1 | // ***************************************************************************** |
---|
2 | // Copyright (C) 2015, UChicago Argonne, LLC |
---|
3 | // All Rights Reserved |
---|
4 | // memlog (ANL-SF-15-081) |
---|
5 | // Hal Finkel, Argonne National Laboratory |
---|
6 | // |
---|
7 | // OPEN SOURCE LICENSE |
---|
8 | // |
---|
9 | // Under the terms of Contract No. DE-AC02-06CH11357 with UChicago Argonne, LLC, |
---|
10 | // the U.S. Government retains certain rights in this software. |
---|
11 | // |
---|
12 | // Redistribution and use in source and binary forms, with or without |
---|
13 | // modification, are permitted provided that the following conditions are met: |
---|
14 | // |
---|
15 | // 1. Redistributions of source code must retain the above copyright notice, this |
---|
16 | // list of conditions and the following disclaimer. |
---|
17 | // |
---|
18 | // 2. Redistributions in binary form must reproduce the above copyright notice, |
---|
19 | // this list of conditions and the following disclaimer in the documentation |
---|
20 | // and/or other materials provided with the distribution. |
---|
21 | // |
---|
22 | // 3. Neither the names of UChicago Argonne, LLC or the Department of Energy nor |
---|
23 | // the names of its contributors may be used to endorse or promote products |
---|
24 | // derived from this software without specific prior written permission. |
---|
25 | // |
---|
26 | // ***************************************************************************** |
---|
27 | // DISCLAIMER |
---|
28 | // |
---|
29 | // THE SOFTWARE IS SUPPLIED "AS IS" WITHOUT WARRANTY OF ANY KIND. |
---|
30 | // |
---|
31 | // NEITHER THE UNTED STATES GOVERNMENT, NOR THE UNITED STATES DEPARTMENT OF |
---|
32 | // ENERGY, NOR UCHICAGO ARGONNE, LLC, NOR ANY OF THEIR EMPLOYEES, MAKES ANY |
---|
33 | // WARRANTY, EXPRESS OR IMPLIED, OR ASSUMES ANY LEGAL LIABILITY OR RESPONSIBILITY |
---|
34 | // FOR THE ACCURACY, COMPLETENESS, OR USEFULNESS OF ANY INFORMATION, DATA, |
---|
35 | // APPARATUS, PRODUCT, OR PROCESS DISCLOSED, OR REPRESENTS THAT ITS USE WOULD NOT |
---|
36 | // INFRINGE PRIVATELY OWNED RIGHTS. |
---|
37 | // |
---|
38 | // ***************************************************************************** |
---|
39 | |
---|
40 | #ifndef _GNU_SOURCE |
---|
41 | #define _GNU_SOURCE |
---|
42 | #endif |
---|
43 | |
---|
44 | #include <cstdlib> |
---|
45 | #include <cstdio> |
---|
46 | #include <cstring> |
---|
47 | #include <cstdint> |
---|
48 | |
---|
49 | // NOTE: This source makes very minimal use of C++11 features. It can still be |
---|
50 | // compiled by g++ 4.4.7 with -std=gnu++0x. |
---|
51 | #include <unordered_map> |
---|
52 | #include <utility> |
---|
53 | |
---|
54 | #include <limits.h> |
---|
55 | #include <errno.h> |
---|
56 | #include <malloc.h> |
---|
57 | #include <execinfo.h> |
---|
58 | #include <sys/mman.h> |
---|
59 | #include <sys/syscall.h> |
---|
60 | #include <sys/time.h> |
---|
61 | #include <sys/resource.h> |
---|
62 | #include <sys/types.h> |
---|
63 | #include <sys/stat.h> |
---|
64 | #include <sys/utsname.h> |
---|
65 | #include <fcntl.h> |
---|
66 | #include <unistd.h> |
---|
67 | |
---|
68 | #include <pthread.h> |
---|
69 | #include <dlfcn.h> |
---|
70 | |
---|
71 | #ifdef __bgq__ |
---|
72 | #include <spi/include/kernel/location.h> |
---|
73 | #include <spi/include/kernel/memory.h> |
---|
74 | #endif |
---|
75 | |
---|
76 | using namespace std; |
---|
77 | |
---|
78 | // NOTE: When static linking, this depends on linker wrapping. |
---|
79 | // Add to your LDFLAGS: |
---|
80 | // -Wl,--wrap,malloc,--wrap,free,--wrap,realloc,--wrap,calloc,--wrap,memalign /path/to/memlog_s.o -lpthread -ldl |
---|
81 | |
---|
82 | static FILE *log_file = 0; |
---|
83 | static pthread_mutex_t log_mutex = PTHREAD_MUTEX_INITIALIZER; |
---|
84 | |
---|
85 | // The malloc hook might use functions that call malloc, and we need to make |
---|
86 | // sure this does not cause an infinite loop. |
---|
87 | static __thread int in_malloc = 0; |
---|
88 | static char self_path[PATH_MAX+1] = { '\0' }; |
---|
89 | |
---|
90 | #ifdef __bgq__ |
---|
91 | static int on_bgq = 0; |
---|
92 | #endif |
---|
93 | |
---|
94 | static void *initial_brk = 0; |
---|
95 | |
---|
96 | static unordered_map<void *, Dl_info> *dladdr_cache = 0; |
---|
97 | |
---|
98 | __attribute__((__constructor__)) |
---|
99 | static void record_init() { |
---|
100 | struct utsname u; |
---|
101 | uname(&u); |
---|
102 | |
---|
103 | int id = (int) getpid(); |
---|
104 | #ifdef __bgq__ |
---|
105 | // If we're really running on a BG/Q compute node, use the job rank instead |
---|
106 | // of the pid because the node name might not really be globally unique. |
---|
107 | if (!strcmp(u.sysname, "CNK") && !strcmp(u.machine, "BGQ")) { |
---|
108 | id = (int) Kernel_GetRank(); |
---|
109 | on_bgq = 1; |
---|
110 | } |
---|
111 | #endif |
---|
112 | |
---|
113 | // If we're running under a common batch system, add the job id to the output |
---|
114 | // file names (add it as a prefix so that sorting the files will sort by job |
---|
115 | // first). |
---|
116 | char *job_id = 0; |
---|
117 | const char *job_id_vars[] = |
---|
118 | { "COBALT_JOBID", "PBS_JOBID", "SLURM_JOB_ID", "JOB_ID" }; |
---|
119 | for (int i = 0; i < sizeof(job_id_vars)/sizeof(job_id_vars[0]); ++i) { |
---|
120 | job_id = getenv(job_id_vars[i]); |
---|
121 | if (job_id) |
---|
122 | break; |
---|
123 | } |
---|
124 | |
---|
125 | char log_name[PATH_MAX+1]; |
---|
126 | if (job_id) |
---|
127 | snprintf(log_name, PATH_MAX+1, "%s.%s.%d.memlog", job_id, u.nodename, id); |
---|
128 | else |
---|
129 | snprintf(log_name, PATH_MAX+1, "%s.%d.memlog", u.nodename, id); |
---|
130 | log_file = fopen(log_name, "w"); |
---|
131 | if (!log_file) |
---|
132 | fprintf(stderr, "fopen failed for '%s': %m\n", log_name); |
---|
133 | |
---|
134 | const char *link_name = "/proc/self/exe"; |
---|
135 | readlink(link_name, self_path, PATH_MAX); |
---|
136 | |
---|
137 | initial_brk = sbrk(0); |
---|
138 | } |
---|
139 | |
---|
140 | __attribute__((__destructor__)) |
---|
141 | static void record_cleanup() { |
---|
142 | if (!log_file) |
---|
143 | return; |
---|
144 | |
---|
145 | // These functions might call free, but we're shutting down, so don't try to |
---|
146 | // unwind the stack from here... |
---|
147 | in_malloc = 1; |
---|
148 | |
---|
149 | // Avoid any racing by obtaining the lock. |
---|
150 | if (pthread_mutex_lock(&log_mutex)) |
---|
151 | return; |
---|
152 | |
---|
153 | (void) fflush(log_file); |
---|
154 | (void) fclose(log_file); |
---|
155 | |
---|
156 | if (dladdr_cache) |
---|
157 | delete dladdr_cache; |
---|
158 | } |
---|
159 | |
---|
160 | // dladdr is, relatively, quit slow. For this to work on a large application, |
---|
161 | // we need to cache the lookup results. |
---|
162 | static int dladdr_cached(void * addr, Dl_info *info) { |
---|
163 | if (!dladdr_cache) |
---|
164 | dladdr_cache = new unordered_map<void *, Dl_info>; |
---|
165 | |
---|
166 | auto I = dladdr_cache->find(addr); |
---|
167 | if (I == dladdr_cache->end()) { |
---|
168 | int r; |
---|
169 | if (!(r = dladdr(addr, info))) |
---|
170 | memset(info, 0, sizeof(Dl_info)); |
---|
171 | |
---|
172 | dladdr_cache->insert(make_pair(addr, *info)); |
---|
173 | return r; |
---|
174 | } |
---|
175 | |
---|
176 | memcpy(info, &I->second, sizeof(Dl_info)); |
---|
177 | return 1; |
---|
178 | } |
---|
179 | |
---|
180 | static void print_context(const void *caller, int show_backtrace) { |
---|
181 | struct rusage usage; |
---|
182 | if (getrusage(RUSAGE_SELF, &usage)) { |
---|
183 | fprintf(stderr, "getrusage failed: %m\n"); |
---|
184 | return; |
---|
185 | } |
---|
186 | |
---|
187 | fprintf(log_file, "\t%ld.%06ld %ld %ld", usage.ru_utime.tv_sec, |
---|
188 | usage.ru_utime.tv_usec, usage.ru_maxrss, syscall(SYS_gettid)); |
---|
189 | |
---|
190 | // Some other memory stats (like with maxrss, report these in KB). |
---|
191 | size_t arena_size = ((size_t) sbrk(0)) - (size_t) initial_brk; |
---|
192 | |
---|
193 | uint64_t mmap_size = 0; |
---|
194 | #ifdef __bgq__ |
---|
195 | if (on_bgq) |
---|
196 | (void) Kernel_GetMemorySize(KERNEL_MEMSIZE_MMAP, &mmap_size); |
---|
197 | #endif |
---|
198 | |
---|
199 | fprintf(log_file, " %ld %ld", arena_size >> 10, mmap_size >> 10); |
---|
200 | |
---|
201 | if (!show_backtrace) |
---|
202 | return; |
---|
203 | |
---|
204 | void *pcs[1024]; |
---|
205 | int num_pcs = backtrace(pcs, 1024); |
---|
206 | |
---|
207 | int found_caller = 0; |
---|
208 | for (int pci = 0; pci < num_pcs; ++pci) { |
---|
209 | intptr_t pc = (intptr_t) pcs[pci]; |
---|
210 | |
---|
211 | if (!pc) |
---|
212 | break; |
---|
213 | |
---|
214 | if (!found_caller) { |
---|
215 | if (pc != (intptr_t) caller) |
---|
216 | continue; |
---|
217 | |
---|
218 | found_caller = 1; |
---|
219 | } |
---|
220 | |
---|
221 | intptr_t off, relpc; |
---|
222 | const char *proc_name; |
---|
223 | const char *file_name; |
---|
224 | Dl_info dlinfo; |
---|
225 | if (dladdr_cached((void *) pc, &dlinfo) && dlinfo.dli_fname && |
---|
226 | *dlinfo.dli_fname) { |
---|
227 | intptr_t saddr = (intptr_t) dlinfo.dli_saddr; |
---|
228 | if (saddr) { |
---|
229 | #if defined(__powerpc64__) && !defined(__powerpc64le__) |
---|
230 | // On PPC64 ELFv1, the symbol address points to the function descriptor, not |
---|
231 | // the actual starting address. |
---|
232 | saddr = *(intptr_t*) saddr; |
---|
233 | #endif |
---|
234 | |
---|
235 | off = pc - saddr; |
---|
236 | relpc = pc - ((intptr_t) dlinfo.dli_fbase); |
---|
237 | } else { |
---|
238 | off = 0; |
---|
239 | relpc = 0; |
---|
240 | } |
---|
241 | |
---|
242 | proc_name = dlinfo.dli_sname; |
---|
243 | if (!proc_name) |
---|
244 | proc_name = "?"; |
---|
245 | |
---|
246 | file_name = dlinfo.dli_fname; |
---|
247 | } else { |
---|
248 | // We don't know these... |
---|
249 | off = 0; |
---|
250 | relpc = 0; |
---|
251 | proc_name = "?"; |
---|
252 | |
---|
253 | // If we can't determine the file, assume it is the base executable |
---|
254 | // (which does the right thing for statically-linked binaries). |
---|
255 | file_name = self_path; |
---|
256 | } |
---|
257 | |
---|
258 | fprintf(log_file, "\t%s (%s+0x%x) [0x%lx (0x%lx)]", file_name, proc_name, (int) off, |
---|
259 | (long) pc, (long) relpc); |
---|
260 | } |
---|
261 | } |
---|
262 | |
---|
263 | static void record_malloc(size_t size, void *ptr, const void *caller) { |
---|
264 | if (!log_file) |
---|
265 | return; |
---|
266 | |
---|
267 | if (pthread_mutex_lock(&log_mutex)) |
---|
268 | return; |
---|
269 | |
---|
270 | fprintf(log_file, "M: %zd %p", size, ptr); |
---|
271 | print_context(caller, 1); |
---|
272 | fprintf(log_file, "\n"); |
---|
273 | |
---|
274 | done: |
---|
275 | pthread_mutex_unlock(&log_mutex); |
---|
276 | } |
---|
277 | |
---|
278 | static void record_free(void *ptr, const void *caller) { |
---|
279 | if (!log_file) |
---|
280 | return; |
---|
281 | |
---|
282 | if (pthread_mutex_lock(&log_mutex)) |
---|
283 | return; |
---|
284 | |
---|
285 | fprintf(log_file, "F: %p", ptr); |
---|
286 | print_context(caller, 0); |
---|
287 | fprintf(log_file, "\n"); |
---|
288 | |
---|
289 | done: |
---|
290 | pthread_mutex_unlock(&log_mutex); |
---|
291 | } |
---|
292 | |
---|
293 | #ifdef __PIC__ |
---|
294 | static int (*__real_posix_memalign)(void **memptr, size_t alignment, |
---|
295 | size_t size) = 0; |
---|
296 | |
---|
297 | static void *(*__real_mmap)(void *addr, size_t length, int prot, int flags, |
---|
298 | int fd, off_t offset) = 0; |
---|
299 | static void *(*__real_mmap64)(void *addr, size_t length, int prot, int flags, |
---|
300 | int fd, off64_t offset) = 0; |
---|
301 | static int (*__real_munmap)(void *addr, size_t length) = 0; |
---|
302 | #else |
---|
303 | extern "C" { |
---|
304 | extern int __real_posix_memalign(void **memptr, size_t alignment, size_t size); |
---|
305 | |
---|
306 | extern void *__real_mmap(void *addr, size_t length, int prot, int flags, |
---|
307 | int fd, off_t offset); |
---|
308 | extern void *__real_mmap64(void *addr, size_t length, int prot, int flags, |
---|
309 | int fd, off64_t offset); |
---|
310 | extern int __real_munmap(void *addr, size_t length); |
---|
311 | } |
---|
312 | #endif |
---|
313 | |
---|
314 | // glibc exports its underlying malloc implementation under the name |
---|
315 | // __libc_malloc so that hooks like this can use it. |
---|
316 | extern "C" { |
---|
317 | extern void *__libc_malloc(size_t size); |
---|
318 | extern void *__libc_valloc(size_t size); |
---|
319 | extern void *__libc_realloc(void *ptr, size_t size); |
---|
320 | extern void *__libc_calloc(size_t nmemb, size_t size); |
---|
321 | extern void *__libc_memalign(size_t boundary, size_t size); |
---|
322 | extern void __libc_free(void *ptr); |
---|
323 | |
---|
324 | #ifdef __PIC__ |
---|
325 | #define FUNC(x) x |
---|
326 | #else |
---|
327 | #define FUNC(x) __wrap_ ## x |
---|
328 | #endif |
---|
329 | |
---|
330 | void *FUNC(malloc)(size_t size) { |
---|
331 | const void *caller = |
---|
332 | __builtin_extract_return_addr(__builtin_return_address(0)); |
---|
333 | |
---|
334 | if (in_malloc) |
---|
335 | return __libc_malloc(size); |
---|
336 | |
---|
337 | in_malloc = 1; |
---|
338 | |
---|
339 | void *ptr = __libc_malloc(size); |
---|
340 | if (ptr) |
---|
341 | record_malloc(size, ptr, caller); |
---|
342 | |
---|
343 | in_malloc = 0; |
---|
344 | return ptr; |
---|
345 | } |
---|
346 | |
---|
347 | void *FUNC(valloc)(size_t size) { |
---|
348 | const void *caller = |
---|
349 | __builtin_extract_return_addr(__builtin_return_address(0)); |
---|
350 | |
---|
351 | if (in_malloc) |
---|
352 | return __libc_valloc(size); |
---|
353 | |
---|
354 | in_malloc = 1; |
---|
355 | |
---|
356 | void *ptr = __libc_valloc(size); |
---|
357 | if (ptr) |
---|
358 | record_malloc(size, ptr, caller); |
---|
359 | |
---|
360 | in_malloc = 0; |
---|
361 | return ptr; |
---|
362 | } |
---|
363 | |
---|
364 | void *FUNC(realloc)(void *ptr, size_t size) { |
---|
365 | const void *caller = |
---|
366 | __builtin_extract_return_addr(__builtin_return_address(0)); |
---|
367 | |
---|
368 | if (in_malloc) |
---|
369 | return __libc_realloc(ptr, size); |
---|
370 | |
---|
371 | in_malloc = 1; |
---|
372 | |
---|
373 | void *nptr = __libc_realloc(ptr, size); |
---|
374 | |
---|
375 | if (ptr) |
---|
376 | record_free(ptr, caller); |
---|
377 | if (nptr) |
---|
378 | record_malloc(size, nptr, caller); |
---|
379 | |
---|
380 | in_malloc = 0; |
---|
381 | |
---|
382 | return nptr; |
---|
383 | } |
---|
384 | |
---|
385 | void *FUNC(calloc)(size_t nmemb, size_t size) { |
---|
386 | const void *caller = |
---|
387 | __builtin_extract_return_addr(__builtin_return_address(0)); |
---|
388 | |
---|
389 | if (in_malloc) |
---|
390 | return __libc_calloc(nmemb, size); |
---|
391 | |
---|
392 | in_malloc = 1; |
---|
393 | |
---|
394 | void *ptr = __libc_calloc(nmemb, size); |
---|
395 | |
---|
396 | if (ptr) |
---|
397 | record_malloc(nmemb*size, ptr, caller); |
---|
398 | |
---|
399 | in_malloc = 0; |
---|
400 | |
---|
401 | return ptr; |
---|
402 | } |
---|
403 | |
---|
404 | void *FUNC(memalign)(size_t boundary, size_t size) { |
---|
405 | const void *caller = |
---|
406 | __builtin_extract_return_addr(__builtin_return_address(0)); |
---|
407 | |
---|
408 | if (in_malloc) |
---|
409 | return __libc_memalign(boundary, size); |
---|
410 | |
---|
411 | in_malloc = 1; |
---|
412 | |
---|
413 | void *ptr = __libc_memalign(boundary, size); |
---|
414 | |
---|
415 | if (ptr) |
---|
416 | record_malloc(size, ptr, caller); |
---|
417 | |
---|
418 | in_malloc = 0; |
---|
419 | |
---|
420 | return ptr; |
---|
421 | } |
---|
422 | |
---|
423 | void FUNC(free)(void *ptr) { |
---|
424 | const void *caller = |
---|
425 | __builtin_extract_return_addr(__builtin_return_address(0)); |
---|
426 | |
---|
427 | if (in_malloc || !ptr) |
---|
428 | return __libc_free(ptr); |
---|
429 | |
---|
430 | in_malloc = 1; |
---|
431 | |
---|
432 | record_free(ptr, caller); |
---|
433 | |
---|
434 | __libc_free(ptr); |
---|
435 | |
---|
436 | in_malloc = 0; |
---|
437 | } |
---|
438 | |
---|
439 | int FUNC(posix_memalign)(void **memptr, size_t alignment, size_t size) { |
---|
440 | const void *caller = |
---|
441 | __builtin_extract_return_addr(__builtin_return_address(0)); |
---|
442 | |
---|
443 | #ifdef __PIC__ |
---|
444 | if (!__real_posix_memalign) |
---|
445 | if (!(*(void **) (&__real_posix_memalign) = |
---|
446 | dlsym(RTLD_NEXT, "posix_memalign"))) { |
---|
447 | return ELIBACC; |
---|
448 | } |
---|
449 | #endif |
---|
450 | |
---|
451 | if (in_malloc) |
---|
452 | return __real_posix_memalign(memptr, alignment, size); |
---|
453 | |
---|
454 | in_malloc = 1; |
---|
455 | |
---|
456 | int r = __real_posix_memalign(memptr, alignment, size); |
---|
457 | |
---|
458 | if (!r) |
---|
459 | record_malloc(size, *memptr, caller); |
---|
460 | |
---|
461 | in_malloc = 0; |
---|
462 | |
---|
463 | return r; |
---|
464 | } |
---|
465 | |
---|
466 | void *FUNC(mmap)(void *addr, size_t length, int prot, int flags, |
---|
467 | int fd, off_t offset) { |
---|
468 | const void *caller = |
---|
469 | __builtin_extract_return_addr(__builtin_return_address(0)); |
---|
470 | |
---|
471 | #ifdef __PIC__ |
---|
472 | if (!__real_mmap) |
---|
473 | if (!(*(void **) (&__real_mmap) = dlsym(RTLD_NEXT, "mmap"))) { |
---|
474 | errno = ELIBACC; |
---|
475 | return MAP_FAILED; |
---|
476 | } |
---|
477 | #endif |
---|
478 | |
---|
479 | if (in_malloc) |
---|
480 | return __real_mmap(addr, length, prot, flags, fd, offset); |
---|
481 | |
---|
482 | in_malloc = 1; |
---|
483 | |
---|
484 | void *ptr = __real_mmap(addr, length, prot, flags, fd, offset); |
---|
485 | |
---|
486 | if (ptr != MAP_FAILED) |
---|
487 | record_malloc(length, ptr, caller); |
---|
488 | |
---|
489 | in_malloc = 0; |
---|
490 | |
---|
491 | return ptr; |
---|
492 | } |
---|
493 | |
---|
494 | void *FUNC(mmap64)(void *addr, size_t length, int prot, int flags, |
---|
495 | int fd, off64_t offset) { |
---|
496 | const void *caller = |
---|
497 | __builtin_extract_return_addr(__builtin_return_address(0)); |
---|
498 | |
---|
499 | #ifdef __PIC__ |
---|
500 | if (!__real_mmap64) |
---|
501 | if (!(*(void **) (&__real_mmap64) = dlsym(RTLD_NEXT, "mmap64"))) { |
---|
502 | errno = ELIBACC; |
---|
503 | return MAP_FAILED; |
---|
504 | } |
---|
505 | #endif |
---|
506 | |
---|
507 | if (in_malloc) |
---|
508 | return __real_mmap64(addr, length, prot, flags, fd, offset); |
---|
509 | |
---|
510 | in_malloc = 1; |
---|
511 | |
---|
512 | void *ptr = __real_mmap64(addr, length, prot, flags, fd, offset); |
---|
513 | |
---|
514 | if (ptr != MAP_FAILED) |
---|
515 | record_malloc(length, ptr, caller); |
---|
516 | |
---|
517 | in_malloc = 0; |
---|
518 | |
---|
519 | return ptr; |
---|
520 | } |
---|
521 | |
---|
522 | int FUNC(munmap)(void *addr, size_t length) { |
---|
523 | const void *caller = |
---|
524 | __builtin_extract_return_addr(__builtin_return_address(0)); |
---|
525 | |
---|
526 | #ifdef __PIC__ |
---|
527 | if (!__real_munmap) |
---|
528 | if (!(*(void **) (&__real_munmap) = dlsym(RTLD_NEXT, "munmap"))) { |
---|
529 | errno = ELIBACC; |
---|
530 | return -1; |
---|
531 | } |
---|
532 | #endif |
---|
533 | |
---|
534 | if (in_malloc) |
---|
535 | return __real_munmap(addr, length); |
---|
536 | |
---|
537 | in_malloc = 1; |
---|
538 | |
---|
539 | record_free(addr, caller); |
---|
540 | |
---|
541 | int r = __real_munmap(addr, length); |
---|
542 | |
---|
543 | in_malloc = 0; |
---|
544 | |
---|
545 | return r; |
---|
546 | } |
---|
547 | |
---|
548 | } // extern "C" |
---|
549 | |
---|