void
runtime·notetsleep(Note *n, int64 ns)
{
int64 deadline, now;
if(ns < 0) {
runtime·notesleep(n);
return;
}
if(runtime·atomicload(&n->key) != 0)
return;
if(m->profilehz > 0)
runtime·setprof(false);
deadline = runtime·nanotime() + ns;
for(;;) {
runtime·futexsleep(&n->key, 0, ns);
if(runtime·atomicload(&n->key) != 0)
break;
now = runtime·nanotime();
if(now >= deadline)
break;
ns = deadline - now;
}
if(m->profilehz > 0)
runtime·setprof(true);
}
void
runtime·entersyscall(void)
{
uint32 v;
if(m->profilehz > 0)
runtime·setprof(false);
// Leave SP around for gc and traceback.
runtime·gosave(&g->sched);
g->gcsp = g->sched.sp;
g->gcstack = g->stackbase;
g->gcguard = g->stackguard;
g->status = Gsyscall;
if(g->gcsp < g->gcguard-StackGuard || g->gcstack < g->gcsp) {
// runtime·printf("entersyscall inconsistent %p [%p,%p]\n",
// g->gcsp, g->gcguard-StackGuard, g->gcstack);
runtime·throw("entersyscall");
}
// Fast path.
// The slow path inside the schedlock/schedunlock will get
// through without stopping if it does:
// mcpu--
// gwait not true
// waitstop && mcpu <= mcpumax not true
// If we can do the same with a single atomic add,
// then we can skip the locks.
v = runtime·xadd(&runtime·sched.atomic, -1<<mcpuShift);
if(!atomic_gwaiting(v) && (!atomic_waitstop(v) || atomic_mcpu(v) > atomic_mcpumax(v)))
return;
schedlock();
v = runtime·atomicload(&runtime·sched.atomic);
if(atomic_gwaiting(v)) {
matchmg();
v = runtime·atomicload(&runtime·sched.atomic);
}
if(atomic_waitstop(v) && atomic_mcpu(v) <= atomic_mcpumax(v)) {
runtime·xadd(&runtime·sched.atomic, -1<<waitstopShift);
runtime·notewakeup(&runtime·sched.stopped);
}
// Re-save sched in case one of the calls
// (notewakeup, matchmg) triggered something using it.
runtime·gosave(&g->sched);
schedunlock();
}
void
runtime·notesleep(Note *n)
{
if(m->profilehz > 0)
runtime·setprof(false);
while(runtime·atomicload(&n->key) == 0)
runtime·futexsleep(&n->key, 0, -1);
if(m->profilehz > 0)
runtime·setprof(true);
}
runtime·findfunc(uintptr addr)
{
Func *f;
int32 nf, n;
// Use atomic double-checked locking,
// because when called from pprof signal
// handler, findfunc must run without
// grabbing any locks.
// (Before enabling the signal handler,
// SetCPUProfileRate calls findfunc to trigger
// the initialization outside the handler.)
// Avoid deadlock on fault during malloc
// by not calling buildfuncs if we're already in malloc.
if(!m->mallocing && !m->gcing) {
if(runtime·atomicload(&funcinit) == 0) {
runtime·lock(&funclock);
if(funcinit == 0) {
buildfuncs();
runtime·atomicstore(&funcinit, 1);
}
runtime·unlock(&funclock);
}
}
if(nfunc == 0)
return nil;
if(addr < func[0].entry || addr >= func[nfunc].entry)
return nil;
// binary search to find func with entry <= addr.
f = func;
nf = nfunc;
while(nf > 0) {
n = nf/2;
if(f[n].entry <= addr && addr < f[n+1].entry)
return &f[n];
else if(addr < f[n].entry)
nf = n;
else {
f += n+1;
nf -= n+1;
}
}
// can't get here -- we already checked above
// that the address was in the table bounds.
// this can only happen if the table isn't sorted
// by address or if the binary search above is buggy.
runtime·prints("findfunc unreachable\n");
return nil;
}
// Implementation of runtime.GOMAXPROCS.
// delete when scheduler is stronger
int32
runtime·gomaxprocsfunc(int32 n)
{
int32 ret;
uint32 v;
schedlock();
ret = runtime·gomaxprocs;
if(n <= 0)
n = ret;
if(n > maxgomaxprocs)
n = maxgomaxprocs;
runtime·gomaxprocs = n;
if(runtime·gomaxprocs > 1)
runtime·singleproc = false;
if(runtime·gcwaiting != 0) {
if(atomic_mcpumax(runtime·sched.atomic) != 1)
runtime·throw("invalid mcpumax during gc");
schedunlock();
return ret;
}
setmcpumax(n);
// If there are now fewer allowed procs
// than procs running, stop.
v = runtime·atomicload(&runtime·sched.atomic);
if(atomic_mcpu(v) > n) {
schedunlock();
runtime·gosched();
return ret;
}
// handle more procs
matchmg();
schedunlock();
return ret;
}
// Get the next goroutine that m should run.
// Sched must be locked on entry, is unlocked on exit.
// Makes sure that at most $GOMAXPROCS g's are
// running on cpus (not in system calls) at any given time.
static G*
nextgandunlock(void)
{
G *gp;
uint32 v;
top:
if(atomic_mcpu(runtime·sched.atomic) >= maxgomaxprocs)
runtime·throw("negative mcpu");
// If there is a g waiting as m->nextg, the mcpu++
// happened before it was passed to mnextg.
if(m->nextg != nil) {
gp = m->nextg;
m->nextg = nil;
schedunlock();
return gp;
}
if(m->lockedg != nil) {
// We can only run one g, and it's not available.
// Make sure some other cpu is running to handle
// the ordinary run queue.
if(runtime·sched.gwait != 0) {
matchmg();
// m->lockedg might have been on the queue.
if(m->nextg != nil) {
gp = m->nextg;
m->nextg = nil;
schedunlock();
return gp;
}
}
} else {
// Look for work on global queue.
while(haveg() && canaddmcpu()) {
gp = gget();
if(gp == nil)
runtime·throw("gget inconsistency");
if(gp->lockedm) {
mnextg(gp->lockedm, gp);
continue;
}
runtime·sched.grunning++;
schedunlock();
return gp;
}
// The while loop ended either because the g queue is empty
// or because we have maxed out our m procs running go
// code (mcpu >= mcpumax). We need to check that
// concurrent actions by entersyscall/exitsyscall cannot
// invalidate the decision to end the loop.
//
// We hold the sched lock, so no one else is manipulating the
// g queue or changing mcpumax. Entersyscall can decrement
// mcpu, but if does so when there is something on the g queue,
// the gwait bit will be set, so entersyscall will take the slow path
// and use the sched lock. So it cannot invalidate our decision.
//
// Wait on global m queue.
mput(m);
}
// Look for deadlock situation.
// There is a race with the scavenger that causes false negatives:
// if the scavenger is just starting, then we have
// scvg != nil && grunning == 0 && gwait == 0
// and we do not detect a deadlock. It is possible that we should
// add that case to the if statement here, but it is too close to Go 1
// to make such a subtle change. Instead, we work around the
// false negative in trivial programs by calling runtime.gosched
// from the main goroutine just before main.main.
// See runtime·main above.
//
// On a related note, it is also possible that the scvg == nil case is
// wrong and should include gwait, but that does not happen in
// standard Go programs, which all start the scavenger.
//
if((scvg == nil && runtime·sched.grunning == 0) ||
(scvg != nil && runtime·sched.grunning == 1 && runtime·sched.gwait == 0 &&
(scvg->status == Grunning || scvg->status == Gsyscall))) {
runtime·throw("all goroutines are asleep - deadlock!");
}
m->nextg = nil;
m->waitnextg = 1;
runtime·noteclear(&m->havenextg);
// Stoptheworld is waiting for all but its cpu to go to stop.
// Entersyscall might have decremented mcpu too, but if so
// it will see the waitstop and take the slow path.
// Exitsyscall never increments mcpu beyond mcpumax.
v = runtime·atomicload(&runtime·sched.atomic);
if(atomic_waitstop(v) && atomic_mcpu(v) <= atomic_mcpumax(v)) {
// set waitstop = 0 (known to be 1)
runtime·xadd(&runtime·sched.atomic, -1<<waitstopShift);
runtime·notewakeup(&runtime·sched.stopped);
}
schedunlock();
runtime·notesleep(&m->havenextg);
if(m->helpgc) {
runtime·gchelper();
m->helpgc = 0;
runtime·lock(&runtime·sched);
goto top;
}
if((gp = m->nextg) == nil)
runtime·throw("bad m->nextg in nextgoroutine");
m->nextg = nil;
return gp;
}
void
runtime·notesleep(Note *n)
{
while(runtime·atomicload(&n->state) == 0)
futexsleep(&n->state, 0);
}
// Get the next goroutine that m should run.
// Sched must be locked on entry, is unlocked on exit.
// Makes sure that at most $GOMAXPROCS g's are
// running on cpus (not in system calls) at any given time.
static G*
nextgandunlock(void)
{
G *gp;
uint32 v;
top:
if(atomic_mcpu(runtime·sched.atomic) >= maxgomaxprocs)
runtime·throw("negative mcpu");
// If there is a g waiting as m->nextg, the mcpu++
// happened before it was passed to mnextg.
if(m->nextg != nil) {
gp = m->nextg;
m->nextg = nil;
schedunlock();
return gp;
}
if(m->lockedg != nil) {
// We can only run one g, and it's not available.
// Make sure some other cpu is running to handle
// the ordinary run queue.
if(runtime·sched.gwait != 0) {
matchmg();
// m->lockedg might have been on the queue.
if(m->nextg != nil) {
gp = m->nextg;
m->nextg = nil;
schedunlock();
return gp;
}
}
} else {
// Look for work on global queue.
while(haveg() && canaddmcpu()) {
gp = gget();
if(gp == nil)
runtime·throw("gget inconsistency");
if(gp->lockedm) {
mnextg(gp->lockedm, gp);
continue;
}
runtime·sched.grunning++;
schedunlock();
return gp;
}
// The while loop ended either because the g queue is empty
// or because we have maxed out our m procs running go
// code (mcpu >= mcpumax). We need to check that
// concurrent actions by entersyscall/exitsyscall cannot
// invalidate the decision to end the loop.
//
// We hold the sched lock, so no one else is manipulating the
// g queue or changing mcpumax. Entersyscall can decrement
// mcpu, but if does so when there is something on the g queue,
// the gwait bit will be set, so entersyscall will take the slow path
// and use the sched lock. So it cannot invalidate our decision.
//
// Wait on global m queue.
mput(m);
}
// Look for deadlock situation: one single active g which happens to be scvg.
if(runtime·sched.grunning == 1 && runtime·sched.gwait == 0) {
if(scvg->status == Grunning || scvg->status == Gsyscall)
runtime·throw("all goroutines are asleep - deadlock!");
}
m->nextg = nil;
m->waitnextg = 1;
runtime·noteclear(&m->havenextg);
// Stoptheworld is waiting for all but its cpu to go to stop.
// Entersyscall might have decremented mcpu too, but if so
// it will see the waitstop and take the slow path.
// Exitsyscall never increments mcpu beyond mcpumax.
v = runtime·atomicload(&runtime·sched.atomic);
if(atomic_waitstop(v) && atomic_mcpu(v) <= atomic_mcpumax(v)) {
// set waitstop = 0 (known to be 1)
runtime·xadd(&runtime·sched.atomic, -1<<waitstopShift);
runtime·notewakeup(&runtime·sched.stopped);
}
schedunlock();
runtime·notesleep(&m->havenextg);
if(m->helpgc) {
runtime·gchelper();
m->helpgc = 0;
runtime·lock(&runtime·sched);
goto top;
}
if((gp = m->nextg) == nil)
runtime·throw("bad m->nextg in nextgoroutine");
m->nextg = nil;
return gp;
}
