agent/src/os/linux/ps_core.c

changeset 435
a61af66fc99e
child 485
485d403e94e1
equal deleted inserted replaced
-1:000000000000 435:a61af66fc99e
1 /*
2 * Copyright 2003-2007 Sun Microsystems, Inc. All Rights Reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
20 * CA 95054 USA or visit www.sun.com if you need additional information or
21 * have any questions.
22 *
23 */
24
25 #include <jni.h>
26 #include <unistd.h>
27 #include <fcntl.h>
28 #include <string.h>
29 #include <stdlib.h>
30 #include <stddef.h>
31 #include <elf.h>
32 #include <link.h>
33 #include "libproc_impl.h"
34 #include "salibelf.h"
35
36 // This file has the libproc implementation to read core files.
37 // For live processes, refer to ps_proc.c. Portions of this is adapted
38 // /modelled after Solaris libproc.so (in particular Pcore.c)
39
40 //----------------------------------------------------------------------
41 // ps_prochandle cleanup helper functions
42
43 // close all file descriptors
44 static void close_elf_files(struct ps_prochandle* ph) {
45 lib_info* lib = NULL;
46
47 // close core file descriptor
48 if (ph->core->core_fd >= 0)
49 close(ph->core->core_fd);
50
51 // close exec file descriptor
52 if (ph->core->exec_fd >= 0)
53 close(ph->core->exec_fd);
54
55 // close interp file descriptor
56 if (ph->core->interp_fd >= 0)
57 close(ph->core->interp_fd);
58
59 // close class share archive file
60 if (ph->core->classes_jsa_fd >= 0)
61 close(ph->core->classes_jsa_fd);
62
63 // close all library file descriptors
64 lib = ph->libs;
65 while (lib) {
66 int fd = lib->fd;
67 if (fd >= 0 && fd != ph->core->exec_fd) close(fd);
68 lib = lib->next;
69 }
70 }
71
72 // clean all map_info stuff
73 static void destroy_map_info(struct ps_prochandle* ph) {
74 map_info* map = ph->core->maps;
75 while (map) {
76 map_info* next = map->next;
77 free(map);
78 map = next;
79 }
80
81 if (ph->core->map_array) {
82 free(ph->core->map_array);
83 }
84
85 // Part of the class sharing workaround
86 map = ph->core->class_share_maps;
87 while (map) {
88 map_info* next = map->next;
89 free(map);
90 map = next;
91 }
92 }
93
94 // ps_prochandle operations
95 static void core_release(struct ps_prochandle* ph) {
96 if (ph->core) {
97 close_elf_files(ph);
98 destroy_map_info(ph);
99 free(ph->core);
100 }
101 }
102
103 static map_info* allocate_init_map(int fd, off_t offset, uintptr_t vaddr, size_t memsz) {
104 map_info* map;
105 if ( (map = (map_info*) calloc(1, sizeof(map_info))) == NULL) {
106 print_debug("can't allocate memory for map_info\n");
107 return NULL;
108 }
109
110 // initialize map
111 map->fd = fd;
112 map->offset = offset;
113 map->vaddr = vaddr;
114 map->memsz = memsz;
115 return map;
116 }
117
118 // add map info with given fd, offset, vaddr and memsz
119 static map_info* add_map_info(struct ps_prochandle* ph, int fd, off_t offset,
120 uintptr_t vaddr, size_t memsz) {
121 map_info* map;
122 if ((map = allocate_init_map(fd, offset, vaddr, memsz)) == NULL) {
123 return NULL;
124 }
125
126 // add this to map list
127 map->next = ph->core->maps;
128 ph->core->maps = map;
129 ph->core->num_maps++;
130
131 return map;
132 }
133
134 // Part of the class sharing workaround
135 static map_info* add_class_share_map_info(struct ps_prochandle* ph, off_t offset,
136 uintptr_t vaddr, size_t memsz) {
137 map_info* map;
138 if ((map = allocate_init_map(ph->core->classes_jsa_fd,
139 offset, vaddr, memsz)) == NULL) {
140 return NULL;
141 }
142
143 map->next = ph->core->class_share_maps;
144 ph->core->class_share_maps = map;
145 }
146
147 // Return the map_info for the given virtual address. We keep a sorted
148 // array of pointers in ph->map_array, so we can binary search.
149 static map_info* core_lookup(struct ps_prochandle *ph, uintptr_t addr)
150 {
151 int mid, lo = 0, hi = ph->core->num_maps - 1;
152 map_info *mp;
153
154 while (hi - lo > 1) {
155 mid = (lo + hi) / 2;
156 if (addr >= ph->core->map_array[mid]->vaddr)
157 lo = mid;
158 else
159 hi = mid;
160 }
161
162 if (addr < ph->core->map_array[hi]->vaddr)
163 mp = ph->core->map_array[lo];
164 else
165 mp = ph->core->map_array[hi];
166
167 if (addr >= mp->vaddr && addr < mp->vaddr + mp->memsz)
168 return (mp);
169
170
171 // Part of the class sharing workaround
172 // Unfortunately, we have no way of detecting -Xshare state.
173 // Check out the share maps atlast, if we don't find anywhere.
174 // This is done this way so to avoid reading share pages
175 // ahead of other normal maps. For eg. with -Xshare:off we don't
176 // want to prefer class sharing data to data from core.
177 mp = ph->core->class_share_maps;
178 if (mp) {
179 print_debug("can't locate map_info at 0x%lx, trying class share maps\n",
180 addr);
181 }
182 while (mp) {
183 if (addr >= mp->vaddr && addr < mp->vaddr + mp->memsz) {
184 print_debug("located map_info at 0x%lx from class share maps\n",
185 addr);
186 return (mp);
187 }
188 mp = mp->next;
189 }
190
191 print_debug("can't locate map_info at 0x%lx\n", addr);
192 return (NULL);
193 }
194
195 //---------------------------------------------------------------
196 // Part of the class sharing workaround:
197 //
198 // With class sharing, pages are mapped from classes[_g].jsa file.
199 // The read-only class sharing pages are mapped as MAP_SHARED,
200 // PROT_READ pages. These pages are not dumped into core dump.
201 // With this workaround, these pages are read from classes[_g].jsa.
202
203 // FIXME: !HACK ALERT!
204 // The format of sharing achive file header is needed to read shared heap
205 // file mappings. For now, I am hard coding portion of FileMapHeader here.
206 // Refer to filemap.hpp.
207
208 // FileMapHeader describes the shared space data in the file to be
209 // mapped. This structure gets written to a file. It is not a class,
210 // so that the compilers don't add any compiler-private data to it.
211
212 // Refer to CompactingPermGenGen::n_regions in compactingPermGenGen.hpp
213 #define NUM_SHARED_MAPS 4
214
215 // Refer to FileMapInfo::_current_version in filemap.hpp
216 #define CURRENT_ARCHIVE_VERSION 1
217
218 struct FileMapHeader {
219 int _magic; // identify file type.
220 int _version; // (from enum, above.)
221 size_t _alignment; // how shared archive should be aligned
222
223 struct space_info {
224 int _file_offset; // sizeof(this) rounded to vm page size
225 char* _base; // copy-on-write base address
226 size_t _capacity; // for validity checking
227 size_t _used; // for setting space top on read
228
229 // 4991491 NOTICE These are C++ bool's in filemap.hpp and must match up with
230 // the C type matching the C++ bool type on any given platform. For
231 // Hotspot on Linux we assume the corresponding C type is char but
232 // licensees on Linux versions may need to adjust the type of these fields.
233 char _read_only; // read only space?
234 char _allow_exec; // executable code in space?
235
236 } _space[NUM_SHARED_MAPS]; // was _space[CompactingPermGenGen::n_regions];
237
238 // Ignore the rest of the FileMapHeader. We don't need those fields here.
239 };
240
241 static bool read_int(struct ps_prochandle* ph, uintptr_t addr, int* pvalue) {
242 int i;
243 if (ps_pdread(ph, (psaddr_t) addr, &i, sizeof(i)) == PS_OK) {
244 *pvalue = i;
245 return true;
246 } else {
247 return false;
248 }
249 }
250
251 static bool read_pointer(struct ps_prochandle* ph, uintptr_t addr, uintptr_t* pvalue) {
252 uintptr_t uip;
253 if (ps_pdread(ph, (psaddr_t) addr, &uip, sizeof(uip)) == PS_OK) {
254 *pvalue = uip;
255 return true;
256 } else {
257 return false;
258 }
259 }
260
261 // used to read strings from debuggee
262 static bool read_string(struct ps_prochandle* ph, uintptr_t addr, char* buf, size_t size) {
263 size_t i = 0;
264 char c = ' ';
265
266 while (c != '\0') {
267 if (ps_pdread(ph, (psaddr_t) addr, &c, sizeof(char)) != PS_OK)
268 return false;
269 if (i < size - 1)
270 buf[i] = c;
271 else // smaller buffer
272 return false;
273 i++; addr++;
274 }
275
276 buf[i] = '\0';
277 return true;
278 }
279
280 #define USE_SHARED_SPACES_SYM "UseSharedSpaces"
281 // mangled name of Arguments::SharedArchivePath
282 #define SHARED_ARCHIVE_PATH_SYM "_ZN9Arguments17SharedArchivePathE"
283
284 static bool init_classsharing_workaround(struct ps_prochandle* ph) {
285 lib_info* lib = ph->libs;
286 while (lib != NULL) {
287 // we are iterating over shared objects from the core dump. look for
288 // libjvm[_g].so.
289 const char *jvm_name = 0;
290 if ((jvm_name = strstr(lib->name, "/libjvm.so")) != 0 ||
291 (jvm_name = strstr(lib->name, "/libjvm_g.so")) != 0) {
292 char classes_jsa[PATH_MAX];
293 struct FileMapHeader header;
294 size_t n = 0;
295 int fd = -1, m = 0;
296 uintptr_t base = 0, useSharedSpacesAddr = 0;
297 uintptr_t sharedArchivePathAddrAddr = 0, sharedArchivePathAddr = 0;
298 int useSharedSpaces = 0;
299 map_info* mi = 0;
300
301 memset(classes_jsa, 0, sizeof(classes_jsa));
302 jvm_name = lib->name;
303 useSharedSpacesAddr = lookup_symbol(ph, jvm_name, USE_SHARED_SPACES_SYM);
304 if (useSharedSpacesAddr == 0) {
305 print_debug("can't lookup 'UseSharedSpaces' flag\n");
306 return false;
307 }
308
309 if (read_int(ph, useSharedSpacesAddr, &useSharedSpaces) != true) {
310 print_debug("can't read the value of 'UseSharedSpaces' flag\n");
311 return false;
312 }
313
314 if (useSharedSpaces == 0) {
315 print_debug("UseSharedSpaces is false, assuming -Xshare:off!\n");
316 return true;
317 }
318
319 sharedArchivePathAddrAddr = lookup_symbol(ph, jvm_name, SHARED_ARCHIVE_PATH_SYM);
320 if (sharedArchivePathAddrAddr == 0) {
321 print_debug("can't lookup shared archive path symbol\n");
322 return false;
323 }
324
325 if (read_pointer(ph, sharedArchivePathAddrAddr, &sharedArchivePathAddr) != true) {
326 print_debug("can't read shared archive path pointer\n");
327 return false;
328 }
329
330 if (read_string(ph, sharedArchivePathAddr, classes_jsa, sizeof(classes_jsa)) != true) {
331 print_debug("can't read shared archive path value\n");
332 return false;
333 }
334
335 print_debug("looking for %s\n", classes_jsa);
336 // open the class sharing archive file
337 fd = pathmap_open(classes_jsa);
338 if (fd < 0) {
339 print_debug("can't open %s!\n", classes_jsa);
340 ph->core->classes_jsa_fd = -1;
341 return false;
342 } else {
343 print_debug("opened %s\n", classes_jsa);
344 }
345
346 // read FileMapHeader from the file
347 memset(&header, 0, sizeof(struct FileMapHeader));
348 if ((n = read(fd, &header, sizeof(struct FileMapHeader)))
349 != sizeof(struct FileMapHeader)) {
350 print_debug("can't read shared archive file map header from %s\n", classes_jsa);
351 close(fd);
352 return false;
353 }
354
355 // check file magic
356 if (header._magic != 0xf00baba2) {
357 print_debug("%s has bad shared archive file magic number 0x%x, expecing 0xf00baba2\n",
358 classes_jsa, header._magic);
359 close(fd);
360 return false;
361 }
362
363 // check version
364 if (header._version != CURRENT_ARCHIVE_VERSION) {
365 print_debug("%s has wrong shared archive file version %d, expecting %d\n",
366 classes_jsa, header._version, CURRENT_ARCHIVE_VERSION);
367 close(fd);
368 return false;
369 }
370
371 ph->core->classes_jsa_fd = fd;
372 // add read-only maps from classes[_g].jsa to the list of maps
373 for (m = 0; m < NUM_SHARED_MAPS; m++) {
374 if (header._space[m]._read_only) {
375 base = (uintptr_t) header._space[m]._base;
376 // no need to worry about the fractional pages at-the-end.
377 // possible fractional pages are handled by core_read_data.
378 add_class_share_map_info(ph, (off_t) header._space[m]._file_offset,
379 base, (size_t) header._space[m]._used);
380 print_debug("added a share archive map at 0x%lx\n", base);
381 }
382 }
383 return true;
384 }
385 lib = lib->next;
386 }
387 return true;
388 }
389
390
391 //---------------------------------------------------------------------------
392 // functions to handle map_info
393
394 // Order mappings based on virtual address. We use this function as the
395 // callback for sorting the array of map_info pointers.
396 static int core_cmp_mapping(const void *lhsp, const void *rhsp)
397 {
398 const map_info *lhs = *((const map_info **)lhsp);
399 const map_info *rhs = *((const map_info **)rhsp);
400
401 if (lhs->vaddr == rhs->vaddr)
402 return (0);
403
404 return (lhs->vaddr < rhs->vaddr ? -1 : 1);
405 }
406
407 // we sort map_info by starting virtual address so that we can do
408 // binary search to read from an address.
409 static bool sort_map_array(struct ps_prochandle* ph) {
410 size_t num_maps = ph->core->num_maps;
411 map_info* map = ph->core->maps;
412 int i = 0;
413
414 // allocate map_array
415 map_info** array;
416 if ( (array = (map_info**) malloc(sizeof(map_info*) * num_maps)) == NULL) {
417 print_debug("can't allocate memory for map array\n");
418 return false;
419 }
420
421 // add maps to array
422 while (map) {
423 array[i] = map;
424 i++;
425 map = map->next;
426 }
427
428 // sort is called twice. If this is second time, clear map array
429 if (ph->core->map_array) free(ph->core->map_array);
430 ph->core->map_array = array;
431 // sort the map_info array by base virtual address.
432 qsort(ph->core->map_array, ph->core->num_maps, sizeof (map_info*),
433 core_cmp_mapping);
434
435 // print map
436 if (is_debug()) {
437 int j = 0;
438 print_debug("---- sorted virtual address map ----\n");
439 for (j = 0; j < ph->core->num_maps; j++) {
440 print_debug("base = 0x%lx\tsize = %d\n", ph->core->map_array[j]->vaddr,
441 ph->core->map_array[j]->memsz);
442 }
443 }
444
445 return true;
446 }
447
448 #ifndef MIN
449 #define MIN(x, y) (((x) < (y))? (x): (y))
450 #endif
451
452 static bool core_read_data(struct ps_prochandle* ph, uintptr_t addr, char *buf, size_t size) {
453 ssize_t resid = size;
454 int page_size=sysconf(_SC_PAGE_SIZE);
455 while (resid != 0) {
456 map_info *mp = core_lookup(ph, addr);
457 uintptr_t mapoff;
458 ssize_t len, rem;
459 off_t off;
460 int fd;
461
462 if (mp == NULL)
463 break; /* No mapping for this address */
464
465 fd = mp->fd;
466 mapoff = addr - mp->vaddr;
467 len = MIN(resid, mp->memsz - mapoff);
468 off = mp->offset + mapoff;
469
470 if ((len = pread(fd, buf, len, off)) <= 0)
471 break;
472
473 resid -= len;
474 addr += len;
475 buf = (char *)buf + len;
476
477 // mappings always start at page boundary. But, may end in fractional
478 // page. fill zeros for possible fractional page at the end of a mapping.
479 rem = mp->memsz % page_size;
480 if (rem > 0) {
481 rem = page_size - rem;
482 len = MIN(resid, rem);
483 resid -= len;
484 addr += len;
485 // we are not assuming 'buf' to be zero initialized.
486 memset(buf, 0, len);
487 buf += len;
488 }
489 }
490
491 if (resid) {
492 print_debug("core read failed for %d byte(s) @ 0x%lx (%d more bytes)\n",
493 size, addr, resid);
494 return false;
495 } else {
496 return true;
497 }
498 }
499
500 // null implementation for write
501 static bool core_write_data(struct ps_prochandle* ph,
502 uintptr_t addr, const char *buf , size_t size) {
503 return false;
504 }
505
506 static bool core_get_lwp_regs(struct ps_prochandle* ph, lwpid_t lwp_id,
507 struct user_regs_struct* regs) {
508 // for core we have cached the lwp regs from NOTE section
509 thread_info* thr = ph->threads;
510 while (thr) {
511 if (thr->lwp_id == lwp_id) {
512 memcpy(regs, &thr->regs, sizeof(struct user_regs_struct));
513 return true;
514 }
515 thr = thr->next;
516 }
517 return false;
518 }
519
520 static ps_prochandle_ops core_ops = {
521 release: core_release,
522 p_pread: core_read_data,
523 p_pwrite: core_write_data,
524 get_lwp_regs: core_get_lwp_regs
525 };
526
527 // read regs and create thread from NT_PRSTATUS entries from core file
528 static bool core_handle_prstatus(struct ps_prochandle* ph, const char* buf, size_t nbytes) {
529 // we have to read prstatus_t from buf
530 // assert(nbytes == sizeof(prstaus_t), "size mismatch on prstatus_t");
531 prstatus_t* prstat = (prstatus_t*) buf;
532 thread_info* newthr;
533 print_debug("got integer regset for lwp %d\n", prstat->pr_pid);
534 // we set pthread_t to -1 for core dump
535 if((newthr = add_thread_info(ph, (pthread_t) -1, prstat->pr_pid)) == NULL)
536 return false;
537
538 // copy regs
539 memcpy(&newthr->regs, prstat->pr_reg, sizeof(struct user_regs_struct));
540
541 if (is_debug()) {
542 print_debug("integer regset\n");
543 #ifdef i386
544 // print the regset
545 print_debug("\teax = 0x%x\n", newthr->regs.eax);
546 print_debug("\tebx = 0x%x\n", newthr->regs.ebx);
547 print_debug("\tecx = 0x%x\n", newthr->regs.ecx);
548 print_debug("\tedx = 0x%x\n", newthr->regs.edx);
549 print_debug("\tesp = 0x%x\n", newthr->regs.esp);
550 print_debug("\tebp = 0x%x\n", newthr->regs.ebp);
551 print_debug("\tesi = 0x%x\n", newthr->regs.esi);
552 print_debug("\tedi = 0x%x\n", newthr->regs.edi);
553 print_debug("\teip = 0x%x\n", newthr->regs.eip);
554 #endif
555
556 #if defined(amd64) || defined(x86_64)
557 // print the regset
558 print_debug("\tr15 = 0x%lx\n", newthr->regs.r15);
559 print_debug("\tr14 = 0x%lx\n", newthr->regs.r14);
560 print_debug("\tr13 = 0x%lx\n", newthr->regs.r13);
561 print_debug("\tr12 = 0x%lx\n", newthr->regs.r12);
562 print_debug("\trbp = 0x%lx\n", newthr->regs.rbp);
563 print_debug("\trbx = 0x%lx\n", newthr->regs.rbx);
564 print_debug("\tr11 = 0x%lx\n", newthr->regs.r11);
565 print_debug("\tr10 = 0x%lx\n", newthr->regs.r10);
566 print_debug("\tr9 = 0x%lx\n", newthr->regs.r9);
567 print_debug("\tr8 = 0x%lx\n", newthr->regs.r8);
568 print_debug("\trax = 0x%lx\n", newthr->regs.rax);
569 print_debug("\trcx = 0x%lx\n", newthr->regs.rcx);
570 print_debug("\trdx = 0x%lx\n", newthr->regs.rdx);
571 print_debug("\trsi = 0x%lx\n", newthr->regs.rsi);
572 print_debug("\trdi = 0x%lx\n", newthr->regs.rdi);
573 print_debug("\torig_rax = 0x%lx\n", newthr->regs.orig_rax);
574 print_debug("\trip = 0x%lx\n", newthr->regs.rip);
575 print_debug("\tcs = 0x%lx\n", newthr->regs.cs);
576 print_debug("\teflags = 0x%lx\n", newthr->regs.eflags);
577 print_debug("\trsp = 0x%lx\n", newthr->regs.rsp);
578 print_debug("\tss = 0x%lx\n", newthr->regs.ss);
579 print_debug("\tfs_base = 0x%lx\n", newthr->regs.fs_base);
580 print_debug("\tgs_base = 0x%lx\n", newthr->regs.gs_base);
581 print_debug("\tds = 0x%lx\n", newthr->regs.ds);
582 print_debug("\tes = 0x%lx\n", newthr->regs.es);
583 print_debug("\tfs = 0x%lx\n", newthr->regs.fs);
584 print_debug("\tgs = 0x%lx\n", newthr->regs.gs);
585 #endif
586 }
587
588 return true;
589 }
590
591 #define ROUNDUP(x, y) ((((x)+((y)-1))/(y))*(y))
592
593 // read NT_PRSTATUS entries from core NOTE segment
594 static bool core_handle_note(struct ps_prochandle* ph, ELF_PHDR* note_phdr) {
595 char* buf = NULL;
596 char* p = NULL;
597 size_t size = note_phdr->p_filesz;
598
599 // we are interested in just prstatus entries. we will ignore the rest.
600 // Advance the seek pointer to the start of the PT_NOTE data
601 if (lseek(ph->core->core_fd, note_phdr->p_offset, SEEK_SET) == (off_t)-1) {
602 print_debug("failed to lseek to PT_NOTE data\n");
603 return false;
604 }
605
606 // Now process the PT_NOTE structures. Each one is preceded by
607 // an Elf{32/64}_Nhdr structure describing its type and size.
608 if ( (buf = (char*) malloc(size)) == NULL) {
609 print_debug("can't allocate memory for reading core notes\n");
610 goto err;
611 }
612
613 // read notes into buffer
614 if (read(ph->core->core_fd, buf, size) != size) {
615 print_debug("failed to read notes, core file must have been truncated\n");
616 goto err;
617 }
618
619 p = buf;
620 while (p < buf + size) {
621 ELF_NHDR* notep = (ELF_NHDR*) p;
622 char* descdata = p + sizeof(ELF_NHDR) + ROUNDUP(notep->n_namesz, 4);
623 print_debug("Note header with n_type = %d and n_descsz = %u\n",
624 notep->n_type, notep->n_descsz);
625
626 if (notep->n_type == NT_PRSTATUS) {
627 if (core_handle_prstatus(ph, descdata, notep->n_descsz) != true)
628 return false;
629 }
630 p = descdata + ROUNDUP(notep->n_descsz, 4);
631 }
632
633 free(buf);
634 return true;
635
636 err:
637 if (buf) free(buf);
638 return false;
639 }
640
641 // read all segments from core file
642 static bool read_core_segments(struct ps_prochandle* ph, ELF_EHDR* core_ehdr) {
643 int i = 0;
644 ELF_PHDR* phbuf = NULL;
645 ELF_PHDR* core_php = NULL;
646
647 if ((phbuf = read_program_header_table(ph->core->core_fd, core_ehdr)) == NULL)
648 return false;
649
650 /*
651 * Now iterate through the program headers in the core file.
652 * We're interested in two types of Phdrs: PT_NOTE (which
653 * contains a set of saved /proc structures), and PT_LOAD (which
654 * represents a memory mapping from the process's address space).
655 *
656 * Difference b/w Solaris PT_NOTE and Linux PT_NOTE:
657 *
658 * In Solaris there are two PT_NOTE segments the first PT_NOTE (if present)
659 * contains /proc structs in the pre-2.6 unstructured /proc format. the last
660 * PT_NOTE has data in new /proc format.
661 *
662 * In Solaris, there is only one pstatus (process status). pstatus contains
663 * integer register set among other stuff. For each LWP, we have one lwpstatus
664 * entry that has integer regset for that LWP.
665 *
666 * Linux threads are actually 'clone'd processes. To support core analysis
667 * of "multithreaded" process, Linux creates more than one pstatus (called
668 * "prstatus") entry in PT_NOTE. Each prstatus entry has integer regset for one
669 * "thread". Please refer to Linux kernel src file 'fs/binfmt_elf.c', in particular
670 * function "elf_core_dump".
671 */
672
673 for (core_php = phbuf, i = 0; i < core_ehdr->e_phnum; i++) {
674 switch (core_php->p_type) {
675 case PT_NOTE:
676 if (core_handle_note(ph, core_php) != true) goto err;
677 break;
678
679 case PT_LOAD: {
680 if (core_php->p_filesz != 0) {
681 if (add_map_info(ph, ph->core->core_fd, core_php->p_offset,
682 core_php->p_vaddr, core_php->p_filesz) == NULL) goto err;
683 }
684 break;
685 }
686 }
687
688 core_php++;
689 }
690
691 free(phbuf);
692 return true;
693 err:
694 free(phbuf);
695 return false;
696 }
697
698 // read segments of a shared object
699 static bool read_lib_segments(struct ps_prochandle* ph, int lib_fd, ELF_EHDR* lib_ehdr, uintptr_t lib_base) {
700 int i = 0;
701 ELF_PHDR* phbuf;
702 ELF_PHDR* lib_php = NULL;
703
704 if ((phbuf = read_program_header_table(lib_fd, lib_ehdr)) == NULL)
705 return false;
706
707 // we want to process only PT_LOAD segments that are not writable.
708 // i.e., text segments. The read/write/exec (data) segments would
709 // have been already added from core file segments.
710 for (lib_php = phbuf, i = 0; i < lib_ehdr->e_phnum; i++) {
711 if ((lib_php->p_type == PT_LOAD) && !(lib_php->p_flags & PF_W) && (lib_php->p_filesz != 0)) {
712 if (add_map_info(ph, lib_fd, lib_php->p_offset, lib_php->p_vaddr + lib_base, lib_php->p_filesz) == NULL)
713 goto err;
714 }
715 lib_php++;
716 }
717
718 free(phbuf);
719 return true;
720 err:
721 free(phbuf);
722 return false;
723 }
724
725 // process segments from interpreter (ld.so or ld-linux.so)
726 static bool read_interp_segments(struct ps_prochandle* ph) {
727 ELF_EHDR interp_ehdr;
728
729 if (read_elf_header(ph->core->interp_fd, &interp_ehdr) != true) {
730 print_debug("interpreter is not a valid ELF file\n");
731 return false;
732 }
733
734 if (read_lib_segments(ph, ph->core->interp_fd, &interp_ehdr, ph->core->ld_base_addr) != true) {
735 print_debug("can't read segments of interpreter\n");
736 return false;
737 }
738
739 return true;
740 }
741
742 // process segments of a a.out
743 static bool read_exec_segments(struct ps_prochandle* ph, ELF_EHDR* exec_ehdr) {
744 int i = 0;
745 ELF_PHDR* phbuf = NULL;
746 ELF_PHDR* exec_php = NULL;
747
748 if ((phbuf = read_program_header_table(ph->core->exec_fd, exec_ehdr)) == NULL)
749 return false;
750
751 for (exec_php = phbuf, i = 0; i < exec_ehdr->e_phnum; i++) {
752 switch (exec_php->p_type) {
753
754 // add mappings for PT_LOAD segments
755 case PT_LOAD: {
756 // add only non-writable segments of non-zero filesz
757 if (!(exec_php->p_flags & PF_W) && exec_php->p_filesz != 0) {
758 if (add_map_info(ph, ph->core->exec_fd, exec_php->p_offset, exec_php->p_vaddr, exec_php->p_filesz) == NULL) goto err;
759 }
760 break;
761 }
762
763 // read the interpreter and it's segments
764 case PT_INTERP: {
765 char interp_name[BUF_SIZE];
766
767 pread(ph->core->exec_fd, interp_name, MIN(exec_php->p_filesz, BUF_SIZE), exec_php->p_offset);
768 print_debug("ELF interpreter %s\n", interp_name);
769 // read interpreter segments as well
770 if ((ph->core->interp_fd = pathmap_open(interp_name)) < 0) {
771 print_debug("can't open runtime loader\n");
772 goto err;
773 }
774 break;
775 }
776
777 // from PT_DYNAMIC we want to read address of first link_map addr
778 case PT_DYNAMIC: {
779 ph->core->dynamic_addr = exec_php->p_vaddr;
780 print_debug("address of _DYNAMIC is 0x%lx\n", ph->core->dynamic_addr);
781 break;
782 }
783
784 } // switch
785 exec_php++;
786 } // for
787
788 free(phbuf);
789 return true;
790 err:
791 free(phbuf);
792 return false;
793 }
794
795
796 #define FIRST_LINK_MAP_OFFSET offsetof(struct r_debug, r_map)
797 #define LD_BASE_OFFSET offsetof(struct r_debug, r_ldbase)
798 #define LINK_MAP_ADDR_OFFSET offsetof(struct link_map, l_addr)
799 #define LINK_MAP_NAME_OFFSET offsetof(struct link_map, l_name)
800 #define LINK_MAP_NEXT_OFFSET offsetof(struct link_map, l_next)
801
802 // read shared library info from runtime linker's data structures.
803 // This work is done by librtlb_db in Solaris
804 static bool read_shared_lib_info(struct ps_prochandle* ph) {
805 uintptr_t addr = ph->core->dynamic_addr;
806 uintptr_t debug_base;
807 uintptr_t first_link_map_addr;
808 uintptr_t ld_base_addr;
809 uintptr_t link_map_addr;
810 uintptr_t lib_base_diff;
811 uintptr_t lib_base;
812 uintptr_t lib_name_addr;
813 char lib_name[BUF_SIZE];
814 ELF_DYN dyn;
815 ELF_EHDR elf_ehdr;
816 int lib_fd;
817
818 // _DYNAMIC has information of the form
819 // [tag] [data] [tag] [data] .....
820 // Both tag and data are pointer sized.
821 // We look for dynamic info with DT_DEBUG. This has shared object info.
822 // refer to struct r_debug in link.h
823
824 dyn.d_tag = DT_NULL;
825 while (dyn.d_tag != DT_DEBUG) {
826 if (ps_pdread(ph, (psaddr_t) addr, &dyn, sizeof(ELF_DYN)) != PS_OK) {
827 print_debug("can't read debug info from _DYNAMIC\n");
828 return false;
829 }
830 addr += sizeof(ELF_DYN);
831 }
832
833 // we have got Dyn entry with DT_DEBUG
834 debug_base = dyn.d_un.d_ptr;
835 // at debug_base we have struct r_debug. This has first link map in r_map field
836 if (ps_pdread(ph, (psaddr_t) debug_base + FIRST_LINK_MAP_OFFSET,
837 &first_link_map_addr, sizeof(uintptr_t)) != PS_OK) {
838 print_debug("can't read first link map address\n");
839 return false;
840 }
841
842 // read ld_base address from struct r_debug
843 if (ps_pdread(ph, (psaddr_t) debug_base + LD_BASE_OFFSET, &ld_base_addr,
844 sizeof(uintptr_t)) != PS_OK) {
845 print_debug("can't read ld base address\n");
846 return false;
847 }
848 ph->core->ld_base_addr = ld_base_addr;
849
850 print_debug("interpreter base address is 0x%lx\n", ld_base_addr);
851
852 // now read segments from interp (i.e ld.so or ld-linux.so)
853 if (read_interp_segments(ph) != true)
854 return false;
855
856 // after adding interpreter (ld.so) mappings sort again
857 if (sort_map_array(ph) != true)
858 return false;
859
860 print_debug("first link map is at 0x%lx\n", first_link_map_addr);
861
862 link_map_addr = first_link_map_addr;
863 while (link_map_addr != 0) {
864 // read library base address of the .so. Note that even though <sys/link.h> calls
865 // link_map->l_addr as "base address", this is * not * really base virtual
866 // address of the shared object. This is actually the difference b/w the virtual
867 // address mentioned in shared object and the actual virtual base where runtime
868 // linker loaded it. We use "base diff" in read_lib_segments call below.
869
870 if (ps_pdread(ph, (psaddr_t) link_map_addr + LINK_MAP_ADDR_OFFSET,
871 &lib_base_diff, sizeof(uintptr_t)) != PS_OK) {
872 print_debug("can't read shared object base address diff\n");
873 return false;
874 }
875
876 // read address of the name
877 if (ps_pdread(ph, (psaddr_t) link_map_addr + LINK_MAP_NAME_OFFSET,
878 &lib_name_addr, sizeof(uintptr_t)) != PS_OK) {
879 print_debug("can't read address of shared object name\n");
880 return false;
881 }
882
883 // read name of the shared object
884 if (read_string(ph, (uintptr_t) lib_name_addr, lib_name, sizeof(lib_name)) != true) {
885 print_debug("can't read shared object name\n");
886 return false;
887 }
888
889 if (lib_name[0] != '\0') {
890 // ignore empty lib names
891 lib_fd = pathmap_open(lib_name);
892
893 if (lib_fd < 0) {
894 print_debug("can't open shared object %s\n", lib_name);
895 // continue with other libraries...
896 } else {
897 if (read_elf_header(lib_fd, &elf_ehdr)) {
898 lib_base = lib_base_diff + find_base_address(lib_fd, &elf_ehdr);
899 print_debug("reading library %s @ 0x%lx [ 0x%lx ]\n",
900 lib_name, lib_base, lib_base_diff);
901 // while adding library mappings we need to use "base difference".
902 if (! read_lib_segments(ph, lib_fd, &elf_ehdr, lib_base_diff)) {
903 print_debug("can't read shared object's segments\n");
904 close(lib_fd);
905 return false;
906 }
907 add_lib_info_fd(ph, lib_name, lib_fd, lib_base);
908 // Map info is added for the library (lib_name) so
909 // we need to re-sort it before calling the p_pdread.
910 if (sort_map_array(ph) != true)
911 return false;
912 } else {
913 print_debug("can't read ELF header for shared object %s\n", lib_name);
914 close(lib_fd);
915 // continue with other libraries...
916 }
917 }
918 }
919
920 // read next link_map address
921 if (ps_pdread(ph, (psaddr_t) link_map_addr + LINK_MAP_NEXT_OFFSET,
922 &link_map_addr, sizeof(uintptr_t)) != PS_OK) {
923 print_debug("can't read next link in link_map\n");
924 return false;
925 }
926 }
927
928 return true;
929 }
930
931 // the one and only one exposed stuff from this file
932 struct ps_prochandle* Pgrab_core(const char* exec_file, const char* core_file) {
933 ELF_EHDR core_ehdr;
934 ELF_EHDR exec_ehdr;
935 ELF_EHDR lib_ehdr;
936
937 struct ps_prochandle* ph = (struct ps_prochandle*) calloc(1, sizeof(struct ps_prochandle));
938 if (ph == NULL) {
939 print_debug("can't allocate ps_prochandle\n");
940 return NULL;
941 }
942
943 if ((ph->core = (struct core_data*) calloc(1, sizeof(struct core_data))) == NULL) {
944 free(ph);
945 print_debug("can't allocate ps_prochandle\n");
946 return NULL;
947 }
948
949 // initialize ph
950 ph->ops = &core_ops;
951 ph->core->core_fd = -1;
952 ph->core->exec_fd = -1;
953 ph->core->interp_fd = -1;
954
955 // open the core file
956 if ((ph->core->core_fd = open(core_file, O_RDONLY)) < 0) {
957 print_debug("can't open core file\n");
958 goto err;
959 }
960
961 // read core file ELF header
962 if (read_elf_header(ph->core->core_fd, &core_ehdr) != true || core_ehdr.e_type != ET_CORE) {
963 print_debug("core file is not a valid ELF ET_CORE file\n");
964 goto err;
965 }
966
967 if ((ph->core->exec_fd = open(exec_file, O_RDONLY)) < 0) {
968 print_debug("can't open executable file\n");
969 goto err;
970 }
971
972 if (read_elf_header(ph->core->exec_fd, &exec_ehdr) != true || exec_ehdr.e_type != ET_EXEC) {
973 print_debug("executable file is not a valid ELF ET_EXEC file\n");
974 goto err;
975 }
976
977 // process core file segments
978 if (read_core_segments(ph, &core_ehdr) != true)
979 goto err;
980
981 // process exec file segments
982 if (read_exec_segments(ph, &exec_ehdr) != true)
983 goto err;
984
985 // exec file is also treated like a shared object for symbol search
986 if (add_lib_info_fd(ph, exec_file, ph->core->exec_fd,
987 (uintptr_t)0 + find_base_address(ph->core->exec_fd, &exec_ehdr)) == NULL)
988 goto err;
989
990 // allocate and sort maps into map_array, we need to do this
991 // here because read_shared_lib_info needs to read from debuggee
992 // address space
993 if (sort_map_array(ph) != true)
994 goto err;
995
996 if (read_shared_lib_info(ph) != true)
997 goto err;
998
999 // sort again because we have added more mappings from shared objects
1000 if (sort_map_array(ph) != true)
1001 goto err;
1002
1003 if (init_classsharing_workaround(ph) != true)
1004 goto err;
1005
1006 return ph;
1007
1008 err:
1009 Prelease(ph);
1010 return NULL;
1011 }

mercurial