• Home
  • History
  • Annotate
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /******************************************************************************/
2 #ifdef JEMALLOC_H_TYPES
3 
4 typedef struct prof_bt_s prof_bt_t;
5 typedef struct prof_cnt_s prof_cnt_t;
6 typedef struct prof_tctx_s prof_tctx_t;
7 typedef struct prof_gctx_s prof_gctx_t;
8 typedef struct prof_tdata_s prof_tdata_t;
9 
10 /* Option defaults. */
11 #ifdef JEMALLOC_PROF
12 #  define PROF_PREFIX_DEFAULT		"jeprof"
13 #else
14 #  define PROF_PREFIX_DEFAULT		""
15 #endif
16 #define	LG_PROF_SAMPLE_DEFAULT		19
17 #define	LG_PROF_INTERVAL_DEFAULT	-1
18 
19 /*
20  * Hard limit on stack backtrace depth.  The version of prof_backtrace() that
21  * is based on __builtin_return_address() necessarily has a hard-coded number
22  * of backtrace frame handlers, and should be kept in sync with this setting.
23  */
24 #define	PROF_BT_MAX			128
25 
26 /* Initial hash table size. */
27 #define	PROF_CKH_MINITEMS		64
28 
29 /* Size of memory buffer to use when writing dump files. */
30 #define	PROF_DUMP_BUFSIZE		65536
31 
32 /* Size of stack-allocated buffer used by prof_printf(). */
33 #define	PROF_PRINTF_BUFSIZE		128
34 
35 /*
36  * Number of mutexes shared among all gctx's.  No space is allocated for these
37  * unless profiling is enabled, so it's okay to over-provision.
38  */
39 #define	PROF_NCTX_LOCKS			1024
40 
41 /*
42  * Number of mutexes shared among all tdata's.  No space is allocated for these
43  * unless profiling is enabled, so it's okay to over-provision.
44  */
45 #define	PROF_NTDATA_LOCKS		256
46 
47 /*
48  * prof_tdata pointers close to NULL are used to encode state information that
49  * is used for cleaning up during thread shutdown.
50  */
51 #define	PROF_TDATA_STATE_REINCARNATED	((prof_tdata_t *)(uintptr_t)1)
52 #define	PROF_TDATA_STATE_PURGATORY	((prof_tdata_t *)(uintptr_t)2)
53 #define	PROF_TDATA_STATE_MAX		PROF_TDATA_STATE_PURGATORY
54 
55 #endif /* JEMALLOC_H_TYPES */
56 /******************************************************************************/
57 #ifdef JEMALLOC_H_STRUCTS
58 
59 struct prof_bt_s {
60 	/* Backtrace, stored as len program counters. */
61 	void		**vec;
62 	unsigned	len;
63 };
64 
65 #ifdef JEMALLOC_PROF_LIBGCC
66 /* Data structure passed to libgcc _Unwind_Backtrace() callback functions. */
67 typedef struct {
68 	prof_bt_t	*bt;
69 	unsigned	max;
70 } prof_unwind_data_t;
71 #endif
72 
73 struct prof_cnt_s {
74 	/* Profiling counters. */
75 	uint64_t	curobjs;
76 	uint64_t	curbytes;
77 	uint64_t	accumobjs;
78 	uint64_t	accumbytes;
79 };
80 
81 typedef enum {
82 	prof_tctx_state_initializing,
83 	prof_tctx_state_nominal,
84 	prof_tctx_state_dumping,
85 	prof_tctx_state_purgatory /* Dumper must finish destroying. */
86 } prof_tctx_state_t;
87 
88 struct prof_tctx_s {
89 	/* Thread data for thread that performed the allocation. */
90 	prof_tdata_t		*tdata;
91 
92 	/*
93 	 * Copy of tdata->thr_uid, necessary because tdata may be defunct during
94 	 * teardown.
95 	 */
96 	uint64_t		thr_uid;
97 
98 	/* Profiling counters, protected by tdata->lock. */
99 	prof_cnt_t		cnts;
100 
101 	/* Associated global context. */
102 	prof_gctx_t		*gctx;
103 
104 	/*
105 	 * UID that distinguishes multiple tctx's created by the same thread,
106 	 * but coexisting in gctx->tctxs.  There are two ways that such
107 	 * coexistence can occur:
108 	 * - A dumper thread can cause a tctx to be retained in the purgatory
109 	 *   state.
110 	 * - Although a single "producer" thread must create all tctx's which
111 	 *   share the same thr_uid, multiple "consumers" can each concurrently
112 	 *   execute portions of prof_tctx_destroy().  prof_tctx_destroy() only
113 	 *   gets called once each time cnts.cur{objs,bytes} drop to 0, but this
114 	 *   threshold can be hit again before the first consumer finishes
115 	 *   executing prof_tctx_destroy().
116 	 */
117 	uint64_t		tctx_uid;
118 
119 	/* Linkage into gctx's tctxs. */
120 	rb_node(prof_tctx_t)	tctx_link;
121 
122 	/*
123 	 * True during prof_alloc_prep()..prof_malloc_sample_object(), prevents
124 	 * sample vs destroy race.
125 	 */
126 	bool			prepared;
127 
128 	/* Current dump-related state, protected by gctx->lock. */
129 	prof_tctx_state_t	state;
130 
131 	/*
132 	 * Copy of cnts snapshotted during early dump phase, protected by
133 	 * dump_mtx.
134 	 */
135 	prof_cnt_t		dump_cnts;
136 };
137 typedef rb_tree(prof_tctx_t) prof_tctx_tree_t;
138 
139 struct prof_gctx_s {
140 	/* Protects nlimbo, cnt_summed, and tctxs. */
141 	malloc_mutex_t		*lock;
142 
143 	/*
144 	 * Number of threads that currently cause this gctx to be in a state of
145 	 * limbo due to one of:
146 	 *   - Initializing this gctx.
147 	 *   - Initializing per thread counters associated with this gctx.
148 	 *   - Preparing to destroy this gctx.
149 	 *   - Dumping a heap profile that includes this gctx.
150 	 * nlimbo must be 1 (single destroyer) in order to safely destroy the
151 	 * gctx.
152 	 */
153 	unsigned		nlimbo;
154 
155 	/*
156 	 * Tree of profile counters, one for each thread that has allocated in
157 	 * this context.
158 	 */
159 	prof_tctx_tree_t	tctxs;
160 
161 	/* Linkage for tree of contexts to be dumped. */
162 	rb_node(prof_gctx_t)	dump_link;
163 
164 	/* Temporary storage for summation during dump. */
165 	prof_cnt_t		cnt_summed;
166 
167 	/* Associated backtrace. */
168 	prof_bt_t		bt;
169 
170 	/* Backtrace vector, variable size, referred to by bt. */
171 	void			*vec[1];
172 };
173 typedef rb_tree(prof_gctx_t) prof_gctx_tree_t;
174 
175 struct prof_tdata_s {
176 	malloc_mutex_t		*lock;
177 
178 	/* Monotonically increasing unique thread identifier. */
179 	uint64_t		thr_uid;
180 
181 	/*
182 	 * Monotonically increasing discriminator among tdata structures
183 	 * associated with the same thr_uid.
184 	 */
185 	uint64_t		thr_discrim;
186 
187 	/* Included in heap profile dumps if non-NULL. */
188 	char			*thread_name;
189 
190 	bool			attached;
191 	bool			expired;
192 
193 	rb_node(prof_tdata_t)	tdata_link;
194 
195 	/*
196 	 * Counter used to initialize prof_tctx_t's tctx_uid.  No locking is
197 	 * necessary when incrementing this field, because only one thread ever
198 	 * does so.
199 	 */
200 	uint64_t		tctx_uid_next;
201 
202 	/*
203 	 * Hash of (prof_bt_t *)-->(prof_tctx_t *).  Each thread tracks
204 	 * backtraces for which it has non-zero allocation/deallocation counters
205 	 * associated with thread-specific prof_tctx_t objects.  Other threads
206 	 * may write to prof_tctx_t contents when freeing associated objects.
207 	 */
208 	ckh_t			bt2tctx;
209 
210 	/* Sampling state. */
211 	uint64_t		prng_state;
212 	uint64_t		bytes_until_sample;
213 
214 	/* State used to avoid dumping while operating on prof internals. */
215 	bool			enq;
216 	bool			enq_idump;
217 	bool			enq_gdump;
218 
219 	/*
220 	 * Set to true during an early dump phase for tdata's which are
221 	 * currently being dumped.  New threads' tdata's have this initialized
222 	 * to false so that they aren't accidentally included in later dump
223 	 * phases.
224 	 */
225 	bool			dumping;
226 
227 	/*
228 	 * True if profiling is active for this tdata's thread
229 	 * (thread.prof.active mallctl).
230 	 */
231 	bool			active;
232 
233 	/* Temporary storage for summation during dump. */
234 	prof_cnt_t		cnt_summed;
235 
236 	/* Backtrace vector, used for calls to prof_backtrace(). */
237 	void			*vec[PROF_BT_MAX];
238 };
239 typedef rb_tree(prof_tdata_t) prof_tdata_tree_t;
240 
241 #endif /* JEMALLOC_H_STRUCTS */
242 /******************************************************************************/
243 #ifdef JEMALLOC_H_EXTERNS
244 
245 extern bool	opt_prof;
246 extern bool	opt_prof_active;
247 extern bool	opt_prof_thread_active_init;
248 extern size_t	opt_lg_prof_sample;   /* Mean bytes between samples. */
249 extern ssize_t	opt_lg_prof_interval; /* lg(prof_interval). */
250 extern bool	opt_prof_gdump;       /* High-water memory dumping. */
251 extern bool	opt_prof_final;       /* Final profile dumping. */
252 extern bool	opt_prof_leak;        /* Dump leak summary at exit. */
253 extern bool	opt_prof_accum;       /* Report cumulative bytes. */
254 extern char	opt_prof_prefix[
255     /* Minimize memory bloat for non-prof builds. */
256 #ifdef JEMALLOC_PROF
257     PATH_MAX +
258 #endif
259     1];
260 
261 /* Accessed via prof_active_[gs]et{_unlocked,}(). */
262 extern bool	prof_active;
263 
264 /* Accessed via prof_gdump_[gs]et{_unlocked,}(). */
265 extern bool	prof_gdump_val;
266 
267 /*
268  * Profile dump interval, measured in bytes allocated.  Each arena triggers a
269  * profile dump when it reaches this threshold.  The effect is that the
270  * interval between profile dumps averages prof_interval, though the actual
271  * interval between dumps will tend to be sporadic, and the interval will be a
272  * maximum of approximately (prof_interval * narenas).
273  */
274 extern uint64_t	prof_interval;
275 
276 /*
277  * Initialized as opt_lg_prof_sample, and potentially modified during profiling
278  * resets.
279  */
280 extern size_t	lg_prof_sample;
281 
282 void	prof_alloc_rollback(tsd_t *tsd, prof_tctx_t *tctx, bool updated);
283 void	prof_malloc_sample_object(const void *ptr, size_t usize,
284     prof_tctx_t *tctx);
285 void	prof_free_sampled_object(tsd_t *tsd, size_t usize, prof_tctx_t *tctx);
286 void	bt_init(prof_bt_t *bt, void **vec);
287 void	prof_backtrace(prof_bt_t *bt);
288 prof_tctx_t	*prof_lookup(tsd_t *tsd, prof_bt_t *bt);
289 #ifdef JEMALLOC_JET
290 size_t	prof_tdata_count(void);
291 size_t	prof_bt_count(void);
292 const prof_cnt_t *prof_cnt_all(void);
293 typedef int (prof_dump_open_t)(bool, const char *);
294 extern prof_dump_open_t *prof_dump_open;
295 typedef bool (prof_dump_header_t)(bool, const prof_cnt_t *);
296 extern prof_dump_header_t *prof_dump_header;
297 #endif
298 void	prof_idump(void);
299 bool	prof_mdump(const char *filename);
300 void	prof_gdump(void);
301 prof_tdata_t	*prof_tdata_init(tsd_t *tsd);
302 prof_tdata_t	*prof_tdata_reinit(tsd_t *tsd, prof_tdata_t *tdata);
303 void	prof_reset(tsd_t *tsd, size_t lg_sample);
304 void	prof_tdata_cleanup(tsd_t *tsd);
305 const char	*prof_thread_name_get(void);
306 bool	prof_active_get(void);
307 bool	prof_active_set(bool active);
308 int	prof_thread_name_set(tsd_t *tsd, const char *thread_name);
309 bool	prof_thread_active_get(void);
310 bool	prof_thread_active_set(bool active);
311 bool	prof_thread_active_init_get(void);
312 bool	prof_thread_active_init_set(bool active_init);
313 bool	prof_gdump_get(void);
314 bool	prof_gdump_set(bool active);
315 void	prof_boot0(void);
316 void	prof_boot1(void);
317 bool	prof_boot2(void);
318 void	prof_prefork(void);
319 void	prof_postfork_parent(void);
320 void	prof_postfork_child(void);
321 void	prof_sample_threshold_update(prof_tdata_t *tdata);
322 
323 #endif /* JEMALLOC_H_EXTERNS */
324 /******************************************************************************/
325 #ifdef JEMALLOC_H_INLINES
326 
327 #ifndef JEMALLOC_ENABLE_INLINE
328 bool	prof_active_get_unlocked(void);
329 bool	prof_gdump_get_unlocked(void);
330 prof_tdata_t	*prof_tdata_get(tsd_t *tsd, bool create);
331 bool	prof_sample_accum_update(tsd_t *tsd, size_t usize, bool commit,
332     prof_tdata_t **tdata_out);
333 prof_tctx_t	*prof_alloc_prep(tsd_t *tsd, size_t usize, bool update);
334 prof_tctx_t	*prof_tctx_get(const void *ptr);
335 void	prof_tctx_set(const void *ptr, prof_tctx_t *tctx);
336 void	prof_malloc_sample_object(const void *ptr, size_t usize,
337     prof_tctx_t *tctx);
338 void	prof_malloc(const void *ptr, size_t usize, prof_tctx_t *tctx);
339 void	prof_realloc(tsd_t *tsd, const void *ptr, size_t usize,
340     prof_tctx_t *tctx, bool updated, size_t old_usize, prof_tctx_t *old_tctx);
341 void	prof_free(tsd_t *tsd, const void *ptr, size_t usize);
342 #endif
343 
344 #if (defined(JEMALLOC_ENABLE_INLINE) || defined(JEMALLOC_PROF_C_))
345 JEMALLOC_ALWAYS_INLINE bool
prof_active_get_unlocked(void)346 prof_active_get_unlocked(void)
347 {
348 
349 	/*
350 	 * Even if opt_prof is true, sampling can be temporarily disabled by
351 	 * setting prof_active to false.  No locking is used when reading
352 	 * prof_active in the fast path, so there are no guarantees regarding
353 	 * how long it will take for all threads to notice state changes.
354 	 */
355 	return (prof_active);
356 }
357 
358 JEMALLOC_ALWAYS_INLINE bool
prof_gdump_get_unlocked(void)359 prof_gdump_get_unlocked(void)
360 {
361 
362 	/*
363 	 * No locking is used when reading prof_gdump_val in the fast path, so
364 	 * there are no guarantees regarding how long it will take for all
365 	 * threads to notice state changes.
366 	 */
367 	return (prof_gdump_val);
368 }
369 
370 JEMALLOC_ALWAYS_INLINE prof_tdata_t *
prof_tdata_get(tsd_t * tsd,bool create)371 prof_tdata_get(tsd_t *tsd, bool create)
372 {
373 	prof_tdata_t *tdata;
374 
375 	cassert(config_prof);
376 
377 	tdata = tsd_prof_tdata_get(tsd);
378 	if (create) {
379 		if (unlikely(tdata == NULL)) {
380 			if (tsd_nominal(tsd)) {
381 				tdata = prof_tdata_init(tsd);
382 				tsd_prof_tdata_set(tsd, tdata);
383 			}
384 		} else if (unlikely(tdata->expired)) {
385 			tdata = prof_tdata_reinit(tsd, tdata);
386 			tsd_prof_tdata_set(tsd, tdata);
387 		}
388 		assert(tdata == NULL || tdata->attached);
389 	}
390 
391 	return (tdata);
392 }
393 
394 JEMALLOC_ALWAYS_INLINE prof_tctx_t *
prof_tctx_get(const void * ptr)395 prof_tctx_get(const void *ptr)
396 {
397 
398 	cassert(config_prof);
399 	assert(ptr != NULL);
400 
401 	return (arena_prof_tctx_get(ptr));
402 }
403 
404 JEMALLOC_ALWAYS_INLINE void
prof_tctx_set(const void * ptr,prof_tctx_t * tctx)405 prof_tctx_set(const void *ptr, prof_tctx_t *tctx)
406 {
407 
408 	cassert(config_prof);
409 	assert(ptr != NULL);
410 
411 	arena_prof_tctx_set(ptr, tctx);
412 }
413 
414 JEMALLOC_ALWAYS_INLINE bool
prof_sample_accum_update(tsd_t * tsd,size_t usize,bool update,prof_tdata_t ** tdata_out)415 prof_sample_accum_update(tsd_t *tsd, size_t usize, bool update,
416     prof_tdata_t **tdata_out)
417 {
418 	prof_tdata_t *tdata;
419 
420 	cassert(config_prof);
421 
422 	tdata = prof_tdata_get(tsd, true);
423 	if ((uintptr_t)tdata <= (uintptr_t)PROF_TDATA_STATE_MAX)
424 		tdata = NULL;
425 
426 	if (tdata_out != NULL)
427 		*tdata_out = tdata;
428 
429 	if (tdata == NULL)
430 		return (true);
431 
432 	if (tdata->bytes_until_sample >= usize) {
433 		if (update)
434 			tdata->bytes_until_sample -= usize;
435 		return (true);
436 	} else {
437 		/* Compute new sample threshold. */
438 		if (update)
439 			prof_sample_threshold_update(tdata);
440 		return (!tdata->active);
441 	}
442 }
443 
444 JEMALLOC_ALWAYS_INLINE prof_tctx_t *
prof_alloc_prep(tsd_t * tsd,size_t usize,bool update)445 prof_alloc_prep(tsd_t *tsd, size_t usize, bool update)
446 {
447 	prof_tctx_t *ret;
448 	prof_tdata_t *tdata;
449 	prof_bt_t bt;
450 
451 	assert(usize == s2u(usize));
452 
453 	if (!prof_active_get_unlocked() || likely(prof_sample_accum_update(tsd,
454 	    usize, update, &tdata)))
455 		ret = (prof_tctx_t *)(uintptr_t)1U;
456 	else {
457 		bt_init(&bt, tdata->vec);
458 		prof_backtrace(&bt);
459 		ret = prof_lookup(tsd, &bt);
460 	}
461 
462 	return (ret);
463 }
464 
465 JEMALLOC_ALWAYS_INLINE void
prof_malloc(const void * ptr,size_t usize,prof_tctx_t * tctx)466 prof_malloc(const void *ptr, size_t usize, prof_tctx_t *tctx)
467 {
468 
469 	cassert(config_prof);
470 	assert(ptr != NULL);
471 	assert(usize == isalloc(ptr, true));
472 
473 	if (unlikely((uintptr_t)tctx > (uintptr_t)1U))
474 		prof_malloc_sample_object(ptr, usize, tctx);
475 	else
476 		prof_tctx_set(ptr, (prof_tctx_t *)(uintptr_t)1U);
477 }
478 
479 JEMALLOC_ALWAYS_INLINE void
prof_realloc(tsd_t * tsd,const void * ptr,size_t usize,prof_tctx_t * tctx,bool updated,size_t old_usize,prof_tctx_t * old_tctx)480 prof_realloc(tsd_t *tsd, const void *ptr, size_t usize, prof_tctx_t *tctx,
481     bool updated, size_t old_usize, prof_tctx_t *old_tctx)
482 {
483 
484 	cassert(config_prof);
485 	assert(ptr != NULL || (uintptr_t)tctx <= (uintptr_t)1U);
486 
487 	if (!updated && ptr != NULL) {
488 		assert(usize == isalloc(ptr, true));
489 		if (prof_sample_accum_update(tsd, usize, true, NULL)) {
490 			/*
491 			 * Don't sample.  The usize passed to PROF_ALLOC_PREP()
492 			 * was larger than what actually got allocated, so a
493 			 * backtrace was captured for this allocation, even
494 			 * though its actual usize was insufficient to cross the
495 			 * sample threshold.
496 			 */
497 			tctx = (prof_tctx_t *)(uintptr_t)1U;
498 		}
499 	}
500 
501 	if (unlikely((uintptr_t)old_tctx > (uintptr_t)1U))
502 		prof_free_sampled_object(tsd, old_usize, old_tctx);
503 	if (unlikely((uintptr_t)tctx > (uintptr_t)1U))
504 		prof_malloc_sample_object(ptr, usize, tctx);
505 	else
506 		prof_tctx_set(ptr, (prof_tctx_t *)(uintptr_t)1U);
507 }
508 
509 JEMALLOC_ALWAYS_INLINE void
prof_free(tsd_t * tsd,const void * ptr,size_t usize)510 prof_free(tsd_t *tsd, const void *ptr, size_t usize)
511 {
512 	prof_tctx_t *tctx = prof_tctx_get(ptr);
513 
514 	cassert(config_prof);
515 	assert(usize == isalloc(ptr, true));
516 
517 	if (unlikely((uintptr_t)tctx > (uintptr_t)1U))
518 		prof_free_sampled_object(tsd, usize, tctx);
519 }
520 #endif
521 
522 #endif /* JEMALLOC_H_INLINES */
523 /******************************************************************************/
524