1 /*
2 * Copyright (c) 2003, 2013, Oracle and/or its affiliates. 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 Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * 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_jboolean(struct ps_prochandle* ph, uintptr_t addr, jboolean* pvalue) {
242 jboolean 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 jboolean 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 // Hotspot vm types are not exported to build this library. So
310 // using equivalent type jboolean to read the value of
311 // UseSharedSpaces which is same as hotspot type "bool".
312 if (read_jboolean(ph, useSharedSpacesAddr, &useSharedSpaces) != true) {
313 print_debug("can't read the value of 'UseSharedSpaces' flag\n");
314 return false;
315 }
316
317 if ((int)useSharedSpaces == 0) {
318 print_debug("UseSharedSpaces is false, assuming -Xshare:off!\n");
319 return true;
320 }
321
322 sharedArchivePathAddrAddr = lookup_symbol(ph, jvm_name, SHARED_ARCHIVE_PATH_SYM);
323 if (sharedArchivePathAddrAddr == 0) {
324 print_debug("can't lookup shared archive path symbol\n");
325 return false;
326 }
327
328 if (read_pointer(ph, sharedArchivePathAddrAddr, &sharedArchivePathAddr) != true) {
329 print_debug("can't read shared archive path pointer\n");
330 return false;
331 }
332
333 if (read_string(ph, sharedArchivePathAddr, classes_jsa, sizeof(classes_jsa)) != true) {
334 print_debug("can't read shared archive path value\n");
335 return false;
336 }
337
338 print_debug("looking for %s\n", classes_jsa);
339 // open the class sharing archive file
340 fd = pathmap_open(classes_jsa);
341 if (fd < 0) {
342 print_debug("can't open %s!\n", classes_jsa);
343 ph->core->classes_jsa_fd = -1;
344 return false;
345 } else {
346 print_debug("opened %s\n", classes_jsa);
347 }
348
349 // read FileMapHeader from the file
350 memset(&header, 0, sizeof(struct FileMapHeader));
351 if ((n = read(fd, &header, sizeof(struct FileMapHeader)))
352 != sizeof(struct FileMapHeader)) {
353 print_debug("can't read shared archive file map header from %s\n", classes_jsa);
354 close(fd);
355 return false;
356 }
357
358 // check file magic
359 if (header._magic != 0xf00baba2) {
360 print_debug("%s has bad shared archive file magic number 0x%x, expecing 0xf00baba2\n",
361 classes_jsa, header._magic);
362 close(fd);
363 return false;
364 }
365
366 // check version
367 if (header._version != CURRENT_ARCHIVE_VERSION) {
368 print_debug("%s has wrong shared archive file version %d, expecting %d\n",
369 classes_jsa, header._version, CURRENT_ARCHIVE_VERSION);
370 close(fd);
371 return false;
372 }
373
374 ph->core->classes_jsa_fd = fd;
375 // add read-only maps from classes[_g].jsa to the list of maps
376 for (m = 0; m < NUM_SHARED_MAPS; m++) {
377 if (header._space[m]._read_only) {
378 base = (uintptr_t) header._space[m]._base;
379 // no need to worry about the fractional pages at-the-end.
380 // possible fractional pages are handled by core_read_data.
381 add_class_share_map_info(ph, (off_t) header._space[m]._file_offset,
382 base, (size_t) header._space[m]._used);
383 print_debug("added a share archive map at 0x%lx\n", base);
384 }
385 }
386 return true;
387 }
388 lib = lib->next;
389 }
390 return true;
391 }
392
393
394 //---------------------------------------------------------------------------
395 // functions to handle map_info
396
397 // Order mappings based on virtual address. We use this function as the
398 // callback for sorting the array of map_info pointers.
399 static int core_cmp_mapping(const void *lhsp, const void *rhsp)
400 {
401 const map_info *lhs = *((const map_info **)lhsp);
402 const map_info *rhs = *((const map_info **)rhsp);
403
404 if (lhs->vaddr == rhs->vaddr)
405 return (0);
406
407 return (lhs->vaddr < rhs->vaddr ? -1 : 1);
408 }
409
410 // we sort map_info by starting virtual address so that we can do
411 // binary search to read from an address.
412 static bool sort_map_array(struct ps_prochandle* ph) {
413 size_t num_maps = ph->core->num_maps;
414 map_info* map = ph->core->maps;
415 int i = 0;
416
417 // allocate map_array
418 map_info** array;
419 if ( (array = (map_info**) malloc(sizeof(map_info*) * num_maps)) == NULL) {
420 print_debug("can't allocate memory for map array\n");
421 return false;
422 }
423
424 // add maps to array
425 while (map) {
426 array[i] = map;
427 i++;
428 map = map->next;
429 }
430
431 // sort is called twice. If this is second time, clear map array
432 if (ph->core->map_array) free(ph->core->map_array);
433 ph->core->map_array = array;
434 // sort the map_info array by base virtual address.
435 qsort(ph->core->map_array, ph->core->num_maps, sizeof (map_info*),
436 core_cmp_mapping);
437
438 // print map
439 if (is_debug()) {
440 int j = 0;
441 print_debug("---- sorted virtual address map ----\n");
442 for (j = 0; j < ph->core->num_maps; j++) {
443 print_debug("base = 0x%lx\tsize = %zu\n", ph->core->map_array[j]->vaddr,
444 ph->core->map_array[j]->memsz);
445 }
446 }
447
448 return true;
449 }
450
451 #ifndef MIN
452 #define MIN(x, y) (((x) < (y))? (x): (y))
453 #endif
454
455 static bool core_read_data(struct ps_prochandle* ph, uintptr_t addr, char *buf, size_t size) {
456 ssize_t resid = size;
457 int page_size=sysconf(_SC_PAGE_SIZE);
458 while (resid != 0) {
459 map_info *mp = core_lookup(ph, addr);
460 uintptr_t mapoff;
461 ssize_t len, rem;
462 off_t off;
463 int fd;
464
465 if (mp == NULL)
466 break; /* No mapping for this address */
467
468 fd = mp->fd;
469 mapoff = addr - mp->vaddr;
470 len = MIN(resid, mp->memsz - mapoff);
471 off = mp->offset + mapoff;
472
473 if ((len = pread(fd, buf, len, off)) <= 0)
474 break;
475
476 resid -= len;
477 addr += len;
478 buf = (char *)buf + len;
479
480 // mappings always start at page boundary. But, may end in fractional
481 // page. fill zeros for possible fractional page at the end of a mapping.
482 rem = mp->memsz % page_size;
483 if (rem > 0) {
484 rem = page_size - rem;
485 len = MIN(resid, rem);
486 resid -= len;
487 addr += len;
488 // we are not assuming 'buf' to be zero initialized.
489 memset(buf, 0, len);
490 buf += len;
491 }
492 }
493
494 if (resid) {
495 print_debug("core read failed for %d byte(s) @ 0x%lx (%d more bytes)\n",
496 size, addr, resid);
497 return false;
498 } else {
499 return true;
500 }
501 }
502
503 // null implementation for write
504 static bool core_write_data(struct ps_prochandle* ph,
505 uintptr_t addr, const char *buf , size_t size) {
506 return false;
507 }
508
509 static bool core_get_lwp_regs(struct ps_prochandle* ph, lwpid_t lwp_id,
510 struct user_regs_struct* regs) {
511 // for core we have cached the lwp regs from NOTE section
512 thread_info* thr = ph->threads;
513 while (thr) {
514 if (thr->lwp_id == lwp_id) {
515 memcpy(regs, &thr->regs, sizeof(struct user_regs_struct));
516 return true;
517 }
518 thr = thr->next;
519 }
520 return false;
521 }
522
523 static ps_prochandle_ops core_ops = {
524 .release= core_release,
525 .p_pread= core_read_data,
526 .p_pwrite= core_write_data,
527 .get_lwp_regs= core_get_lwp_regs
528 };
529
530 // read regs and create thread from NT_PRSTATUS entries from core file
531 static bool core_handle_prstatus(struct ps_prochandle* ph, const char* buf, size_t nbytes) {
532 // we have to read prstatus_t from buf
533 // assert(nbytes == sizeof(prstaus_t), "size mismatch on prstatus_t");
534 prstatus_t* prstat = (prstatus_t*) buf;
535 thread_info* newthr;
536 print_debug("got integer regset for lwp %d\n", prstat->pr_pid);
537 // we set pthread_t to -1 for core dump
538 if((newthr = add_thread_info(ph, (pthread_t) -1, prstat->pr_pid)) == NULL)
539 return false;
540
541 // copy regs
542 memcpy(&newthr->regs, prstat->pr_reg, sizeof(struct user_regs_struct));
543
544 if (is_debug()) {
545 print_debug("integer regset\n");
546 #ifdef i386
547 // print the regset
548 print_debug("\teax = 0x%x\n", newthr->regs.eax);
549 print_debug("\tebx = 0x%x\n", newthr->regs.ebx);
550 print_debug("\tecx = 0x%x\n", newthr->regs.ecx);
551 print_debug("\tedx = 0x%x\n", newthr->regs.edx);
552 print_debug("\tesp = 0x%x\n", newthr->regs.esp);
553 print_debug("\tebp = 0x%x\n", newthr->regs.ebp);
554 print_debug("\tesi = 0x%x\n", newthr->regs.esi);
555 print_debug("\tedi = 0x%x\n", newthr->regs.edi);
556 print_debug("\teip = 0x%x\n", newthr->regs.eip);
557 #endif
558
559 #if defined(amd64) || defined(x86_64)
560 // print the regset
561 print_debug("\tr15 = 0x%lx\n", newthr->regs.r15);
562 print_debug("\tr14 = 0x%lx\n", newthr->regs.r14);
563 print_debug("\tr13 = 0x%lx\n", newthr->regs.r13);
564 print_debug("\tr12 = 0x%lx\n", newthr->regs.r12);
565 print_debug("\trbp = 0x%lx\n", newthr->regs.rbp);
566 print_debug("\trbx = 0x%lx\n", newthr->regs.rbx);
567 print_debug("\tr11 = 0x%lx\n", newthr->regs.r11);
568 print_debug("\tr10 = 0x%lx\n", newthr->regs.r10);
569 print_debug("\tr9 = 0x%lx\n", newthr->regs.r9);
570 print_debug("\tr8 = 0x%lx\n", newthr->regs.r8);
571 print_debug("\trax = 0x%lx\n", newthr->regs.rax);
572 print_debug("\trcx = 0x%lx\n", newthr->regs.rcx);
573 print_debug("\trdx = 0x%lx\n", newthr->regs.rdx);
574 print_debug("\trsi = 0x%lx\n", newthr->regs.rsi);
575 print_debug("\trdi = 0x%lx\n", newthr->regs.rdi);
576 print_debug("\torig_rax = 0x%lx\n", newthr->regs.orig_rax);
577 print_debug("\trip = 0x%lx\n", newthr->regs.rip);
578 print_debug("\tcs = 0x%lx\n", newthr->regs.cs);
579 print_debug("\teflags = 0x%lx\n", newthr->regs.eflags);
580 print_debug("\trsp = 0x%lx\n", newthr->regs.rsp);
581 print_debug("\tss = 0x%lx\n", newthr->regs.ss);
582 print_debug("\tfs_base = 0x%lx\n", newthr->regs.fs_base);
583 print_debug("\tgs_base = 0x%lx\n", newthr->regs.gs_base);
584 print_debug("\tds = 0x%lx\n", newthr->regs.ds);
585 print_debug("\tes = 0x%lx\n", newthr->regs.es);
586 print_debug("\tfs = 0x%lx\n", newthr->regs.fs);
587 print_debug("\tgs = 0x%lx\n", newthr->regs.gs);
588 #endif
589 }
590
591 return true;
592 }
593
594 #define ROUNDUP(x, y) ((((x)+((y)-1))/(y))*(y))
595
596 // read NT_PRSTATUS entries from core NOTE segment
597 static bool core_handle_note(struct ps_prochandle* ph, ELF_PHDR* note_phdr) {
598 char* buf = NULL;
599 char* p = NULL;
600 size_t size = note_phdr->p_filesz;
601
602 // we are interested in just prstatus entries. we will ignore the rest.
603 // Advance the seek pointer to the start of the PT_NOTE data
604 if (lseek(ph->core->core_fd, note_phdr->p_offset, SEEK_SET) == (off_t)-1) {
605 print_debug("failed to lseek to PT_NOTE data\n");
606 return false;
607 }
608
609 // Now process the PT_NOTE structures. Each one is preceded by
610 // an Elf{32/64}_Nhdr structure describing its type and size.
611 if ( (buf = (char*) malloc(size)) == NULL) {
612 print_debug("can't allocate memory for reading core notes\n");
613 goto err;
614 }
615
616 // read notes into buffer
617 if (read(ph->core->core_fd, buf, size) != size) {
618 print_debug("failed to read notes, core file must have been truncated\n");
619 goto err;
620 }
621
622 p = buf;
623 while (p < buf + size) {
624 ELF_NHDR* notep = (ELF_NHDR*) p;
625 char* descdata = p + sizeof(ELF_NHDR) + ROUNDUP(notep->n_namesz, 4);
626 print_debug("Note header with n_type = %d and n_descsz = %u\n",
627 notep->n_type, notep->n_descsz);
628
629 if (notep->n_type == NT_PRSTATUS) {
630 if (core_handle_prstatus(ph, descdata, notep->n_descsz) != true)
631 return false;
632 }
633 p = descdata + ROUNDUP(notep->n_descsz, 4);
634 }
635
636 free(buf);
637 return true;
638
639 err:
640 if (buf) free(buf);
641 return false;
642 }
643
644 // read all segments from core file
645 static bool read_core_segments(struct ps_prochandle* ph, ELF_EHDR* core_ehdr) {
646 int i = 0;
647 ELF_PHDR* phbuf = NULL;
648 ELF_PHDR* core_php = NULL;
649
650 if ((phbuf = read_program_header_table(ph->core->core_fd, core_ehdr)) == NULL)
651 return false;
652
653 /*
654 * Now iterate through the program headers in the core file.
655 * We're interested in two types of Phdrs: PT_NOTE (which
656 * contains a set of saved /proc structures), and PT_LOAD (which
657 * represents a memory mapping from the process's address space).
658 *
659 * Difference b/w Solaris PT_NOTE and Linux PT_NOTE:
660 *
661 * In Solaris there are two PT_NOTE segments the first PT_NOTE (if present)
662 * contains /proc structs in the pre-2.6 unstructured /proc format. the last
663 * PT_NOTE has data in new /proc format.
664 *
665 * In Solaris, there is only one pstatus (process status). pstatus contains
666 * integer register set among other stuff. For each LWP, we have one lwpstatus
667 * entry that has integer regset for that LWP.
668 *
669 * Linux threads are actually 'clone'd processes. To support core analysis
670 * of "multithreaded" process, Linux creates more than one pstatus (called
671 * "prstatus") entry in PT_NOTE. Each prstatus entry has integer regset for one
672 * "thread". Please refer to Linux kernel src file 'fs/binfmt_elf.c', in particular
673 * function "elf_core_dump".
674 */
675
676 for (core_php = phbuf, i = 0; i < core_ehdr->e_phnum; i++) {
677 switch (core_php->p_type) {
678 case PT_NOTE:
679 if (core_handle_note(ph, core_php) != true) goto err;
680 break;
681
682 case PT_LOAD: {
683 if (core_php->p_filesz != 0) {
684 if (add_map_info(ph, ph->core->core_fd, core_php->p_offset,
685 core_php->p_vaddr, core_php->p_filesz) == NULL) goto err;
686 }
687 break;
688 }
689 }
690
691 core_php++;
692 }
693
694 free(phbuf);
695 return true;
696 err:
697 free(phbuf);
698 return false;
699 }
700
701 // read segments of a shared object
702 static bool read_lib_segments(struct ps_prochandle* ph, int lib_fd, ELF_EHDR* lib_ehdr, uintptr_t lib_base) {
703 int i = 0;
704 ELF_PHDR* phbuf;
705 ELF_PHDR* lib_php = NULL;
706
707 int page_size=sysconf(_SC_PAGE_SIZE);
708
709 if ((phbuf = read_program_header_table(lib_fd, lib_ehdr)) == NULL)
710 return false;
711
712 // we want to process only PT_LOAD segments that are not writable.
713 // i.e., text segments. The read/write/exec (data) segments would
714 // have been already added from core file segments.
715 for (lib_php = phbuf, i = 0; i < lib_ehdr->e_phnum; i++) {
716 if ((lib_php->p_type == PT_LOAD) && !(lib_php->p_flags & PF_W) && (lib_php->p_filesz != 0)) {
717 uintptr_t target_vaddr = lib_php->p_vaddr + lib_base;
718 map_info *existing_map = core_lookup(ph, target_vaddr);
719
720 if (existing_map == NULL) {
721 if (add_map_info(ph, lib_fd, lib_php->p_offset,
722 target_vaddr, lib_php->p_filesz) == NULL) {
723 goto err;
724 }
725 } else {
726 if ((existing_map->memsz != page_size) &&
727 (existing_map->fd != lib_fd) &&
728 (existing_map->memsz != lib_php->p_filesz)) {
729
730 print_debug("address conflict @ 0x%lx (size = %ld, flags = %d\n)",
731 target_vaddr, lib_php->p_filesz, lib_php->p_flags);
732 goto err;
733 }
734
735 /* replace PT_LOAD segment with library segment */
736 print_debug("overwrote with new address mapping (memsz %ld -> %ld)\n",
737 existing_map->memsz, lib_php->p_filesz);
738
739 existing_map->fd = lib_fd;
740 existing_map->offset = lib_php->p_offset;
741 existing_map->memsz = lib_php->p_filesz;
742 }
743 }
744 lib_php++;
745 }
746
747 free(phbuf);
748 return true;
749 err:
750 free(phbuf);
751 return false;
752 }
753
754 // process segments from interpreter (ld.so or ld-linux.so)
755 static bool read_interp_segments(struct ps_prochandle* ph) {
756 ELF_EHDR interp_ehdr;
757
758 if (read_elf_header(ph->core->interp_fd, &interp_ehdr) != true) {
759 print_debug("interpreter is not a valid ELF file\n");
760 return false;
761 }
762
763 if (read_lib_segments(ph, ph->core->interp_fd, &interp_ehdr, ph->core->ld_base_addr) != true) {
764 print_debug("can't read segments of interpreter\n");
765 return false;
766 }
767
768 return true;
769 }
770
771 // process segments of a a.out
772 static bool read_exec_segments(struct ps_prochandle* ph, ELF_EHDR* exec_ehdr) {
773 int i = 0;
774 ELF_PHDR* phbuf = NULL;
775 ELF_PHDR* exec_php = NULL;
776
777 if ((phbuf = read_program_header_table(ph->core->exec_fd, exec_ehdr)) == NULL)
778 return false;
779
780 for (exec_php = phbuf, i = 0; i < exec_ehdr->e_phnum; i++) {
781 switch (exec_php->p_type) {
782
783 // add mappings for PT_LOAD segments
784 case PT_LOAD: {
785 // add only non-writable segments of non-zero filesz
786 if (!(exec_php->p_flags & PF_W) && exec_php->p_filesz != 0) {
787 if (add_map_info(ph, ph->core->exec_fd, exec_php->p_offset, exec_php->p_vaddr, exec_php->p_filesz) == NULL) goto err;
788 }
789 break;
790 }
791
792 // read the interpreter and it's segments
793 case PT_INTERP: {
794 char interp_name[BUF_SIZE];
795
796 pread(ph->core->exec_fd, interp_name, MIN(exec_php->p_filesz, BUF_SIZE), exec_php->p_offset);
797 print_debug("ELF interpreter %s\n", interp_name);
798 // read interpreter segments as well
799 if ((ph->core->interp_fd = pathmap_open(interp_name)) < 0) {
800 print_debug("can't open runtime loader\n");
801 goto err;
802 }
803 break;
804 }
805
806 // from PT_DYNAMIC we want to read address of first link_map addr
807 case PT_DYNAMIC: {
808 ph->core->dynamic_addr = exec_php->p_vaddr;
809 print_debug("address of _DYNAMIC is 0x%lx\n", ph->core->dynamic_addr);
810 break;
811 }
812
813 } // switch
814 exec_php++;
815 } // for
816
817 free(phbuf);
818 return true;
819 err:
820 free(phbuf);
821 return false;
822 }
823
824
825 #define FIRST_LINK_MAP_OFFSET offsetof(struct r_debug, r_map)
826 #define LD_BASE_OFFSET offsetof(struct r_debug, r_ldbase)
827 #define LINK_MAP_ADDR_OFFSET offsetof(struct link_map, l_addr)
828 #define LINK_MAP_NAME_OFFSET offsetof(struct link_map, l_name)
829 #define LINK_MAP_NEXT_OFFSET offsetof(struct link_map, l_next)
830
831 // read shared library info from runtime linker's data structures.
832 // This work is done by librtlb_db in Solaris
833 static bool read_shared_lib_info(struct ps_prochandle* ph) {
834 uintptr_t addr = ph->core->dynamic_addr;
835 uintptr_t debug_base;
836 uintptr_t first_link_map_addr;
837 uintptr_t ld_base_addr;
838 uintptr_t link_map_addr;
839 uintptr_t lib_base_diff;
840 uintptr_t lib_base;
841 uintptr_t lib_name_addr;
842 char lib_name[BUF_SIZE];
843 ELF_DYN dyn;
844 ELF_EHDR elf_ehdr;
845 int lib_fd;
846
847 // _DYNAMIC has information of the form
848 // [tag] [data] [tag] [data] .....
849 // Both tag and data are pointer sized.
850 // We look for dynamic info with DT_DEBUG. This has shared object info.
851 // refer to struct r_debug in link.h
852
853 dyn.d_tag = DT_NULL;
854 while (dyn.d_tag != DT_DEBUG) {
855 if (ps_pdread(ph, (psaddr_t) addr, &dyn, sizeof(ELF_DYN)) != PS_OK) {
856 print_debug("can't read debug info from _DYNAMIC\n");
857 return false;
858 }
859 addr += sizeof(ELF_DYN);
860 }
861
862 // we have got Dyn entry with DT_DEBUG
863 debug_base = dyn.d_un.d_ptr;
864 // at debug_base we have struct r_debug. This has first link map in r_map field
865 if (ps_pdread(ph, (psaddr_t) debug_base + FIRST_LINK_MAP_OFFSET,
866 &first_link_map_addr, sizeof(uintptr_t)) != PS_OK) {
867 print_debug("can't read first link map address\n");
868 return false;
869 }
870
871 // read ld_base address from struct r_debug
872 if (ps_pdread(ph, (psaddr_t) debug_base + LD_BASE_OFFSET, &ld_base_addr,
873 sizeof(uintptr_t)) != PS_OK) {
874 print_debug("can't read ld base address\n");
875 return false;
876 }
877 ph->core->ld_base_addr = ld_base_addr;
878
879 print_debug("interpreter base address is 0x%lx\n", ld_base_addr);
880
881 // now read segments from interp (i.e ld.so or ld-linux.so)
882 if (read_interp_segments(ph) != true)
883 return false;
884
885 // after adding interpreter (ld.so) mappings sort again
886 if (sort_map_array(ph) != true)
887 return false;
888
889 print_debug("first link map is at 0x%lx\n", first_link_map_addr);
890
891 link_map_addr = first_link_map_addr;
892 while (link_map_addr != 0) {
893 // read library base address of the .so. Note that even though <sys/link.h> calls
894 // link_map->l_addr as "base address", this is * not * really base virtual
895 // address of the shared object. This is actually the difference b/w the virtual
896 // address mentioned in shared object and the actual virtual base where runtime
897 // linker loaded it. We use "base diff" in read_lib_segments call below.
898
899 if (ps_pdread(ph, (psaddr_t) link_map_addr + LINK_MAP_ADDR_OFFSET,
900 &lib_base_diff, sizeof(uintptr_t)) != PS_OK) {
901 print_debug("can't read shared object base address diff\n");
902 return false;
903 }
904
905 // read address of the name
906 if (ps_pdread(ph, (psaddr_t) link_map_addr + LINK_MAP_NAME_OFFSET,
907 &lib_name_addr, sizeof(uintptr_t)) != PS_OK) {
908 print_debug("can't read address of shared object name\n");
909 return false;
910 }
911
912 // read name of the shared object
913 lib_name[0] = '\0';
914 if (lib_name_addr != 0 &&
915 read_string(ph, (uintptr_t) lib_name_addr, lib_name, sizeof(lib_name)) != true) {
916 print_debug("can't read shared object name\n");
917 // don't let failure to read the name stop opening the file. If something is really wrong
918 // it will fail later.
919 }
920
921 if (lib_name[0] != '\0') {
922 // ignore empty lib names
923 lib_fd = pathmap_open(lib_name);
924
925 if (lib_fd < 0) {
926 print_debug("can't open shared object %s\n", lib_name);
927 // continue with other libraries...
928 } else {
929 if (read_elf_header(lib_fd, &elf_ehdr)) {
930 lib_base = lib_base_diff + find_base_address(lib_fd, &elf_ehdr);
931 print_debug("reading library %s @ 0x%lx [ 0x%lx ]\n",
932 lib_name, lib_base, lib_base_diff);
933 // while adding library mappings we need to use "base difference".
934 if (! read_lib_segments(ph, lib_fd, &elf_ehdr, lib_base_diff)) {
935 print_debug("can't read shared object's segments\n");
936 close(lib_fd);
937 return false;
938 }
939 add_lib_info_fd(ph, lib_name, lib_fd, lib_base);
940 // Map info is added for the library (lib_name) so
941 // we need to re-sort it before calling the p_pdread.
942 if (sort_map_array(ph) != true)
943 return false;
944 } else {
945 print_debug("can't read ELF header for shared object %s\n", lib_name);
946 close(lib_fd);
947 // continue with other libraries...
948 }
949 }
950 }
951
952 // read next link_map address
953 if (ps_pdread(ph, (psaddr_t) link_map_addr + LINK_MAP_NEXT_OFFSET,
954 &link_map_addr, sizeof(uintptr_t)) != PS_OK) {
955 print_debug("can't read next link in link_map\n");
956 return false;
957 }
958 }
959
960 return true;
961 }
962
963 // the one and only one exposed stuff from this file
964 struct ps_prochandle* Pgrab_core(const char* exec_file, const char* core_file) {
965 ELF_EHDR core_ehdr;
966 ELF_EHDR exec_ehdr;
967 ELF_EHDR lib_ehdr;
968
969 struct ps_prochandle* ph = (struct ps_prochandle*) calloc(1, sizeof(struct ps_prochandle));
970 if (ph == NULL) {
971 print_debug("can't allocate ps_prochandle\n");
972 return NULL;
973 }
974
975 if ((ph->core = (struct core_data*) calloc(1, sizeof(struct core_data))) == NULL) {
976 free(ph);
977 print_debug("can't allocate ps_prochandle\n");
978 return NULL;
979 }
980
981 // initialize ph
982 ph->ops = &core_ops;
983 ph->core->core_fd = -1;
984 ph->core->exec_fd = -1;
985 ph->core->interp_fd = -1;
986
987 // open the core file
988 if ((ph->core->core_fd = open(core_file, O_RDONLY)) < 0) {
989 print_debug("can't open core file\n");
990 goto err;
991 }
992
993 // read core file ELF header
994 if (read_elf_header(ph->core->core_fd, &core_ehdr) != true || core_ehdr.e_type != ET_CORE) {
995 print_debug("core file is not a valid ELF ET_CORE file\n");
996 goto err;
997 }
998
999 if ((ph->core->exec_fd = open(exec_file, O_RDONLY)) < 0) {
1000 print_debug("can't open executable file\n");
1001 goto err;
1002 }
1003
1004 if (read_elf_header(ph->core->exec_fd, &exec_ehdr) != true || exec_ehdr.e_type != ET_EXEC) {
1005 print_debug("executable file is not a valid ELF ET_EXEC file\n");
1006 goto err;
1007 }
1008
1009 // process core file segments
1010 if (read_core_segments(ph, &core_ehdr) != true)
1011 goto err;
1012
1013 // process exec file segments
1014 if (read_exec_segments(ph, &exec_ehdr) != true)
1015 goto err;
1016
1017 // exec file is also treated like a shared object for symbol search
1018 if (add_lib_info_fd(ph, exec_file, ph->core->exec_fd,
1019 (uintptr_t)0 + find_base_address(ph->core->exec_fd, &exec_ehdr)) == NULL)
1020 goto err;
1021
1022 // allocate and sort maps into map_array, we need to do this
1023 // here because read_shared_lib_info needs to read from debuggee
1024 // address space
1025 if (sort_map_array(ph) != true)
1026 goto err;
1027
1028 if (read_shared_lib_info(ph) != true)
1029 goto err;
1030
1031 // sort again because we have added more mappings from shared objects
1032 if (sort_map_array(ph) != true)
1033 goto err;
1034
1035 if (init_classsharing_workaround(ph) != true)
1036 goto err;
1037
1038 return ph;
1039
1040 err:
1041 Prelease(ph);
1042 return NULL;
1043 }