tools/trace-malloc/lib/nsTraceMalloc.c@b8a032363ba2
tools/trace-malloc/lib/nsTraceMalloc.c
Thu, 22 Jan 2015 13:21:57 +0100
- author
- Michael Schloh von Bennewitz <michael@schloh.com>
- date
- Thu, 22 Jan 2015 13:21:57 +0100
- branch
- TOR_BUG_9701
- changeset 15
- b8a032363ba2
- permissions
- -rw-r--r--
Incorporate requested changes from Mozilla in review:
https://bugzilla.mozilla.org/show_bug.cgi?id=1123480#c6
1 /* -*- Mode: C; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 4 -*-
2 * vim:cindent:ts=8:et:sw=4:
3 *
4 * This Source Code Form is subject to the terms of the Mozilla Public
5 * License, v. 2.0. If a copy of the MPL was not distributed with this
6 * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
7 #ifdef NS_TRACE_MALLOC
8 /*
9 * TODO:
10 * - FIXME https://bugzilla.mozilla.org/show_bug.cgi?id=392008
11 * - extend logfile so 'F' record tells free stack
12 */
13 #include <errno.h>
14 #include <fcntl.h>
15 #include <stdio.h>
16 #include <string.h>
17 #ifdef XP_UNIX
18 #include <unistd.h>
19 #include <sys/stat.h>
20 #include <sys/time.h>
21 #endif
22 #include "plhash.h"
23 #include "pratom.h"
24 #include "prlog.h"
25 #include "prlock.h"
26 #include "prmon.h"
27 #include "prprf.h"
28 #include "prenv.h"
29 #include "prnetdb.h"
30 #include "nsTraceMalloc.h"
31 #include "nscore.h"
32 #include "prinit.h"
33 #include "prthread.h"
34 #include "plstr.h"
35 #include "nsStackWalk.h"
36 #include "nsTraceMallocCallbacks.h"
37 #include "nsTypeInfo.h"
38 #include "mozilla/PoisonIOInterposer.h"
40 #if defined(XP_MACOSX)
42 #include <malloc/malloc.h>
44 #define WRITE_FLAGS "w"
46 #define __libc_malloc(x) malloc(x)
47 #define __libc_realloc(x, y) realloc(x, y)
48 #define __libc_free(x) free(x)
50 #elif defined(XP_UNIX)
52 #include <malloc.h>
54 #define WRITE_FLAGS "w"
56 #ifdef WRAP_SYSTEM_INCLUDES
57 #pragma GCC visibility push(default)
58 #endif
59 extern __ptr_t __libc_malloc(size_t);
60 extern __ptr_t __libc_calloc(size_t, size_t);
61 extern __ptr_t __libc_realloc(__ptr_t, size_t);
62 extern void __libc_free(__ptr_t);
63 extern __ptr_t __libc_memalign(size_t, size_t);
64 extern __ptr_t __libc_valloc(size_t);
65 #ifdef WRAP_SYSTEM_INCLUDES
66 #pragma GCC visibility pop
67 #endif
69 #elif defined(XP_WIN32)
71 #include <sys/timeb.h> /* for timeb */
72 #include <sys/stat.h> /* for fstat */
74 #include <io.h> /*for write*/
76 #define WRITE_FLAGS "w"
78 #define __libc_malloc(x) dhw_orig_malloc(x)
79 #define __libc_realloc(x, y) dhw_orig_realloc(x,y)
80 #define __libc_free(x) dhw_orig_free(x)
82 #else /* not XP_MACOSX, XP_UNIX, or XP_WIN32 */
84 # error "Unknown build configuration!"
86 #endif
88 typedef struct logfile logfile;
90 #define STARTUP_TMBUFSIZE (64 * 1024)
91 #define LOGFILE_TMBUFSIZE (16 * 1024)
93 struct logfile {
94 int fd;
95 int lfd; /* logical fd, dense among all logfiles */
96 char *buf;
97 int bufsize;
98 int pos;
99 uint32_t size;
100 uint32_t simsize;
101 logfile *next;
102 logfile **prevp;
103 };
105 static char default_buf[STARTUP_TMBUFSIZE];
106 static logfile default_logfile =
107 {-1, 0, default_buf, STARTUP_TMBUFSIZE, 0, 0, 0, NULL, NULL};
108 static logfile *logfile_list = NULL;
109 static logfile **logfile_tail = &logfile_list;
110 static logfile *logfp = &default_logfile;
111 static PRLock *tmlock = NULL;
112 #ifndef PATH_MAX
113 #define PATH_MAX 4096
114 #endif
115 static char sdlogname[PATH_MAX] = ""; /* filename for shutdown leak log */
117 /*
118 * This enables/disables trace-malloc logging.
119 *
120 * It is separate from suppress_tracing so that we do not have to pay
121 * the performance cost of repeated TM_TLS_GET_DATA calls when
122 * trace-malloc is disabled (which is not as bad as the locking we used
123 * to have).
124 *
125 * It must default to zero, since it can be tested by the Linux malloc
126 * hooks before NS_TraceMallocStartup sets it.
127 */
128 static uint32_t tracing_enabled = 0;
130 /*
131 * Control whether we should log stacks
132 */
133 static uint32_t stacks_enabled = 1;
135 /*
136 * This lock must be held while manipulating the calltree, the
137 * allocations table, the log, or the tmstats.
138 *
139 * Callers should not *enter* the lock without checking suppress_tracing
140 * first; otherwise they risk trying to re-enter on the same thread.
141 */
142 #define TM_ENTER_LOCK(t) \
143 PR_BEGIN_MACRO \
144 PR_ASSERT(t->suppress_tracing != 0); \
145 if (tmlock) \
146 PR_Lock(tmlock); \
147 PR_END_MACRO
149 #define TM_EXIT_LOCK(t) \
150 PR_BEGIN_MACRO \
151 PR_ASSERT(t->suppress_tracing != 0); \
152 if (tmlock) \
153 PR_Unlock(tmlock); \
154 PR_END_MACRO
156 #define TM_SUPPRESS_TRACING_AND_ENTER_LOCK(t) \
157 PR_BEGIN_MACRO \
158 t->suppress_tracing++; \
159 TM_ENTER_LOCK(t); \
160 PR_END_MACRO
162 #define TM_EXIT_LOCK_AND_UNSUPPRESS_TRACING(t) \
163 PR_BEGIN_MACRO \
164 TM_EXIT_LOCK(t); \
165 t->suppress_tracing--; \
166 PR_END_MACRO
169 /*
170 * Thread-local storage.
171 *
172 * We can't use NSPR thread-local storage for this because it mallocs
173 * within PR_GetThreadPrivate (the first time) and PR_SetThreadPrivate
174 * (which can be worked around by protecting all uses of those functions
175 * with a monitor, ugh) and because it calls malloc/free when the
176 * thread-local storage is in an inconsistent state within
177 * PR_SetThreadPrivate (when expanding the thread-local storage array)
178 * and _PRI_DetachThread (when and after deleting the thread-local
179 * storage array).
180 */
182 #ifdef XP_WIN32
184 #include <windows.h>
186 #define TM_TLS_INDEX_TYPE DWORD
187 #define TM_CREATE_TLS_INDEX(i_) PR_BEGIN_MACRO \
188 (i_) = TlsAlloc(); \
189 PR_END_MACRO
190 #define TM_DESTROY_TLS_INDEX(i_) TlsFree((i_))
191 #define TM_GET_TLS_DATA(i_) TlsGetValue((i_))
192 #define TM_SET_TLS_DATA(i_, v_) TlsSetValue((i_), (v_))
194 #else
196 #include <pthread.h>
198 #define TM_TLS_INDEX_TYPE pthread_key_t
199 #define TM_CREATE_TLS_INDEX(i_) pthread_key_create(&(i_), NULL)
200 #define TM_DESTROY_TLS_INDEX(i_) pthread_key_delete((i_))
201 #define TM_GET_TLS_DATA(i_) pthread_getspecific((i_))
202 #define TM_SET_TLS_DATA(i_, v_) pthread_setspecific((i_), (v_))
204 #endif
206 static TM_TLS_INDEX_TYPE tls_index;
207 static PRBool tls_index_initialized = PR_FALSE;
209 /* FIXME (maybe): This is currently unused; we leak the thread-local data. */
210 #if 0
211 static void
212 free_tm_thread(void *priv)
213 {
214 tm_thread *t = (tm_thread*) priv;
216 PR_ASSERT(t->suppress_tracing == 0);
218 if (t->in_heap) {
219 t->suppress_tracing = 1;
220 if (t->backtrace_buf.buffer)
221 __libc_free(t->backtrace_buf.buffer);
223 __libc_free(t);
224 }
225 }
226 #endif
228 tm_thread *
229 tm_get_thread(void)
230 {
231 tm_thread *t;
232 tm_thread stack_tm_thread;
234 if (!tls_index_initialized) {
235 /**
236 * Assume that the first call to |malloc| will occur before
237 * there are multiple threads. (If that's not the case, we
238 * probably need to do the necessary synchronization without
239 * using NSPR primitives. See discussion in
240 * https://bugzilla.mozilla.org/show_bug.cgi?id=442192
241 */
242 TM_CREATE_TLS_INDEX(tls_index);
243 tls_index_initialized = PR_TRUE;
244 }
246 t = TM_GET_TLS_DATA(tls_index);
248 if (!t) {
249 /*
250 * First, store a tm_thread on the stack to suppress for the
251 * malloc below
252 */
253 stack_tm_thread.suppress_tracing = 1;
254 stack_tm_thread.backtrace_buf.buffer = NULL;
255 stack_tm_thread.backtrace_buf.size = 0;
256 stack_tm_thread.backtrace_buf.entries = 0;
257 TM_SET_TLS_DATA(tls_index, &stack_tm_thread);
259 t = (tm_thread*) __libc_malloc(sizeof(tm_thread));
260 t->suppress_tracing = 0;
261 t->backtrace_buf = stack_tm_thread.backtrace_buf;
262 TM_SET_TLS_DATA(tls_index, t);
264 PR_ASSERT(stack_tm_thread.suppress_tracing == 1); /* balanced */
265 }
267 return t;
268 }
270 /* We don't want more than 32 logfiles open at once, ok? */
271 typedef uint32_t lfd_set;
273 #define LFD_SET_STATIC_INITIALIZER 0
274 #define LFD_SET_SIZE 32
276 #define LFD_ZERO(s) (*(s) = 0)
277 #define LFD_BIT(i) ((uint32_t)1 << (i))
278 #define LFD_TEST(i,s) (LFD_BIT(i) & *(s))
279 #define LFD_SET(i,s) (*(s) |= LFD_BIT(i))
280 #define LFD_CLR(i,s) (*(s) &= ~LFD_BIT(i))
282 static logfile *get_logfile(int fd)
283 {
284 logfile *fp;
285 int lfd;
287 for (fp = logfile_list; fp; fp = fp->next) {
288 if (fp->fd == fd)
289 return fp;
290 }
291 lfd = 0;
292 retry:
293 for (fp = logfile_list; fp; fp = fp->next) {
294 if (fp->fd == lfd) {
295 if (++lfd >= LFD_SET_SIZE)
296 return NULL;
297 goto retry;
298 }
299 }
300 fp = __libc_malloc(sizeof(logfile) + LOGFILE_TMBUFSIZE);
301 if (!fp)
302 return NULL;
303 fp->fd = fd;
304 fp->lfd = lfd;
305 fp->buf = (char*) (fp + 1);
306 fp->bufsize = LOGFILE_TMBUFSIZE;
307 fp->pos = 0;
308 fp->size = fp->simsize = 0;
309 fp->next = NULL;
310 fp->prevp = logfile_tail;
311 *logfile_tail = fp;
312 logfile_tail = &fp->next;
313 return fp;
314 }
316 static void flush_logfile(logfile *fp)
317 {
318 int len, cnt, fd;
319 char *bp;
321 len = fp->pos;
322 if (len == 0)
323 return;
324 fp->pos = 0;
325 fd = fp->fd;
326 if (fd >= 0) {
327 fp->size += len;
328 bp = fp->buf;
329 do {
330 cnt = write(fd, bp, len);
331 if (cnt <= 0) {
332 printf("### nsTraceMalloc: write failed or wrote 0 bytes!\n");
333 return;
334 }
335 bp += cnt;
336 len -= cnt;
337 } while (len > 0);
338 }
339 fp->simsize += len;
340 }
342 static void log_byte(logfile *fp, char byte)
343 {
344 if (fp->pos == fp->bufsize)
345 flush_logfile(fp);
346 fp->buf[fp->pos++] = byte;
347 }
349 static void log_string(logfile *fp, const char *str)
350 {
351 int len, rem, cnt;
353 len = strlen(str) + 1; /* include null terminator */
354 while ((rem = fp->pos + len - fp->bufsize) > 0) {
355 cnt = len - rem;
356 memcpy(&fp->buf[fp->pos], str, cnt);
357 str += cnt;
358 fp->pos += cnt;
359 flush_logfile(fp);
360 len = rem;
361 }
362 memcpy(&fp->buf[fp->pos], str, len);
363 fp->pos += len;
364 }
366 static void log_filename(logfile* fp, const char* filename)
367 {
368 if (strlen(filename) < 512) {
369 char *bp, *cp, buf[512];
371 bp = strstr(strcpy(buf, filename), "mozilla");
372 if (!bp)
373 bp = buf;
375 for (cp = bp; *cp; cp++) {
376 if (*cp == '\\')
377 *cp = '/';
378 }
380 filename = bp;
381 }
382 log_string(fp, filename);
383 }
385 static void log_uint32(logfile *fp, uint32_t ival)
386 {
387 if (ival < 0x80) {
388 /* 0xxx xxxx */
389 log_byte(fp, (char) ival);
390 } else if (ival < 0x4000) {
391 /* 10xx xxxx xxxx xxxx */
392 log_byte(fp, (char) ((ival >> 8) | 0x80));
393 log_byte(fp, (char) (ival & 0xff));
394 } else if (ival < 0x200000) {
395 /* 110x xxxx xxxx xxxx xxxx xxxx */
396 log_byte(fp, (char) ((ival >> 16) | 0xc0));
397 log_byte(fp, (char) ((ival >> 8) & 0xff));
398 log_byte(fp, (char) (ival & 0xff));
399 } else if (ival < 0x10000000) {
400 /* 1110 xxxx xxxx xxxx xxxx xxxx xxxx xxxx */
401 log_byte(fp, (char) ((ival >> 24) | 0xe0));
402 log_byte(fp, (char) ((ival >> 16) & 0xff));
403 log_byte(fp, (char) ((ival >> 8) & 0xff));
404 log_byte(fp, (char) (ival & 0xff));
405 } else {
406 /* 1111 0000 xxxx xxxx xxxx xxxx xxxx xxxx xxxx xxxx */
407 log_byte(fp, (char) 0xf0);
408 log_byte(fp, (char) ((ival >> 24) & 0xff));
409 log_byte(fp, (char) ((ival >> 16) & 0xff));
410 log_byte(fp, (char) ((ival >> 8) & 0xff));
411 log_byte(fp, (char) (ival & 0xff));
412 }
413 }
415 static void log_event1(logfile *fp, char event, uint32_t serial)
416 {
417 log_byte(fp, event);
418 log_uint32(fp, (uint32_t) serial);
419 }
421 static void log_event2(logfile *fp, char event, uint32_t serial, size_t size)
422 {
423 log_event1(fp, event, serial);
424 log_uint32(fp, (uint32_t) size);
425 }
427 static void log_event3(logfile *fp, char event, uint32_t serial, size_t oldsize,
428 size_t size)
429 {
430 log_event2(fp, event, serial, oldsize);
431 log_uint32(fp, (uint32_t) size);
432 }
434 static void log_event4(logfile *fp, char event, uint32_t serial, uint32_t ui2,
435 uint32_t ui3, uint32_t ui4)
436 {
437 log_event3(fp, event, serial, ui2, ui3);
438 log_uint32(fp, ui4);
439 }
441 static void log_event5(logfile *fp, char event, uint32_t serial, uint32_t ui2,
442 uint32_t ui3, uint32_t ui4, uint32_t ui5)
443 {
444 log_event4(fp, event, serial, ui2, ui3, ui4);
445 log_uint32(fp, ui5);
446 }
448 static void log_event6(logfile *fp, char event, uint32_t serial, uint32_t ui2,
449 uint32_t ui3, uint32_t ui4, uint32_t ui5, uint32_t ui6)
450 {
451 log_event5(fp, event, serial, ui2, ui3, ui4, ui5);
452 log_uint32(fp, ui6);
453 }
455 static void log_event7(logfile *fp, char event, uint32_t serial, uint32_t ui2,
456 uint32_t ui3, uint32_t ui4, uint32_t ui5, uint32_t ui6,
457 uint32_t ui7)
458 {
459 log_event6(fp, event, serial, ui2, ui3, ui4, ui5, ui6);
460 log_uint32(fp, ui7);
461 }
463 static void log_event8(logfile *fp, char event, uint32_t serial, uint32_t ui2,
464 uint32_t ui3, uint32_t ui4, uint32_t ui5, uint32_t ui6,
465 uint32_t ui7, uint32_t ui8)
466 {
467 log_event7(fp, event, serial, ui2, ui3, ui4, ui5, ui6, ui7);
468 log_uint32(fp, ui8);
469 }
471 typedef struct callsite callsite;
473 struct callsite {
474 void* pc;
475 uint32_t serial;
476 lfd_set lfdset;
477 const char *name; /* pointer to string owned by methods table */
478 const char *library; /* pointer to string owned by libraries table */
479 int offset;
480 callsite *parent;
481 callsite *siblings;
482 callsite *kids;
483 };
485 /* NB: these counters are incremented and decremented only within tmlock. */
486 static uint32_t library_serial_generator = 0;
487 static uint32_t method_serial_generator = 0;
488 static uint32_t callsite_serial_generator = 0;
489 static uint32_t tmstats_serial_generator = 0;
490 static uint32_t filename_serial_generator = 0;
492 /* Root of the tree of callsites, the sum of all (cycle-compressed) stacks. */
493 static callsite calltree_root =
494 {0, 0, LFD_SET_STATIC_INITIALIZER, NULL, NULL, 0, NULL, NULL, NULL};
496 /* a fake pc for when stacks are disabled; must be different from the
497 pc in calltree_root */
498 #define STACK_DISABLED_PC ((void*)1)
500 /* Basic instrumentation. */
501 static nsTMStats tmstats = NS_TMSTATS_STATIC_INITIALIZER;
503 /* Parent with the most kids (tmstats.calltree_maxkids). */
504 static callsite *calltree_maxkids_parent;
506 /* Calltree leaf for path with deepest stack backtrace. */
507 static callsite *calltree_maxstack_top;
509 /* Last site (i.e., calling pc) that recurred during a backtrace. */
510 static callsite *last_callsite_recurrence;
512 static void log_tmstats(logfile *fp)
513 {
514 log_event1(fp, TM_EVENT_STATS, ++tmstats_serial_generator);
515 log_uint32(fp, tmstats.calltree_maxstack);
516 log_uint32(fp, tmstats.calltree_maxdepth);
517 log_uint32(fp, tmstats.calltree_parents);
518 log_uint32(fp, tmstats.calltree_maxkids);
519 log_uint32(fp, tmstats.calltree_kidhits);
520 log_uint32(fp, tmstats.calltree_kidmisses);
521 log_uint32(fp, tmstats.calltree_kidsteps);
522 log_uint32(fp, tmstats.callsite_recurrences);
523 log_uint32(fp, tmstats.backtrace_calls);
524 log_uint32(fp, tmstats.backtrace_failures);
525 log_uint32(fp, tmstats.btmalloc_failures);
526 log_uint32(fp, tmstats.dladdr_failures);
527 log_uint32(fp, tmstats.malloc_calls);
528 log_uint32(fp, tmstats.malloc_failures);
529 log_uint32(fp, tmstats.calloc_calls);
530 log_uint32(fp, tmstats.calloc_failures);
531 log_uint32(fp, tmstats.realloc_calls);
532 log_uint32(fp, tmstats.realloc_failures);
533 log_uint32(fp, tmstats.free_calls);
534 log_uint32(fp, tmstats.null_free_calls);
535 log_uint32(fp, calltree_maxkids_parent ? calltree_maxkids_parent->serial
536 : 0);
537 log_uint32(fp, calltree_maxstack_top ? calltree_maxstack_top->serial : 0);
538 }
540 static void *generic_alloctable(void *pool, size_t size)
541 {
542 return __libc_malloc(size);
543 }
545 static void generic_freetable(void *pool, void *item)
546 {
547 __libc_free(item);
548 }
550 typedef struct lfdset_entry {
551 PLHashEntry base;
552 lfd_set lfdset;
553 } lfdset_entry;
555 static PLHashEntry *lfdset_allocentry(void *pool, const void *key)
556 {
557 lfdset_entry *le = __libc_malloc(sizeof *le);
558 if (le)
559 LFD_ZERO(&le->lfdset);
560 return &le->base;
561 }
563 static void lfdset_freeentry(void *pool, PLHashEntry *he, unsigned flag)
564 {
565 lfdset_entry *le;
567 if (flag != HT_FREE_ENTRY)
568 return;
569 le = (lfdset_entry*) he;
570 __libc_free((void*) le);
571 }
573 static PLHashAllocOps lfdset_hashallocops = {
574 generic_alloctable, generic_freetable,
575 lfdset_allocentry, lfdset_freeentry
576 };
578 /* Table of library pathnames mapped to to logged 'L' record serial numbers. */
579 static PLHashTable *libraries = NULL;
581 /* Table of filename pathnames mapped to logged 'G' record serial numbers. */
582 static PLHashTable *filenames = NULL;
584 /* Table mapping method names to logged 'N' record serial numbers. */
585 static PLHashTable *methods = NULL;
587 /*
588 * Presumes that its caller is holding tmlock, but may temporarily exit
589 * the lock.
590 */
591 static callsite *
592 calltree(void **stack, size_t num_stack_entries, tm_thread *t)
593 {
594 logfile *fp = logfp;
595 void *pc;
596 uint32_t nkids;
597 callsite *parent, *site, **csp, *tmp;
598 int maxstack;
599 uint32_t library_serial, method_serial, filename_serial;
600 const char *library, *method, *filename;
601 char *slash;
602 PLHashNumber hash;
603 PLHashEntry **hep, *he;
604 lfdset_entry *le;
605 size_t stack_index;
606 nsCodeAddressDetails details;
607 nsresult rv;
609 maxstack = (num_stack_entries > tmstats.calltree_maxstack);
610 if (maxstack) {
611 /* these two are the same, although that used to be less clear */
612 tmstats.calltree_maxstack = num_stack_entries;
613 tmstats.calltree_maxdepth = num_stack_entries;
614 }
616 /* Reverse the stack again, finding and building a path in the tree. */
617 parent = &calltree_root;
618 stack_index = num_stack_entries;
619 do {
620 --stack_index;
621 pc = stack[stack_index];
623 csp = &parent->kids;
624 while ((site = *csp) != NULL) {
625 if (site->pc == pc) {
626 tmstats.calltree_kidhits++;
628 /* Put the most recently used site at the front of siblings. */
629 *csp = site->siblings;
630 site->siblings = parent->kids;
631 parent->kids = site;
633 /* Check whether we've logged for this site and logfile yet. */
634 if (!LFD_TEST(fp->lfd, &site->lfdset)) {
635 /*
636 * Some other logfile put this site in the calltree. We
637 * must log an event for site, and possibly first for its
638 * method and/or library. Note the code after the while
639 * loop that tests if (!site).
640 */
641 break;
642 }
644 /* Site already built and logged to fp -- go up the stack. */
645 goto upward;
646 }
647 tmstats.calltree_kidsteps++;
648 csp = &site->siblings;
649 }
651 if (!site) {
652 tmstats.calltree_kidmisses++;
654 /* Check for recursion: see if pc is on our ancestor line. */
655 for (site = parent; site; site = site->parent) {
656 if (site->pc == pc) {
657 tmstats.callsite_recurrences++;
658 last_callsite_recurrence = site;
659 goto upward;
660 }
661 }
662 }
664 /*
665 * Not in tree at all, or not logged to fp: let's find our symbolic
666 * callsite info.
667 */
669 if (!stacks_enabled) {
670 /*
671 * Fake the necessary information for our single fake stack
672 * frame.
673 */
674 PL_strncpyz(details.library, "stacks_disabled",
675 sizeof(details.library));
676 details.loffset = 0;
677 details.filename[0] = '\0';
678 details.lineno = 0;
679 details.function[0] = '\0';
680 details.foffset = 0;
681 } else {
682 /*
683 * NS_DescribeCodeAddress can (on Linux) acquire a lock inside
684 * the shared library loader. Another thread might call malloc
685 * while holding that lock (when loading a shared library). So
686 * we have to exit tmlock around this call. For details, see
687 * https://bugzilla.mozilla.org/show_bug.cgi?id=363334#c3
688 *
689 * We could be more efficient by building the nodes in the
690 * calltree, exiting the monitor once to describe all of them,
691 * and then filling in the descriptions for any that hadn't been
692 * described already. But this is easier for now.
693 */
694 TM_EXIT_LOCK(t);
695 rv = NS_DescribeCodeAddress(pc, &details);
696 TM_ENTER_LOCK(t);
697 if (NS_FAILED(rv)) {
698 tmstats.dladdr_failures++;
699 goto fail;
700 }
701 }
703 /* Check whether we need to emit a library trace record. */
704 library_serial = 0;
705 library = NULL;
706 if (details.library[0]) {
707 if (!libraries) {
708 libraries = PL_NewHashTable(100, PL_HashString,
709 PL_CompareStrings, PL_CompareValues,
710 &lfdset_hashallocops, NULL);
711 if (!libraries) {
712 tmstats.btmalloc_failures++;
713 goto fail;
714 }
715 }
716 hash = PL_HashString(details.library);
717 hep = PL_HashTableRawLookup(libraries, hash, details.library);
718 he = *hep;
719 if (he) {
720 library = (char*) he->key;
721 library_serial = (uint32_t) NS_PTR_TO_INT32(he->value);
722 le = (lfdset_entry *) he;
723 if (LFD_TEST(fp->lfd, &le->lfdset)) {
724 /* We already logged an event on fp for this library. */
725 le = NULL;
726 }
727 } else {
728 library = strdup(details.library);
729 if (library) {
730 library_serial = ++library_serial_generator;
731 he = PL_HashTableRawAdd(libraries, hep, hash, library,
732 NS_INT32_TO_PTR(library_serial));
733 }
734 if (!he) {
735 tmstats.btmalloc_failures++;
736 goto fail;
737 }
738 le = (lfdset_entry *) he;
739 }
740 if (le) {
741 /* Need to log an event to fp for this lib. */
742 slash = strrchr(library, '/');
743 log_event1(fp, TM_EVENT_LIBRARY, library_serial);
744 log_string(fp, slash ? slash + 1 : library);
745 LFD_SET(fp->lfd, &le->lfdset);
746 }
747 }
749 /* For compatibility with current log format, always emit a
750 * filename trace record, using "noname" / 0 when no file name
751 * is available. */
752 filename_serial = 0;
753 filename = details.filename[0] ? details.filename : "noname";
754 if (!filenames) {
755 filenames = PL_NewHashTable(100, PL_HashString,
756 PL_CompareStrings, PL_CompareValues,
757 &lfdset_hashallocops, NULL);
758 if (!filenames) {
759 tmstats.btmalloc_failures++;
760 return NULL;
761 }
762 }
763 hash = PL_HashString(filename);
764 hep = PL_HashTableRawLookup(filenames, hash, filename);
765 he = *hep;
766 if (he) {
767 filename = (char*) he->key;
768 filename_serial = (uint32_t) NS_PTR_TO_INT32(he->value);
769 le = (lfdset_entry *) he;
770 if (LFD_TEST(fp->lfd, &le->lfdset)) {
771 /* We already logged an event on fp for this filename. */
772 le = NULL;
773 }
774 } else {
775 filename = strdup(filename);
776 if (filename) {
777 filename_serial = ++filename_serial_generator;
778 he = PL_HashTableRawAdd(filenames, hep, hash, filename,
779 NS_INT32_TO_PTR(filename_serial));
780 }
781 if (!he) {
782 tmstats.btmalloc_failures++;
783 return NULL;
784 }
785 le = (lfdset_entry *) he;
786 }
787 if (le) {
788 /* Need to log an event to fp for this filename. */
789 log_event1(fp, TM_EVENT_FILENAME, filename_serial);
790 log_filename(fp, filename);
791 LFD_SET(fp->lfd, &le->lfdset);
792 }
794 if (!details.function[0]) {
795 PR_snprintf(details.function, sizeof(details.function),
796 "%s+%X", library ? library : "main", details.loffset);
797 }
799 /* Emit an 'N' (for New method, 'M' is for malloc!) event if needed. */
800 method_serial = 0;
801 if (!methods) {
802 methods = PL_NewHashTable(10000, PL_HashString,
803 PL_CompareStrings, PL_CompareValues,
804 &lfdset_hashallocops, NULL);
805 if (!methods) {
806 tmstats.btmalloc_failures++;
807 goto fail;
808 }
809 }
810 hash = PL_HashString(details.function);
811 hep = PL_HashTableRawLookup(methods, hash, details.function);
812 he = *hep;
813 if (he) {
814 method = (char*) he->key;
815 method_serial = (uint32_t) NS_PTR_TO_INT32(he->value);
816 le = (lfdset_entry *) he;
817 if (LFD_TEST(fp->lfd, &le->lfdset)) {
818 /* We already logged an event on fp for this method. */
819 le = NULL;
820 }
821 } else {
822 method = strdup(details.function);
823 if (method) {
824 method_serial = ++method_serial_generator;
825 he = PL_HashTableRawAdd(methods, hep, hash, method,
826 NS_INT32_TO_PTR(method_serial));
827 }
828 if (!he) {
829 tmstats.btmalloc_failures++;
830 return NULL;
831 }
832 le = (lfdset_entry *) he;
833 }
834 if (le) {
835 log_event4(fp, TM_EVENT_METHOD, method_serial, library_serial,
836 filename_serial, details.lineno);
837 log_string(fp, method);
838 LFD_SET(fp->lfd, &le->lfdset);
839 }
841 /* Create a new callsite record. */
842 if (!site) {
843 site = __libc_malloc(sizeof(callsite));
844 if (!site) {
845 tmstats.btmalloc_failures++;
846 goto fail;
847 }
849 /* Update parent and max-kids-per-parent stats. */
850 if (!parent->kids)
851 tmstats.calltree_parents++;
852 nkids = 1;
853 for (tmp = parent->kids; tmp; tmp = tmp->siblings)
854 nkids++;
855 if (nkids > tmstats.calltree_maxkids) {
856 tmstats.calltree_maxkids = nkids;
857 calltree_maxkids_parent = parent;
858 }
860 /* Insert the new site into the tree. */
861 site->pc = pc;
862 site->serial = ++callsite_serial_generator;
863 LFD_ZERO(&site->lfdset);
864 site->name = method;
865 site->library = library;
866 site->offset = details.loffset;
867 site->parent = parent;
868 site->siblings = parent->kids;
869 parent->kids = site;
870 site->kids = NULL;
871 }
873 /* Log the site with its parent, method, and offset. */
874 log_event4(fp, TM_EVENT_CALLSITE, site->serial, parent->serial,
875 method_serial, details.foffset);
876 LFD_SET(fp->lfd, &site->lfdset);
878 upward:
879 parent = site;
880 } while (stack_index > 0);
882 if (maxstack)
883 calltree_maxstack_top = site;
885 return site;
887 fail:
888 return NULL;
889 }
891 /*
892 * Buffer the stack from top at low index to bottom at high, so that we can
893 * reverse it in calltree.
894 */
895 static void
896 stack_callback(void *pc, void *sp, void *closure)
897 {
898 stack_buffer_info *info = (stack_buffer_info*) closure;
900 /*
901 * If we run out of buffer, keep incrementing entries so that
902 * backtrace can call us again with a bigger buffer.
903 */
904 if (info->entries < info->size)
905 info->buffer[info->entries] = pc;
906 ++info->entries;
907 }
909 /*
910 * The caller MUST NOT be holding tmlock when calling backtrace.
911 * On return, if *immediate_abort is set, then the return value is NULL
912 * and the thread is in a very dangerous situation (e.g. holding
913 * sem_pool_lock in Mac OS X pthreads); the caller should bail out
914 * without doing anything (such as acquiring locks).
915 */
916 static callsite *
917 backtrace(tm_thread *t, int skipFrames, int *immediate_abort)
918 {
919 callsite *site;
920 stack_buffer_info *info = &t->backtrace_buf;
921 void ** new_stack_buffer;
922 size_t new_stack_buffer_size;
923 nsresult rv;
925 t->suppress_tracing++;
927 if (!stacks_enabled) {
928 #if defined(XP_MACOSX)
929 /* Walk the stack, even if stacks_enabled is false. We do this to
930 check if we must set immediate_abort. */
931 info->entries = 0;
932 rv = NS_StackWalk(stack_callback, skipFrames, /* maxFrames */ 0, info,
933 0, NULL);
934 *immediate_abort = rv == NS_ERROR_UNEXPECTED;
935 if (rv == NS_ERROR_UNEXPECTED || info->entries == 0) {
936 t->suppress_tracing--;
937 return NULL;
938 }
939 #endif
941 /*
942 * Create a single fake stack frame so that all the tools get
943 * data in the correct format.
944 */
945 *immediate_abort = 0;
946 if (info->size < 1) {
947 PR_ASSERT(!info->buffer); /* !info->size == !info->buffer */
948 info->buffer = __libc_malloc(1 * sizeof(void*));
949 if (!info->buffer)
950 return NULL;
951 info->size = 1;
952 }
954 info->entries = 1;
955 info->buffer[0] = STACK_DISABLED_PC;
956 } else {
957 /*
958 * NS_StackWalk can (on Windows) acquire a lock the shared library
959 * loader. Another thread might call malloc while holding that lock
960 * (when loading a shared library). So we can't be in tmlock during
961 * this call. For details, see
962 * https://bugzilla.mozilla.org/show_bug.cgi?id=374829#c8
963 */
965 /*
966 * skipFrames == 0 means |backtrace| should show up, so don't use
967 * skipFrames + 1.
968 * NB: this call is repeated below if the buffer is too small.
969 */
970 info->entries = 0;
971 rv = NS_StackWalk(stack_callback, skipFrames, /* maxFrames */ 0, info,
972 0, NULL);
973 *immediate_abort = rv == NS_ERROR_UNEXPECTED;
974 if (rv == NS_ERROR_UNEXPECTED || info->entries == 0) {
975 t->suppress_tracing--;
976 return NULL;
977 }
979 /*
980 * To avoid allocating in stack_callback (which, on Windows, is
981 * called on a different thread from the one we're running on here),
982 * reallocate here if it didn't have a big enough buffer (which
983 * includes the first call on any thread), and call it again.
984 */
985 if (info->entries > info->size) {
986 new_stack_buffer_size = 2 * info->entries;
987 new_stack_buffer = __libc_realloc(info->buffer,
988 new_stack_buffer_size * sizeof(void*));
989 if (!new_stack_buffer)
990 return NULL;
991 info->buffer = new_stack_buffer;
992 info->size = new_stack_buffer_size;
994 /* and call NS_StackWalk again */
995 info->entries = 0;
996 NS_StackWalk(stack_callback, skipFrames, /* maxFrames */ 0, info,
997 0, NULL);
999 /* same stack */
1000 PR_ASSERT(info->entries * 2 == new_stack_buffer_size);
1001 }
1002 }
1004 TM_ENTER_LOCK(t);
1006 site = calltree(info->buffer, info->entries, t);
1008 tmstats.backtrace_calls++;
1009 if (!site) {
1010 tmstats.backtrace_failures++;
1011 PR_ASSERT(tmstats.backtrace_failures < 100);
1012 }
1013 TM_EXIT_LOCK(t);
1015 t->suppress_tracing--;
1016 return site;
1017 }
1019 typedef struct allocation {
1020 PLHashEntry entry;
1021 size_t size;
1022 FILE *trackfp; /* for allocation tracking */
1023 } allocation;
1025 #define ALLOC_HEAP_SIZE 150000
1027 static allocation alloc_heap[ALLOC_HEAP_SIZE];
1028 static allocation *alloc_freelist = NULL;
1029 static int alloc_heap_initialized = 0;
1031 static PLHashEntry *alloc_allocentry(void *pool, const void *key)
1032 {
1033 allocation **listp, *alloc;
1034 int n;
1036 if (!alloc_heap_initialized) {
1037 n = ALLOC_HEAP_SIZE;
1038 listp = &alloc_freelist;
1039 for (alloc = alloc_heap; --n >= 0; alloc++) {
1040 *listp = alloc;
1041 listp = (allocation**) &alloc->entry.next;
1042 }
1043 *listp = NULL;
1044 alloc_heap_initialized = 1;
1045 }
1047 listp = &alloc_freelist;
1048 alloc = *listp;
1049 if (!alloc)
1050 return __libc_malloc(sizeof(allocation));
1051 *listp = (allocation*) alloc->entry.next;
1052 return &alloc->entry;
1053 }
1055 static void alloc_freeentry(void *pool, PLHashEntry *he, unsigned flag)
1056 {
1057 allocation *alloc;
1059 if (flag != HT_FREE_ENTRY)
1060 return;
1061 alloc = (allocation*) he;
1062 if ((ptrdiff_t)(alloc - alloc_heap) < (ptrdiff_t)ALLOC_HEAP_SIZE) {
1063 alloc->entry.next = &alloc_freelist->entry;
1064 alloc_freelist = alloc;
1065 } else {
1066 __libc_free((void*) alloc);
1067 }
1068 }
1070 static PLHashAllocOps alloc_hashallocops = {
1071 generic_alloctable, generic_freetable,
1072 alloc_allocentry, alloc_freeentry
1073 };
1075 static PLHashNumber hash_pointer(const void *key)
1076 {
1077 return (PLHashNumber) key;
1078 }
1080 static PLHashTable *allocations = NULL;
1082 static PLHashTable *new_allocations(void)
1083 {
1084 allocations = PL_NewHashTable(200000, hash_pointer,
1085 PL_CompareValues, PL_CompareValues,
1086 &alloc_hashallocops, NULL);
1087 return allocations;
1088 }
1090 #define get_allocations() (allocations ? allocations : new_allocations())
1092 #if defined(XP_MACOSX)
1094 /* from malloc.c in Libc */
1095 typedef void
1096 malloc_logger_t(uint32_t type, uintptr_t arg1, uintptr_t arg2, uintptr_t arg3,
1097 uintptr_t result, uint32_t num_hot_frames_to_skip);
1099 extern malloc_logger_t *malloc_logger;
1101 #define MALLOC_LOG_TYPE_ALLOCATE 2
1102 #define MALLOC_LOG_TYPE_DEALLOCATE 4
1103 #define MALLOC_LOG_TYPE_HAS_ZONE 8
1104 #define MALLOC_LOG_TYPE_CLEARED 64
1106 static void
1107 my_malloc_logger(uint32_t type, uintptr_t arg1, uintptr_t arg2, uintptr_t arg3,
1108 uintptr_t result, uint32_t num_hot_frames_to_skip)
1109 {
1110 uintptr_t all_args[3] = { arg1, arg2, arg3 };
1111 uintptr_t *args = all_args + ((type & MALLOC_LOG_TYPE_HAS_ZONE) ? 1 : 0);
1113 uint32_t alloc_type =
1114 type & (MALLOC_LOG_TYPE_ALLOCATE | MALLOC_LOG_TYPE_DEALLOCATE);
1115 tm_thread *t = tm_get_thread();
1117 if (alloc_type == (MALLOC_LOG_TYPE_ALLOCATE | MALLOC_LOG_TYPE_DEALLOCATE)) {
1118 ReallocCallback((void*)args[0], (void*)result, args[1], 0, 0, t);
1119 } else if (alloc_type == MALLOC_LOG_TYPE_ALLOCATE) {
1120 /*
1121 * We don't get size/count information for calloc, so just use
1122 * MallocCallback.
1123 */
1124 MallocCallback((void*)result, args[0], 0, 0, t);
1125 } else if (alloc_type == MALLOC_LOG_TYPE_DEALLOCATE) {
1126 FreeCallback((void*)args[0], 0, 0, t);
1127 }
1128 }
1130 static void
1131 StartupHooker(void)
1132 {
1133 PR_ASSERT(!malloc_logger);
1134 malloc_logger = my_malloc_logger;
1135 }
1137 static void
1138 ShutdownHooker(void)
1139 {
1140 PR_ASSERT(malloc_logger == my_malloc_logger);
1141 malloc_logger = NULL;
1142 }
1144 #elif defined(XP_UNIX)
1146 /*
1147 * We can't use glibc's malloc hooks because they can't be used in a
1148 * threadsafe manner. They require unsetting the hooks to call into the
1149 * original malloc implementation, and then resetting them when the
1150 * original implementation returns. If another thread calls the same
1151 * allocation function while the hooks are unset, we have no chance to
1152 * intercept the call.
1153 */
1155 NS_EXTERNAL_VIS_(__ptr_t)
1156 malloc(size_t size)
1157 {
1158 uint32_t start, end;
1159 __ptr_t ptr;
1160 tm_thread *t;
1162 if (!tracing_enabled || !PR_Initialized() ||
1163 (t = tm_get_thread())->suppress_tracing != 0) {
1164 return __libc_malloc(size);
1165 }
1167 t->suppress_tracing++;
1168 start = PR_IntervalNow();
1169 ptr = __libc_malloc(size);
1170 end = PR_IntervalNow();
1171 t->suppress_tracing--;
1173 MallocCallback(ptr, size, start, end, t);
1175 return ptr;
1176 }
1178 NS_EXTERNAL_VIS_(__ptr_t)
1179 calloc(size_t count, size_t size)
1180 {
1181 uint32_t start, end;
1182 __ptr_t ptr;
1183 tm_thread *t;
1185 if (!tracing_enabled || !PR_Initialized() ||
1186 (t = tm_get_thread())->suppress_tracing != 0) {
1187 return __libc_calloc(count, size);
1188 }
1190 t->suppress_tracing++;
1191 start = PR_IntervalNow();
1192 ptr = __libc_calloc(count, size);
1193 end = PR_IntervalNow();
1194 t->suppress_tracing--;
1196 CallocCallback(ptr, count, size, start, end, t);
1198 return ptr;
1199 }
1201 NS_EXTERNAL_VIS_(__ptr_t)
1202 realloc(__ptr_t oldptr, size_t size)
1203 {
1204 uint32_t start, end;
1205 __ptr_t ptr;
1206 tm_thread *t;
1208 if (!tracing_enabled || !PR_Initialized() ||
1209 (t = tm_get_thread())->suppress_tracing != 0) {
1210 return __libc_realloc(oldptr, size);
1211 }
1213 t->suppress_tracing++;
1214 start = PR_IntervalNow();
1215 ptr = __libc_realloc(oldptr, size);
1216 end = PR_IntervalNow();
1217 t->suppress_tracing--;
1219 /* FIXME bug 392008: We could race with reallocation of oldptr. */
1220 ReallocCallback(oldptr, ptr, size, start, end, t);
1222 return ptr;
1223 }
1225 NS_EXTERNAL_VIS_(void*)
1226 valloc(size_t size)
1227 {
1228 uint32_t start, end;
1229 __ptr_t ptr;
1230 tm_thread *t;
1232 if (!tracing_enabled || !PR_Initialized() ||
1233 (t = tm_get_thread())->suppress_tracing != 0) {
1234 return __libc_valloc(size);
1235 }
1237 t->suppress_tracing++;
1238 start = PR_IntervalNow();
1239 ptr = __libc_valloc(size);
1240 end = PR_IntervalNow();
1241 t->suppress_tracing--;
1243 MallocCallback(ptr, size, start, end, t);
1245 return ptr;
1246 }
1248 NS_EXTERNAL_VIS_(void*)
1249 memalign(size_t boundary, size_t size)
1250 {
1251 uint32_t start, end;
1252 __ptr_t ptr;
1253 tm_thread *t;
1255 if (!tracing_enabled || !PR_Initialized() ||
1256 (t = tm_get_thread())->suppress_tracing != 0) {
1257 return __libc_memalign(boundary, size);
1258 }
1260 t->suppress_tracing++;
1261 start = PR_IntervalNow();
1262 ptr = __libc_memalign(boundary, size);
1263 end = PR_IntervalNow();
1264 t->suppress_tracing--;
1266 MallocCallback(ptr, size, start, end, t);
1268 return ptr;
1269 }
1271 NS_EXTERNAL_VIS_(int)
1272 posix_memalign(void **memptr, size_t alignment, size_t size)
1273 {
1274 __ptr_t ptr = memalign(alignment, size);
1275 if (!ptr)
1276 return ENOMEM;
1277 *memptr = ptr;
1278 return 0;
1279 }
1281 NS_EXTERNAL_VIS_(void)
1282 free(__ptr_t ptr)
1283 {
1284 uint32_t start, end;
1285 tm_thread *t;
1287 if (!tracing_enabled || !PR_Initialized() ||
1288 (t = tm_get_thread())->suppress_tracing != 0) {
1289 __libc_free(ptr);
1290 return;
1291 }
1293 t->suppress_tracing++;
1294 start = PR_IntervalNow();
1295 __libc_free(ptr);
1296 end = PR_IntervalNow();
1297 t->suppress_tracing--;
1299 /* FIXME bug 392008: We could race with reallocation of ptr. */
1301 FreeCallback(ptr, start, end, t);
1302 }
1304 NS_EXTERNAL_VIS_(void)
1305 cfree(void *ptr)
1306 {
1307 free(ptr);
1308 }
1310 #define StartupHooker() PR_BEGIN_MACRO PR_END_MACRO
1311 #define ShutdownHooker() PR_BEGIN_MACRO PR_END_MACRO
1313 #elif defined(XP_WIN32)
1315 /* See nsWinTraceMalloc.cpp. */
1317 #endif
1319 static const char magic[] = NS_TRACE_MALLOC_MAGIC;
1321 static void
1322 log_header(int logfd)
1323 {
1324 uint32_t ticksPerSec = PR_htonl(PR_TicksPerSecond());
1325 (void) write(logfd, magic, NS_TRACE_MALLOC_MAGIC_SIZE);
1326 (void) write(logfd, &ticksPerSec, sizeof ticksPerSec);
1327 }
1329 PR_IMPLEMENT(void)
1330 NS_TraceMallocStartup(int logfd)
1331 {
1332 const char* stack_disable_env;
1334 /* We must be running on the primordial thread. */
1335 PR_ASSERT(tracing_enabled == 0);
1336 PR_ASSERT(logfp == &default_logfile);
1337 tracing_enabled = (logfd >= 0);
1339 if (logfd >= 3)
1340 MozillaRegisterDebugFD(logfd);
1342 /* stacks are disabled if this env var is set to a non-empty value */
1343 stack_disable_env = PR_GetEnv("NS_TRACE_MALLOC_DISABLE_STACKS");
1344 stacks_enabled = !stack_disable_env || !*stack_disable_env;
1346 if (tracing_enabled) {
1347 PR_ASSERT(logfp->simsize == 0); /* didn't overflow startup buffer */
1349 /* Log everything in logfp (aka default_logfile)'s buffer to logfd. */
1350 logfp->fd = logfd;
1351 logfile_list = &default_logfile;
1352 logfp->prevp = &logfile_list;
1353 logfile_tail = &logfp->next;
1354 log_header(logfd);
1355 }
1357 RegisterTraceMallocShutdown();
1359 tmlock = PR_NewLock();
1360 (void) tm_get_thread(); /* ensure index initialization while it's easy */
1362 if (tracing_enabled)
1363 StartupHooker();
1364 }
1366 /*
1367 * Options for log files, with the log file name either as the next option
1368 * or separated by '=' (e.g. "./mozilla --trace-malloc * malloc.log" or
1369 * "./mozilla --trace-malloc=malloc.log").
1370 */
1371 static const char TMLOG_OPTION[] = "--trace-malloc";
1372 static const char SDLOG_OPTION[] = "--shutdown-leaks";
1374 #define SHOULD_PARSE_ARG(name_, log_, arg_) \
1375 (0 == strncmp(arg_, name_, sizeof(name_) - 1))
1377 #define PARSE_ARG(name_, log_, argv_, i_, consumed_) \
1378 PR_BEGIN_MACRO \
1379 char _nextchar = argv_[i_][sizeof(name_) - 1]; \
1380 if (_nextchar == '=') { \
1381 log_ = argv_[i_] + sizeof(name_); \
1382 consumed_ = 1; \
1383 } else if (_nextchar == '\0') { \
1384 log_ = argv_[i_+1]; \
1385 consumed_ = 2; \
1386 } \
1387 PR_END_MACRO
1389 PR_IMPLEMENT(int)
1390 NS_TraceMallocStartupArgs(int argc, char **argv)
1391 {
1392 int i, logfd = -1, consumed, logflags;
1393 char *tmlogname = NULL, *sdlogname_local = NULL;
1395 /*
1396 * Look for the --trace-malloc <logfile> option early, to avoid missing
1397 * early mallocs (we miss static constructors whose output overflows the
1398 * log file's static 16K output buffer).
1399 */
1400 for (i = 1; i < argc; i += consumed) {
1401 consumed = 0;
1402 if (SHOULD_PARSE_ARG(TMLOG_OPTION, tmlogname, argv[i]))
1403 PARSE_ARG(TMLOG_OPTION, tmlogname, argv, i, consumed);
1404 else if (SHOULD_PARSE_ARG(SDLOG_OPTION, sdlogname_local, argv[i]))
1405 PARSE_ARG(SDLOG_OPTION, sdlogname_local, argv, i, consumed);
1407 if (consumed) {
1408 #ifndef XP_WIN32 /* If we don't comment this out, it will crash Windows. */
1409 int j;
1410 /* Now remove --trace-malloc and its argument from argv. */
1411 argc -= consumed;
1412 for (j = i; j < argc; ++j)
1413 argv[j] = argv[j+consumed];
1414 argv[argc] = NULL;
1415 consumed = 0; /* don't advance next iteration */
1416 #endif
1417 } else {
1418 consumed = 1;
1419 }
1420 }
1422 if (tmlogname) {
1423 #ifdef XP_UNIX
1424 int pipefds[2];
1425 #endif
1427 switch (*tmlogname) {
1428 #ifdef XP_UNIX
1429 case '|':
1430 if (pipe(pipefds) == 0) {
1431 pid_t pid = fork();
1432 if (pid == 0) {
1433 /* In child: set up stdin, parse args, and exec. */
1434 int maxargc, nargc;
1435 char **nargv, *token;
1437 if (pipefds[0] != 0) {
1438 dup2(pipefds[0], 0);
1439 close(pipefds[0]);
1440 }
1441 close(pipefds[1]);
1443 tmlogname = strtok(tmlogname + 1, " \t");
1444 maxargc = 3;
1445 nargv = (char **) malloc((maxargc+1) * sizeof(char *));
1446 if (!nargv) exit(1);
1447 nargc = 0;
1448 nargv[nargc++] = tmlogname;
1449 while ((token = strtok(NULL, " \t")) != NULL) {
1450 if (nargc == maxargc) {
1451 maxargc *= 2;
1452 nargv = (char**)
1453 realloc(nargv, (maxargc+1) * sizeof(char*));
1454 if (!nargv) exit(1);
1455 }
1456 nargv[nargc++] = token;
1457 }
1458 nargv[nargc] = NULL;
1460 (void) setsid();
1461 execvp(tmlogname, nargv);
1462 exit(127);
1463 }
1465 if (pid > 0) {
1466 /* In parent: set logfd to the pipe's write side. */
1467 close(pipefds[0]);
1468 logfd = pipefds[1];
1469 }
1470 }
1471 if (logfd < 0) {
1472 fprintf(stderr,
1473 "%s: can't pipe to trace-malloc child process %s: %s\n",
1474 argv[0], tmlogname, strerror(errno));
1475 exit(1);
1476 }
1477 break;
1478 #endif /*XP_UNIX*/
1479 case '-':
1480 /* Don't log from startup, but do prepare to log later. */
1481 /* XXX traditional meaning of '-' as option argument is "stdin" or "stdout" */
1482 if (tmlogname[1] == '\0')
1483 break;
1484 /* FALL THROUGH */
1486 default:
1487 logflags = O_CREAT | O_WRONLY | O_TRUNC;
1488 #if defined(XP_WIN32)
1489 /*
1490 * Avoid translations on WIN32.
1491 */
1492 logflags |= O_BINARY;
1493 #endif
1494 logfd = open(tmlogname, logflags, 0644);
1495 if (logfd < 0) {
1496 fprintf(stderr,
1497 "%s: can't create trace-malloc log named %s: %s\n",
1498 argv[0], tmlogname, strerror(errno));
1499 exit(1);
1500 }
1501 break;
1502 }
1503 }
1505 if (sdlogname_local) {
1506 strncpy(sdlogname, sdlogname_local, sizeof(sdlogname));
1507 sdlogname[sizeof(sdlogname) - 1] = '\0';
1508 }
1510 NS_TraceMallocStartup(logfd);
1511 return argc;
1512 }
1514 PR_IMPLEMENT(PRBool)
1515 NS_TraceMallocHasStarted(void)
1516 {
1517 return tmlock ? PR_TRUE : PR_FALSE;
1518 }
1520 PR_IMPLEMENT(void)
1521 NS_TraceMallocShutdown(void)
1522 {
1523 logfile *fp;
1525 if (sdlogname[0])
1526 NS_TraceMallocDumpAllocations(sdlogname);
1528 if (tmstats.backtrace_failures) {
1529 fprintf(stderr,
1530 "TraceMalloc backtrace failures: %lu (malloc %lu dladdr %lu)\n",
1531 (unsigned long) tmstats.backtrace_failures,
1532 (unsigned long) tmstats.btmalloc_failures,
1533 (unsigned long) tmstats.dladdr_failures);
1534 }
1535 while ((fp = logfile_list) != NULL) {
1536 logfile_list = fp->next;
1537 log_tmstats(fp);
1538 flush_logfile(fp);
1539 if (fp->fd >= 0) {
1540 MozillaUnRegisterDebugFD(fp->fd);
1541 close(fp->fd);
1542 fp->fd = -1;
1543 }
1544 if (fp != &default_logfile) {
1545 if (fp == logfp)
1546 logfp = &default_logfile;
1547 free((void*) fp);
1548 }
1549 }
1550 if (tmlock) {
1551 PRLock *lock = tmlock;
1552 tmlock = NULL;
1553 PR_DestroyLock(lock);
1554 }
1555 if (tracing_enabled) {
1556 tracing_enabled = 0;
1557 ShutdownHooker();
1558 }
1559 }
1561 PR_IMPLEMENT(void)
1562 NS_TraceMallocDisable(void)
1563 {
1564 tm_thread *t = tm_get_thread();
1565 logfile *fp;
1566 uint32_t sample;
1568 /* Robustify in case of duplicate call. */
1569 PR_ASSERT(tracing_enabled);
1570 if (tracing_enabled == 0)
1571 return;
1573 TM_SUPPRESS_TRACING_AND_ENTER_LOCK(t);
1574 for (fp = logfile_list; fp; fp = fp->next)
1575 flush_logfile(fp);
1576 sample = --tracing_enabled;
1577 TM_EXIT_LOCK_AND_UNSUPPRESS_TRACING(t);
1578 if (sample == 0)
1579 ShutdownHooker();
1580 }
1582 PR_IMPLEMENT(void)
1583 NS_TraceMallocEnable(void)
1584 {
1585 tm_thread *t = tm_get_thread();
1586 uint32_t sample;
1588 TM_SUPPRESS_TRACING_AND_ENTER_LOCK(t);
1589 sample = ++tracing_enabled;
1590 TM_EXIT_LOCK_AND_UNSUPPRESS_TRACING(t);
1591 if (sample == 1)
1592 StartupHooker();
1593 }
1595 PR_IMPLEMENT(int)
1596 NS_TraceMallocChangeLogFD(int fd)
1597 {
1598 logfile *oldfp, *fp;
1599 struct stat sb;
1600 tm_thread *t = tm_get_thread();
1602 TM_SUPPRESS_TRACING_AND_ENTER_LOCK(t);
1603 oldfp = logfp;
1604 if (oldfp->fd != fd) {
1605 flush_logfile(oldfp);
1606 fp = get_logfile(fd);
1607 if (!fp) {
1608 TM_EXIT_LOCK_AND_UNSUPPRESS_TRACING(t);
1609 return -2;
1610 }
1611 if (fd >= 0 && fstat(fd, &sb) == 0 && sb.st_size == 0)
1612 log_header(fd);
1613 logfp = fp;
1614 }
1615 TM_EXIT_LOCK_AND_UNSUPPRESS_TRACING(t);
1616 return oldfp->fd;
1617 }
1619 static int
1620 lfd_clr_enumerator(PLHashEntry *he, int i, void *arg)
1621 {
1622 lfdset_entry *le = (lfdset_entry*) he;
1623 logfile *fp = (logfile*) arg;
1625 LFD_CLR(fp->lfd, &le->lfdset);
1626 return HT_ENUMERATE_NEXT;
1627 }
1629 static void
1630 lfd_clr_walk(callsite *site, logfile *fp)
1631 {
1632 callsite *kid;
1634 LFD_CLR(fp->lfd, &site->lfdset);
1635 for (kid = site->kids; kid; kid = kid->siblings)
1636 lfd_clr_walk(kid, fp);
1637 }
1639 PR_IMPLEMENT(void)
1640 NS_TraceMallocCloseLogFD(int fd)
1641 {
1642 logfile *fp;
1643 tm_thread *t = tm_get_thread();
1645 TM_SUPPRESS_TRACING_AND_ENTER_LOCK(t);
1647 fp = get_logfile(fd);
1648 if (fp) {
1649 flush_logfile(fp);
1650 if (fp == &default_logfile) {
1651 /* Leave default_logfile in logfile_list with an fd of -1. */
1652 fp->fd = -1;
1654 /* NB: we can never free lfd 0, it belongs to default_logfile. */
1655 PR_ASSERT(fp->lfd == 0);
1656 } else {
1657 /* Clear fp->lfd in all possible lfdsets. */
1658 PL_HashTableEnumerateEntries(libraries, lfd_clr_enumerator, fp);
1659 PL_HashTableEnumerateEntries(methods, lfd_clr_enumerator, fp);
1660 lfd_clr_walk(&calltree_root, fp);
1662 /* Unlink fp from logfile_list, freeing lfd for reallocation. */
1663 *fp->prevp = fp->next;
1664 if (!fp->next) {
1665 PR_ASSERT(logfile_tail == &fp->next);
1666 logfile_tail = fp->prevp;
1667 }
1669 /* Reset logfp if we must, then free fp. */
1670 if (fp == logfp)
1671 logfp = &default_logfile;
1672 free((void*) fp);
1673 }
1674 }
1676 TM_EXIT_LOCK_AND_UNSUPPRESS_TRACING(t);
1677 MozillaUnRegisterDebugFD(fd);
1678 close(fd);
1679 }
1681 PR_IMPLEMENT(void)
1682 NS_TraceMallocLogTimestamp(const char *caption)
1683 {
1684 logfile *fp;
1685 #ifdef XP_UNIX
1686 struct timeval tv;
1687 #endif
1688 #ifdef XP_WIN32
1689 struct _timeb tb;
1690 #endif
1691 tm_thread *t = tm_get_thread();
1693 TM_SUPPRESS_TRACING_AND_ENTER_LOCK(t);
1695 fp = logfp;
1696 log_byte(fp, TM_EVENT_TIMESTAMP);
1698 #ifdef XP_UNIX
1699 gettimeofday(&tv, NULL);
1700 log_uint32(fp, (uint32_t) tv.tv_sec);
1701 log_uint32(fp, (uint32_t) tv.tv_usec);
1702 #endif
1703 #ifdef XP_WIN32
1704 _ftime(&tb);
1705 log_uint32(fp, (uint32_t) tb.time);
1706 log_uint32(fp, (uint32_t) tb.millitm);
1707 #endif
1708 log_string(fp, caption);
1710 TM_EXIT_LOCK_AND_UNSUPPRESS_TRACING(t);
1711 }
1713 static void
1714 print_stack(FILE *ofp, callsite *site)
1715 {
1716 while (site) {
1717 if (site->name || site->parent) {
1718 fprintf(ofp, "%s[%s +0x%X]\n",
1719 site->name, site->library, site->offset);
1720 }
1721 site = site->parent;
1722 }
1723 }
1725 static const char *allocation_format =
1726 (sizeof(void*) == 4) ? "\t0x%08zX\n" :
1727 (sizeof(void*) == 8) ? "\t0x%016zX\n" :
1728 "UNEXPECTED sizeof(void*)";
1730 static int
1731 allocation_enumerator(PLHashEntry *he, int i, void *arg)
1732 {
1733 allocation *alloc = (allocation*) he;
1734 FILE *ofp = (FILE*) arg;
1735 callsite *site = (callsite*) he->value;
1737 size_t *p, *end;
1739 fprintf(ofp, "%p <%s> (%lu)\n",
1740 he->key,
1741 nsGetTypeName(he->key),
1742 (unsigned long) alloc->size);
1744 for (p = (size_t*) he->key,
1745 end = (size_t*) ((char*)he->key + alloc->size);
1746 p < end; ++p) {
1747 fprintf(ofp, allocation_format, *p);
1748 }
1750 print_stack(ofp, site);
1751 fputc('\n', ofp);
1752 return HT_ENUMERATE_NEXT;
1753 }
1755 PR_IMPLEMENT(void)
1756 NS_TraceStack(int skip, FILE *ofp)
1757 {
1758 callsite *site;
1759 tm_thread *t = tm_get_thread();
1760 int immediate_abort;
1762 site = backtrace(t, skip + 1, &immediate_abort);
1763 while (site) {
1764 if (site->name || site->parent) {
1765 fprintf(ofp, "%s[%s +0x%X]\n",
1766 site->name, site->library, site->offset);
1767 }
1768 site = site->parent;
1769 }
1770 }
1772 PR_IMPLEMENT(int)
1773 NS_TraceMallocDumpAllocations(const char *pathname)
1774 {
1775 FILE *ofp;
1776 int rv;
1777 int fd;
1779 tm_thread *t = tm_get_thread();
1781 TM_SUPPRESS_TRACING_AND_ENTER_LOCK(t);
1783 ofp = fopen(pathname, WRITE_FLAGS);
1784 if (ofp) {
1785 MozillaRegisterDebugFD(fileno(ofp));
1786 if (allocations) {
1787 PL_HashTableEnumerateEntries(allocations, allocation_enumerator,
1788 ofp);
1789 }
1790 rv = ferror(ofp) ? -1 : 0;
1791 MozillaUnRegisterDebugFILE(ofp);
1792 fclose(ofp);
1793 } else {
1794 rv = -1;
1795 }
1797 TM_EXIT_LOCK_AND_UNSUPPRESS_TRACING(t);
1799 return rv;
1800 }
1802 PR_IMPLEMENT(void)
1803 NS_TraceMallocFlushLogfiles(void)
1804 {
1805 logfile *fp;
1806 tm_thread *t = tm_get_thread();
1808 TM_SUPPRESS_TRACING_AND_ENTER_LOCK(t);
1810 for (fp = logfile_list; fp; fp = fp->next)
1811 flush_logfile(fp);
1813 TM_EXIT_LOCK_AND_UNSUPPRESS_TRACING(t);
1814 }
1816 PR_IMPLEMENT(void)
1817 NS_TrackAllocation(void* ptr, FILE *ofp)
1818 {
1819 allocation *alloc;
1820 tm_thread *t = tm_get_thread();
1822 fprintf(ofp, "Trying to track %p\n", (void*) ptr);
1823 setlinebuf(ofp);
1825 TM_SUPPRESS_TRACING_AND_ENTER_LOCK(t);
1826 if (get_allocations()) {
1827 alloc = (allocation*)
1828 *PL_HashTableRawLookup(allocations, hash_pointer(ptr), ptr);
1829 if (alloc) {
1830 fprintf(ofp, "Tracking %p\n", (void*) ptr);
1831 alloc->trackfp = ofp;
1832 } else {
1833 fprintf(ofp, "Not tracking %p\n", (void*) ptr);
1834 }
1835 }
1836 TM_EXIT_LOCK_AND_UNSUPPRESS_TRACING(t);
1837 }
1839 PR_IMPLEMENT(void)
1840 MallocCallback(void *ptr, size_t size, uint32_t start, uint32_t end, tm_thread *t)
1841 {
1842 callsite *site;
1843 PLHashEntry *he;
1844 allocation *alloc;
1845 int immediate_abort;
1847 if (!tracing_enabled || t->suppress_tracing != 0)
1848 return;
1850 site = backtrace(t, 2, &immediate_abort);
1851 if (immediate_abort)
1852 return;
1854 TM_SUPPRESS_TRACING_AND_ENTER_LOCK(t);
1855 tmstats.malloc_calls++;
1856 if (!ptr) {
1857 tmstats.malloc_failures++;
1858 } else {
1859 if (site) {
1860 log_event5(logfp, TM_EVENT_MALLOC,
1861 site->serial, start, end - start,
1862 (uint32_t)NS_PTR_TO_INT32(ptr), size);
1863 }
1864 if (get_allocations()) {
1865 he = PL_HashTableAdd(allocations, ptr, site);
1866 if (he) {
1867 alloc = (allocation*) he;
1868 alloc->size = size;
1869 alloc->trackfp = NULL;
1870 }
1871 }
1872 }
1873 TM_EXIT_LOCK_AND_UNSUPPRESS_TRACING(t);
1874 }
1876 PR_IMPLEMENT(void)
1877 CallocCallback(void *ptr, size_t count, size_t size, uint32_t start, uint32_t end, tm_thread *t)
1878 {
1879 callsite *site;
1880 PLHashEntry *he;
1881 allocation *alloc;
1882 int immediate_abort;
1884 if (!tracing_enabled || t->suppress_tracing != 0)
1885 return;
1887 site = backtrace(t, 2, &immediate_abort);
1888 if (immediate_abort)
1889 return;
1891 TM_SUPPRESS_TRACING_AND_ENTER_LOCK(t);
1892 tmstats.calloc_calls++;
1893 if (!ptr) {
1894 tmstats.calloc_failures++;
1895 } else {
1896 size *= count;
1897 if (site) {
1898 log_event5(logfp, TM_EVENT_CALLOC,
1899 site->serial, start, end - start,
1900 (uint32_t)NS_PTR_TO_INT32(ptr), size);
1901 }
1902 if (get_allocations()) {
1903 he = PL_HashTableAdd(allocations, ptr, site);
1904 if (he) {
1905 alloc = (allocation*) he;
1906 alloc->size = size;
1907 alloc->trackfp = NULL;
1908 }
1909 }
1910 }
1911 TM_EXIT_LOCK_AND_UNSUPPRESS_TRACING(t);
1912 }
1914 PR_IMPLEMENT(void)
1915 ReallocCallback(void * oldptr, void *ptr, size_t size,
1916 uint32_t start, uint32_t end, tm_thread *t)
1917 {
1918 callsite *oldsite, *site;
1919 size_t oldsize;
1920 PLHashNumber hash;
1921 PLHashEntry **hep, *he;
1922 allocation *alloc;
1923 FILE *trackfp = NULL;
1924 int immediate_abort;
1926 if (!tracing_enabled || t->suppress_tracing != 0)
1927 return;
1929 site = backtrace(t, 2, &immediate_abort);
1930 if (immediate_abort)
1931 return;
1933 TM_SUPPRESS_TRACING_AND_ENTER_LOCK(t);
1934 tmstats.realloc_calls++;
1935 oldsite = NULL;
1936 oldsize = 0;
1937 hep = NULL;
1938 he = NULL;
1939 if (oldptr && get_allocations()) {
1940 hash = hash_pointer(oldptr);
1941 hep = PL_HashTableRawLookup(allocations, hash, oldptr);
1942 he = *hep;
1943 if (he) {
1944 oldsite = (callsite*) he->value;
1945 alloc = (allocation*) he;
1946 oldsize = alloc->size;
1947 trackfp = alloc->trackfp;
1948 if (trackfp) {
1949 fprintf(alloc->trackfp,
1950 "\nrealloc(%p, %lu), oldsize %lu, alloc site %p\n",
1951 (void*) ptr, (unsigned long) size,
1952 (unsigned long) oldsize, (void*) oldsite);
1953 NS_TraceStack(1, trackfp);
1954 }
1955 }
1956 }
1957 if (!ptr && size) {
1958 /*
1959 * When realloc() fails, the original block is not freed or moved, so
1960 * we'll leave the allocation entry untouched.
1961 */
1962 tmstats.realloc_failures++;
1963 } else {
1964 if (site) {
1965 log_event8(logfp, TM_EVENT_REALLOC,
1966 site->serial, start, end - start,
1967 (uint32_t)NS_PTR_TO_INT32(ptr), size,
1968 oldsite ? oldsite->serial : 0,
1969 (uint32_t)NS_PTR_TO_INT32(oldptr), oldsize);
1970 }
1971 if (ptr && allocations) {
1972 if (ptr != oldptr) {
1973 /*
1974 * If we're reallocating (not allocating new space by passing
1975 * null to realloc) and realloc moved the block, free oldptr.
1976 */
1977 if (he)
1978 PL_HashTableRawRemove(allocations, hep, he);
1980 /* Record the new allocation now, setting he. */
1981 he = PL_HashTableAdd(allocations, ptr, site);
1982 } else {
1983 /*
1984 * If we haven't yet recorded an allocation (possibly due to a
1985 * temporary memory shortage), do it now.
1986 */
1987 if (!he)
1988 he = PL_HashTableAdd(allocations, ptr, site);
1989 }
1990 if (he) {
1991 alloc = (allocation*) he;
1992 alloc->size = size;
1993 alloc->trackfp = trackfp;
1994 }
1995 }
1996 }
1997 TM_EXIT_LOCK_AND_UNSUPPRESS_TRACING(t);
1998 }
2000 PR_IMPLEMENT(void)
2001 FreeCallback(void * ptr, uint32_t start, uint32_t end, tm_thread *t)
2002 {
2003 PLHashEntry **hep, *he;
2004 callsite *site;
2005 allocation *alloc;
2007 if (!tracing_enabled || t->suppress_tracing != 0)
2008 return;
2010 /*
2011 * FIXME: Perhaps we should call backtrace() so we can check for
2012 * immediate_abort. However, the only current contexts where
2013 * immediate_abort will be true do not call free(), so for now,
2014 * let's avoid the cost of backtrace(). See bug 478195.
2015 */
2016 TM_SUPPRESS_TRACING_AND_ENTER_LOCK(t);
2017 tmstats.free_calls++;
2018 if (!ptr) {
2019 tmstats.null_free_calls++;
2020 } else {
2021 if (get_allocations()) {
2022 hep = PL_HashTableRawLookup(allocations, hash_pointer(ptr), ptr);
2023 he = *hep;
2024 if (he) {
2025 site = (callsite*) he->value;
2026 if (site) {
2027 alloc = (allocation*) he;
2028 if (alloc->trackfp) {
2029 fprintf(alloc->trackfp, "\nfree(%p), alloc site %p\n",
2030 (void*) ptr, (void*) site);
2031 NS_TraceStack(1, alloc->trackfp);
2032 }
2033 log_event5(logfp, TM_EVENT_FREE,
2034 site->serial, start, end - start,
2035 (uint32_t)NS_PTR_TO_INT32(ptr), alloc->size);
2036 }
2037 PL_HashTableRawRemove(allocations, hep, he);
2038 }
2039 }
2040 }
2041 TM_EXIT_LOCK_AND_UNSUPPRESS_TRACING(t);
2042 }
2044 PR_IMPLEMENT(nsTMStackTraceID)
2045 NS_TraceMallocGetStackTrace(void)
2046 {
2047 callsite *site;
2048 int dummy;
2049 tm_thread *t = tm_get_thread();
2051 PR_ASSERT(t->suppress_tracing == 0);
2053 site = backtrace(t, 2, &dummy);
2054 return (nsTMStackTraceID) site;
2055 }
2057 PR_IMPLEMENT(void)
2058 NS_TraceMallocPrintStackTrace(FILE *ofp, nsTMStackTraceID id)
2059 {
2060 print_stack(ofp, (callsite *)id);
2061 }
2063 #endif /* NS_TRACE_MALLOC */
