Stats.c 56.5 KB
Newer Older
1
/* -----------------------------------------------------------------------------
2
 *
3
 * (c) The GHC Team, 1998-2005
4 5 6 7 8
 *
 * Statistics and timing-related functions.
 *
 * ---------------------------------------------------------------------------*/

Simon Marlow's avatar
Simon Marlow committed
9
#include "PosixSource.h"
10
#include "Rts.h"
Simon Marlow's avatar
Simon Marlow committed
11

12
#include "RtsFlags.h"
13
#include "RtsUtils.h"
14
#include "Schedule.h"
15
#include "Stats.h"
16
#include "Profiling.h"
17
#include "GetTime.h"
Simon Marlow's avatar
Simon Marlow committed
18
#include "sm/Storage.h"
Simon Marlow's avatar
Simon Marlow committed
19 20
#include "sm/GCThread.h"
#include "sm/BlockAlloc.h"
21

22 23 24 25 26
// for spin/yield counters
#include "sm/GC.h"
#include "ThreadPaused.h"
#include "Messages.h"

27
#include <string.h> // for memset
28

Simon Marlow's avatar
Simon Marlow committed
29
#define TimeToSecondsDbl(t) ((double)(t) / TIME_RESOLUTION)
30

Simon Marlow's avatar
Simon Marlow committed
31
static Time
Simon Marlow's avatar
Simon Marlow committed
32 33 34
    start_init_cpu, start_init_elapsed,
    end_init_cpu,   end_init_elapsed,
    start_exit_cpu, start_exit_elapsed,
Simon Marlow's avatar
Simon Marlow committed
35
    start_exit_gc_elapsed, start_exit_gc_cpu,
Simon Marlow's avatar
Simon Marlow committed
36
    end_exit_cpu,   end_exit_elapsed;
37

Ben Gamari's avatar
Ben Gamari committed
38
#if defined(PROFILING)
Simon Marlow's avatar
Simon Marlow committed
39 40
static Time RP_start_time  = 0, RP_tot_time  = 0;  // retainer prof user time
static Time RPe_start_time = 0, RPe_tot_time = 0;  // retainer prof elap time
sof's avatar
sof committed
41

Simon Marlow's avatar
Simon Marlow committed
42 43
static Time HC_start_time, HC_tot_time = 0;     // heap census prof user time
static Time HCe_start_time, HCe_tot_time = 0;   // heap census prof elap time
sof's avatar
sof committed
44
#endif
45

46 47 48 49 50 51 52
#if defined(PROF_SPIN)
volatile StgWord64 whitehole_lockClosure_spin = 0;
volatile StgWord64 whitehole_lockClosure_yield = 0;
volatile StgWord64 whitehole_threadPaused_spin = 0;
volatile StgWord64 whitehole_executeMessage_spin = 0;
#endif

Simon Marlow's avatar
Simon Marlow committed
53 54 55 56 57 58 59
//
// All the stats!
//
// This is where we accumulate all the stats during execution, and it's also
// in a convenient form that we can copy over to a caller of getRTSStats().
//
static RTSStats stats;
60

61
static W_ GC_end_faults = 0;
62

Simon Marlow's avatar
Simon Marlow committed
63 64 65
static Time *GC_coll_cpu = NULL;
static Time *GC_coll_elapsed = NULL;
static Time *GC_coll_max_pause = NULL;
66

Ben Gamari's avatar
Ben Gamari committed
67
static void statsPrintf( char *s, ... ) GNUC3_ATTRIBUTE(format (PRINTF, 1, 2));
68 69 70
static void statsFlush( void );
static void statsClose( void );

Simon Marlow's avatar
Simon Marlow committed
71 72 73
/* -----------------------------------------------------------------------------
   Current elapsed time
   ------------------------------------------------------------------------- */
sof's avatar
sof committed
74

Simon Marlow's avatar
Simon Marlow committed
75
Time stat_getElapsedTime(void)
Simon Marlow's avatar
Simon Marlow committed
76
{
Simon Marlow's avatar
Simon Marlow committed
77
    return getProcessElapsedTime() - start_init_elapsed;
Simon Marlow's avatar
Simon Marlow committed
78 79
}

Simon Marlow's avatar
Simon Marlow committed
80 81 82
/* ---------------------------------------------------------------------------
   Measure the current MUT time, for profiling
   ------------------------------------------------------------------------ */
83

Ian Lynagh's avatar
Ian Lynagh committed
84
double
Simon Marlow's avatar
Simon Marlow committed
85
mut_user_time_until( Time t )
Ian Lynagh's avatar
Ian Lynagh committed
86
{
Simon Marlow's avatar
Simon Marlow committed
87
    return TimeToSecondsDbl(t - stats.gc_cpu_ns);
88 89
    // heapCensus() time is included in GC_tot_cpu, so we don't need
    // to subtract it here.
90 91 92

    // TODO: This seems wrong to me. Surely we should be subtracting
    // (at least) start_init_cpu?
Ian Lynagh's avatar
Ian Lynagh committed
93 94
}

95
double
96
mut_user_time( void )
97
{
Simon Marlow's avatar
Simon Marlow committed
98
    Time cpu;
Simon Marlow's avatar
Simon Marlow committed
99
    cpu = getProcessCPUTime();
Ian Lynagh's avatar
Ian Lynagh committed
100
    return mut_user_time_until(cpu);
101 102
}

Ben Gamari's avatar
Ben Gamari committed
103
#if defined(PROFILING)
104
/*
Simon Marlow's avatar
Simon Marlow committed
105
  mut_user_time_during_RP() returns the MUT time during retainer profiling.
106
  The same is for mut_user_time_during_HC();
107
 */
108
static double
109 110
mut_user_time_during_RP( void )
{
Simon Marlow's avatar
Simon Marlow committed
111
    return TimeToSecondsDbl(RP_start_time - stats.gc_cpu_ns - RP_tot_time);
112 113
}

sof's avatar
sof committed
114
#endif /* PROFILING */
115

Simon Marlow's avatar
Simon Marlow committed
116 117 118 119
/* ---------------------------------------------------------------------------
   initStats0() has no dependencies, it can be called right at the beginning
   ------------------------------------------------------------------------ */

120
void
121 122
initStats0(void)
{
Simon Marlow's avatar
Simon Marlow committed
123 124 125 126
    start_init_cpu    = 0;
    start_init_elapsed = 0;
    end_init_cpu     = 0;
    end_init_elapsed  = 0;
127

Simon Marlow's avatar
Simon Marlow committed
128 129
    start_exit_cpu    = 0;
    start_exit_elapsed = 0;
Simon Marlow's avatar
Simon Marlow committed
130 131
    start_exit_gc_cpu    = 0;
    start_exit_gc_elapsed = 0;
Simon Marlow's avatar
Simon Marlow committed
132 133
    end_exit_cpu     = 0;
    end_exit_elapsed  = 0;
134

Ben Gamari's avatar
Ben Gamari committed
135
#if defined(PROFILING)
136 137 138 139 140 141 142 143 144 145 146 147
    RP_start_time  = 0;
    RP_tot_time  = 0;
    RPe_start_time = 0;
    RPe_tot_time = 0;

    HC_start_time = 0;
    HC_tot_time = 0;
    HCe_start_time = 0;
    HCe_tot_time = 0;
#endif

    GC_end_faults = 0;
Simon Marlow's avatar
Simon Marlow committed
148 149 150 151 152 153 154 155 156 157 158 159 160 161

    stats = (RTSStats) {
        .gcs = 0,
        .major_gcs = 0,
        .allocated_bytes = 0,
        .max_live_bytes = 0,
        .max_large_objects_bytes = 0,
        .max_compact_bytes = 0,
        .max_slop_bytes = 0,
        .max_mem_in_use_bytes = 0,
        .cumulative_live_bytes = 0,
        .copied_bytes = 0,
        .par_copied_bytes = 0,
        .cumulative_par_max_copied_bytes = 0,
162
        .cumulative_par_balanced_copied_bytes = 0,
163 164 165 166 167 168 169
        .gc_spin_spin = 0,
        .gc_spin_yield = 0,
        .mut_spin_spin = 0,
        .mut_spin_yield = 0,
        .any_work = 0,
        .no_work = 0,
        .scav_find_work = 0,
170 171
        .init_cpu_ns = 0,
        .init_elapsed_ns = 0,
Simon Marlow's avatar
Simon Marlow committed
172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188
        .mutator_cpu_ns = 0,
        .mutator_elapsed_ns = 0,
        .gc_cpu_ns = 0,
        .gc_elapsed_ns = 0,
        .cpu_ns = 0,
        .elapsed_ns = 0,
        .gc = {
            .gen = 0,
            .threads = 0,
            .allocated_bytes = 0,
            .live_bytes = 0,
            .large_objects_bytes = 0,
            .compact_bytes = 0,
            .slop_bytes = 0,
            .mem_in_use_bytes = 0,
            .copied_bytes = 0,
            .par_max_copied_bytes = 0,
189
            .par_balanced_copied_bytes = 0,
Simon Marlow's avatar
Simon Marlow committed
190 191 192 193 194
            .sync_elapsed_ns = 0,
            .cpu_ns = 0,
            .elapsed_ns = 0
        }
    };
195
}
196

Simon Marlow's avatar
Simon Marlow committed
197 198 199 200
/* ---------------------------------------------------------------------------
   initStats1() can be called after setupRtsFlags()
   ------------------------------------------------------------------------ */

201 202
void initGenerationStats(void);

203 204
void
initStats1 (void)
205 206
{
    if (RtsFlags.GcFlags.giveStats >= VERBOSE_GC_STATS) {
207 208
        statsPrintf("    Alloc    Copied     Live     GC     GC      TOT      TOT  Page Flts\n");
        statsPrintf("    bytes     bytes     bytes   user   elap     user     elap\n");
209
    }
210 211
    GC_coll_cpu =
        (Time *)stgMallocBytes(
Simon Marlow's avatar
Simon Marlow committed
212
            sizeof(Time)*RtsFlags.GcFlags.generations,
213 214 215 216 217
            "initStats");
    GC_coll_elapsed =
        (Time *)stgMallocBytes(
            sizeof(Time)*RtsFlags.GcFlags.generations,
            "initStats");
Simon Marlow's avatar
Simon Marlow committed
218
    GC_coll_max_pause =
219 220 221
        (Time *)stgMallocBytes(
            sizeof(Time)*RtsFlags.GcFlags.generations,
            "initStats");
222 223 224 225 226 227 228
    initGenerationStats();
}

void
initGenerationStats()
{
    for (uint32_t i = 0; i < RtsFlags.GcFlags.generations; i++) {
229
        GC_coll_cpu[i] = 0;
Simon Marlow's avatar
Simon Marlow committed
230 231
        GC_coll_elapsed[i] = 0;
        GC_coll_max_pause[i] = 0;
232
    }
233
}
234

235 236 237 238 239 240 241 242 243 244
/* ---------------------------------------------------------------------------
   Reset stats of child process after fork()
   ------------------------------------------------------------------------ */

void resetChildProcessStats()
{
    initStats0();
    initGenerationStats();
}

245 246 247
/* -----------------------------------------------------------------------------
   Initialisation time...
   -------------------------------------------------------------------------- */
248 249

void
250
stat_startInit(void)
251
{
Simon Marlow's avatar
Simon Marlow committed
252
    getProcessTimes(&start_init_cpu, &start_init_elapsed);
253 254
}

255
void
256
stat_endInit(void)
257
{
Simon Marlow's avatar
Simon Marlow committed
258
    getProcessTimes(&end_init_cpu, &end_init_elapsed);
259 260
    stats.init_cpu_ns = end_init_cpu - start_init_cpu;
    stats.init_elapsed_ns = end_init_elapsed - start_init_elapsed;
261 262
}

263 264
/* -----------------------------------------------------------------------------
   stat_startExit and stat_endExit
265

266 267 268
   These two measure the time taken in shutdownHaskell().
   -------------------------------------------------------------------------- */

269 270 271
void
stat_startExit(void)
{
Simon Marlow's avatar
Simon Marlow committed
272
    getProcessTimes(&start_exit_cpu, &start_exit_elapsed);
Simon Marlow's avatar
Simon Marlow committed
273 274
    start_exit_gc_elapsed = stats.gc_elapsed_ns;
    start_exit_gc_cpu = stats.gc_cpu_ns;
275 276 277 278 279
}

void
stat_endExit(void)
{
Simon Marlow's avatar
Simon Marlow committed
280
    getProcessTimes(&end_exit_cpu, &end_exit_elapsed);
281 282
}

283 284 285 286 287 288
void
stat_startGCSync (gc_thread *gct)
{
    gct->gc_sync_start_elapsed = getProcessElapsedTime();
}

289 290 291 292 293
/* -----------------------------------------------------------------------------
   Called at the beginning of each GC
   -------------------------------------------------------------------------- */

void
294
stat_startGC (Capability *cap, gc_thread *gct)
295
{
Ben Gamari's avatar
Ben Gamari committed
296 297
    if (RtsFlags.GcFlags.ringBell) {
        debugBelch("\007");
298 299
    }

Simon Marlow's avatar
Simon Marlow committed
300
    getProcessTimes(&gct->gc_start_cpu, &gct->gc_start_elapsed);
301 302 303 304 305 306 307 308 309

    // Post EVENT_GC_START with the same timestamp as used for stats
    // (though converted from Time=StgInt64 to EventTimestamp=StgWord64).
    // Here, as opposed to other places, the event is emitted on the cap
    // that initiates the GC and external tools expect it to have the same
    // timestamp as used in +RTS -s calculcations.
    traceEventGcStartAtT(cap,
                         TimeToNS(gct->gc_start_elapsed - start_init_elapsed));

Simon Marlow's avatar
Simon Marlow committed
310 311 312 313
    if (RtsFlags.GcFlags.giveStats != NO_GC_STATS)
    {
        gct->gc_start_faults = getPageFaults();
    }
314

Simon Marlow's avatar
Simon Marlow committed
315
    updateNurseriesStats();
316 317
}

318 319 320 321 322
/* -----------------------------------------------------------------------------
   Called at the end of each GC
   -------------------------------------------------------------------------- */

void
323 324 325 326 327
stat_endGC (Capability *cap, gc_thread *gct, W_ live, W_ copied, W_ slop,
            uint32_t gen, uint32_t par_n_threads, W_ par_max_copied,
            W_ par_balanced_copied, W_ gc_spin_spin, W_ gc_spin_yield,
            W_ mut_spin_spin, W_ mut_spin_yield, W_ any_work, W_ no_work,
            W_ scav_find_work)
328
{
Ben Gamari's avatar
Ben Gamari committed
329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348
    // -------------------------------------------------
    // Collect all the stats about this GC in stats.gc. We always do this since
    // it's relatively cheap and we need allocated_bytes to catch heap
    // overflows.

    stats.gc.gen = gen;
    stats.gc.threads = par_n_threads;

    uint64_t tot_alloc_bytes = calcTotalAllocated() * sizeof(W_);

    // allocated since the last GC
    stats.gc.allocated_bytes = tot_alloc_bytes - stats.allocated_bytes;

    stats.gc.live_bytes = live * sizeof(W_);
    stats.gc.large_objects_bytes = calcTotalLargeObjectsW() * sizeof(W_);
    stats.gc.compact_bytes = calcTotalCompactW() * sizeof(W_);
    stats.gc.slop_bytes = slop * sizeof(W_);
    stats.gc.mem_in_use_bytes = mblocks_allocated * MBLOCK_SIZE;
    stats.gc.copied_bytes = copied * sizeof(W_);
    stats.gc.par_max_copied_bytes = par_max_copied * sizeof(W_);
349
    stats.gc.par_balanced_copied_bytes = par_balanced_copied * sizeof(W_);
Ben Gamari's avatar
Ben Gamari committed
350

duog's avatar
duog committed
351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370
    bool stats_enabled =
        RtsFlags.GcFlags.giveStats != NO_GC_STATS ||
        rtsConfig.gcDoneHook != NULL;

    if (stats_enabled
      || RtsFlags.ProfFlags.doHeapProfile) // heap profiling needs GC_tot_time
    {
        // We only update the times when stats are explicitly enabled since
        // getProcessTimes (e.g. requiring a system call) can be expensive on
        // some platforms.
        Time current_cpu, current_elapsed;
        getProcessTimes(&current_cpu, &current_elapsed);
        stats.cpu_ns = current_cpu - start_init_cpu;
        stats.elapsed_ns = current_elapsed - start_init_elapsed;

        stats.gc.sync_elapsed_ns =
            gct->gc_start_elapsed - gct->gc_sync_start_elapsed;
        stats.gc.elapsed_ns = current_elapsed - gct->gc_start_elapsed;
        stats.gc.cpu_ns = current_cpu - gct->gc_start_cpu;
    }
Ben Gamari's avatar
Ben Gamari committed
371 372 373 374 375 376 377 378 379 380 381 382
    // -------------------------------------------------
    // Update the cumulative stats

    stats.gcs++;
    stats.allocated_bytes = tot_alloc_bytes;
    stats.max_mem_in_use_bytes = peak_mblocks_allocated * MBLOCK_SIZE;

    GC_coll_cpu[gen] += stats.gc.cpu_ns;
    GC_coll_elapsed[gen] += stats.gc.elapsed_ns;
    if (GC_coll_max_pause[gen] < stats.gc.elapsed_ns) {
        GC_coll_max_pause[gen] = stats.gc.elapsed_ns;
    }
Simon Marlow's avatar
Simon Marlow committed
383

Ben Gamari's avatar
Ben Gamari committed
384 385 386 387 388
    stats.copied_bytes += stats.gc.copied_bytes;
    if (par_n_threads > 1) {
        stats.par_copied_bytes += stats.gc.copied_bytes;
        stats.cumulative_par_max_copied_bytes +=
            stats.gc.par_max_copied_bytes;
389 390
        stats.cumulative_par_balanced_copied_bytes +=
            stats.gc.par_balanced_copied_bytes;
391 392 393 394 395 396 397
        stats.any_work += any_work;
        stats.no_work += no_work;
        stats.scav_find_work += scav_find_work;
        stats.gc_spin_spin += gc_spin_spin;
        stats.gc_spin_yield += gc_spin_yield;
        stats.mut_spin_spin += mut_spin_spin;
        stats.mut_spin_yield += mut_spin_yield;
Ben Gamari's avatar
Ben Gamari committed
398 399 400
    }
    stats.gc_cpu_ns += stats.gc.cpu_ns;
    stats.gc_elapsed_ns += stats.gc.elapsed_ns;
Simon Marlow's avatar
Simon Marlow committed
401

Ben Gamari's avatar
Ben Gamari committed
402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417
    if (gen == RtsFlags.GcFlags.generations-1) { // major GC?
        stats.major_gcs++;
        if (stats.gc.live_bytes > stats.max_live_bytes) {
            stats.max_live_bytes = stats.gc.live_bytes;
        }
        if (stats.gc.large_objects_bytes > stats.max_large_objects_bytes) {
            stats.max_large_objects_bytes = stats.gc.large_objects_bytes;
        }
        if (stats.gc.compact_bytes > stats.max_compact_bytes) {
            stats.max_compact_bytes = stats.gc.compact_bytes;
        }
        if (stats.gc.slop_bytes > stats.max_slop_bytes) {
            stats.max_slop_bytes = stats.gc.slop_bytes;
        }
        stats.cumulative_live_bytes += stats.gc.live_bytes;
    }
Simon Marlow's avatar
Simon Marlow committed
418

Ben Gamari's avatar
Ben Gamari committed
419 420
    // -------------------------------------------------
    // Do the more expensive bits only when stats are enabled.
Simon Marlow's avatar
Simon Marlow committed
421

duog's avatar
duog committed
422
    if (stats_enabled)
Ben Gamari's avatar
Ben Gamari committed
423
    {
Simon Marlow's avatar
Simon Marlow committed
424 425
        // -------------------------------------------------
        // Emit events to the event log
426

427 428 429 430 431
        // Has to be emitted while all caps stopped for GC, but before GC_END.
        // See trac.haskell.org/ThreadScope/wiki/RTSsummaryEvents
        // for a detailed design rationale of the current setup
        // of GC eventlog events.
        traceEventGcGlobalSync(cap);
432

433
        // Emitted before GC_END on all caps, which simplifies tools code.
434 435
        traceEventGcStats(cap,
                          CAPSET_HEAP_DEFAULT,
Simon Marlow's avatar
Simon Marlow committed
436 437 438
                          stats.gc.gen,
                          stats.gc.copied_bytes,
                          stats.gc.slop_bytes,
439
                          /* current loss due to fragmentation */
440 441
                          (mblocks_allocated * BLOCKS_PER_MBLOCK
                           - n_alloc_blocks) * BLOCK_SIZE,
442
                          par_n_threads,
Simon Marlow's avatar
Simon Marlow committed
443
                          stats.gc.par_max_copied_bytes,
444 445
                          stats.gc.copied_bytes,
                          stats.gc.par_balanced_copied_bytes);
Simon Marlow's avatar
Simon Marlow committed
446

447 448 449 450 451
        // Post EVENT_GC_END with the same timestamp as used for stats
        // (though converted from Time=StgInt64 to EventTimestamp=StgWord64).
        // Here, as opposed to other places, the event is emitted on the cap
        // that initiates the GC and external tools expect it to have the same
        // timestamp as used in +RTS -s calculcations.
Simon Marlow's avatar
Simon Marlow committed
452
        traceEventGcEndAtT(cap, TimeToNS(stats.elapsed_ns));
Simon Marlow's avatar
Simon Marlow committed
453

Simon Marlow's avatar
Simon Marlow committed
454 455 456 457 458
        if (gen == RtsFlags.GcFlags.generations-1) { // major GC?
            traceEventHeapLive(cap,
                               CAPSET_HEAP_DEFAULT,
                               stats.gc.live_bytes);
        }
459

Simon Marlow's avatar
Simon Marlow committed
460 461
        // -------------------------------------------------
        // Print GC stats to stdout or a file (+RTS -S/-s)
462

Simon Marlow's avatar
Simon Marlow committed
463
        if (RtsFlags.GcFlags.giveStats == VERBOSE_GC_STATS) {
464 465
            W_ faults = getPageFaults();

466
            statsPrintf("%9" FMT_Word64 " %9" FMT_Word64 " %9" FMT_Word64,
Simon Marlow's avatar
Simon Marlow committed
467 468 469 470 471 472 473 474 475
                        stats.gc.allocated_bytes, stats.gc.copied_bytes,
                        stats.gc.live_bytes);

            statsPrintf(" %6.3f %6.3f %8.3f %8.3f %4"
                        FMT_Word " %4" FMT_Word "  (Gen: %2d)\n",
                    TimeToSecondsDbl(stats.gc.cpu_ns),
                    TimeToSecondsDbl(stats.gc.elapsed_ns),
                    TimeToSecondsDbl(stats.cpu_ns),
                    TimeToSecondsDbl(stats.elapsed_ns),
476
                    faults - gct->gc_start_faults,
Simon Marlow's avatar
Simon Marlow committed
477 478 479 480
                        gct->gc_start_faults - GC_end_faults,
                    gen);

            GC_end_faults = faults;
481 482
            statsFlush();
        }
483

484 485

        if (rtsConfig.gcDoneHook != NULL) {
Simon Marlow's avatar
Simon Marlow committed
486
            rtsConfig.gcDoneHook(&stats.gc);
Simon Marlow's avatar
Simon Marlow committed
487
        }
488

489
        traceEventHeapSize(cap,
490
                           CAPSET_HEAP_DEFAULT,
Simon Marlow's avatar
Simon Marlow committed
491
                           mblocks_allocated * MBLOCK_SIZE);
492 493 494
    }
}

495 496 497
/* -----------------------------------------------------------------------------
   Called at the beginning of each Retainer Profiliing
   -------------------------------------------------------------------------- */
Ben Gamari's avatar
Ben Gamari committed
498
#if defined(PROFILING)
499 500
void
stat_startRP(void)
501
{
Simon Marlow's avatar
Simon Marlow committed
502
    Time user, elapsed;
503
    getProcessTimes( &user, &elapsed );
504 505 506

    RP_start_time = user;
    RPe_start_time = elapsed;
507
}
sof's avatar
sof committed
508
#endif /* PROFILING */
509 510 511 512

/* -----------------------------------------------------------------------------
   Called at the end of each Retainer Profiliing
   -------------------------------------------------------------------------- */
513

Ben Gamari's avatar
Ben Gamari committed
514
#if defined(PROFILING)
515 516
void
stat_endRP(
517
  uint32_t retainerGeneration,
518
  int maxStackSize,
519
  double averageNumVisit)
520
{
Simon Marlow's avatar
Simon Marlow committed
521
    Time user, elapsed;
522
    getProcessTimes( &user, &elapsed );
523 524 525

    RP_tot_time += user - RP_start_time;
    RPe_tot_time += elapsed - RPe_start_time;
526

527 528 529 530 531
    fprintf(prof_file, "Retainer Profiling: %d, at %f seconds\n",
      retainerGeneration, mut_user_time_during_RP());
    fprintf(prof_file, "\tMax auxiliary stack size = %u\n", maxStackSize);
    fprintf(prof_file, "\tAverage number of visits per object = %f\n",
            averageNumVisit);
532
}
sof's avatar
sof committed
533
#endif /* PROFILING */
534 535

/* -----------------------------------------------------------------------------
536
   Called at the beginning of each heap census
537
   -------------------------------------------------------------------------- */
Ben Gamari's avatar
Ben Gamari committed
538
#if defined(PROFILING)
539
void
540
stat_startHeapCensus(void)
541
{
Simon Marlow's avatar
Simon Marlow committed
542
    Time user, elapsed;
543
    getProcessTimes( &user, &elapsed );
544 545 546

    HC_start_time = user;
    HCe_start_time = elapsed;
547
}
sof's avatar
sof committed
548
#endif /* PROFILING */
549 550

/* -----------------------------------------------------------------------------
551
   Called at the end of each heap census
552
   -------------------------------------------------------------------------- */
Ben Gamari's avatar
Ben Gamari committed
553
#if defined(PROFILING)
554
void
555
stat_endHeapCensus(void)
556
{
Simon Marlow's avatar
Simon Marlow committed
557
    Time user, elapsed;
558
    getProcessTimes( &user, &elapsed );
559 560 561

    HC_tot_time += user - HC_start_time;
    HCe_tot_time += elapsed - HCe_start_time;
562
}
sof's avatar
sof committed
563
#endif /* PROFILING */
564

565 566 567 568 569 570 571 572
/* -----------------------------------------------------------------------------
   Called at the end of execution

   NOTE: number of allocations is not entirely accurate: it doesn't
   take into account the few bytes at the end of the heap that
   were left unused when the heap-check failed.
   -------------------------------------------------------------------------- */

Ben Gamari's avatar
Ben Gamari committed
573
#if defined(DEBUG)
Simon Marlow's avatar
Simon Marlow committed
574 575 576 577 578 579 580 581 582 583 584 585 586 587
#define TICK_VAR_INI(arity) \
  StgInt SLOW_CALLS_##arity = 1; \
  StgInt RIGHT_ARITY_##arity = 1; \
  StgInt TAGGED_PTR_##arity = 0;

TICK_VAR_INI(1)
TICK_VAR_INI(2)

StgInt TOTAL_CALLS=1;
#endif

/* Report the value of a counter */
#define REPORT(counter) \
  { \
Ben Gamari's avatar
Ben Gamari committed
588
    showStgWord64(counter,temp,true/*commas*/); \
589
    statsPrintf("  (" #counter ")  : %s\n",temp); \
Simon Marlow's avatar
Simon Marlow committed
590 591 592 593 594
  }

/* Report the value of a counter as a percentage of another counter */
#define REPORT_PCT(counter,countertot) \
  statsPrintf("  (" #counter ") %% of (" #countertot ") : %.1f%%\n", \
595
              counter*100.0/countertot)
Simon Marlow's avatar
Simon Marlow committed
596 597 598 599 600 601 602 603 604 605

#define TICK_PRINT(arity) \
  REPORT(SLOW_CALLS_##arity); \
  REPORT_PCT(RIGHT_ARITY_##arity,SLOW_CALLS_##arity); \
  REPORT_PCT(TAGGED_PTR_##arity,RIGHT_ARITY_##arity); \
  REPORT(RIGHT_ARITY_##arity); \
  REPORT(TAGGED_PTR_##arity)

#define TICK_PRINT_TOT(arity) \
  statsPrintf("  (SLOW_CALLS_" #arity ") %% of (TOTAL_CALLS) : %.1f%%\n", \
606
              SLOW_CALLS_##arity * 100.0/TOTAL_CALLS)
Simon Marlow's avatar
Simon Marlow committed
607

608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638
/*
Note [RTS Stats Reporting]
==========================

There are currently three reporting functions:
  * report_summary:
      Responsible for producing '+RTS -s' output.
      Will report internal counters if the RTS flag --internal-counters is
      passed. See [Internal Counters Stats]
  * report_machine_readable:
      Responsible for producing '+RTS -t --machine-readable' output.
  * report_one_line:
      Responsible for productin '+RTS -t' output

Stats are accumulated into the global variable 'stats' as the program runs, then
in 'stat_exit' we do the following:
  * Finalise 'stats'. This involves setting final running times and allocations
    that have not yet been accounted for.
  * Create a RTSSummaryStats. This contains all data for reports that is not
    included in stats (because they do not make sense before the program has
    completed) or in a global variable.
  * call the appropriate report_* function, passing the newly constructed
    RTSSummaryStats.

To ensure that the data in the different reports is kept consistent, the
report_* functions should not do any calculation, excepting unit changes and
formatting. If you need to add a new calculated field, add it to
RTSSummaryStats.

*/

639

640 641 642 643 644 645 646 647 648 649
static void init_RTSSummaryStats(RTSSummaryStats* sum)
{
    const size_t sizeof_gc_summary_stats =
      RtsFlags.GcFlags.generations * sizeof(GenerationSummaryStats);

    memset(sum, 0, sizeof(RTSSummaryStats));
    sum->gc_summary_stats =
      stgMallocBytes(sizeof_gc_summary_stats,
                     "alloc_RTSSummaryStats.gc_summary_stats");
    memset(sum->gc_summary_stats, 0, sizeof_gc_summary_stats);
650
}
651

652 653
static void free_RTSSummaryStats(RTSSummaryStats * sum)
{
654 655
    stgFree(sum->gc_summary_stats);
    sum->gc_summary_stats = NULL;
656
}
Simon Marlow's avatar
Simon Marlow committed
657

658
static void report_summary(const RTSSummaryStats* sum)
659
{
660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677
    // We should do no calculation, other than unit changes and formatting, and
    // we should not not use any data from outside of globals, sum and stats
    // here. See Note [RTS Stats Reporting]

    uint32_t g;
    char temp[512];
    showStgWord64(stats.allocated_bytes, temp, true/*commas*/);
    statsPrintf("%16s bytes allocated in the heap\n", temp);

    showStgWord64(stats.copied_bytes, temp, true/*commas*/);
    statsPrintf("%16s bytes copied during GC\n", temp);

    if ( stats.major_gcs > 0 ) {
        showStgWord64(stats.max_live_bytes, temp, true/*commas*/);
        statsPrintf("%16s bytes maximum residency (%" FMT_Word32
                    " sample(s))\n",
                    temp, stats.major_gcs);
    }
678

679 680 681
    showStgWord64(stats.max_slop_bytes, temp, true/*commas*/);
    statsPrintf("%16s bytes maximum slop\n", temp);

682
    statsPrintf("%16" FMT_Word64 " MiB total memory in use (%"
683
                FMT_Word64 " MB lost due to fragmentation)\n\n",
684
                stats.max_mem_in_use_bytes  / (1024 * 1024),
685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702
                sum->fragmentation_bytes / (1024 * 1024));

    /* Print garbage collections in each gen */
    statsPrintf("                                     Tot time (elapsed)  Avg pause  Max pause\n");
    for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
        const GenerationSummaryStats * gen_stats =
            &sum->gc_summary_stats[g];
        statsPrintf("  Gen %2d     %5d colls"
                    ", %5d par   %6.3fs  %6.3fs     %3.4fs    %3.4fs\n",
                    g, // REVIEWERS: this used to be gen->no
                    //, this can't ever be different right?
                    gen_stats->collections,
                    gen_stats->par_collections,
                    TimeToSecondsDbl(gen_stats->cpu_ns),
                    TimeToSecondsDbl(gen_stats->elapsed_ns),
                    TimeToSecondsDbl(gen_stats->avg_pause_ns),
                    TimeToSecondsDbl(gen_stats->max_pause_ns));
    }
703

704
    statsPrintf("\n");
705

706 707 708 709 710 711 712
#if defined(THREADED_RTS)
    if (RtsFlags.ParFlags.parGcEnabled && sum->work_balance > 0) {
        // See Note [Work Balance]
        statsPrintf("  Parallel GC work balance: "
                    "%.2f%% (serial 0%%, perfect 100%%)\n\n",
                    sum->work_balance * 100);
    }
713

714 715 716 717 718 719
    statsPrintf("  TASKS: %d "
                "(%d bound, %d peak workers (%d total), using -N%d)\n\n",
                taskCount, sum->bound_task_count,
                peakWorkerCount, workerCount,
                n_capabilities);

720
    statsPrintf("  SPARKS: %" FMT_Word64
721
                " (%" FMT_Word " converted, %" FMT_Word " overflowed, %"
722 723 724 725 726 727
                FMT_Word " dud, %" FMT_Word " GC'd, %" FMT_Word " fizzled)\n\n",
                sum->sparks_count,
                sum->sparks.converted, sum->sparks.overflowed,
                sum->sparks.dud, sum->sparks.gcd,
                sum->sparks.fizzled);
#endif
728

729 730 731
    statsPrintf("  INIT    time  %7.3fs  (%7.3fs elapsed)\n",
                TimeToSecondsDbl(stats.init_cpu_ns),
                TimeToSecondsDbl(stats.init_elapsed_ns));
732

733 734 735 736 737 738
    statsPrintf("  MUT     time  %7.3fs  (%7.3fs elapsed)\n",
                TimeToSecondsDbl(stats.mutator_cpu_ns),
                TimeToSecondsDbl(stats.mutator_elapsed_ns));
    statsPrintf("  GC      time  %7.3fs  (%7.3fs elapsed)\n",
                TimeToSecondsDbl(stats.gc_cpu_ns),
                TimeToSecondsDbl(stats.gc_elapsed_ns));
739

740 741
#if defined(PROFILING)
    statsPrintf("  RP      time  %7.3fs  (%7.3fs elapsed)\n",
742 743
                TimeToSecondsDbl(sum->rp_cpu_ns),
                TimeToSecondsDbl(sum->rp_elapsed_ns));
744 745 746 747 748 749 750 751 752 753 754 755 756 757 758
    statsPrintf("  PROF    time  %7.3fs  (%7.3fs elapsed)\n",
                TimeToSecondsDbl(sum->hc_cpu_ns),
                TimeToSecondsDbl(sum->hc_elapsed_ns));
#endif
    statsPrintf("  EXIT    time  %7.3fs  (%7.3fs elapsed)\n",
                TimeToSecondsDbl(sum->exit_cpu_ns),
                TimeToSecondsDbl(sum->exit_elapsed_ns));
    statsPrintf("  Total   time  %7.3fs  (%7.3fs elapsed)\n\n",
                TimeToSecondsDbl(stats.cpu_ns),
                TimeToSecondsDbl(stats.elapsed_ns));
#if !defined(THREADED_RTS)
    statsPrintf("  %%GC     time     %5.1f%%  (%.1f%% elapsed)\n\n",
                sum->gc_cpu_percent * 100,
                sum->gc_elapsed_percent * 100);
#endif
dmp's avatar
dmp committed
759

760
    showStgWord64(sum->alloc_rate, temp, true/*commas*/);
dmp's avatar
dmp committed
761

762
    statsPrintf("  Alloc rate    %s bytes per MUT second\n\n", temp);
Simon Marlow's avatar
Simon Marlow committed
763

764 765 766 767
    statsPrintf("  Productivity %5.1f%% of total user, "
                "%.1f%% of total elapsed\n\n",
                sum->productivity_cpu_percent * 100,
                sum->productivity_elapsed_percent * 100);
Simon Marlow's avatar
Simon Marlow committed
768

769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858
    // See Note [Internal Counter Stats] for a description of the
    // following counters. If you add a counter here, please remember
    // to update the Note.
    if (RtsFlags.MiscFlags.internalCounters) {
#if defined(THREADED_RTS) && defined(PROF_SPIN)
        const int32_t col_width[] = {4, -30, 14, 14};
        statsPrintf("Internal Counters:\n");
        statsPrintf("%*s" "%*s" "%*s" "%*s" "\n"
                    , col_width[0], ""
                    , col_width[1], "SpinLock"
                    , col_width[2], "Spins"
                    , col_width[3], "Yields");
        statsPrintf("%*s" "%*s" "%*" FMT_Word64 "%*" FMT_Word64 "\n"
                    , col_width[0], ""
                    , col_width[1], "gc_alloc_block_sync"
                    , col_width[2], gc_alloc_block_sync.spin
                    , col_width[3], gc_alloc_block_sync.yield);
        statsPrintf("%*s" "%*s" "%*" FMT_Word64 "%*" FMT_Word64 "\n"
                    , col_width[0], ""
                    , col_width[1], "gc_spin"
                    , col_width[2], stats.gc_spin_spin
                    , col_width[3], stats.gc_spin_yield);
        statsPrintf("%*s" "%*s" "%*" FMT_Word64 "%*" FMT_Word64 "\n"
                    , col_width[0], ""
                    , col_width[1], "mut_spin"
                    , col_width[2], stats.mut_spin_spin
                    , col_width[3], stats.mut_spin_yield);
        statsPrintf("%*s" "%*s" "%*" FMT_Word64 "%*s\n"
                    , col_width[0], ""
                    , col_width[1], "whitehole_gc"
                    , col_width[2], whitehole_gc_spin
                    , col_width[3], "n/a");
        statsPrintf("%*s" "%*s" "%*" FMT_Word64 "%*s\n"
                    , col_width[0], ""
                    , col_width[1], "whitehole_threadPaused"
                    , col_width[2], whitehole_threadPaused_spin
                    , col_width[3], "n/a");
        statsPrintf("%*s" "%*s" "%*" FMT_Word64 "%*s\n"
                    , col_width[0], ""
                    , col_width[1], "whitehole_executeMessage"
                    , col_width[2], whitehole_executeMessage_spin
                    , col_width[3], "n/a");
        statsPrintf("%*s" "%*s" "%*" FMT_Word64 "%*" FMT_Word64 "\n"
                    , col_width[0], ""
                    , col_width[1], "whitehole_lockClosure"
                    , col_width[2], whitehole_lockClosure_spin
                    , col_width[3], whitehole_lockClosure_yield);
        // waitForGcThreads isn't really spin-locking(see the function)
        // but these numbers still seem useful.
        statsPrintf("%*s" "%*s" "%*" FMT_Word64 "%*" FMT_Word64 "\n"
                    , col_width[0], ""
                    , col_width[1], "waitForGcThreads"
                    , col_width[2], waitForGcThreads_spin
                    , col_width[3], waitForGcThreads_yield);

        for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
            int prefix_length = 0;
            statsPrintf("%*s" "gen[%" FMT_Word32 "%n",
                        col_width[0], "", g, &prefix_length);
            prefix_length -= col_width[0];
            int suffix_length = col_width[1] + prefix_length;
            suffix_length =
                  suffix_length > 0 ? col_width[1] : suffix_length;

            statsPrintf("%*s" "%*" FMT_Word64 "%*" FMT_Word64 "\n"
                        , suffix_length, "].sync"
                        , col_width[2], generations[g].sync.spin
                        , col_width[3], generations[g].sync.yield);
        }
        statsPrintf("\n");
        statsPrintf("%*s" "%*s" "%*" FMT_Word64 "\n"
                    , col_width[0], ""
                    , col_width[1], "any_work"
                    , col_width[2], stats.any_work);
        statsPrintf("%*s" "%*s" "%*" FMT_Word64 "\n"
                    , col_width[0], ""
                    , col_width[1], "no_work"
                    , col_width[2], stats.no_work);
        statsPrintf("%*s" "%*s" "%*" FMT_Word64 "\n"
                    , col_width[0], ""
                    , col_width[1], "scav_find_work"
                    , col_width[2], stats.scav_find_work);
#elif defined(THREADED_RTS) // THREADED_RTS && PROF_SPIN
        statsPrintf("Internal Counters require the RTS to be built "
                "with PROF_SPIN"); // PROF_SPIN is not #defined here
#else // THREADED_RTS
        statsPrintf("Internal Counters require the threaded RTS");
#endif
    }
}
859

860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904
static void report_machine_readable (const RTSSummaryStats * sum)
{
    // We should do no calculation, other than unit changes and formatting, and
    // we should not not use any data from outside of globals, sum and stats
    // here. See Note [RTS Stats Reporting]
    uint32_t g;

#define MR_STAT(field_name,format,value) \
    statsPrintf(" ,(\"" field_name "\", \"%" format "\")\n", value)
#define MR_STAT_GEN(gen,field_name,format,value) \
    statsPrintf(" ,(\"gen_%" FMT_Word32 "_" field_name "\", \"%" \
      format "\")\n", g, value)

    // These first values are for backwards compatibility.
    // Some of these first fields are duplicated with more machine-readable
    // names, or to match the name in RtsStats.

    // we don't use for the first field helper macro here because the prefix is
    // different
    statsPrintf(" [(\"%s\", \"%" FMT_Word64 "\")\n", "bytes allocated",
                stats.allocated_bytes);
    MR_STAT("num_GCs", FMT_Word32, stats.gcs);
    MR_STAT("average_bytes_used", FMT_Word64, sum->average_bytes_used);
    MR_STAT("max_bytes_used", FMT_Word64, stats.max_live_bytes);
    MR_STAT("num_byte_usage_samples", FMT_Word32, stats.major_gcs);
    MR_STAT("peak_megabytes_allocated", FMT_Word64,
      stats.max_mem_in_use_bytes / (1024 * 1024));

    MR_STAT("init_cpu_seconds", "f", TimeToSecondsDbl(stats.init_cpu_ns));
    MR_STAT("init_wall_seconds", "f", TimeToSecondsDbl(stats.init_elapsed_ns));
    MR_STAT("mut_cpu_seconds", "f", TimeToSecondsDbl(stats.mutator_cpu_ns));
    MR_STAT("mut_wall_seconds", "f",
            TimeToSecondsDbl(stats.mutator_elapsed_ns));
    MR_STAT("GC_cpu_seconds", "f", TimeToSecondsDbl(stats.gc_cpu_ns));
    MR_STAT("GC_wall_seconds", "f", TimeToSecondsDbl(stats.gc_elapsed_ns));

    // end backward compatibility

    // First, the rest of the times

    MR_STAT("exit_cpu_seconds", "f", TimeToSecondsDbl(sum->exit_cpu_ns));
    MR_STAT("exit_wall_seconds", "f", TimeToSecondsDbl(sum->exit_elapsed_ns));
#if defined(PROFILING)
    MR_STAT("rp_cpu_seconds", "f", TimeToSecondsDbl(sum->rp_cpu_ns));
    MR_STAT("rp_wall_seconds", "f", TimeToSecondsDbl(sum->rp_elapsed_ns));
905 906
    MR_STAT("hc_cpu_seconds", "f", TimeToSecondsDbl(sum->hc_cpu_ns));
    MR_STAT("hc_wall_seconds", "f", TimeToSecondsDbl(sum->hc_elapsed_ns));
907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946
#endif
    MR_STAT("total_cpu_seconds", "f", TimeToSecondsDbl(stats.cpu_ns));
    MR_STAT("total_wall_seconds", "f",
            TimeToSecondsDbl(stats.elapsed_ns));

    // next, the remainder of the fields of RTSStats, except internal counters

    // The first two are duplicates of those above, but have more machine
    // readable names that match the field names in RTSStats.


    // gcs has been done as num_GCs above
    MR_STAT("major_gcs", FMT_Word32, stats.major_gcs);
    MR_STAT("allocated_bytes", FMT_Word64, stats.allocated_bytes);
    MR_STAT("max_live_bytes", FMT_Word64, stats.max_live_bytes);
    MR_STAT("max_large_objects_bytes", FMT_Word64,
            stats.max_large_objects_bytes);
    MR_STAT("max_compact_bytes", FMT_Word64, stats.max_compact_bytes);
    MR_STAT("max_slop_bytes", FMT_Word64, stats.max_slop_bytes);
    // This duplicates, except for unit, peak_megabytes_allocated above
    MR_STAT("max_mem_in_use_bytes", FMT_Word64, stats.max_mem_in_use_bytes);
    MR_STAT("cumulative_live_bytes", FMT_Word64, stats.cumulative_live_bytes);
    MR_STAT("copied_bytes", FMT_Word64, stats.copied_bytes);
    MR_STAT("par_copied_bytes", FMT_Word64, stats.par_copied_bytes);
    MR_STAT("cumulative_par_max_copied_bytes", FMT_Word64,
            stats.cumulative_par_max_copied_bytes);
    MR_STAT("cumulative_par_balanced_copied_bytes", FMT_Word64,
            stats.cumulative_par_balanced_copied_bytes);

    // next, the computed fields in RTSSummaryStats
#if !defined(THREADED_RTS) // THREADED_RTS
    MR_STAT("gc_cpu_percent", "f", sum->gc_cpu_percent);
    MR_STAT("gc_wall_percent", "f", sum->gc_cpu_percent);
#endif
    MR_STAT("fragmentation_bytes", FMT_Word64, sum->fragmentation_bytes);
    // average_bytes_used is done above
    MR_STAT("alloc_rate", FMT_Word64, sum->alloc_rate);
    MR_STAT("productivity_cpu_percent", "f", sum->productivity_cpu_percent);
    MR_STAT("productivity_wall_percent", "f",
            sum->productivity_elapsed_percent);
947

948
    // next, the THREADED_RTS fields in RTSSummaryStats
949

950 951 952
#if defined(THREADED_RTS)
    MR_STAT("bound_task_count", FMT_Word32, sum->bound_task_count);
    MR_STAT("sparks_count", FMT_Word64, sum->sparks_count);
Ben Gamari's avatar
Ben Gamari committed
953 954 955 956 957
    MR_STAT("sparks_converted", FMT_Word, sum->sparks.converted);
    MR_STAT("sparks_overflowed", FMT_Word, sum->sparks.overflowed);
    MR_STAT("sparks_dud ", FMT_Word, sum->sparks.dud);
    MR_STAT("sparks_gcd", FMT_Word, sum->sparks.gcd);
    MR_STAT("sparks_fizzled", FMT_Word, sum->sparks.fizzled);
958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014
    MR_STAT("work_balance", "f", sum->work_balance);

    // next, globals (other than internal counters)
    MR_STAT("n_capabilities", FMT_Word32, n_capabilities);
    MR_STAT("task_count", FMT_Word32, taskCount);
    MR_STAT("peak_worker_count", FMT_Word32, peakWorkerCount);
    MR_STAT("worker_count", FMT_Word32, workerCount);

    // next, internal counters
#if defined(PROF_SPIN)
    MR_STAT("gc_alloc_block_sync_spin", FMT_Word64, gc_alloc_block_sync.spin);
    MR_STAT("gc_alloc_block_sync_yield", FMT_Word64,
            gc_alloc_block_sync.yield);
    MR_STAT("gc_alloc_block_sync_spin", FMT_Word64, gc_alloc_block_sync.spin);
    MR_STAT("gc_spin_spin", FMT_Word64, stats.gc_spin_spin);
    MR_STAT("gc_spin_yield", FMT_Word64, stats.gc_spin_yield);
    MR_STAT("mut_spin_spin", FMT_Word64, stats.mut_spin_spin);
    MR_STAT("mut_spin_yield", FMT_Word64, stats.mut_spin_yield);
    MR_STAT("waitForGcThreads_spin", FMT_Word64, waitForGcThreads_spin);
    MR_STAT("waitForGcThreads_yield", FMT_Word64,
            waitForGcThreads_yield);
    MR_STAT("whitehole_gc_spin", FMT_Word64, whitehole_gc_spin);
    MR_STAT("whitehole_lockClosure_spin", FMT_Word64,
            whitehole_lockClosure_spin);
    MR_STAT("whitehole_lockClosure_yield", FMT_Word64,
            whitehole_lockClosure_yield);
    MR_STAT("whitehole_executeMessage_spin", FMT_Word64,
            whitehole_executeMessage_spin);
    MR_STAT("whitehole_threadPaused_spin", FMT_Word64,
            whitehole_threadPaused_spin);
    MR_STAT("any_work", FMT_Word64,
            stats.any_work);
    MR_STAT("no_work", FMT_Word64,
            stats.no_work);
    MR_STAT("scav_find_work", FMT_Word64,
            stats.scav_find_work);
#endif // PROF_SPIN
#endif // THREADED_RTS

    // finally, per-generation stats. Named as, for example for generation 0,
    // gen_0_collections
    for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
        const GenerationSummaryStats* gc_sum = &sum->gc_summary_stats[g];
        MR_STAT_GEN(g, "collections", FMT_Word32, gc_sum->collections);
        MR_STAT_GEN(g, "par_collections", FMT_Word32, gc_sum->par_collections);
        MR_STAT_GEN(g, "cpu_seconds", "f", TimeToSecondsDbl(gc_sum->cpu_ns));
        MR_STAT_GEN(g, "wall_seconds", "f",
                    TimeToSecondsDbl(gc_sum->elapsed_ns));
        MR_STAT_GEN(g, "max_pause_seconds", "f",
                    TimeToSecondsDbl(gc_sum->max_pause_ns));
        MR_STAT_GEN(g, "avg_pause_seconds", "f",
                    TimeToSecondsDbl(gc_sum->avg_pause_ns));
#if defined(THREADED_RTS) && defined(PROF_SPIN)
        MR_STAT_GEN(g, "sync_spin", FMT_Word64, gc_sum->sync_spin);
        MR_STAT_GEN(g, "sync_yield", FMT_Word64, gc_sum->sync_yield);
#endif
    }
1015

1016 1017
    statsPrintf(" ]\n");
}
1018

1019 1020 1021 1022 1023
static void report_one_line(const RTSSummaryStats * sum)
{
    // We should do no calculation, other than unit changes and formatting, and
    // we should not not use any data from outside of globals, sum and stats
    // here. See Note [RTS Stats Reporting]
Simon Marlow's avatar
Simon Marlow committed
1024

1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047
    /* print the long long separately to avoid bugginess on mingwin (2001-07-02,
    mingw-0.5) */
    statsPrintf("<<ghc: %" FMT_Word64 " bytes, "
                "%" FMT_Word32 " GCs, "
                "%" FMT_Word64 "/%" FMT_Word64 " avg/max bytes residency "
                "(%" FMT_Word32 " samples), "
                "%" FMT_Word64 "M in use, "
                "%.3f INIT (%.3f elapsed), "
                "%.3f MUT (%.3f elapsed), "
                "%.3f GC (%.3f elapsed) :ghc>>\n",
                stats.allocated_bytes,
                stats.gcs,
                sum->average_bytes_used,
                stats.max_live_bytes,
                stats.major_gcs,
                stats.max_mem_in_use_bytes / (1024 * 1024),
                TimeToSecondsDbl(stats.init_cpu_ns),
                TimeToSecondsDbl(stats.init_elapsed_ns),
                TimeToSecondsDbl(stats.mutator_cpu_ns),
                TimeToSecondsDbl(stats.mutator_elapsed_ns),
                TimeToSecondsDbl(stats.gc_cpu_ns),
                TimeToSecondsDbl(stats.gc_elapsed_ns));
}
Simon Marlow's avatar
Simon Marlow committed
1048

1049 1050 1051 1052 1053
void
stat_exit (void)
{
    RTSSummaryStats sum;
    init_RTSSummaryStats(&sum);
1054

1055 1056 1057 1058 1059 1060
    if (RtsFlags.GcFlags.giveStats != NO_GC_STATS) {
        // First we tidy the times in stats, and populate the times in sum.
        // In particular, we adjust the gc_* time in stats to remove
        // profiling times.
        {
            Time now_cpu_ns, now_elapsed_ns;
1061
            getProcessTimes(&now_cpu_ns, &now_elapsed_ns);
1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079

            stats.cpu_ns = now_cpu_ns - start_init_cpu;
            stats.elapsed_ns = now_elapsed_ns - start_init_elapsed;
            /* avoid divide by zero if stats.total_cpu_ns is measured as 0.00
               seconds -- SDM */
            if (stats.cpu_ns <= 0) { stats.cpu_ns = 1; }
            if (stats.elapsed_ns <= 0) { stats.elapsed_ns = 1; }

#if defined(PROFILING)
            sum.rp_cpu_ns = RP_tot_time;
            sum.rp_elapsed_ns = RPe_tot_time;
            sum.hc_cpu_ns = HC_tot_time;
            sum.hc_elapsed_ns = HCe_tot_time;
#endif // PROFILING

            // We do a GC during the EXIT phase. We'll attribute the cost of
            // that to GC instead of EXIT, so carefully subtract it from the
            // EXIT time.
1080 1081 1082 1083 1084
            // Note that exit_gc includes RP and HC for the exit GC too.
            Time exit_gc_cpu     = stats.gc_cpu_ns - start_exit_gc_cpu;
            Time exit_gc_elapsed = stats.gc_elapsed_ns - start_exit_gc_elapsed;

            ASSERT(exit_gc_elapsed > 0);
1085 1086 1087 1088 1089 1090 1091 1092

            sum.exit_cpu_ns     = end_exit_cpu
                                      - start_exit_cpu
                                      - exit_gc_cpu;
            sum.exit_elapsed_ns = end_exit_elapsed
                                       - start_exit_elapsed
                                       - exit_gc_elapsed;

1093 1094
            ASSERT(sum.exit_elapsed_ns >= 0);

1095 1096
            stats.mutator_cpu_ns     = start_exit_cpu
                                 - end_init_cpu
1097
                                 - (stats.gc_cpu_ns - exit_gc_cpu);
1098 1099
            stats.mutator_elapsed_ns = start_exit_elapsed
                                 - end_init_elapsed
1100 1101 1102
                                 - (stats.gc_elapsed_ns - exit_gc_elapsed);

            ASSERT(stats.mutator_elapsed_ns >= 0);
1103 1104 1105 1106 1107 1108 1109

            if (stats.mutator_cpu_ns < 0) { stats.mutator_cpu_ns = 0; }

            // The subdivision of runtime into INIT/EXIT/GC/MUT is just adding
            // and subtracting, so the parts should add up to the total exactly.
            // Note that stats->total_ns is captured a tiny bit later than
            // end_exit_elapsed, so we don't use it here.
1110 1111 1112 1113
            ASSERT(stats.init_elapsed_ns // INIT
                   + stats.mutator_elapsed_ns // MUT
                   + stats.gc_elapsed_ns // GC
                   + sum.exit_elapsed_ns // EXIT
1114
                   == end_exit_elapsed - start_init_elapsed);
1115

1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132
            // heapCensus() is called by the GC, so RP and HC time are
            // included in the GC stats.  We therefore subtract them to
            // obtain the actual GC cpu time.
            Time prof_cpu     = sum.rp_cpu_ns + sum.hc_cpu_ns;
            Time prof_elapsed = sum.rp_elapsed_ns + sum.hc_elapsed_ns;

            stats.gc_cpu_ns      -=  prof_cpu;
            stats.gc_elapsed_ns  -=  prof_elapsed;

            // This assertion is probably not necessary; make sure the
            // subdivision with PROF also makes sense
            ASSERT(stats.init_elapsed_ns // INIT
                   + stats.mutator_elapsed_ns // MUT
                   + stats.gc_elapsed_ns // GC
                   + sum.exit_elapsed_ns // EXIT
                   + (sum.rp_elapsed_ns + sum.hc_elapsed_ns) // PROF
                   == end_exit_elapsed - start_init_elapsed);
1133
        }
1134

1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148
        // REVIEWERS: it's not clear to me why the following isn't done in
        // stat_endGC of the last garbage collection?

        // We account for the last garbage collection
        {
            uint64_t tot_alloc_bytes = calcTotalAllocated() * sizeof(W_);
            stats.gc.allocated_bytes = tot_alloc_bytes - stats.allocated_bytes;
            stats.allocated_bytes = tot_alloc_bytes;
            if (RtsFlags.GcFlags.giveStats >= VERBOSE_GC_STATS) {
                statsPrintf("%9" FMT_Word " %9.9s %9.9s",
                            (W_)stats.gc.allocated_bytes, "", "");
                statsPrintf(" %6.3f %6.3f\n\n", 0.0, 0.0);
            }
        }
Simon Marlow's avatar
Simon Marlow committed
1149

1150 1151 1152 1153 1154
        // We populate the remainder (non-time elements) of sum
        {
    #if defined(THREADED_RTS)
            sum.bound_task_count = taskCount - workerCount;

1155
            for (uint32_t i = 0; i < n_capabilities; i++) {
1156 1157 1158 1159 1160 1161 1162 1163 1164
                sum.sparks.created   += capabilities[i]->spark_stats.created;
                sum.sparks.dud       += capabilities[i]->spark_stats.dud;
                sum.sparks.overflowed+=
                  capabilities[i]->spark_stats.overflowed;
                sum.sparks.converted +=
                  capabilities[i]->spark_stats.converted;
                sum.sparks.gcd       += capabilities[i]->spark_stats.gcd;
                sum.sparks.fizzled   += capabilities[i]->spark_stats.fizzled;
            }
1165

1166 1167 1168 1169 1170 1171 1172 1173 1174
            sum.sparks_count = sum.sparks.created
                + sum.sparks.dud
                + sum.sparks.overflowed;

            if (RtsFlags.ParFlags.parGcEnabled && stats.par_copied_bytes > 0) {
                // See Note [Work Balance]
                sum.work_balance =
                    (double)stats.cumulative_par_balanced_copied_bytes
                    / (double)stats.par_copied_bytes;
1175
            } else {
1176
                sum.work_balance = 0;
1177
            }
1178

1179

1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208
    #else // THREADED_RTS
            sum.gc_cpu_percent     = stats.gc_cpu_ns
                                  / stats.cpu_ns;
            sum.gc_elapsed_percent = stats.gc_elapsed_ns
                                  / stats.elapsed_ns;
    #endif // THREADED_RTS

            sum.fragmentation_bytes =
                (uint64_t)(peak_mblocks_allocated
                         * BLOCKS_PER_MBLOCK
                         * BLOCK_SIZE_W
                         - hw_alloc_blocks * BLOCK_SIZE_W)
                / (uint64_t)sizeof(W_);

            sum.average_bytes_used = stats.major_gcs == 0 ? 0 :
                 stats.cumulative_live_bytes/stats.major_gcs,

            sum.alloc_rate = stats.mutator_cpu_ns == 0 ? 0 :
                (uint64_t)((double)stats.allocated_bytes
                / TimeToSecondsDbl(stats.mutator_cpu_ns));

            // REVIEWERS: These two values didn't used to include the exit times
            sum.productivity_cpu_percent =
                TimeToSecondsDbl(stats.cpu_ns
                                - stats.gc_cpu_ns
                                - stats.init_cpu_ns
                                - sum.exit_cpu_ns)
                / TimeToSecondsDbl(stats.cpu_ns);

1209 1210
            ASSERT(sum.productivity_cpu_percent >= 0);

1211 1212 1213 1214 1215 1216 1217
            sum.productivity_elapsed_percent =
                TimeToSecondsDbl(stats.elapsed_ns
                                - stats.gc_elapsed_ns
                                - stats.init_elapsed_ns
                                - sum.exit_elapsed_ns)
                / TimeToSecondsDbl(stats.elapsed_ns);

1218 1219 1220
            ASSERT(sum.productivity_elapsed_percent >= 0);

            for(uint32_t g = 0; g < RtsFlags.GcFlags.generations; ++g) {
1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233
                const generation* gen = &generations[g];
                GenerationSummaryStats* gen_stats = &sum.gc_summary_stats[g];
                gen_stats->collections = gen->collections;
                gen_stats->par_collections = gen->par_collections;
                gen_stats->cpu_ns = GC_coll_cpu[g];
                gen_stats->elapsed_ns = GC_coll_elapsed[g];
                gen_stats->max_pause_ns = GC_coll_max_pause[g];
                gen_stats->avg_pause_ns = gen->collections == 0 ?
                    0 : (GC_coll_elapsed[g] / gen->collections);
    #if defined(THREADED_RTS) && defined(PROF_SPIN)
                gen_stats->sync_spin = gen->sync.spin;
                gen_stats->sync_yield = gen->sync.yield;
    #endif // PROF_SPIN
1234
            }
1235 1236
        }

1237 1238 1239 1240 1241
        // Now we generate the report
        if (RtsFlags.GcFlags.giveStats >= SUMMARY_GC_STATS) {
            report_summary(&sum);
        }

1242
        if (RtsFlags.GcFlags.giveStats == ONELINE_GC_STATS) {
1243 1244 1245 1246 1247 1248
            if (RtsFlags.MiscFlags.machineReadable) {
                report_machine_readable(&sum);
            }
            else {
                report_one_line(&sum);
            }
1249 1250 1251 1252
        }

        statsFlush();
        statsClose();
1253
    }
simonmarhaskell@gmail.com's avatar
simonmarhaskell@gmail.com committed
1254

1255 1256
    free_RTSSummaryStats(&sum);

Simon Marlow's avatar
Simon Marlow committed
1257
    if (GC_coll_cpu) {
Simon Marlow's avatar
Simon Marlow committed
1258
      stgFree(GC_coll_cpu);
Simon Marlow's avatar
Simon Marlow committed
1259 1260 1261
      GC_coll_cpu = NULL;
    }
    if (GC_coll_elapsed) {
Simon Marlow's avatar
Simon Marlow committed
1262
      stgFree(GC_coll_elapsed);
Simon Marlow's avatar
Simon Marlow committed
1263 1264 1265 1266 1267 1268
      GC_coll_elapsed = NULL;
    }
    if (GC_coll_max_pause) {
      stgFree(GC_coll_max_pause);
      GC_coll_max_pause = NULL;
    }
1269
}
1270

1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348
/* Note [Work Balance]
----------------------
Work balance is a measure of how evenly the work done during parallel garbage
collection is spread across threads. To compute work balance we must take care
to account for the number of GC threads changing between GCs. The statistics we
track must have the number of GC threads "integrated out".

We accumulate two values from each garbage collection:
* par_copied: is a measure of the total work done during parallel garbage
  collection
* par_balanced_copied: is a measure of the balanced work done
  during parallel garbage collection.

par_copied is simple to compute, but par_balanced_copied_bytes is somewhat more
complicated:

For a given garbage collection:
Let gc_copied := total copies during the gc
    gc_copied_i := copies by the ith thread during the gc
    num_gc_threads := the number of threads participating in the gc
    balance_limit := (gc_copied / num_gc_threads)

If we were to graph gc_copied_i, sorted from largest to smallest we would see
something like:

       |X
  ^    |X X
  |    |X X X            X: unbalanced copies
copies |-----------      Y: balanced copies by the busiest GC thread
       |Y Z Z            Z: other balanced copies
       |Y Z Z Z
       |Y Z Z Z Z
       |Y Z Z Z Z Z
       |===========
       |1 2 3 4 5 6
          i ->

where the --- line is at balance_limit. Balanced copies are those under the ---
line, i.e. the area of the Ys and Zs. Note that the area occupied by the Ys will
always equal balance_limit. Completely balanced gc has every thread copying
balance_limit and a completely unbalanced gc has a single thread copying
gc_copied.

One could define par_balance_copied as the areas of the Ys and Zs in the graph
above, however we would like the ratio of (par_balance_copied / gc_copied) to
range from 0 to 1, so that work_balance will be a nice percentage, also ranging
from 0 to 1. We therefore define par_balanced_copied as:

                                                        (  num_gc_threads  )
{Sum[Min(gc_copied_i,balance_limit)] - balance_limit} * (------------------)
  i                                                     (num_gc_threads - 1)
                                          vvv                  vvv
                                           S                    T

Where the S and T terms serve to remove the area of the Ys, and
to normalize the result to lie between 0 and gc_copied.

Note that the implementation orders these operations differently to minimize
error due to integer rounding.

Then cumulative work balance is computed as
(cumulative_par_balanced_copied_bytes / par_copied_byes)

Previously, cumulative work balance was computed as:

(cumulative_par_max_copied_bytes)
(-------------------------------) - 1
(       par_copied_bytes        )
-------------------------------------
        (n_capabilities - 1)

This was less accurate than the current method, and invalid whenever a garbage
collection had occurred with num_gc_threads /= n_capabilities; which can happen
when setNumCapabilities is called, when -qn is passed as an RTS option, or when
the number of gc threads is limited to the number of cores.
See #13830
*/

1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381
/*
Note [Internal Counter Stats]
-----------------------------
What do the counts at the end of a '+RTS -s --internal-counters' report mean?
They are detailed below. Most of these counters are used by multiple threads
with no attempt at synchronisation. This means that reported values  may be
lower than the true value and this becomes more likely and more severe as
contention increases.

The first counters are for various SpinLock-like constructs in the RTS. See
Spinlock.h for the definition of a SpinLock. We maintain up two counters per
SpinLock:
* spin: The number of busy-spins over the length of the program.
* yield: The number of times the SpinLock spun SPIN_COUNT times without success
         and called yieldThread().
Not all of these are actual SpinLocks, see the details below.

Actual SpinLocks:
* gc_alloc_block:
    This SpinLock protects the block allocator and free list manager. See
    BlockAlloc.c.
* gc_spin and mut_spin:
    These SpinLocks are used to herd gc worker threads during parallel garbage
    collection. See gcWorkerThread, wakeup_gc_threads and releaseGCThreads.
* gen[g].sync:
    These SpinLocks, one per generation, protect the generations[g] data
    structure during garbage collection.

waitForGcThreads:
  These counters are incremented while we wait for all threads to be ready
  for a parallel garbage collection. We yield more than we spin in this case.

In several places in the runtime we must take a lock on a closure. To do this,
1382
we replace its info table with stg_WHITEHOLE_info, spinning if it is already
1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410
a white-hole. Sometimes we yieldThread() if we spin too long, sometimes we
don't. We count these white-hole spins and include them in the SpinLocks table.
If a particular loop does not yield, we put "n/a" in the table. They are named
for the function that has the spinning loop except that several loops in the
garbage collector accumulate into whitehole_gc.
TODO: Should these counters be more or less granular?

white-hole spin counters:
* whitehole_gc
* whitehole_lockClosure
* whitehole_executeMessage
* whitehole_threadPaused


We count the number of calls of several functions in the parallel garbage
collector.

Parallel garbage collector counters:
* any_work:
    A cheap function called whenever a gc_thread is ready for work. Does
    not do any work.
* no_work:
    Incremented whenever any_work finds no work.
* scav_find_work:
    Called to do work when any_work return true.

*/

1411 1412 1413 1414 1415 1416
/* -----------------------------------------------------------------------------
   stat_describe_gens

   Produce some detailed info on the state of the generational GC.
   -------------------------------------------------------------------------- */
void
1417
statDescribeGens(void)
1418
{
1419
  uint32_t g, mut, lge, compacts, i;
1420 1421 1422
  W_ gen_slop;
  W_ tot_live, tot_slop;
  W_ gen_live, gen_blocks;
1423
  bdescr *bd;
1424
  generation *gen;
1425

1426
  debugBelch(
1427 1428 1429 1430
"----------------------------------------------------------------------\n"
"  Gen     Max  Mut-list  Blocks    Large  Compacts      Live      Slop\n"
"       Blocks     Bytes          Objects                              \n"
"----------------------------------------------------------------------\n");
1431

1432 1433
  tot_live = 0;
  tot_slop = 0;
Simon Marlow's avatar
Simon Marlow committed
1434

1435
  for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
Simon Marlow's avatar
Simon Marlow committed
1436 1437 1438 1439 1440 1441
      gen = &generations[g];

      for (bd = gen->large_objects, lge = 0; bd; bd = bd->link) {
          lge++;
      }

1442 1443 1444 1445
      for (bd = gen->compact_objects, compacts = 0; bd; bd = bd->link) {
          compacts++;
      }

Simon Marlow's avatar
Simon Marlow committed
1446 1447