summaryrefslogtreecommitdiffstats
path: root/src/activity.c
blob: 65d7032dee26de3c4b7932408da84008c51ef16c (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
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
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
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
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
905
906
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
947
948
949
950
951
952
953
954
955
956
957
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
/*
 * activity measurement functions.
 *
 * Copyright 2000-2018 Willy Tarreau <w@1wt.eu>
 *
 * This program is free software; you can redistribute it and/or
 * modify it under the terms of the GNU General Public License
 * as published by the Free Software Foundation; either version
 * 2 of the License, or (at your option) any later version.
 *
 */

#include <haproxy/activity-t.h>
#include <haproxy/api.h>
#include <haproxy/applet.h>
#include <haproxy/cfgparse.h>
#include <haproxy/clock.h>
#include <haproxy/channel.h>
#include <haproxy/cli.h>
#include <haproxy/freq_ctr.h>
#include <haproxy/sc_strm.h>
#include <haproxy/stconn.h>
#include <haproxy/tools.h>
#include <haproxy/xxhash.h>

/* CLI context for the "show profiling" command */
struct show_prof_ctx {
	int dump_step;  /* 0,1,2,4,5,6; see cli_iohandler_show_profiling() */
	int linenum;    /* next line to be dumped (starts at 0) */
	int maxcnt;     /* max line count per step (0=not set)  */
	int by_addr;    /* 0=sort by usage, 1=sort by address   */
};

#if defined(DEBUG_MEM_STATS)
/* these ones are macros in bug.h when DEBUG_MEM_STATS is set, and will
 * prevent the new ones from being redefined.
 */
#undef calloc
#undef malloc
#undef realloc
#endif

/* bit field of profiling options. Beware, may be modified at runtime! */
unsigned int profiling __read_mostly = HA_PROF_TASKS_AOFF;
unsigned long task_profiling_mask __read_mostly = 0;

/* One struct per thread containing all collected measurements */
struct activity activity[MAX_THREADS] __attribute__((aligned(64))) = { };

/* One struct per function pointer hash entry (256 values, 0=collision) */
struct sched_activity sched_activity[256] __attribute__((aligned(64))) = { };


#ifdef USE_MEMORY_PROFILING
/* determine the number of buckets to store stats */
#define MEMPROF_HASH_BITS 10
#define MEMPROF_HASH_BUCKETS (1U << MEMPROF_HASH_BITS)

enum memprof_method {
	MEMPROF_METH_UNKNOWN = 0,
	MEMPROF_METH_MALLOC,
	MEMPROF_METH_CALLOC,
	MEMPROF_METH_REALLOC,
	MEMPROF_METH_FREE,
	MEMPROF_METH_METHODS /* count, must be last */
};

static const char *const memprof_methods[MEMPROF_METH_METHODS] = {
	"unknown", "malloc", "calloc", "realloc", "free",
};

/* stats:
 *   - malloc increases alloc
 *   - free increases free (if non null)
 *   - realloc increases either depending on the size change.
 * when the real size is known (malloc_usable_size()), it's used in free_tot
 * and alloc_tot, otherwise the requested size is reported in alloc_tot and
 * zero in free_tot.
 */
struct memprof_stats {
	const void *caller;
	enum memprof_method method;
	/* 4-7 bytes hole here */
	unsigned long long alloc_calls;
	unsigned long long free_calls;
	unsigned long long alloc_tot;
	unsigned long long free_tot;
};

/* last one is for hash collisions ("others") and has no caller address */
struct memprof_stats memprof_stats[MEMPROF_HASH_BUCKETS + 1] = { };

/* used to detect recursive calls */
static THREAD_LOCAL int in_memprof = 0;

/* perform a pointer hash by scrambling its bits and retrieving the most
 * mixed ones (topmost ones in 32-bit, middle ones in 64-bit).
 */
static unsigned int memprof_hash_ptr(const void *p)
{
	unsigned long long x = (unsigned long)p;

	x = 0xcbda9653U * x;
	if (sizeof(long) == 4)
		x >>= 32;
	else
		x >>= 33 - MEMPROF_HASH_BITS / 2;
	return x & (MEMPROF_HASH_BUCKETS - 1);
}

/* These ones are used by glibc and will be called early. They are in charge of
 * initializing the handlers with the original functions.
 */
static void *memprof_malloc_initial_handler(size_t size);
static void *memprof_calloc_initial_handler(size_t nmemb, size_t size);
static void *memprof_realloc_initial_handler(void *ptr, size_t size);
static void  memprof_free_initial_handler(void *ptr);

/* Fallback handlers for the main alloc/free functions. They are preset to
 * the initializer in order to save a test in the functions's critical path.
 */
static void *(*memprof_malloc_handler)(size_t size)               = memprof_malloc_initial_handler;
static void *(*memprof_calloc_handler)(size_t nmemb, size_t size) = memprof_calloc_initial_handler;
static void *(*memprof_realloc_handler)(void *ptr, size_t size)   = memprof_realloc_initial_handler;
static void  (*memprof_free_handler)(void *ptr)                   = memprof_free_initial_handler;

/* Used to force to die if it's not possible to retrieve the allocation
 * functions. We cannot even use stdio in this case.
 */
static __attribute__((noreturn)) void memprof_die(const char *msg)
{
	DISGUISE(write(2, msg, strlen(msg)));
	exit(1);
}

/* Resolve original allocation functions and initialize all handlers.
 * This must be called very early at boot, before the very first malloc()
 * call, and is not thread-safe! It's not even possible to use stdio there.
 * Worse, we have to account for the risk of reentrance from dlsym() when
 * it tries to prepare its error messages. Here its ahndled by in_memprof
 * that makes allocators return NULL. dlsym() handles it gracefully. An
 * alternate approach consists in calling aligned_alloc() from these places
 * but that would mean not being able to intercept it later if considered
 * useful to do so.
 */
static void memprof_init()
{
	in_memprof++;
	memprof_malloc_handler  = get_sym_next_addr("malloc");
	if (!memprof_malloc_handler)
		memprof_die("FATAL: malloc() function not found.\n");

	memprof_calloc_handler  = get_sym_next_addr("calloc");
	if (!memprof_calloc_handler)
		memprof_die("FATAL: calloc() function not found.\n");

	memprof_realloc_handler = get_sym_next_addr("realloc");
	if (!memprof_realloc_handler)
		memprof_die("FATAL: realloc() function not found.\n");

	memprof_free_handler    = get_sym_next_addr("free");
	if (!memprof_free_handler)
		memprof_die("FATAL: free() function not found.\n");
	in_memprof--;
}

/* the initial handlers will initialize all regular handlers and will call the
 * one they correspond to. A single one of these functions will typically be
 * called, though it's unknown which one (as any might be called before main).
 */
static void *memprof_malloc_initial_handler(size_t size)
{
	if (in_memprof) {
		/* it's likely that dlsym() needs malloc(), let's fail */
		return NULL;
	}

	memprof_init();
	return memprof_malloc_handler(size);
}

static void *memprof_calloc_initial_handler(size_t nmemb, size_t size)
{
	if (in_memprof) {
		/* it's likely that dlsym() needs calloc(), let's fail */
		return NULL;
	}
	memprof_init();
	return memprof_calloc_handler(nmemb, size);
}

static void *memprof_realloc_initial_handler(void *ptr, size_t size)
{
	if (in_memprof) {
		/* it's likely that dlsym() needs realloc(), let's fail */
		return NULL;
	}

	memprof_init();
	return memprof_realloc_handler(ptr, size);
}

static void  memprof_free_initial_handler(void *ptr)
{
	memprof_init();
	memprof_free_handler(ptr);
}

/* Assign a bin for the memprof_stats to the return address. May perform a few
 * attempts before finding the right one, but always succeeds (in the worst
 * case, returns a default bin). The caller address is atomically set except
 * for the default one which is never set.
 */
static struct memprof_stats *memprof_get_bin(const void *ra, enum memprof_method meth)
{
	int retries = 16; // up to 16 consecutive entries may be tested.
	const void *old;
	unsigned int bin;

	bin = memprof_hash_ptr(ra);
	for (; memprof_stats[bin].caller != ra; bin = (bin + 1) & (MEMPROF_HASH_BUCKETS - 1)) {
		if (!--retries) {
			bin = MEMPROF_HASH_BUCKETS;
			break;
		}

		old = NULL;
		if (!memprof_stats[bin].caller &&
		    HA_ATOMIC_CAS(&memprof_stats[bin].caller, &old, ra)) {
			memprof_stats[bin].method = meth;
			break;
		}
	}
	return &memprof_stats[bin];
}

/* This is the new global malloc() function. It must optimize for the normal
 * case (i.e. profiling disabled) hence the first test to permit a direct jump.
 * It must remain simple to guarantee the lack of reentrance. stdio is not
 * possible there even for debugging. The reported size is the really allocated
 * one as returned by malloc_usable_size(), because this will allow it to be
 * compared to the one before realloc() or free(). This is a GNU and jemalloc
 * extension but other systems may also store this size in ptr[-1].
 */
void *malloc(size_t size)
{
	struct memprof_stats *bin;
	void *ret;

	if (likely(!(profiling & HA_PROF_MEMORY)))
		return memprof_malloc_handler(size);

	ret = memprof_malloc_handler(size);
	size = malloc_usable_size(ret) + sizeof(void *);

	bin = memprof_get_bin(__builtin_return_address(0), MEMPROF_METH_MALLOC);
	_HA_ATOMIC_ADD(&bin->alloc_calls, 1);
	_HA_ATOMIC_ADD(&bin->alloc_tot, size);
	return ret;
}

/* This is the new global calloc() function. It must optimize for the normal
 * case (i.e. profiling disabled) hence the first test to permit a direct jump.
 * It must remain simple to guarantee the lack of reentrance. stdio is not
 * possible there even for debugging. The reported size is the really allocated
 * one as returned by malloc_usable_size(), because this will allow it to be
 * compared to the one before realloc() or free(). This is a GNU and jemalloc
 * extension but other systems may also store this size in ptr[-1].
 */
void *calloc(size_t nmemb, size_t size)
{
	struct memprof_stats *bin;
	void *ret;

	if (likely(!(profiling & HA_PROF_MEMORY)))
		return memprof_calloc_handler(nmemb, size);

	ret = memprof_calloc_handler(nmemb, size);
	size = malloc_usable_size(ret) + sizeof(void *);

	bin = memprof_get_bin(__builtin_return_address(0), MEMPROF_METH_CALLOC);
	_HA_ATOMIC_ADD(&bin->alloc_calls, 1);
	_HA_ATOMIC_ADD(&bin->alloc_tot, size);
	return ret;
}

/* This is the new global realloc() function. It must optimize for the normal
 * case (i.e. profiling disabled) hence the first test to permit a direct jump.
 * It must remain simple to guarantee the lack of reentrance. stdio is not
 * possible there even for debugging. The reported size is the really allocated
 * one as returned by malloc_usable_size(), because this will allow it to be
 * compared to the one before realloc() or free(). This is a GNU and jemalloc
 * extension but other systems may also store this size in ptr[-1].
 * Depending on the old vs new size, it's considered as an allocation or a free
 * (or neither if the size remains the same).
 */
void *realloc(void *ptr, size_t size)
{
	struct memprof_stats *bin;
	size_t size_before;
	void *ret;

	if (likely(!(profiling & HA_PROF_MEMORY)))
		return memprof_realloc_handler(ptr, size);

	size_before = malloc_usable_size(ptr);
	ret = memprof_realloc_handler(ptr, size);
	size = malloc_usable_size(ret);

	/* only count the extra link for new allocations */
	if (!ptr)
		size += sizeof(void *);

	bin = memprof_get_bin(__builtin_return_address(0), MEMPROF_METH_REALLOC);
	if (size > size_before) {
		_HA_ATOMIC_ADD(&bin->alloc_calls, 1);
		_HA_ATOMIC_ADD(&bin->alloc_tot, size - size_before);
	} else if (size < size_before) {
		_HA_ATOMIC_ADD(&bin->free_calls, 1);
		_HA_ATOMIC_ADD(&bin->free_tot, size_before - size);
	}
	return ret;
}

/* This is the new global free() function. It must optimize for the normal
 * case (i.e. profiling disabled) hence the first test to permit a direct jump.
 * It must remain simple to guarantee the lack of reentrance. stdio is not
 * possible there even for debugging. The reported size is the really allocated
 * one as returned by malloc_usable_size(), because this will allow it to be
 * compared to the one before realloc() or free(). This is a GNU and jemalloc
 * extension but other systems may also store this size in ptr[-1]. Since
 * free() is often called on NULL pointers to collect garbage at the end of
 * many functions or during config parsing, as a special case free(NULL)
 * doesn't update any stats.
 */
void free(void *ptr)
{
	struct memprof_stats *bin;
	size_t size_before;

	if (likely(!(profiling & HA_PROF_MEMORY) || !ptr)) {
		memprof_free_handler(ptr);
		return;
	}

	size_before = malloc_usable_size(ptr) + sizeof(void *);
	memprof_free_handler(ptr);

	bin = memprof_get_bin(__builtin_return_address(0), MEMPROF_METH_FREE);
	_HA_ATOMIC_ADD(&bin->free_calls, 1);
	_HA_ATOMIC_ADD(&bin->free_tot, size_before);
}

#endif // USE_MEMORY_PROFILING

/* Updates the current thread's statistics about stolen CPU time. The unit for
 * <stolen> is half-milliseconds.
 */
void report_stolen_time(uint64_t stolen)
{
	activity[tid].cpust_total += stolen;
	update_freq_ctr(&activity[tid].cpust_1s, stolen);
	update_freq_ctr_period(&activity[tid].cpust_15s, 15000, stolen);
}

/* Update avg_loop value for the current thread and possibly decide to enable
 * task-level profiling on the current thread based on its average run time.
 * The <run_time> argument is the number of microseconds elapsed since the
 * last time poll() returned.
 */
void activity_count_runtime(uint32_t run_time)
{
	uint32_t up, down;

	/* 1 millisecond per loop on average over last 1024 iterations is
	 * enough to turn on profiling.
	 */
	up = 1000;
	down = up * 99 / 100;

	run_time = swrate_add(&activity[tid].avg_loop_us, TIME_STATS_SAMPLES, run_time);

	/* In automatic mode, reaching the "up" threshold on average switches
	 * profiling to "on" when automatic, and going back below the "down"
	 * threshold switches to off. The forced modes don't check the load.
	 */
	if (!(task_profiling_mask & tid_bit)) {
		if (unlikely((profiling & HA_PROF_TASKS_MASK) == HA_PROF_TASKS_ON ||
		             ((profiling & HA_PROF_TASKS_MASK) == HA_PROF_TASKS_AON &&
		             swrate_avg(run_time, TIME_STATS_SAMPLES) >= up)))
			_HA_ATOMIC_OR(&task_profiling_mask, tid_bit);
	} else {
		if (unlikely((profiling & HA_PROF_TASKS_MASK) == HA_PROF_TASKS_OFF ||
		             ((profiling & HA_PROF_TASKS_MASK) == HA_PROF_TASKS_AOFF &&
		             swrate_avg(run_time, TIME_STATS_SAMPLES) <= down)))
			_HA_ATOMIC_AND(&task_profiling_mask, ~tid_bit);
	}
}

#ifdef USE_MEMORY_PROFILING
/* config parser for global "profiling.memory", accepts "on" or "off" */
static int cfg_parse_prof_memory(char **args, int section_type, struct proxy *curpx,
                                const struct proxy *defpx, const char *file, int line,
                                char **err)
{
	if (too_many_args(1, args, err, NULL))
		return -1;

	if (strcmp(args[1], "on") == 0)
		profiling |= HA_PROF_MEMORY;
	else if (strcmp(args[1], "off") == 0)
		profiling &= ~HA_PROF_MEMORY;
	else {
		memprintf(err, "'%s' expects either 'on' or 'off' but got '%s'.", args[0], args[1]);
		return -1;
	}
	return 0;
}
#endif // USE_MEMORY_PROFILING

/* config parser for global "profiling.tasks", accepts "on" or "off" */
static int cfg_parse_prof_tasks(char **args, int section_type, struct proxy *curpx,
                                const struct proxy *defpx, const char *file, int line,
                                char **err)
{
	if (too_many_args(1, args, err, NULL))
		return -1;

	if (strcmp(args[1], "on") == 0)
		profiling = (profiling & ~HA_PROF_TASKS_MASK) | HA_PROF_TASKS_ON;
	else if (strcmp(args[1], "auto") == 0)
		profiling = (profiling & ~HA_PROF_TASKS_MASK) | HA_PROF_TASKS_AOFF;
	else if (strcmp(args[1], "off") == 0)
		profiling = (profiling & ~HA_PROF_TASKS_MASK) | HA_PROF_TASKS_OFF;
	else {
		memprintf(err, "'%s' expects either 'on', 'auto', or 'off' but got '%s'.", args[0], args[1]);
		return -1;
	}
	return 0;
}

/* parse a "set profiling" command. It always returns 1. */
static int cli_parse_set_profiling(char **args, char *payload, struct appctx *appctx, void *private)
{
	if (!cli_has_level(appctx, ACCESS_LVL_ADMIN))
		return 1;

	if (strcmp(args[2], "memory") == 0) {
#ifdef USE_MEMORY_PROFILING
		if (strcmp(args[3], "on") == 0) {
			unsigned int old = profiling;
			int i;

			while (!_HA_ATOMIC_CAS(&profiling, &old, old | HA_PROF_MEMORY))
				;

			/* also flush current profiling stats */
			for (i = 0; i < sizeof(memprof_stats) / sizeof(memprof_stats[0]); i++) {
				HA_ATOMIC_STORE(&memprof_stats[i].alloc_calls, 0);
				HA_ATOMIC_STORE(&memprof_stats[i].free_calls, 0);
				HA_ATOMIC_STORE(&memprof_stats[i].alloc_tot, 0);
				HA_ATOMIC_STORE(&memprof_stats[i].free_tot, 0);
				HA_ATOMIC_STORE(&memprof_stats[i].caller, NULL);
			}
		}
		else if (strcmp(args[3], "off") == 0) {
			unsigned int old = profiling;

			while (!_HA_ATOMIC_CAS(&profiling, &old, old & ~HA_PROF_MEMORY))
				;
		}
		else
			return cli_err(appctx, "Expects either 'on' or 'off'.\n");
		return 1;
#else
		return cli_err(appctx, "Memory profiling not compiled in.\n");
#endif
	}

	if (strcmp(args[2], "tasks") != 0)
		return cli_err(appctx, "Expects either 'tasks' or 'memory'.\n");

	if (strcmp(args[3], "on") == 0) {
		unsigned int old = profiling;
		int i;

		while (!_HA_ATOMIC_CAS(&profiling, &old, (old & ~HA_PROF_TASKS_MASK) | HA_PROF_TASKS_ON))
			;
		/* also flush current profiling stats */
		for (i = 0; i < 256; i++) {
			HA_ATOMIC_STORE(&sched_activity[i].calls, 0);
			HA_ATOMIC_STORE(&sched_activity[i].cpu_time, 0);
			HA_ATOMIC_STORE(&sched_activity[i].lat_time, 0);
			HA_ATOMIC_STORE(&sched_activity[i].func, NULL);
		}
	}
	else if (strcmp(args[3], "auto") == 0) {
		unsigned int old = profiling;
		unsigned int new;

		do {
			if ((old & HA_PROF_TASKS_MASK) >= HA_PROF_TASKS_AON)
				new = (old & ~HA_PROF_TASKS_MASK) | HA_PROF_TASKS_AON;
			else
				new = (old & ~HA_PROF_TASKS_MASK) | HA_PROF_TASKS_AOFF;
		} while (!_HA_ATOMIC_CAS(&profiling, &old, new));
	}
	else if (strcmp(args[3], "off") == 0) {
		unsigned int old = profiling;
		while (!_HA_ATOMIC_CAS(&profiling, &old, (old & ~HA_PROF_TASKS_MASK) | HA_PROF_TASKS_OFF))
			;
	}
	else
		return cli_err(appctx, "Expects 'on', 'auto', or 'off'.\n");

	return 1;
}

static int cmp_sched_activity_calls(const void *a, const void *b)
{
	const struct sched_activity *l = (const struct sched_activity *)a;
	const struct sched_activity *r = (const struct sched_activity *)b;

	if (l->calls > r->calls)
		return -1;
	else if (l->calls < r->calls)
		return 1;
	else
		return 0;
}

static int cmp_sched_activity_addr(const void *a, const void *b)
{
	const struct sched_activity *l = (const struct sched_activity *)a;
	const struct sched_activity *r = (const struct sched_activity *)b;

	if (l->func > r->func)
		return -1;
	else if (l->func < r->func)
		return 1;
	else
		return 0;
}

#ifdef USE_MEMORY_PROFILING
/* used by qsort below */
static int cmp_memprof_stats(const void *a, const void *b)
{
	const struct memprof_stats *l = (const struct memprof_stats *)a;
	const struct memprof_stats *r = (const struct memprof_stats *)b;

	if (l->alloc_tot + l->free_tot > r->alloc_tot + r->free_tot)
		return -1;
	else if (l->alloc_tot + l->free_tot < r->alloc_tot + r->free_tot)
		return 1;
	else
		return 0;
}

static int cmp_memprof_addr(const void *a, const void *b)
{
	const struct memprof_stats *l = (const struct memprof_stats *)a;
	const struct memprof_stats *r = (const struct memprof_stats *)b;

	if (l->caller > r->caller)
		return -1;
	else if (l->caller < r->caller)
		return 1;
	else
		return 0;
}
#endif // USE_MEMORY_PROFILING

/* Computes the index of function pointer <func> for use with sched_activity[]
 * or any other similar array passed in <array>, and returns a pointer to the
 * entry after having atomically assigned it to this function pointer. Note
 * that in case of collision, the first entry is returned instead ("other").
 */
struct sched_activity *sched_activity_entry(struct sched_activity *array, const void *func)
{
	uint64_t hash = XXH64_avalanche(XXH64_mergeRound((size_t)func, (size_t)func));
	struct sched_activity *ret;
	const void *old = NULL;

	hash ^= (hash >> 32);
	hash ^= (hash >> 16);
	hash ^= (hash >> 8);
	hash &= 0xff;
	ret = &array[hash];

	if (likely(ret->func == func))
		return ret;

	if (HA_ATOMIC_CAS(&ret->func, &old, func))
		return ret;

	return array;
}

/* This function dumps all profiling settings. It returns 0 if the output
 * buffer is full and it needs to be called again, otherwise non-zero.
 * It dumps some parts depending on the following states from show_prof_ctx:
 *    dump_step:
 *       0, 4: dump status, then jump to 1 if 0
 *       1, 5: dump tasks, then jump to 2 if 1
 *       2, 6: dump memory, then stop
 *    linenum:
 *       restart line for each step (starts at zero)
 *    maxcnt:
 *       may contain a configured max line count for each step (0=not set)
 *    byaddr:
 *       0: sort by usage
 *       1: sort by address
 */
static int cli_io_handler_show_profiling(struct appctx *appctx)
{
	struct show_prof_ctx *ctx = appctx->svcctx;
	struct sched_activity tmp_activity[256] __attribute__((aligned(64)));
#ifdef USE_MEMORY_PROFILING
	struct memprof_stats tmp_memstats[MEMPROF_HASH_BUCKETS + 1];
	unsigned long long tot_alloc_calls, tot_free_calls;
	unsigned long long tot_alloc_bytes, tot_free_bytes;
#endif
	struct stconn *sc = appctx_sc(appctx);
	struct buffer *name_buffer = get_trash_chunk();
	const char *str;
	int max_lines;
	int i, max;

	if (unlikely(sc_ic(sc)->flags & (CF_WRITE_ERROR|CF_SHUTW)))
		return 1;

	chunk_reset(&trash);

	switch (profiling & HA_PROF_TASKS_MASK) {
	case HA_PROF_TASKS_AOFF: str="auto-off"; break;
	case HA_PROF_TASKS_AON:  str="auto-on"; break;
	case HA_PROF_TASKS_ON:   str="on"; break;
	default:                 str="off"; break;
	}

	if ((ctx->dump_step & 3) != 0)
		goto skip_status;

	chunk_printf(&trash,
	             "Per-task CPU profiling              : %-8s      # set profiling tasks {on|auto|off}\n"
	             "Memory usage profiling              : %-8s      # set profiling memory {on|off}\n",
	             str, (profiling & HA_PROF_MEMORY) ? "on" : "off");

	if (applet_putchk(appctx, &trash) == -1) {
		/* failed, try again */
		return 0;
	}

	ctx->linenum = 0; // reset first line to dump
	if ((ctx->dump_step & 4) == 0)
		ctx->dump_step++; // next step

 skip_status:
	if ((ctx->dump_step & 3) != 1)
		goto skip_tasks;

	memcpy(tmp_activity, sched_activity, sizeof(tmp_activity));
	if (ctx->by_addr)
		qsort(tmp_activity, 256, sizeof(tmp_activity[0]), cmp_sched_activity_addr);
	else
		qsort(tmp_activity, 256, sizeof(tmp_activity[0]), cmp_sched_activity_calls);

	if (!ctx->linenum)
		chunk_appendf(&trash, "Tasks activity:\n"
		                      "  function                      calls   cpu_tot   cpu_avg   lat_tot   lat_avg\n");

	max_lines = ctx->maxcnt;
	if (!max_lines)
		max_lines = 256;

	for (i = ctx->linenum; i < max_lines && tmp_activity[i].calls; i++) {
		ctx->linenum = i;
		chunk_reset(name_buffer);

		if (!tmp_activity[i].func)
			chunk_printf(name_buffer, "other");
		else
			resolve_sym_name(name_buffer, "", tmp_activity[i].func);

		/* reserve 35 chars for name+' '+#calls, knowing that longer names
		 * are often used for less often called functions.
		 */
		max = 35 - name_buffer->data;
		if (max < 1)
			max = 1;
		chunk_appendf(&trash, "  %s%*llu", name_buffer->area, max, (unsigned long long)tmp_activity[i].calls);

		print_time_short(&trash, "   ", tmp_activity[i].cpu_time, "");
		print_time_short(&trash, "   ", tmp_activity[i].cpu_time / tmp_activity[i].calls, "");
		print_time_short(&trash, "   ", tmp_activity[i].lat_time, "");
		print_time_short(&trash, "   ", tmp_activity[i].lat_time / tmp_activity[i].calls, "\n");

		if (applet_putchk(appctx, &trash) == -1) {
			/* failed, try again */
			return 0;
		}
	}

	if (applet_putchk(appctx, &trash) == -1) {
		/* failed, try again */
		return 0;
	}

	ctx->linenum = 0; // reset first line to dump
	if ((ctx->dump_step & 4) == 0)
		ctx->dump_step++; // next step

 skip_tasks:

#ifdef USE_MEMORY_PROFILING
	if ((ctx->dump_step & 3) != 2)
		goto skip_mem;

	memcpy(tmp_memstats, memprof_stats, sizeof(tmp_memstats));
	if (ctx->by_addr)
		qsort(tmp_memstats, MEMPROF_HASH_BUCKETS+1, sizeof(tmp_memstats[0]), cmp_memprof_addr);
	else
		qsort(tmp_memstats, MEMPROF_HASH_BUCKETS+1, sizeof(tmp_memstats[0]), cmp_memprof_stats);

	if (!ctx->linenum)
		chunk_appendf(&trash,
		              "Alloc/Free statistics by call place:\n"
		              "         Calls         |         Tot Bytes           |       Caller and method\n"
		              "<- alloc -> <- free  ->|<-- alloc ---> <-- free ---->|\n");

	max_lines = ctx->maxcnt;
	if (!max_lines)
		max_lines = MEMPROF_HASH_BUCKETS + 1;

	for (i = ctx->linenum; i < max_lines; i++) {
		struct memprof_stats *entry = &tmp_memstats[i];

		ctx->linenum = i;
		if (!entry->alloc_calls && !entry->free_calls)
			continue;
		chunk_appendf(&trash, "%11llu %11llu %14llu %14llu| %16p ",
			      entry->alloc_calls, entry->free_calls,
			      entry->alloc_tot, entry->free_tot,
			      entry->caller);

		if (entry->caller)
			resolve_sym_name(&trash, NULL, entry->caller);
		else
			chunk_appendf(&trash, "[other]");

		chunk_appendf(&trash," %s(%lld)", memprof_methods[entry->method],
			      (long long)(entry->alloc_tot - entry->free_tot) / (long long)(entry->alloc_calls + entry->free_calls));

		if (entry->alloc_tot && entry->free_tot) {
			/* that's a realloc, show the total diff to help spot leaks */
			chunk_appendf(&trash," [delta=%lld]", (long long)(entry->alloc_tot - entry->free_tot));
		}

		chunk_appendf(&trash, "\n");

		if (applet_putchk(appctx, &trash) == -1)
			return 0;
	}

	if (applet_putchk(appctx, &trash) == -1)
		return 0;

	tot_alloc_calls = tot_free_calls = tot_alloc_bytes = tot_free_bytes = 0;
	for (i = 0; i < max_lines; i++) {
		tot_alloc_calls += tmp_memstats[i].alloc_calls;
		tot_free_calls  += tmp_memstats[i].free_calls;
		tot_alloc_bytes += tmp_memstats[i].alloc_tot;
		tot_free_bytes  += tmp_memstats[i].free_tot;
	}

	chunk_appendf(&trash,
	              "-----------------------|-----------------------------|\n"
		      "%11llu %11llu %14llu %14llu| <- Total; Delta_calls=%lld; Delta_bytes=%lld\n",
		      tot_alloc_calls, tot_free_calls,
		      tot_alloc_bytes, tot_free_bytes,
		      tot_alloc_calls - tot_free_calls,
		      tot_alloc_bytes - tot_free_bytes);

	if (applet_putchk(appctx, &trash) == -1)
		return 0;

	ctx->linenum = 0; // reset first line to dump
	if ((ctx->dump_step & 4) == 0)
		ctx->dump_step++; // next step

 skip_mem:
#endif // USE_MEMORY_PROFILING

	return 1;
}

/* parse a "show profiling" command. It returns 1 on failure, 0 if it starts to dump.
 *  - cli.i0 is set to the first state (0=all, 4=status, 5=tasks, 6=memory)
 *  - cli.o1 is set to 1 if the output must be sorted by addr instead of usage
 *  - cli.o0 is set to the number of lines of output
 */
static int cli_parse_show_profiling(char **args, char *payload, struct appctx *appctx, void *private)
{
	struct show_prof_ctx *ctx = applet_reserve_svcctx(appctx, sizeof(*ctx));
	int arg;

	if (!cli_has_level(appctx, ACCESS_LVL_ADMIN))
		return 1;

	for (arg = 2; *args[arg]; arg++) {
		if (strcmp(args[arg], "all") == 0) {
			ctx->dump_step = 0; // will cycle through 0,1,2; default
		}
		else if (strcmp(args[arg], "status") == 0) {
			ctx->dump_step = 4; // will visit status only
		}
		else if (strcmp(args[arg], "tasks") == 0) {
			ctx->dump_step = 5; // will visit tasks only
		}
		else if (strcmp(args[arg], "memory") == 0) {
			ctx->dump_step = 6; // will visit memory only
		}
		else if (strcmp(args[arg], "byaddr") == 0) {
			ctx->by_addr = 1; // sort output by address instead of usage
		}
		else if (isdigit((unsigned char)*args[arg])) {
			ctx->maxcnt = atoi(args[arg]); // number of entries to dump
		}
		else
			return cli_err(appctx, "Expects either 'all', 'status', 'tasks', 'memory', 'byaddr' or a max number of output lines.\n");
	}
	return 0;
}

/* This function scans all threads' run queues and collects statistics about
 * running tasks. It returns 0 if the output buffer is full and it needs to be
 * called again, otherwise non-zero.
 */
static int cli_io_handler_show_tasks(struct appctx *appctx)
{
	struct sched_activity tmp_activity[256] __attribute__((aligned(64)));
	struct stconn *sc = appctx_sc(appctx);
	struct buffer *name_buffer = get_trash_chunk();
	struct sched_activity *entry;
	const struct tasklet *tl;
	const struct task *t;
	uint64_t now_ns, lat;
	struct eb32sc_node *rqnode;
	uint64_t tot_calls;
	int thr, queue;
	int i, max;

	if (unlikely(sc_ic(sc)->flags & (CF_WRITE_ERROR|CF_SHUTW)))
		return 1;

	/* It's not possible to scan queues in small chunks and yield in the
	 * middle of the dump and come back again. So what we're doing instead
	 * is to freeze all threads and inspect their queues at once as fast as
	 * possible, using a sched_activity array to collect metrics with
	 * limited collision, then we'll report statistics only. The tasks'
	 * #calls will reflect the number of occurrences, and the lat_time will
	 * reflect the latency when set. We prefer to take the time before
	 * calling thread_isolate() so that the wait time doesn't impact the
	 * measurement accuracy. However this requires to take care of negative
	 * times since tasks might be queued after we retrieve it.
	 */

	now_ns = now_mono_time();
	memset(tmp_activity, 0, sizeof(tmp_activity));

	thread_isolate();

	/* 1. global run queue */

#ifdef USE_THREAD
	rqnode = eb32sc_first(&rqueue, ~0UL);
	while (rqnode) {
		t = eb32sc_entry(rqnode, struct task, rq);
		entry = sched_activity_entry(tmp_activity, t->process);
		if (t->wake_date) {
			lat = now_ns - t->wake_date;
			if ((int64_t)lat > 0)
				entry->lat_time += lat;
		}
		entry->calls++;
		rqnode = eb32sc_next(rqnode, ~0UL);
	}
#endif
	/* 2. all threads's local run queues */
	for (thr = 0; thr < global.nbthread; thr++) {
		/* task run queue */
		rqnode = eb32sc_first(&ha_thread_ctx[thr].rqueue, ~0UL);
		while (rqnode) {
			t = eb32sc_entry(rqnode, struct task, rq);
			entry = sched_activity_entry(tmp_activity, t->process);
			if (t->wake_date) {
				lat = now_ns - t->wake_date;
				if ((int64_t)lat > 0)
					entry->lat_time += lat;
			}
			entry->calls++;
			rqnode = eb32sc_next(rqnode, ~0UL);
		}

		/* shared tasklet list */
		list_for_each_entry(tl, mt_list_to_list(&ha_thread_ctx[thr].shared_tasklet_list), list) {
			t = (const struct task *)tl;
			entry = sched_activity_entry(tmp_activity, t->process);
			if (!TASK_IS_TASKLET(t) && t->wake_date) {
				lat = now_ns - t->wake_date;
				if ((int64_t)lat > 0)
					entry->lat_time += lat;
			}
			entry->calls++;
		}

		/* classful tasklets */
		for (queue = 0; queue < TL_CLASSES; queue++) {
			list_for_each_entry(tl, &ha_thread_ctx[thr].tasklets[queue], list) {
				t = (const struct task *)tl;
				entry = sched_activity_entry(tmp_activity, t->process);
				if (!TASK_IS_TASKLET(t) && t->wake_date) {
					lat = now_ns - t->wake_date;
					if ((int64_t)lat > 0)
						entry->lat_time += lat;
				}
				entry->calls++;
			}
		}
	}

	/* hopefully we're done */
	thread_release();

	chunk_reset(&trash);

	tot_calls = 0;
	for (i = 0; i < 256; i++)
		tot_calls += tmp_activity[i].calls;

	qsort(tmp_activity, 256, sizeof(tmp_activity[0]), cmp_sched_activity_calls);

	chunk_appendf(&trash, "Running tasks: %d (%d threads)\n"
		      "  function                     places     %%    lat_tot   lat_avg\n",
		      (int)tot_calls, global.nbthread);

	for (i = 0; i < 256 && tmp_activity[i].calls; i++) {
		chunk_reset(name_buffer);

		if (!tmp_activity[i].func)
			chunk_printf(name_buffer, "other");
		else
			resolve_sym_name(name_buffer, "", tmp_activity[i].func);

		/* reserve 35 chars for name+' '+#calls, knowing that longer names
		 * are often used for less often called functions.
		 */
		max = 35 - name_buffer->data;
		if (max < 1)
			max = 1;
		chunk_appendf(&trash, "  %s%*llu  %3d.%1d",
		              name_buffer->area, max, (unsigned long long)tmp_activity[i].calls,
		              (int)(100ULL * tmp_activity[i].calls / tot_calls),
		              (int)((1000ULL * tmp_activity[i].calls / tot_calls)%10));
		print_time_short(&trash, "   ", tmp_activity[i].lat_time, "");
		print_time_short(&trash, "   ", tmp_activity[i].lat_time / tmp_activity[i].calls, "\n");
	}

	if (applet_putchk(appctx, &trash) == -1) {
		/* failed, try again */
		return 0;
	}
	return 1;
}

/* config keyword parsers */
static struct cfg_kw_list cfg_kws = {ILH, {
#ifdef USE_MEMORY_PROFILING
	{ CFG_GLOBAL, "profiling.memory",     cfg_parse_prof_memory     },
#endif
	{ CFG_GLOBAL, "profiling.tasks",      cfg_parse_prof_tasks      },
	{ 0, NULL, NULL }
}};

INITCALL1(STG_REGISTER, cfg_register_keywords, &cfg_kws);

/* register cli keywords */
static struct cli_kw_list cli_kws = {{ },{
	{ { "set",  "profiling", NULL }, "set profiling <what> {auto|on|off}      : enable/disable resource profiling (tasks,memory)", cli_parse_set_profiling,  NULL },
	{ { "show", "profiling", NULL }, "show profiling [<what>|<#lines>|byaddr]*: show profiling state (all,status,tasks,memory)",   cli_parse_show_profiling, cli_io_handler_show_profiling, NULL },
	{ { "show", "tasks", NULL },     "show tasks                              : show running tasks",                               NULL, cli_io_handler_show_tasks,     NULL },
	{{},}
}};

INITCALL1(STG_REGISTER, cli_register_kw, &cli_kws);