libgo/runtime/proc.c

   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.
   4
   5 #include <limits.h>
   6 #include <stdlib.h>
   7 #include <pthread.h>
   8 #include <unistd.h>
   9
  10 #include "config.h"
  11 #include "runtime.h"
  12 #include "arch.h"
  13 #include "defs.h"
  14 #include "malloc.h"
  15 #include "go-defer.h"
  16
  17 #ifdef USING_SPLIT_STACK
  18
  19 /* FIXME: These are not declared anywhere.  */
  20
  21 extern void __splitstack_getcontext(void *context[10]);
  22
  23 extern void __splitstack_setcontext(void *context[10]);
  24
  25 extern void *__splitstack_makecontext(size_t, void *context[10], size_t *);
  26
  27 extern void * __splitstack_resetcontext(void *context[10], size_t *);
  28
  29 extern void *__splitstack_find(void *, void *, size_t *, void **, void **,
  30                                void **);
  31
  32 extern void __splitstack_block_signals (int *, int *);
  33
  34 extern void __splitstack_block_signals_context (void *context[10], int *,
  35                                                 int *);
  36
  37 #endif
  38
  39 #if defined(USING_SPLIT_STACK) && defined(LINKER_SUPPORTS_SPLIT_STACK)
  40 # ifdef PTHREAD_STACK_MIN
  41 #  define StackMin PTHREAD_STACK_MIN
  42 # else
  43 #  define StackMin 8192
  44 # endif
  45 #else
  46 # define StackMin 2 * 1024 * 1024
  47 #endif
  48
  49 static void schedule(G*);
  50
  51 typedef struct Sched Sched;
  52
  53 M       runtime_m0;
  54 G       runtime_g0;     // idle goroutine for m0
  55
  56 #ifdef __rtems__
  57 #define __thread
  58 #endif
  59
  60 static __thread G *g;
  61 static __thread M *m;
  62
  63 #ifndef SETCONTEXT_CLOBBERS_TLS
  64
  65 static inline void
  66 initcontext(void)
  67 {
  68 }
  69
  70 static inline void
  71 fixcontext(ucontext_t *c __attribute__ ((unused)))
  72 {
  73 }
  74
  75 # else
  76
  77 # if defined(__x86_64__) && defined(__sun__)
  78
  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
  84 // same.
  85
  86 static __thread greg_t fs;
  87
  88 static inline void
  89 initcontext(void)
  90 {
  91         ucontext_t c;
  92
  93         getcontext(&c);
  94         fs = c.uc_mcontext.gregs[REG_FSBASE];
  95 }
  96
  97 static inline void
  98 fixcontext(ucontext_t* c)
  99 {
 100         c->uc_mcontext.gregs[REG_FSBASE] = fs;
 101 }
 102
 103 # else
 104
 105 #  error unknown case for SETCONTEXT_CLOBBERS_TLS
 106
 107 # endif
 108
 109 #endif
 110
 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
 117 // when necessary.
 118
 119 G* runtime_g(void) __attribute__ ((noinline, no_split_stack));
 120
 121 G*
 122 runtime_g(void)
 123 {
 124         return g;
 125 }
 126
 127 M* runtime_m(void) __attribute__ ((noinline, no_split_stack));
 128
 129 M*
 130 runtime_m(void)
 131 {
 132         return m;
 133 }
 134
 135 int32   runtime_gcwaiting;
 136
 137 // Go scheduler
 138 //
 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.
 144 //
 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.
 151 //
 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.
 159 //
 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
 162 // Linux and OS X.
 163
 164 struct Sched {
 165         Lock;
 166
 167         G *gfree;       // available g's (status == Gdead)
 168         int32 goidgen;
 169
 170         G *ghead;       // g's waiting to run
 171         G *gtail;
 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
 175
 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
 179
 180         volatile uint32 atomic; // atomic scheduling word (see below)
 181
 182         int32 profilehz;        // cpu profiling rate
 183
 184         bool init;  // running initialization
 185         bool lockmain;  // init called runtime.LockOSThread
 186
 187         Note    stopped;        // one g can set waitstop and wait here for m's to stop
 188 };
 189
 190 // The atomic word in sched is an atomic uint32 that
 191 // holds these fields.
 192 //
 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
 197 //
 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.
 204 //
 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).
 210 //
 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.
 217 enum {
 218         mcpuWidth = 15,
 219         mcpuMask = (1<<mcpuWidth) - 1,
 220         mcpuShift = 0,
 221         mcpumaxShift = mcpuShift + mcpuWidth,
 222         waitstopShift = mcpumaxShift + mcpuWidth,
 223         gwaitingShift = waitstopShift+1,
 224
 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
 229         // beyond zero.
 230         maxgomaxprocs = mcpuMask - 10,
 231 };
 232
 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)
 237
 238 Sched runtime_sched;
 239 int32 runtime_gomaxprocs;
 240 bool runtime_singleproc;
 241
 242 static bool canaddmcpu(void);
 243
 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.
 249 static M* mwakeup;
 250
 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
 255 static M* mget(G*);
 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*);
 262
 263 void
 264 setmcpumax(uint32 n)
 265 {
 266         uint32 v, w;
 267
 268         for(;;) {
 269                 v = runtime_sched.atomic;
 270                 w = v;
 271                 w &= ~(mcpuMask<<mcpumaxShift);
 272                 w |= n<<mcpumaxShift;
 273                 if(runtime_cas(&runtime_sched.atomic, v, w))
 274                         break;
 275         }
 276 }
 277
 278 // First function run by a new goroutine.  This replaces gogocall.
 279 static void
 280 kickoff(void)
 281 {
 282         void (*fn)(void*);
 283
 284         fn = (void (*)(void*))(g->entry);
 285         fn(g->param);
 286         runtime_goexit();
 287 }
 288
 289 // Switch context to a different goroutine.  This is like longjmp.
 290 static void runtime_gogo(G*) __attribute__ ((noinline));
 291 static void
 292 runtime_gogo(G* newg)
 293 {
 294 #ifdef USING_SPLIT_STACK
 295         __splitstack_setcontext(&newg->stack_context[0]);
 296 #endif
 297         g = newg;
 298         newg->fromgogo = true;
 299         fixcontext(&newg->context);
 300         setcontext(&newg->context);
 301         runtime_throw("gogo setcontext returned");
 302 }
 303
 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));
 309 static void
 310 runtime_mcall(void (*pfn)(G*))
 311 {
 312         M *mp;
 313         G *gp;
 314 #ifndef USING_SPLIT_STACK
 315         int i;
 316 #endif
 317
 318         // Ensure that all registers are on the stack for the garbage
 319         // collector.
 320         __builtin_unwind_init();
 321
 322         mp = m;
 323         gp = g;
 324         if(gp == mp->g0)
 325                 runtime_throw("runtime: mcall called on m->g0 stack");
 326
 327         if(gp != nil) {
 328
 329 #ifdef USING_SPLIT_STACK
 330                 __splitstack_getcontext(&g->stack_context[0]);
 331 #else
 332                 gp->gcnext_sp = &i;
 333 #endif
 334                 gp->fromgogo = false;
 335                 getcontext(&gp->context);
 336
 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.
 344                 mp = runtime_m();
 345                 gp = runtime_g();
 346         }
 347         if (gp == nil || !gp->fromgogo) {
 348 #ifdef USING_SPLIT_STACK
 349                 __splitstack_setcontext(&mp->g0->stack_context[0]);
 350 #endif
 351                 mp->g0->entry = (byte*)pfn;
 352                 mp->g0->param = gp;
 353
 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.
 357                 g = mp->g0;
 358
 359                 fixcontext(&mp->g0->context);
 360                 setcontext(&mp->g0->context);
 361                 runtime_throw("runtime: mcall function returned");
 362         }
 363 }
 364
 365 // The bootstrap sequence is:
 366 //
 367 //      call osinit
 368 //      call schedinit
 369 //      make & queue new G
 370 //      call runtime_mstart
 371 //
 372 // The new G calls runtime_main.
 373 void
 374 runtime_schedinit(void)
 375 {
 376         int32 n;
 377         const byte *p;
 378
 379         m = &runtime_m0;
 380         g = &runtime_g0;
 381         m->g0 = g;
 382         m->curg = g;
 383         g->m = m;
 384
 385         initcontext();
 386
 387         m->nomemprof++;
 388         runtime_mallocinit();
 389         mcommoninit(m);
 390
 391         runtime_goargs();
 392         runtime_goenvs();
 393
 394         // For debugging:
 395         // Allocate internal symbol table representation now,
 396         // so that we don't need to call malloc when we crash.
 397         // runtime_findfunc(0);
 398
 399         runtime_gomaxprocs = 1;
 400         p = runtime_getenv("GOMAXPROCS");
 401         if(p != nil && (n = runtime_atoi(p)) != 0) {
 402                 if(n > maxgomaxprocs)
 403                         n = maxgomaxprocs;
 404                 runtime_gomaxprocs = n;
 405         }
 406         setmcpumax(runtime_gomaxprocs);
 407         runtime_singleproc = runtime_gomaxprocs == 1;
 408
 409         canaddmcpu();   // mcpu++ to account for bootstrap m
 410         m->helpgc = 1;  // flag to tell schedule() to mcpu--
 411         runtime_sched.grunning++;
 412
 413         // Can not enable GC until all roots are registered.
 414         // mstats.enablegc = 1;
 415         m->nomemprof--;
 416 }
 417
 418 extern void main_init(void) __asm__ ("__go_init_main");
 419 extern void main_main(void) __asm__ ("main.main");
 420
 421 // The main goroutine.
 422 void
 423 runtime_main(void)
 424 {
 425         // Lock the main goroutine onto this, the main OS thread,
 426         // during initialization.  Most programs won't care, but a few
 427         // do require certain calls to be made by the main thread.
 428         // Those can arrange for main.main to run in the main thread
 429         // by calling runtime.LockOSThread during initialization
 430         // to preserve the lock.
 431         runtime_LockOSThread();
 432         runtime_sched.init = true;
 433         main_init();
 434         runtime_sched.init = false;
 435         if(!runtime_sched.lockmain)
 436                 runtime_UnlockOSThread();
 437
 438         // For gccgo we have to wait until after main is initialized
 439         // to enable GC, because initializing main registers the GC
 440         // roots.
 441         mstats.enablegc = 1;
 442
 443         main_main();
 444         runtime_exit(0);
 445         for(;;)
 446                 *(int32*)0 = 0;
 447 }
 448
 449 // Lock the scheduler.
 450 static void
 451 schedlock(void)
 452 {
 453         runtime_lock(&runtime_sched);
 454 }
 455
 456 // Unlock the scheduler.
 457 static void
 458 schedunlock(void)
 459 {
 460         M *m;
 461
 462         m = mwakeup;
 463         mwakeup = nil;
 464         runtime_unlock(&runtime_sched);
 465         if(m != nil)
 466                 runtime_notewakeup(&m->havenextg);
 467 }
 468
 469 void
 470 runtime_goexit(void)
 471 {
 472         g->status = Gmoribund;
 473         runtime_gosched();
 474 }
 475
 476 void
 477 runtime_goroutineheader(G *g)
 478 {
 479         const char *status;
 480
 481         switch(g->status) {
 482         case Gidle:
 483                 status = "idle";
 484                 break;
 485         case Grunnable:
 486                 status = "runnable";
 487                 break;
 488         case Grunning:
 489                 status = "running";
 490                 break;
 491         case Gsyscall:
 492                 status = "syscall";
 493                 break;
 494         case Gwaiting:
 495                 if(g->waitreason)
 496                         status = g->waitreason;
 497                 else
 498                         status = "waiting";
 499                 break;
 500         case Gmoribund:
 501                 status = "moribund";
 502                 break;
 503         default:
 504                 status = "???";
 505                 break;
 506         }
 507         runtime_printf("goroutine %d [%s]:\n", g->goid, status);
 508 }
 509
 510 void
 511 runtime_tracebackothers(G *me)
 512 {
 513         G *g;
 514
 515         for(g = runtime_allg; g != nil; g = g->alllink) {
 516                 if(g == me || g->status == Gdead)
 517                         continue;
 518                 runtime_printf("\n");
 519                 runtime_goroutineheader(g);
 520                 // runtime_traceback(g->sched.pc, g->sched.sp, 0, g);
 521         }
 522 }
 523
 524 // Mark this g as m's idle goroutine.
 525 // This functionality might be used in environments where programs
 526 // are limited to a single thread, to simulate a select-driven
 527 // network server.  It is not exposed via the standard runtime API.
 528 void
 529 runtime_idlegoroutine(void)
 530 {
 531         if(g->idlem != nil)
 532                 runtime_throw("g is already an idle goroutine");
 533         g->idlem = m;
 534 }
 535
 536 static void
 537 mcommoninit(M *m)
 538 {
 539         // Add to runtime_allm so garbage collector doesn't free m
 540         // when it is just in a register or thread-local storage.
 541         m->alllink = runtime_allm;
 542         // runtime_Cgocalls() iterates over allm w/o schedlock,
 543         // so we need to publish it safely.
 544         runtime_atomicstorep((void**)&runtime_allm, m);
 545
 546         m->id = runtime_sched.mcount++;
 547         m->fastrand = 0x49f6428aUL + m->id + runtime_cputicks();
 548
 549         if(m->mcache == nil)
 550                 m->mcache = runtime_allocmcache();
 551 }
 552
 553 // Try to increment mcpu.  Report whether succeeded.
 554 static bool
 555 canaddmcpu(void)
 556 {
 557         uint32 v;
 558
 559         for(;;) {
 560                 v = runtime_sched.atomic;
 561                 if(atomic_mcpu(v) >= atomic_mcpumax(v))
 562                         return 0;
 563                 if(runtime_cas(&runtime_sched.atomic, v, v+(1<<mcpuShift)))
 564                         return 1;
 565         }
 566 }
 567
 568 // Put on `g' queue.  Sched must be locked.
 569 static void
 570 gput(G *g)
 571 {
 572         M *m;
 573
 574         // If g is wired, hand it off directly.
 575         if((m = g->lockedm) != nil && canaddmcpu()) {
 576                 mnextg(m, g);
 577                 return;
 578         }
 579
 580         // If g is the idle goroutine for an m, hand it off.
 581         if(g->idlem != nil) {
 582                 if(g->idlem->idleg != nil) {
 583                         runtime_printf("m%d idle out of sync: g%d g%d\n",
 584                                 g->idlem->id,
 585                                 g->idlem->idleg->goid, g->goid);
 586                         runtime_throw("runtime: double idle");
 587                 }
 588                 g->idlem->idleg = g;
 589                 return;
 590         }
 591
 592         g->schedlink = nil;
 593         if(runtime_sched.ghead == nil)
 594                 runtime_sched.ghead = g;
 595         else
 596                 runtime_sched.gtail->schedlink = g;
 597         runtime_sched.gtail = g;
 598
 599         // increment gwait.
 600         // if it transitions to nonzero, set atomic gwaiting bit.
 601         if(runtime_sched.gwait++ == 0)
 602                 runtime_xadd(&runtime_sched.atomic, 1<<gwaitingShift);
 603 }
 604
 605 // Report whether gget would return something.
 606 static bool
 607 haveg(void)
 608 {
 609         return runtime_sched.ghead != nil || m->idleg != nil;
 610 }
 611
 612 // Get from `g' queue.  Sched must be locked.
 613 static G*
 614 gget(void)
 615 {
 616         G *g;
 617
 618         g = runtime_sched.ghead;
 619         if(g){
 620                 runtime_sched.ghead = g->schedlink;
 621                 if(runtime_sched.ghead == nil)
 622                         runtime_sched.gtail = nil;
 623                 // decrement gwait.
 624                 // if it transitions to zero, clear atomic gwaiting bit.
 625                 if(--runtime_sched.gwait == 0)
 626                         runtime_xadd(&runtime_sched.atomic, -1<<gwaitingShift);
 627         } else if(m->idleg != nil) {
 628                 g = m->idleg;
 629                 m->idleg = nil;
 630         }
 631         return g;
 632 }
 633
 634 // Put on `m' list.  Sched must be locked.
 635 static void
 636 mput(M *m)
 637 {
 638         m->schedlink = runtime_sched.mhead;
 639         runtime_sched.mhead = m;
 640         runtime_sched.mwait++;
 641 }
 642
 643 // Get an `m' to run `g'.  Sched must be locked.
 644 static M*
 645 mget(G *g)
 646 {
 647         M *m;
 648
 649         // if g has its own m, use it.
 650         if(g && (m = g->lockedm) != nil)
 651                 return m;
 652
 653         // otherwise use general m pool.
 654         if((m = runtime_sched.mhead) != nil){
 655                 runtime_sched.mhead = m->schedlink;
 656                 runtime_sched.mwait--;
 657         }
 658         return m;
 659 }
 660
 661 // Mark g ready to run.
 662 void
 663 runtime_ready(G *g)
 664 {
 665         schedlock();
 666         readylocked(g);
 667         schedunlock();
 668 }
 669
 670 // Mark g ready to run.  Sched is already locked.
 671 // G might be running already and about to stop.
 672 // The sched lock protects g->status from changing underfoot.
 673 static void
 674 readylocked(G *g)
 675 {
 676         if(g->m){
 677                 // Running on another machine.
 678                 // Ready it when it stops.
 679                 g->readyonstop = 1;
 680                 return;
 681         }
 682
 683         // Mark runnable.
 684         if(g->status == Grunnable || g->status == Grunning) {
 685                 runtime_printf("goroutine %d has status %d\n", g->goid, g->status);
 686                 runtime_throw("bad g->status in ready");
 687         }
 688         g->status = Grunnable;
 689
 690         gput(g);
 691         matchmg();
 692 }
 693
 694 // Same as readylocked but a different symbol so that
 695 // debuggers can set a breakpoint here and catch all
 696 // new goroutines.
 697 static void
 698 newprocreadylocked(G *g)
 699 {
 700         readylocked(g);
 701 }
 702
 703 // Pass g to m for running.
 704 // Caller has already incremented mcpu.
 705 static void
 706 mnextg(M *m, G *g)
 707 {
 708         runtime_sched.grunning++;
 709         m->nextg = g;
 710         if(m->waitnextg) {
 711                 m->waitnextg = 0;
 712                 if(mwakeup != nil)
 713                         runtime_notewakeup(&mwakeup->havenextg);
 714                 mwakeup = m;
 715         }
 716 }
 717
 718 // Get the next goroutine that m should run.
 719 // Sched must be locked on entry, is unlocked on exit.
 720 // Makes sure that at most $GOMAXPROCS g's are
 721 // running on cpus (not in system calls) at any given time.
 722 static G*
 723 nextgandunlock(void)
 724 {
 725         G *gp;
 726         uint32 v;
 727
 728 top:
 729         if(atomic_mcpu(runtime_sched.atomic) >= maxgomaxprocs)
 730                 runtime_throw("negative mcpu");
 731
 732         // If there is a g waiting as m->nextg, the mcpu++
 733         // happened before it was passed to mnextg.
 734         if(m->nextg != nil) {
 735                 gp = m->nextg;
 736                 m->nextg = nil;
 737                 schedunlock();
 738                 return gp;
 739         }
 740
 741         if(m->lockedg != nil) {
 742                 // We can only run one g, and it's not available.
 743                 // Make sure some other cpu is running to handle
 744                 // the ordinary run queue.
 745                 if(runtime_sched.gwait != 0) {
 746                         matchmg();
 747                         // m->lockedg might have been on the queue.
 748                         if(m->nextg != nil) {
 749                                 gp = m->nextg;
 750                                 m->nextg = nil;
 751                                 schedunlock();
 752                                 return gp;
 753                         }
 754                 }
 755         } else {
 756                 // Look for work on global queue.
 757                 while(haveg() && canaddmcpu()) {
 758                         gp = gget();
 759                         if(gp == nil)
 760                                 runtime_throw("gget inconsistency");
 761
 762                         if(gp->lockedm) {
 763                                 mnextg(gp->lockedm, gp);
 764                                 continue;
 765                         }
 766                         runtime_sched.grunning++;
 767                         schedunlock();
 768                         return gp;
 769                 }
 770
 771                 // The while loop ended either because the g queue is empty
 772                 // or because we have maxed out our m procs running go
 773                 // code (mcpu >= mcpumax).  We need to check that
 774                 // concurrent actions by entersyscall/exitsyscall cannot
 775                 // invalidate the decision to end the loop.
 776                 //
 777                 // We hold the sched lock, so no one else is manipulating the
 778                 // g queue or changing mcpumax.  Entersyscall can decrement
 779                 // mcpu, but if does so when there is something on the g queue,
 780                 // the gwait bit will be set, so entersyscall will take the slow path
 781                 // and use the sched lock.  So it cannot invalidate our decision.
 782                 //
 783                 // Wait on global m queue.
 784                 mput(m);
 785         }
 786
 787         v = runtime_atomicload(&runtime_sched.atomic);
 788         if(runtime_sched.grunning == 0)
 789                 runtime_throw("all goroutines are asleep - deadlock!");
 790         m->nextg = nil;
 791         m->waitnextg = 1;
 792         runtime_noteclear(&m->havenextg);
 793
 794         // Stoptheworld is waiting for all but its cpu to go to stop.
 795         // Entersyscall might have decremented mcpu too, but if so
 796         // it will see the waitstop and take the slow path.
 797         // Exitsyscall never increments mcpu beyond mcpumax.
 798         if(atomic_waitstop(v) && atomic_mcpu(v) <= atomic_mcpumax(v)) {
 799                 // set waitstop = 0 (known to be 1)
 800                 runtime_xadd(&runtime_sched.atomic, -1<<waitstopShift);
 801                 runtime_notewakeup(&runtime_sched.stopped);
 802         }
 803         schedunlock();
 804
 805         runtime_notesleep(&m->havenextg);
 806         if(m->helpgc) {
 807                 runtime_gchelper();
 808                 m->helpgc = 0;
 809                 runtime_lock(&runtime_sched);
 810                 goto top;
 811         }
 812         if((gp = m->nextg) == nil)
 813                 runtime_throw("bad m->nextg in nextgoroutine");
 814         m->nextg = nil;
 815         return gp;
 816 }
 817
 818 int32
 819 runtime_helpgc(bool *extra)
 820 {
 821         M *mp;
 822         int32 n, max;
 823
 824         // Figure out how many CPUs to use.
 825         // Limited by gomaxprocs, number of actual CPUs, and MaxGcproc.
 826         max = runtime_gomaxprocs;
 827         if(max > runtime_ncpu)
 828                 max = runtime_ncpu > 0 ? runtime_ncpu : 1;
 829         if(max > MaxGcproc)
 830                 max = MaxGcproc;
 831
 832         // We're going to use one CPU no matter what.
 833         // Figure out the max number of additional CPUs.
 834         max--;
 835
 836         runtime_lock(&runtime_sched);
 837         n = 0;
 838         while(n < max && (mp = mget(nil)) != nil) {
 839                 n++;
 840                 mp->helpgc = 1;
 841                 mp->waitnextg = 0;
 842                 runtime_notewakeup(&mp->havenextg);
 843         }
 844         runtime_unlock(&runtime_sched);
 845         if(extra)
 846                 *extra = n != max;
 847         return n;
 848 }
 849
 850 void
 851 runtime_stoptheworld(void)
 852 {
 853         uint32 v;
 854
 855         schedlock();
 856         runtime_gcwaiting = 1;
 857
 858         setmcpumax(1);
 859
 860         // while mcpu > 1
 861         for(;;) {
 862                 v = runtime_sched.atomic;
 863                 if(atomic_mcpu(v) <= 1)
 864                         break;
 865
 866                 // It would be unsafe for multiple threads to be using
 867                 // the stopped note at once, but there is only
 868                 // ever one thread doing garbage collection.
 869                 runtime_noteclear(&runtime_sched.stopped);
 870                 if(atomic_waitstop(v))
 871                         runtime_throw("invalid waitstop");
 872
 873                 // atomic { waitstop = 1 }, predicated on mcpu <= 1 check above
 874                 // still being true.
 875                 if(!runtime_cas(&runtime_sched.atomic, v, v+(1<<waitstopShift)))
 876                         continue;
 877
 878                 schedunlock();
 879                 runtime_notesleep(&runtime_sched.stopped);
 880                 schedlock();
 881         }
 882         runtime_singleproc = runtime_gomaxprocs == 1;
 883         schedunlock();
 884 }
 885
 886 void
 887 runtime_starttheworld(bool extra)
 888 {
 889         M *m;
 890
 891         schedlock();
 892         runtime_gcwaiting = 0;
 893         setmcpumax(runtime_gomaxprocs);
 894         matchmg();
 895         if(extra && canaddmcpu()) {
 896                 // Start a new m that will (we hope) be idle
 897                 // and so available to help when the next
 898                 // garbage collection happens.
 899                 // canaddmcpu above did mcpu++
 900                 // (necessary, because m will be doing various
 901                 // initialization work so is definitely running),
 902                 // but m is not running a specific goroutine,
 903                 // so set the helpgc flag as a signal to m's
 904                 // first schedule(nil) to mcpu-- and grunning--.
 905                 m = runtime_newm();
 906                 m->helpgc = 1;
 907                 runtime_sched.grunning++;
 908         }
 909         schedunlock();
 910 }
 911
 912 // Called to start an M.
 913 void*
 914 runtime_mstart(void* mp)
 915 {
 916         m = (M*)mp;
 917         g = m->g0;
 918
 919         initcontext();
 920
 921         g->entry = nil;
 922         g->param = nil;
 923
 924         // Record top of stack for use by mcall.
 925         // Once we call schedule we're never coming back,
 926         // so other calls can reuse this stack space.
 927 #ifdef USING_SPLIT_STACK
 928         __splitstack_getcontext(&g->stack_context[0]);
 929 #else
 930         g->gcinitial_sp = &mp;
 931         // Setting gcstack_size to 0 is a marker meaning that gcinitial_sp
 932         // is the top of the stack, not the bottom.
 933         g->gcstack_size = 0;
 934         g->gcnext_sp = &mp;
 935 #endif
 936         getcontext(&g->context);
 937
 938         if(g->entry != nil) {
 939                 // Got here from mcall.
 940                 void (*pfn)(G*) = (void (*)(G*))g->entry;
 941                 G* gp = (G*)g->param;
 942                 pfn(gp);
 943                 *(int*)0x21 = 0x21;
 944         }
 945         runtime_minit();
 946
 947 #ifdef USING_SPLIT_STACK
 948         {
 949           int dont_block_signals = 0;
 950           __splitstack_block_signals(&dont_block_signals, nil);
 951         }
 952 #endif
 953
 954         schedule(nil);
 955         return nil;
 956 }
 957
 958 typedef struct CgoThreadStart CgoThreadStart;
 959 struct CgoThreadStart
 960 {
 961         M *m;
 962         G *g;
 963         void (*fn)(void);
 964 };
 965
 966 // Kick off new m's as needed (up to mcpumax).
 967 // Sched is locked.
 968 static void
 969 matchmg(void)
 970 {
 971         G *gp;
 972         M *mp;
 973
 974         if(m->mallocing || m->gcing)
 975                 return;
 976
 977         while(haveg() && canaddmcpu()) {
 978                 gp = gget();
 979                 if(gp == nil)
 980                         runtime_throw("gget inconsistency");
 981
 982                 // Find the m that will run gp.
 983                 if((mp = mget(gp)) == nil)
 984                         mp = runtime_newm();
 985                 mnextg(mp, gp);
 986         }
 987 }
 988
 989 // Create a new m.  It will start off with a call to runtime_mstart.
 990 M*
 991 runtime_newm(void)
 992 {
 993         M *m;
 994         pthread_attr_t attr;
 995         pthread_t tid;
 996
 997         m = runtime_malloc(sizeof(M));
 998         mcommoninit(m);
 999         m->g0 = runtime_malg(-1, nil, nil);
1000
1001         if(pthread_attr_init(&attr) != 0)
1002                 runtime_throw("pthread_attr_init");
1003         if(pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_DETACHED) != 0)
1004                 runtime_throw("pthread_attr_setdetachstate");
1005
1006 #ifndef PTHREAD_STACK_MIN
1007 #define PTHREAD_STACK_MIN 8192
1008 #endif
1009         if(pthread_attr_setstacksize(&attr, PTHREAD_STACK_MIN) != 0)
1010                 runtime_throw("pthread_attr_setstacksize");
1011
1012         if(pthread_create(&tid, &attr, runtime_mstart, m) != 0)
1013                 runtime_throw("pthread_create");
1014
1015         return m;
1016 }
1017
1018 // One round of scheduler: find a goroutine and run it.
1019 // The argument is the goroutine that was running before
1020 // schedule was called, or nil if this is the first call.
1021 // Never returns.
1022 static void
1023 schedule(G *gp)
1024 {
1025         int32 hz;
1026         uint32 v;
1027
1028         schedlock();
1029         if(gp != nil) {
1030                 // Just finished running gp.
1031                 gp->m = nil;
1032                 runtime_sched.grunning--;
1033
1034                 // atomic { mcpu-- }
1035                 v = runtime_xadd(&runtime_sched.atomic, -1<<mcpuShift);
1036                 if(atomic_mcpu(v) > maxgomaxprocs)
1037                         runtime_throw("negative mcpu in scheduler");
1038
1039                 switch(gp->status){
1040                 case Grunnable:
1041                 case Gdead:
1042                         // Shouldn't have been running!
1043                         runtime_throw("bad gp->status in sched");
1044                 case Grunning:
1045                         gp->status = Grunnable;
1046                         gput(gp);
1047                         break;
1048                 case Gmoribund:
1049                         gp->status = Gdead;
1050                         if(gp->lockedm) {
1051                                 gp->lockedm = nil;
1052                                 m->lockedg = nil;
1053                         }
1054                         gp->idlem = nil;
1055                         gfput(gp);
1056                         if(--runtime_sched.gcount == 0)
1057                                 runtime_exit(0);
1058                         break;
1059                 }
1060                 if(gp->readyonstop){
1061                         gp->readyonstop = 0;
1062                         readylocked(gp);
1063                 }
1064         } else if(m->helpgc) {
1065                 // Bootstrap m or new m started by starttheworld.
1066                 // atomic { mcpu-- }
1067                 v = runtime_xadd(&runtime_sched.atomic, -1<<mcpuShift);
1068                 if(atomic_mcpu(v) > maxgomaxprocs)
1069                         runtime_throw("negative mcpu in scheduler");
1070                 // Compensate for increment in starttheworld().
1071                 runtime_sched.grunning--;
1072                 m->helpgc = 0;
1073         } else if(m->nextg != nil) {
1074                 // New m started by matchmg.
1075         } else {
1076                 runtime_throw("invalid m state in scheduler");
1077         }
1078
1079         // Find (or wait for) g to run.  Unlocks runtime_sched.
1080         gp = nextgandunlock();
1081         gp->readyonstop = 0;
1082         gp->status = Grunning;
1083         m->curg = gp;
1084         gp->m = m;
1085
1086         // Check whether the profiler needs to be turned on or off.
1087         hz = runtime_sched.profilehz;
1088         if(m->profilehz != hz)
1089                 runtime_resetcpuprofiler(hz);
1090
1091         runtime_gogo(gp);
1092 }
1093
1094 // Enter scheduler.  If g->status is Grunning,
1095 // re-queues g and runs everyone else who is waiting
1096 // before running g again.  If g->status is Gmoribund,
1097 // kills off g.
1098 void
1099 runtime_gosched(void)
1100 {
1101         if(m->locks != 0)
1102                 runtime_throw("gosched holding locks");
1103         if(g == m->g0)
1104                 runtime_throw("gosched of g0");
1105         runtime_mcall(schedule);
1106 }
1107
1108 // The goroutine g is about to enter a system call.
1109 // Record that it's not using the cpu anymore.
1110 // This is called only from the go syscall library and cgocall,
1111 // not from the low-level system calls used by the runtime.
1112 //
1113 // Entersyscall cannot split the stack: the runtime_gosave must
1114 // make g->sched refer to the caller's stack segment, because
1115 // entersyscall is going to return immediately after.
1116 // It's okay to call matchmg and notewakeup even after
1117 // decrementing mcpu, because we haven't released the
1118 // sched lock yet, so the garbage collector cannot be running.
1119
1120 void runtime_entersyscall(void) __attribute__ ((no_split_stack));
1121
1122 void
1123 runtime_entersyscall(void)
1124 {
1125         uint32 v;
1126
1127         // Leave SP around for gc and traceback.
1128 #ifdef USING_SPLIT_STACK
1129         g->gcstack = __splitstack_find(NULL, NULL, &g->gcstack_size,
1130                                        &g->gcnext_segment, &g->gcnext_sp,
1131                                        &g->gcinitial_sp);
1132 #else
1133         g->gcnext_sp = (byte *) &v;
1134 #endif
1135
1136         // Save the registers in the g structure so that any pointers
1137         // held in registers will be seen by the garbage collector.
1138         // We could use getcontext here, but setjmp is more efficient
1139         // because it doesn't need to save the signal mask.
1140         setjmp(g->gcregs);
1141
1142         g->status = Gsyscall;
1143
1144         // Fast path.
1145         // The slow path inside the schedlock/schedunlock will get
1146         // through without stopping if it does:
1147         //      mcpu--
1148         //      gwait not true
1149         //      waitstop && mcpu <= mcpumax not true
1150         // If we can do the same with a single atomic add,
1151         // then we can skip the locks.
1152         v = runtime_xadd(&runtime_sched.atomic, -1<<mcpuShift);
1153         if(!atomic_gwaiting(v) && (!atomic_waitstop(v) || atomic_mcpu(v) > atomic_mcpumax(v)))
1154                 return;
1155
1156         schedlock();
1157         v = runtime_atomicload(&runtime_sched.atomic);
1158         if(atomic_gwaiting(v)) {
1159                 matchmg();
1160                 v = runtime_atomicload(&runtime_sched.atomic);
1161         }
1162         if(atomic_waitstop(v) && atomic_mcpu(v) <= atomic_mcpumax(v)) {
1163                 runtime_xadd(&runtime_sched.atomic, -1<<waitstopShift);
1164                 runtime_notewakeup(&runtime_sched.stopped);
1165         }
1166
1167         schedunlock();
1168 }
1169
1170 // The goroutine g exited its system call.
1171 // Arrange for it to run on a cpu again.
1172 // This is called only from the go syscall library, not
1173 // from the low-level system calls used by the runtime.
1174 void
1175 runtime_exitsyscall(void)
1176 {
1177         G *gp;
1178         uint32 v;
1179
1180         // Fast path.
1181         // If we can do the mcpu++ bookkeeping and
1182         // find that we still have mcpu <= mcpumax, then we can
1183         // start executing Go code immediately, without having to
1184         // schedlock/schedunlock.
1185         gp = g;
1186         v = runtime_xadd(&runtime_sched.atomic, (1<<mcpuShift));
1187         if(m->profilehz == runtime_sched.profilehz && atomic_mcpu(v) <= atomic_mcpumax(v)) {
1188                 // There's a cpu for us, so we can run.
1189                 gp->status = Grunning;
1190                 // Garbage collector isn't running (since we are),
1191                 // so okay to clear gcstack.
1192 #ifdef USING_SPLIT_STACK
1193                 gp->gcstack = nil;
1194 #endif
1195                 gp->gcnext_sp = nil;
1196                 runtime_memclr(gp->gcregs, sizeof gp->gcregs);
1197                 return;
1198         }
1199
1200         // Tell scheduler to put g back on the run queue:
1201         // mostly equivalent to g->status = Grunning,
1202         // but keeps the garbage collector from thinking
1203         // that g is running right now, which it's not.
1204         gp->readyonstop = 1;
1205
1206         // All the cpus are taken.
1207         // The scheduler will ready g and put this m to sleep.
1208         // When the scheduler takes g away from m,
1209         // it will undo the runtime_sched.mcpu++ above.
1210         runtime_gosched();
1211
1212         // Gosched returned, so we're allowed to run now.
1213         // Delete the gcstack information that we left for
1214         // the garbage collector during the system call.
1215         // Must wait until now because until gosched returns
1216         // we don't know for sure that the garbage collector
1217         // is not running.
1218 #ifdef USING_SPLIT_STACK
1219         gp->gcstack = nil;
1220 #endif
1221         gp->gcnext_sp = nil;
1222         runtime_memclr(gp->gcregs, sizeof gp->gcregs);
1223 }
1224
1225 // Allocate a new g, with a stack big enough for stacksize bytes.
1226 G*
1227 runtime_malg(int32 stacksize, byte** ret_stack, size_t* ret_stacksize)
1228 {
1229         G *newg;
1230
1231         newg = runtime_malloc(sizeof(G));
1232         if(stacksize >= 0) {
1233 #if USING_SPLIT_STACK
1234                 int dont_block_signals = 0;
1235
1236                 *ret_stack = __splitstack_makecontext(stacksize,
1237                                                       &newg->stack_context[0],
1238                                                       ret_stacksize);
1239                 __splitstack_block_signals_context(&newg->stack_context[0],
1240                                                    &dont_block_signals, nil);
1241 #else
1242                 *ret_stack = runtime_mallocgc(stacksize, FlagNoProfiling|FlagNoGC, 0, 0);
1243                 *ret_stacksize = stacksize;
1244                 newg->gcinitial_sp = *ret_stack;
1245                 newg->gcstack_size = stacksize;
1246 #endif
1247         }
1248         return newg;
1249 }
1250
1251 /* For runtime package testing.  */
1252
1253 void runtime_testing_entersyscall(void)
1254   __asm__("libgo_runtime.runtime.entersyscall");
1255
1256 void
1257 runtime_testing_entersyscall()
1258 {
1259         runtime_entersyscall();
1260 }
1261
1262 void runtime_testing_exitsyscall(void)
1263   __asm__("libgo_runtime.runtime.exitsyscall");
1264
1265 void
1266 runtime_testing_exitsyscall()
1267 {
1268         runtime_exitsyscall();
1269 }
1270
1271 G*
1272 __go_go(void (*fn)(void*), void* arg)
1273 {
1274         byte *sp;
1275         size_t spsize;
1276         G * volatile newg;      // volatile to avoid longjmp warning
1277
1278         schedlock();
1279
1280         if((newg = gfget()) != nil){
1281 #ifdef USING_SPLIT_STACK
1282                 int dont_block_signals = 0;
1283
1284                 sp = __splitstack_resetcontext(&newg->stack_context[0],
1285                                                &spsize);
1286                 __splitstack_block_signals_context(&newg->stack_context[0],
1287                                                    &dont_block_signals, nil);
1288 #else
1289                 sp = newg->gcinitial_sp;
1290                 spsize = newg->gcstack_size;
1291                 if(spsize == 0)
1292                         runtime_throw("bad spsize in __go_go");
1293                 newg->gcnext_sp = sp;
1294 #endif
1295         } else {
1296                 newg = runtime_malg(StackMin, &sp, &spsize);
1297                 if(runtime_lastg == nil)
1298                         runtime_allg = newg;
1299                 else
1300                         runtime_lastg->alllink = newg;
1301                 runtime_lastg = newg;
1302         }
1303         newg->status = Gwaiting;
1304         newg->waitreason = "new goroutine";
1305
1306         newg->entry = (byte*)fn;
1307         newg->param = arg;
1308         newg->gopc = (uintptr)__builtin_return_address(0);
1309
1310         runtime_sched.gcount++;
1311         runtime_sched.goidgen++;
1312         newg->goid = runtime_sched.goidgen;
1313
1314         if(sp == nil)
1315                 runtime_throw("nil g->stack0");
1316
1317         getcontext(&newg->context);
1318         newg->context.uc_stack.ss_sp = sp;
1319 #ifdef MAKECONTEXT_STACK_TOP
1320         newg->context.uc_stack.ss_sp += spsize;
1321 #endif
1322         newg->context.uc_stack.ss_size = spsize;
1323         makecontext(&newg->context, kickoff, 0);
1324
1325         newprocreadylocked(newg);
1326         schedunlock();
1327
1328         return newg;
1329 //printf(" goid=%d\n", newg->goid);
1330 }
1331
1332 // Put on gfree list.  Sched must be locked.
1333 static void
1334 gfput(G *g)
1335 {
1336         g->schedlink = runtime_sched.gfree;
1337         runtime_sched.gfree = g;
1338 }
1339
1340 // Get from gfree list.  Sched must be locked.
1341 static G*
1342 gfget(void)
1343 {
1344         G *g;
1345
1346         g = runtime_sched.gfree;
1347         if(g)
1348                 runtime_sched.gfree = g->schedlink;
1349         return g;
1350 }
1351
1352 // Run all deferred functions for the current goroutine.
1353 static void
1354 rundefer(void)
1355 {
1356         Defer *d;
1357
1358         while((d = g->defer) != nil) {
1359                 void (*pfn)(void*);
1360
1361                 pfn = d->__pfn;
1362                 d->__pfn = nil;
1363                 if (pfn != nil)
1364                         (*pfn)(d->__arg);
1365                 g->defer = d->__next;
1366                 runtime_free(d);
1367         }
1368 }
1369
1370 void runtime_Goexit (void) asm ("libgo_runtime.runtime.Goexit");
1371
1372 void
1373 runtime_Goexit(void)
1374 {
1375         rundefer();
1376         runtime_goexit();
1377 }
1378
1379 void runtime_Gosched (void) asm ("libgo_runtime.runtime.Gosched");
1380
1381 void
1382 runtime_Gosched(void)
1383 {
1384         runtime_gosched();
1385 }
1386
1387 // Implementation of runtime.GOMAXPROCS.
1388 // delete when scheduler is stronger
1389 int32
1390 runtime_gomaxprocsfunc(int32 n)
1391 {
1392         int32 ret;
1393         uint32 v;
1394
1395         schedlock();
1396         ret = runtime_gomaxprocs;
1397         if(n <= 0)
1398                 n = ret;
1399         if(n > maxgomaxprocs)
1400                 n = maxgomaxprocs;
1401         runtime_gomaxprocs = n;
1402         if(runtime_gomaxprocs > 1)
1403                 runtime_singleproc = false;
1404         if(runtime_gcwaiting != 0) {
1405                 if(atomic_mcpumax(runtime_sched.atomic) != 1)
1406                         runtime_throw("invalid mcpumax during gc");
1407                 schedunlock();
1408                 return ret;
1409         }
1410
1411         setmcpumax(n);
1412
1413         // If there are now fewer allowed procs
1414         // than procs running, stop.
1415         v = runtime_atomicload(&runtime_sched.atomic);
1416         if((int32)atomic_mcpu(v) > n) {
1417                 schedunlock();
1418                 runtime_gosched();
1419                 return ret;
1420         }
1421         // handle more procs
1422         matchmg();
1423         schedunlock();
1424         return ret;
1425 }
1426
1427 void
1428 runtime_LockOSThread(void)
1429 {
1430         if(m == &runtime_m0 && runtime_sched.init) {
1431                 runtime_sched.lockmain = true;
1432                 return;
1433         }
1434         m->lockedg = g;
1435         g->lockedm = m;
1436 }
1437
1438 void
1439 runtime_UnlockOSThread(void)
1440 {
1441         if(m == &runtime_m0 && runtime_sched.init) {
1442                 runtime_sched.lockmain = false;
1443                 return;
1444         }
1445         m->lockedg = nil;
1446         g->lockedm = nil;
1447 }
1448
1449 bool
1450 runtime_lockedOSThread(void)
1451 {
1452         return g->lockedm != nil && m->lockedg != nil;
1453 }
1454
1455 // for testing of callbacks
1456
1457 _Bool runtime_golockedOSThread(void)
1458   asm("libgo_runtime.runtime.golockedOSThread");
1459
1460 _Bool
1461 runtime_golockedOSThread(void)
1462 {
1463         return runtime_lockedOSThread();
1464 }
1465
1466 // for testing of wire, unwire
1467 uint32
1468 runtime_mid()
1469 {
1470         return m->id;
1471 }
1472
1473 int32 runtime_Goroutines (void)
1474   __asm__ ("libgo_runtime.runtime.Goroutines");
1475
1476 int32
1477 runtime_Goroutines()
1478 {
1479         return runtime_sched.gcount;
1480 }
1481
1482 int32
1483 runtime_mcount(void)
1484 {
1485         return runtime_sched.mcount;
1486 }
1487
1488 static struct {
1489         Lock;
1490         void (*fn)(uintptr*, int32);
1491         int32 hz;
1492         uintptr pcbuf[100];
1493 } prof;
1494
1495 // Called if we receive a SIGPROF signal.
1496 void
1497 runtime_sigprof(uint8 *pc __attribute__ ((unused)),
1498                 uint8 *sp __attribute__ ((unused)),
1499                 uint8 *lr __attribute__ ((unused)),
1500                 G *gp __attribute__ ((unused)))
1501 {
1502         // int32 n;
1503
1504         if(prof.fn == nil || prof.hz == 0)
1505                 return;
1506
1507         runtime_lock(&prof);
1508         if(prof.fn == nil) {
1509                 runtime_unlock(&prof);
1510                 return;
1511         }
1512         // n = runtime_gentraceback(pc, sp, lr, gp, 0, prof.pcbuf, nelem(prof.pcbuf));
1513         // if(n > 0)
1514         //      prof.fn(prof.pcbuf, n);
1515         runtime_unlock(&prof);
1516 }
1517
1518 // Arrange to call fn with a traceback hz times a second.
1519 void
1520 runtime_setcpuprofilerate(void (*fn)(uintptr*, int32), int32 hz)
1521 {
1522         // Force sane arguments.
1523         if(hz < 0)
1524                 hz = 0;
1525         if(hz == 0)
1526                 fn = nil;
1527         if(fn == nil)
1528                 hz = 0;
1529
1530         // Stop profiler on this cpu so that it is safe to lock prof.
1531         // if a profiling signal came in while we had prof locked,
1532         // it would deadlock.
1533         runtime_resetcpuprofiler(0);
1534
1535         runtime_lock(&prof);
1536         prof.fn = fn;
1537         prof.hz = hz;
1538         runtime_unlock(&prof);
1539         runtime_lock(&runtime_sched);
1540         runtime_sched.profilehz = hz;
1541         runtime_unlock(&runtime_sched);
1542
1543         if(hz != 0)
1544                 runtime_resetcpuprofiler(hz);
1545 }