeup(&runtime_sched
.stopped
);
839 runtime_notesleep(&m
->havenextg
);
843 runtime_lock(&runtime_sched
);
846 if((gp
= m
->nextg
) == nil
)
847 runtime_throw("bad m->nextg in nextgoroutine");
853 runtime_helpgc(bool *extra
)
858 // Figure out how many CPUs to use.
859 // Limited by gomaxprocs, number of actual CPUs, and MaxGcproc.
860 max
= runtime_gomaxprocs
;
861 if(max
> runtime_ncpu
)
862 max
= runtime_ncpu
> 0 ? runtime_ncpu
: 1;
866 // We're going to use one CPU no matter what.
867 // Figure out the max number of additional CPUs.
870 runtime_lock(&runtime_sched
);
872 while(n
< max
&& (mp
= mget(nil
)) != nil
) {
876 runtime_notewakeup(&mp
->havenextg
);
878 runtime_unlock(&runtime_sched
);
885 runtime_stoptheworld(void)
890 runtime_gcwaiting
= 1;
896 v
= runtime_sched
.atomic
;
897 if(atomic_mcpu(v
) <= 1)
900 // It would be unsafe for multiple threads to be using
901 // the stopped note at once, but there is only
902 // ever one thread doing garbage collection.
903 runtime_noteclear(&runtime_sched
.stopped
);
904 if(atomic_waitstop(v
))
905 runtime_throw("invalid waitstop");
907 // atomic { waitstop = 1 }, predicated on mcpu <= 1 check above
909 if(!runtime_cas(&runtime_sched
.atomic
, v
, v
+(1<<waitstopShift
)))
913 runtime_notesleep(&runtime_sched
.stopped
);
916 runtime_singleproc
= runtime_gomaxprocs
== 1;
921 runtime_starttheworld(bool extra
)
926 runtime_gcwaiting
= 0;
927 setmcpumax(runtime_gomaxprocs
);
929 if(extra
&& canaddmcpu()) {
930 // Start a new m that will (we hope) be idle
931 // and so available to help when the next
932 // garbage collection happens.
933 // canaddmcpu above did mcpu++
934 // (necessary, because m will be doing various
935 // initialization work so is definitely running),
936 // but m is not running a specific goroutine,
937 // so set the helpgc flag as a signal to m's
938 // first schedule(nil) to mcpu-- and grunning--.
941 runtime_sched
.grunning
++;
946 // Called to start an M.
948 runtime_mstart(void* mp
)
958 // Record top of stack for use by mcall.
959 // Once we call schedule we're never coming back,
960 // so other calls can reuse this stack space.
961 #ifdef USING_SPLIT_STACK
962 __splitstack_getcontext(&g
->stack_context
[0]);
964 g
->gcinitial_sp
= &mp
;
965 // Setting gcstack_size to 0 is a marker meaning that gcinitial_sp
966 // is the top of the stack, not the bottom.
970 getcontext(&g
->context
);
972 if(g
->entry
!= nil
) {
973 // Got here from mcall.
974 void (*pfn
)(G
*) = (void (*)(G
*))g
->entry
;
975 G
* gp
= (G
*)g
->param
;
981 #ifdef USING_SPLIT_STACK
983 int dont_block_signals
= 0;
984 __splitstack_block_signals(&dont_block_signals
, nil
);
988 // Install signal handlers; after minit so that minit can
989 // prepare the thread to be able to handle the signals.
997 typedef struct CgoThreadStart CgoThreadStart
;
998 struct CgoThreadStart
1005 // Kick off new m's as needed (up to mcpumax).
1013 if(m
->mallocing
|| m
->gcing
)
1016 while(haveg() && canaddmcpu()) {
1019 runtime_throw("gget inconsistency");
1021 // Find the m that will run gp.
1022 if((mp
= mget(gp
)) == nil
)
1023 mp
= runtime_newm();
1028 // Create a new m. It will start off with a call to runtime_mstart.
1033 pthread_attr_t attr
;
1036 m
= runtime_malloc(sizeof(M
));
1038 m
->g0
= runtime_malg(-1, nil
, nil
);
1040 if(pthread_attr_init(&attr
) != 0)
1041 runtime_throw("pthread_attr_init");
1042 if(pthread_attr_setdetachstate(&attr
, PTHREAD_CREATE_DETACHED
) != 0)
1043 runtime_throw("pthread_attr_setdetachstate");
1045 #ifndef PTHREAD_STACK_MIN
1046 #define PTHREAD_STACK_MIN 8192
1048 if(pthread_attr_setstacksize(&attr
, PTHREAD_STACK_MIN
) != 0)
1049 runtime_throw("pthread_attr_setstacksize");
1051 if(pthread_create(&tid
, &attr
, runtime_mstart
, m
) != 0)
1052 runtime_throw("pthread_create");
1057 // One round of scheduler: find a goroutine and run it.
1058 // The argument is the goroutine that was running before
1059 // schedule was called, or nil if this is the first call.
1069 // Just finished running gp.
1071 runtime_sched
.grunning
--;
1073 // atomic { mcpu-- }
1074 v
= runtime_xadd(&runtime_sched
.atomic
, -1<<mcpuShift
);
1075 if(atomic_mcpu(v
) > maxgomaxprocs
)
1076 runtime_throw("negative mcpu in scheduler");
1081 // Shouldn't have been running!
1082 runtime_throw("bad gp->status in sched");
1084 gp
->status
= Grunnable
;
1095 if(--runtime_sched
.gcount
== 0)
1099 if(gp
->readyonstop
){
1100 gp
->readyonstop
= 0;
1103 } else if(m
->helpgc
) {
1104 // Bootstrap m or new m started by starttheworld.
1105 // atomic { mcpu-- }
1106 v
= runtime_xadd(&runtime_sched
.atomic
, -1<<mcpuShift
);
1107 if(atomic_mcpu(v
) > maxgomaxprocs
)
1108 runtime_throw("negative mcpu in scheduler");
1109 // Compensate for increment in starttheworld().
1110 runtime_sched
.grunning
--;
1112 } else if(m
->nextg
!= nil
) {
1113 // New m started by matchmg.
1115 runtime_throw("invalid m state in scheduler");
1118 // Find (or wait for) g to run. Unlocks runtime_sched.
1119 gp
= nextgandunlock();
1120 gp
->readyonstop
= 0;
1121 gp
->status
= Grunning
;
1125 // Check whether the profiler needs to be turned on or off.
1126 hz
= runtime_sched
.profilehz
;
1127 if(m
->profilehz
!= hz
)
1128 runtime_resetcpuprofiler(hz
);
1133 // Enter scheduler. If g->status is Grunning,
1134 // re-queues g and runs everyone else who is waiting
1135 // before running g again. If g->status is Gmoribund,
1138 runtime_gosched(void)
1141 runtime_throw("gosched holding locks");
1143 runtime_throw("gosched of g0");
1144 runtime_mcall(schedule
);
1147 // The goroutine g is about to enter a system call.
1148 // Record that it's not using the cpu anymore.
1149 // This is called only from the go syscall library and cgocall,
1150 // not from the low-level system calls used by the runtime.
1152 // Entersyscall cannot split the stack: the runtime_gosave must
1153 // make g->sched refer to the caller's stack segment, because
1154 // entersyscall is going to return immediately after.
1155 // It's okay to call matchmg and notewakeup even after
1156 // decrementing mcpu, because we haven't released the
1157 // sched lock yet, so the garbage collector cannot be running.
1159 void runtime_entersyscall(void) __attribute__ ((no_split_stack
));
1162 runtime_entersyscall(void)
1166 if(m
->profilehz
> 0)
1167 runtime_setprof(false);
1169 // Leave SP around for gc and traceback.
1170 #ifdef USING_SPLIT_STACK
1171 g
->gcstack
= __splitstack_find(NULL
, NULL
, &g
->gcstack_size
,
1172 &g
->gcnext_segment
, &g
->gcnext_sp
,
1175 g
->gcnext_sp
= (byte
*) &v
;
1178 // Save the registers in the g structure so that any pointers
1179 // held in registers will be seen by the garbage collector.
1180 // We could use getcontext here, but setjmp is more efficient
1181 // because it doesn't need to save the signal mask.
1184 g
->status
= Gsyscall
;
1187 // The slow path inside the schedlock/schedunlock will get
1188 // through without stopping if it does:
1191 // waitstop && mcpu <= mcpumax not true
1192 // If we can do the same with a single atomic add,
1193 // then we can skip the locks.
1194 v
= runtime_xadd(&runtime_sched
.atomic
, -1<<mcpuShift
);
1195 if(!atomic_gwaiting(v
) && (!atomic_waitstop(v
) || atomic_mcpu(v
) > atomic_mcpumax(v
)))
1199 v
= runtime_atomicload(&runtime_sched
.atomic
);
1200 if(atomic_gwaiting(v
)) {
1202 v
= runtime_atomicload(&runtime_sched
.atomic
);
1204 if(atomic_waitstop(v
) && atomic_mcpu(v
) <= atomic_mcpumax(v
)) {
1205 runtime_xadd(&runtime_sched
.atomic
, -1<<waitstopShift
);
1206 runtime_notewakeup(&runtime_sched
.stopped
);
1212 // The goroutine g exited its system call.
1213 // Arrange for it to run on a cpu again.
1214 // This is called only from the go syscall library, not
1215 // from the low-level system calls used by the runtime.
1217 runtime_exitsyscall(void)
1223 // If we can do the mcpu++ bookkeeping and
1224 // find that we still have mcpu <= mcpumax, then we can
1225 // start executing Go code immediately, without having to
1226 // schedlock/schedunlock.
1228 v
= runtime_xadd(&runtime_sched
.atomic
, (1<<mcpuShift
));
1229 if(m
->profilehz
== runtime_sched
.profilehz
&& atomic_mcpu(v
) <= atomic_mcpumax(v
)) {
1230 // There's a cpu for us, so we can run.
1231 gp
->status
= Grunning
;
1232 // Garbage collector isn't running (since we are),
1233 // so okay to clear gcstack.
1234 #ifdef USING_SPLIT_STACK
1237 gp
->gcnext_sp
= nil
;
1238 runtime_memclr(gp
->gcregs
, sizeof gp
->gcregs
);
1240 if(m
->profilehz
> 0)
1241 runtime_setprof(true);
1245 // Tell scheduler to put g back on the run queue:
1246 // mostly equivalent to g->status = Grunning,
1247 // but keeps the garbage collector from thinking
1248 // that g is running right now, which it's not.
1249 gp
->readyonstop
= 1;
1251 // All the cpus are taken.
1252 // The scheduler will ready g and put this m to sleep.
1253 // When the scheduler takes g away from m,
1254 // it will undo the runtime_sched.mcpu++ above.
1257 // Gosched returned, so we're allowed to run now.
1258 // Delete the gcstack information that we left for
1259 // the garbage collector during the system call.
1260 // Must wait until now because until gosched returns
1261 // we don't know for sure that the garbage collector
1263 #ifdef USING_SPLIT_STACK
1266 gp
->gcnext_sp
= nil
;
1267 runtime_memclr(gp
->gcregs
, sizeof gp
->gcregs
);
1270 // Allocate a new g, with a stack big enough for stacksize bytes.
1272 runtime_malg(int32 stacksize
, byte
** ret_stack
, size_t* ret_stacksize
)
1276 newg
= runtime_malloc(sizeof(G
));
1277 if(stacksize
>= 0) {
1278 #if USING_SPLIT_STACK
1279 int dont_block_signals
= 0;
1281 *ret_stack
= __splitstack_makecontext(stacksize
,
1282 &newg
->stack_context
[0],
1284 __splitstack_block_signals_context(&newg
->stack_context
[0],
1285 &dont_block_signals
, nil
);
1287 *ret_stack
= runtime_mallocgc(stacksize
, FlagNoProfiling
|FlagNoGC
, 0, 0);
1288 *ret_stacksize
= stacksize
;
1289 newg
->gcinitial_sp
= *ret_stack
;
1290 newg
->gcstack_size
= stacksize
;
1296 /* For runtime package testing. */
1298 void runtime_testing_entersyscall(void)
1299 __asm__("libgo_runtime.runtime.entersyscall");
1302 runtime_testing_entersyscall()
1304 runtime_entersyscall();
1307 void runtime_testing_exitsyscall(void)
1308 __asm__("libgo_runtime.runtime.exitsyscall");
1311 runtime_testing_exitsyscall()
1313 runtime_exitsyscall();
1317 __go_go(void (*fn
)(void*), void* arg
)
1321 G
* volatile newg
; // volatile to avoid longjmp warning
1325 if((newg
= gfget()) != nil
){
1326 #ifdef USING_SPLIT_STACK
1327 int dont_block_signals
= 0;
1329 sp
= __splitstack_resetcontext(&newg
->stack_context
[0],
1331 __splitstack_block_signals_context(&newg
->stack_context
[0],
1332 &dont_block_signals
, nil
);
1334 sp
= newg
->gcinitial_sp
;
1335 spsize
= newg
->gcstack_size
;
1337 runtime_throw("bad spsize in __go_go");
1338 newg
->gcnext_sp
= sp
;
1341 newg
= runtime_malg(StackMin
, &sp
, &spsize
);
1342 if(runtime_lastg
== nil
)
1343 runtime_allg
= newg
;
1345 runtime_lastg
->alllink
= newg
;
1346 runtime_lastg
= newg
;
1348 newg
->status
= Gwaiting
;
1349 newg
->waitreason
= "new goroutine";
1351 newg
->entry
= (byte
*)fn
;
1353 newg
->gopc
= (uintptr
)__builtin_return_address(0);
1355 runtime_sched
.gcount
++;
1356 runtime_sched
.goidgen
++;
1357 newg
->goid
= runtime_sched
.goidgen
;
1360 runtime_throw("nil g->stack0");
1362 getcontext(&newg
->context
);
1363 newg
->context
.uc_stack
.ss_sp
= sp
;
1364 #ifdef MAKECONTEXT_STACK_TOP
1365 newg
->context
.uc_stack
.ss_sp
+= spsize
;
1367 newg
->context
.uc_stack
.ss_size
= spsize
;
1368 makecontext(&newg
->context
, kickoff
, 0);
1370 newprocreadylocked(newg
);
1374 //printf(" goid=%d\n", newg->goid);
1377 // Put on gfree list. Sched must be locked.
1381 g
->schedlink
= runtime_sched
.gfree
;
1382 runtime_sched
.gfree
= g
;
1385 // Get from gfree list. Sched must be locked.
1391 g
= runtime_sched
.gfree
;
1393 runtime_sched
.gfree
= g
->schedlink
;
1397 // Run all deferred functions for the current goroutine.
1403 while((d
= g
->defer
) != nil
) {
1410 g
->defer
= d
->__next
;
1415 void runtime_Goexit (void) asm ("libgo_runtime.runtime.Goexit");
1418 runtime_Goexit(void)
1424 void runtime_Gosched (void) asm ("libgo_runtime.runtime.Gosched");
1427 runtime_Gosched(void)
1432 // Implementation of runtime.GOMAXPROCS.
1433 // delete when scheduler is stronger
1435 runtime_gomaxprocsfunc(int32 n
)
1441 ret
= runtime_gomaxprocs
;
1444 if(n
> maxgomaxprocs
)
1446 runtime_gomaxprocs
= n
;
1447 if(runtime_gomaxprocs
> 1)
1448 runtime_singleproc
= false;
1449 if(runtime_gcwaiting
!= 0) {
1450 if(atomic_mcpumax(runtime_sched
.atomic
) != 1)
1451 runtime_throw("invalid mcpumax during gc");
1458 // If there are now fewer allowed procs
1459 // than procs running, stop.
1460 v
= runtime_atomicload(&runtime_sched
.atomic
);
1461 if((int32
)atomic_mcpu(v
) > n
) {
1466 // handle more procs
1473 runtime_LockOSThread(void)
1475 if(m
== &runtime_m0
&& runtime_sched
.init
) {
1476 runtime_sched
.lockmain
= true;
1484 runtime_UnlockOSThread(void)
1486 if(m
== &runtime_m0
&& runtime_sched
.init
) {
1487 runtime_sched
.lockmain
= false;
1495 runtime_lockedOSThread(void)
1497 return g
->lockedm
!= nil
&& m
->lockedg
!= nil
;
1500 // for testing of callbacks
1502 _Bool
runtime_golockedOSThread(void)
1503 asm("libgo_runtime.runtime.golockedOSThread");
1506 runtime_golockedOSThread(void)
1508 return runtime_lockedOSThread();
1511 // for testing of wire, unwire
1518 int32
runtime_NumGoroutine (void)
1519 __asm__ ("libgo_runtime.runtime.NumGoroutine");
1522 runtime_NumGoroutine()
1524 return runtime_sched
.gcount
;
1528 runtime_gcount(void)
1530 return runtime_sched
.gcount
;
1534 runtime_mcount(void)
1536 return runtime_sched
.mcount
;
1541 void (*fn
)(uintptr
*, int32
);
1546 // Called if we receive a SIGPROF signal.
1548 runtime_sigprof(uint8
*pc
__attribute__ ((unused
)),
1549 uint8
*sp
__attribute__ ((unused
)),
1550 uint8
*lr
__attribute__ ((unused
)),
1551 G
*gp
__attribute__ ((unused
)))
1555 if(prof
.fn
== nil
|| prof
.hz
== 0)
1558 runtime_lock(&prof
);
1559 if(prof
.fn
== nil
) {
1560 runtime_unlock(&prof
);
1563 // n = runtime_gentraceback(pc, sp, lr, gp, 0, prof.pcbuf, nelem(prof.pcbuf));
1565 // prof.fn(prof.pcbuf, n);
1566 runtime_unlock(&prof
);
1569 // Arrange to call fn with a traceback hz times a second.
1571 runtime_setcpuprofilerate(void (*fn
)(uintptr
*, int32
), int32 hz
)
1573 // Force sane arguments.
1581 // Stop profiler on this cpu so that it is safe to lock prof.
1582 // if a profiling signal came in while we had prof locked,
1583 // it would deadlock.
1584 runtime_resetcpuprofiler(0);
1586 runtime_lock(&prof
);
1589 runtime_unlock(&prof
);
1590 runtime_lock(&runtime_sched
);
1591 runtime_sched
.profilehz
= hz
;
1592 runtime_unlock(&runtime_sched
);
1595 runtime_resetcpuprofiler(hz
);