Refactor PAPI support, and add profiling of multithreaded GC
[ghc.git] / rts / RtsStartup.c
1 /* -----------------------------------------------------------------------------
2 *
3 * (c) The GHC Team, 1998-2002
4 *
5 * Main function for a standalone Haskell program.
6 *
7 * ---------------------------------------------------------------------------*/
8
9 // PAPI uses caddr_t, which is not POSIX
10 // #include "PosixSource.h"
11
12 #include "Rts.h"
13 #include "RtsAPI.h"
14 #include "RtsUtils.h"
15 #include "RtsFlags.h"
16 #include "OSThreads.h"
17 #include "Schedule.h" /* initScheduler */
18 #include "Stats.h" /* initStats */
19 #include "STM.h" /* initSTM */
20 #include "Signals.h"
21 #include "RtsSignals.h"
22 #include "ThrIOManager.h"
23 #include "Timer.h" /* startTimer, stopTimer */
24 #include "Weak.h"
25 #include "Ticky.h"
26 #include "StgRun.h"
27 #include "Prelude.h" /* fixupRTStoPreludeRefs */
28 #include "HsFFI.h"
29 #include "Linker.h"
30 #include "ThreadLabels.h"
31 #include "BlockAlloc.h"
32 #include "Trace.h"
33 #include "RtsTypeable.h"
34 #include "Stable.h"
35 #include "Hpc.h"
36 #include "FileLock.h"
37
38 #if defined(RTS_GTK_FRONTPANEL)
39 #include "FrontPanel.h"
40 #endif
41
42 # include "Profiling.h"
43
44 #if defined(PROFILING)
45 # include "ProfHeap.h"
46 # include "RetainerProfile.h"
47 #endif
48
49 #if defined(GRAN)
50 # include "GranSimRts.h"
51 #endif
52
53 #if defined(GRAN) || defined(PAR)
54 # include "ParallelRts.h"
55 #endif
56
57 #if defined(PAR)
58 # include "Parallel.h"
59 # include "LLC.h"
60 #endif
61
62 #if defined(mingw32_HOST_OS) && !defined(THREADED_RTS)
63 #include "win32/AsyncIO.h"
64 #endif
65
66 #include <stdlib.h>
67
68 #ifdef HAVE_TERMIOS_H
69 #include <termios.h>
70 #endif
71 #ifdef HAVE_SIGNAL_H
72 #include <signal.h>
73 #endif
74
75 #if USE_PAPI
76 #include "Papi.h"
77 #endif
78
79 // Count of how many outstanding hs_init()s there have been.
80 static int hs_init_count = 0;
81
82 // Here we save the terminal settings on the standard file
83 // descriptors, if we need to change them (eg. to support NoBuffering
84 // input).
85 static void *saved_termios[3] = {NULL,NULL,NULL};
86
87 void*
88 __hscore_get_saved_termios(int fd)
89 {
90 return (0 <= fd && fd < (int)(sizeof(saved_termios) / sizeof(*saved_termios))) ?
91 saved_termios[fd] : NULL;
92 }
93
94 void
95 __hscore_set_saved_termios(int fd, void* ts)
96 {
97 if (0 <= fd && fd < (int)(sizeof(saved_termios) / sizeof(*saved_termios))) {
98 saved_termios[fd] = ts;
99 }
100 }
101
102 /* -----------------------------------------------------------------------------
103 Initialise floating point unit on x86 (currently disabled. why?)
104 (see comment in ghc/compiler/nativeGen/MachInstrs.lhs).
105 -------------------------------------------------------------------------- */
106
107 #define X86_INIT_FPU 0
108
109 #if X86_INIT_FPU
110 static void
111 x86_init_fpu ( void )
112 {
113 __volatile unsigned short int fpu_cw;
114
115 // Grab the control word
116 __asm __volatile ("fnstcw %0" : "=m" (fpu_cw));
117
118 #if 0
119 printf("fpu_cw: %x\n", fpu_cw);
120 #endif
121
122 // Set bits 8-9 to 10 (64-bit precision).
123 fpu_cw = (fpu_cw & 0xfcff) | 0x0200;
124
125 // Store the new control word back
126 __asm __volatile ("fldcw %0" : : "m" (fpu_cw));
127 }
128 #endif
129
130 /* -----------------------------------------------------------------------------
131 Starting up the RTS
132 -------------------------------------------------------------------------- */
133
134 void
135 hs_init(int *argc, char **argv[])
136 {
137 hs_init_count++;
138 if (hs_init_count > 1) {
139 // second and subsequent inits are ignored
140 return;
141 }
142
143 #if defined(DEBUG)
144 /* Start off by initialising the allocator debugging so we can
145 * use it anywhere */
146 initAllocator();
147 #endif
148
149 /* Next we do is grab the start time...just in case we're
150 * collecting timing statistics.
151 */
152 stat_startInit();
153
154 #ifdef PAR
155 /*
156 * The parallel system needs to be initialised and synchronised before
157 * the program is run.
158 */
159 startupParallelSystem(argv);
160
161 if (*argv[0] == '-') { /* Strip off mainPE flag argument */
162 argv++;
163 argc--;
164 }
165
166 argv[1] = argv[0]; /* ignore the nPEs argument */
167 argv++; argc--;
168 #endif
169
170 #ifdef USE_PAPI
171 papi_init();
172 #endif
173
174 /* Set the RTS flags to default values. */
175
176 initRtsFlagsDefaults();
177
178 /* Call the user hook to reset defaults, if present */
179 defaultsHook();
180
181 /* Parse the flags, separating the RTS flags from the programs args */
182 if (argc != NULL && argv != NULL) {
183 setFullProgArgv(*argc,*argv);
184 setupRtsFlags(argc, *argv, &rts_argc, rts_argv);
185 setProgArgv(*argc,*argv);
186 }
187
188 /* initTracing must be after setupRtsFlags() */
189 initTracing();
190
191 #if defined(PAR)
192 /* NB: this really must be done after processing the RTS flags */
193 IF_PAR_DEBUG(verbose,
194 debugBelch("==== Synchronising system (%d PEs)\n", nPEs));
195 synchroniseSystem(); // calls initParallelSystem etc
196 #endif /* PAR */
197
198 /* initialise scheduler data structures (needs to be done before
199 * initStorage()).
200 */
201 initScheduler();
202
203 #if defined(GRAN)
204 /* And start GranSim profiling if required: */
205 if (RtsFlags.GranFlags.GranSimStats.Full)
206 init_gr_simulation(rts_argc, rts_argv, prog_argc, prog_argv);
207 #elif defined(PAR)
208 /* And start GUM profiling if required: */
209 if (RtsFlags.ParFlags.ParStats.Full)
210 init_gr_simulation(rts_argc, rts_argv, prog_argc, prog_argv);
211 #endif /* PAR || GRAN */
212
213 /* initialize the storage manager */
214 initStorage();
215
216 /* initialise the stable pointer table */
217 initStablePtrTable();
218
219 /* initialise the shared Typeable store */
220 initTypeableStore();
221
222 /* initialise file locking, if necessary */
223 #if !defined(mingw32_HOST_OS)
224 initFileLocking();
225 #endif
226
227 #if defined(DEBUG)
228 /* initialise thread label table (tso->char*) */
229 initThreadLabelTable();
230 #endif
231
232 initProfiling1();
233
234 /* start the virtual timer 'subsystem'. */
235 initTimer();
236 startTimer();
237
238 /* Initialise the stats department */
239 initStats();
240
241 #if defined(RTS_USER_SIGNALS)
242 if (RtsFlags.MiscFlags.install_signal_handlers) {
243 /* Initialise the user signal handler set */
244 initUserSignals();
245 /* Set up handler to run on SIGINT, etc. */
246 initDefaultHandlers();
247 }
248 #endif
249
250 #if defined(mingw32_HOST_OS) && !defined(THREADED_RTS)
251 startupAsyncIO();
252 #endif
253
254 #ifdef RTS_GTK_FRONTPANEL
255 if (RtsFlags.GcFlags.frontpanel) {
256 initFrontPanel();
257 }
258 #endif
259
260 #if X86_INIT_FPU
261 x86_init_fpu();
262 #endif
263
264 /* Record initialization times */
265 stat_endInit();
266 }
267
268 // Compatibility interface
269 void
270 startupHaskell(int argc, char *argv[], void (*init_root)(void))
271 {
272 hs_init(&argc, &argv);
273 if(init_root)
274 hs_add_root(init_root);
275 }
276
277
278 /* -----------------------------------------------------------------------------
279 Per-module initialisation
280
281 This process traverses all the compiled modules in the program
282 starting with "Main", and performing per-module initialisation for
283 each one.
284
285 So far, two things happen at initialisation time:
286
287 - we register stable names for each foreign-exported function
288 in that module. This prevents foreign-exported entities, and
289 things they depend on, from being garbage collected.
290
291 - we supply a unique integer to each statically declared cost
292 centre and cost centre stack in the program.
293
294 The code generator inserts a small function "__stginit_<module>" in each
295 module and calls the registration functions in each of the modules it
296 imports.
297
298 The init* functions are compiled in the same way as STG code,
299 i.e. without normal C call/return conventions. Hence we must use
300 StgRun to call this stuff.
301 -------------------------------------------------------------------------- */
302
303 /* The init functions use an explicit stack...
304 */
305 #define INIT_STACK_BLOCKS 4
306 static F_ *init_stack = NULL;
307
308 void
309 hs_add_root(void (*init_root)(void))
310 {
311 bdescr *bd;
312 nat init_sp;
313 Capability *cap;
314
315 cap = rts_lock();
316
317 if (hs_init_count <= 0) {
318 barf("hs_add_root() must be called after hs_init()");
319 }
320
321 /* The initialisation stack grows downward, with sp pointing
322 to the last occupied word */
323 init_sp = INIT_STACK_BLOCKS*BLOCK_SIZE_W;
324 bd = allocGroup_lock(INIT_STACK_BLOCKS);
325 init_stack = (F_ *)bd->start;
326 init_stack[--init_sp] = (F_)stg_init_finish;
327 if (init_root != NULL) {
328 init_stack[--init_sp] = (F_)init_root;
329 }
330
331 cap->r.rSp = (P_)(init_stack + init_sp);
332 StgRun((StgFunPtr)stg_init, &cap->r);
333
334 freeGroup_lock(bd);
335
336 startupHpc();
337
338 // This must be done after module initialisation.
339 // ToDo: make this work in the presence of multiple hs_add_root()s.
340 initProfiling2();
341
342 rts_unlock(cap);
343
344 // ditto.
345 #if defined(THREADED_RTS)
346 ioManagerStart();
347 #endif
348 }
349
350 /* ----------------------------------------------------------------------------
351 * Shutting down the RTS
352 *
353 * The wait_foreign parameter means:
354 * True ==> wait for any threads doing foreign calls now.
355 * False ==> threads doing foreign calls may return in the
356 * future, but will immediately block on a mutex.
357 * (capability->lock).
358 *
359 * If this RTS is a DLL that we're about to unload, then you want
360 * safe=True, otherwise the thread might return to code that has been
361 * unloaded. If this is a standalone program that is about to exit,
362 * then you can get away with safe=False, which is better because we
363 * won't hang on exit if there is a blocked foreign call outstanding.
364 *
365 ------------------------------------------------------------------------- */
366
367 static void
368 hs_exit_(rtsBool wait_foreign)
369 {
370 if (hs_init_count <= 0) {
371 errorBelch("warning: too many hs_exit()s");
372 return;
373 }
374 hs_init_count--;
375 if (hs_init_count > 0) {
376 // ignore until it's the last one
377 return;
378 }
379
380 /* start timing the shutdown */
381 stat_startExit();
382
383 #if defined(RTS_USER_SIGNALS)
384 if (RtsFlags.MiscFlags.install_signal_handlers) {
385 freeSignalHandlers();
386 }
387 #endif
388
389 #if defined(THREADED_RTS)
390 ioManagerDie();
391 #endif
392
393 /* stop all running tasks */
394 exitScheduler(wait_foreign);
395
396 #if defined(GRAN)
397 /* end_gr_simulation prints global stats if requested -- HWL */
398 if (!RtsFlags.GranFlags.GranSimStats.Suppressed)
399 end_gr_simulation();
400 #endif
401
402 /* stop the ticker */
403 stopTimer();
404 exitTimer();
405
406 /* reset the standard file descriptors to blocking mode */
407 resetNonBlockingFd(0);
408 resetNonBlockingFd(1);
409 resetNonBlockingFd(2);
410
411 #if HAVE_TERMIOS_H
412 // Reset the terminal settings on the standard file descriptors,
413 // if we changed them. See System.Posix.Internals.tcSetAttr for
414 // more details, including the reason we termporarily disable
415 // SIGTTOU here.
416 {
417 int fd;
418 sigset_t sigset, old_sigset;
419 sigemptyset(&sigset);
420 sigaddset(&sigset, SIGTTOU);
421 sigprocmask(SIG_BLOCK, &sigset, &old_sigset);
422 for (fd = 0; fd <= 2; fd++) {
423 struct termios* ts = (struct termios*)__hscore_get_saved_termios(fd);
424 if (ts != NULL) {
425 tcsetattr(fd,TCSANOW,ts);
426 }
427 }
428 sigprocmask(SIG_SETMASK, &old_sigset, NULL);
429 }
430 #endif
431
432 #if defined(PAR)
433 /* controlled exit; good thread! */
434 shutdownParallelSystem(0);
435
436 /* global statistics in parallel system */
437 PAR_TICKY_PAR_END();
438 #endif
439
440 /* stop timing the shutdown, we're about to print stats */
441 stat_endExit();
442
443 /* shutdown the hpc support (if needed) */
444 exitHpc();
445
446 // clean up things from the storage manager's point of view.
447 // also outputs the stats (+RTS -s) info.
448 exitStorage();
449
450 /* free the tasks */
451 freeScheduler();
452
453 /* free shared Typeable store */
454 exitTypeableStore();
455
456 /* free file locking tables, if necessary */
457 #if !defined(mingw32_HOST_OS)
458 freeFileLocking();
459 #endif
460
461 /* free the stable pointer table */
462 exitStablePtrTable();
463
464 #if defined(DEBUG)
465 /* free the thread label table */
466 freeThreadLabelTable();
467 #endif
468
469 #ifdef RTS_GTK_FRONTPANEL
470 if (RtsFlags.GcFlags.frontpanel) {
471 stopFrontPanel();
472 }
473 #endif
474
475 #if defined(PROFILING)
476 reportCCSProfiling();
477 #endif
478
479 endProfiling();
480 freeProfiling1();
481
482 #ifdef PROFILING
483 // Originally, this was in report_ccs_profiling(). Now, retainer
484 // profiling might tack some extra stuff on to the end of this file
485 // during endProfiling().
486 if (prof_file != NULL) fclose(prof_file);
487 #endif
488
489 #if defined(TICKY_TICKY)
490 if (RtsFlags.TickyFlags.showTickyStats) PrintTickyInfo();
491 #endif
492
493 #if defined(mingw32_HOST_OS) && !defined(THREADED_RTS)
494 shutdownAsyncIO(wait_foreign);
495 #endif
496
497 /* free hash table storage */
498 exitHashTable();
499
500 // Finally, free all our storage
501 freeStorage();
502
503 #if defined(DEBUG)
504 /* and shut down the allocator debugging */
505 shutdownAllocator();
506 #endif
507
508 }
509
510 // The real hs_exit():
511 void
512 hs_exit(void)
513 {
514 hs_exit_(rtsTrue);
515 // be safe; this might be a DLL
516 }
517
518 // Compatibility interfaces
519 void
520 shutdownHaskell(void)
521 {
522 hs_exit();
523 }
524
525 void
526 shutdownHaskellAndExit(int n)
527 {
528 if (hs_init_count == 1) {
529 OnExitHook();
530 hs_exit_(rtsFalse);
531 // we're about to exit(), no need to wait for foreign calls to return.
532 #if defined(PAR)
533 /* really exit (stg_exit() would call shutdownParallelSystem() again) */
534 exit(n);
535 #else
536 stg_exit(n);
537 #endif
538 }
539 }
540
541 /*
542 * called from STG-land to exit the program
543 */
544
545 #ifdef PAR
546 static int exit_started=rtsFalse;
547 #endif
548
549 void (*exitFn)(int) = 0;
550
551 void
552 stg_exit(int n)
553 {
554 #ifdef PAR
555 /* HACK: avoid a loop when exiting due to a stupid error */
556 if (exit_started)
557 return;
558 exit_started=rtsTrue;
559
560 IF_PAR_DEBUG(verbose, debugBelch("==-- stg_exit %d on [%x]...", n, mytid));
561 shutdownParallelSystem(n);
562 #endif
563 if (exitFn)
564 (*exitFn)(n);
565 exit(n);
566 }