1 /*
2 * Copyright (c) 2003, 2018, 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 #include "cds.h"
36
37 // This file has the libproc implementation to read core files.
38 // For live processes, refer to ps_proc.c. Portions of this is adapted
39 // /modelled after Solaris libproc.so (in particular Pcore.c)
40
41 //----------------------------------------------------------------------
42 // ps_prochandle cleanup helper functions
43
44 // close all file descriptors
45 static void close_files(struct ps_prochandle* ph) {
46 lib_info* lib = NULL;
47
48 // close core file descriptor
49 if (ph->core->core_fd >= 0)
50 close(ph->core->core_fd);
51
52 // close exec file descriptor
53 if (ph->core->exec_fd >= 0)
54 close(ph->core->exec_fd);
55
56 // close interp file descriptor
57 if (ph->core->interp_fd >= 0)
58 close(ph->core->interp_fd);
59
60 // close class share archive file
61 if (ph->core->classes_jsa_fd >= 0)
62 close(ph->core->classes_jsa_fd);
63
64 // close all library file descriptors
65 lib = ph->libs;
66 while (lib) {
67 int fd = lib->fd;
68 if (fd >= 0 && fd != ph->core->exec_fd) {
69 close(fd);
70 }
71 lib = lib->next;
72 }
73 }
74
75 // clean all map_info stuff
76 static void destroy_map_info(struct ps_prochandle* ph) {
77 map_info* map = ph->core->maps;
78 while (map) {
79 map_info* next = map->next;
80 free(map);
81 map = next;
82 }
83
84 if (ph->core->map_array) {
85 free(ph->core->map_array);
86 }
87
88 // Part of the class sharing workaround
89 map = ph->core->class_share_maps;
90 while (map) {
91 map_info* next = map->next;
92 free(map);
93 map = next;
94 }
95 }
96
97 // ps_prochandle operations
98 static void core_release(struct ps_prochandle* ph) {
99 if (ph->core) {
100 close_files(ph);
101 destroy_map_info(ph);
102 free(ph->core);
103 }
104 }
105
106 static map_info* allocate_init_map(int fd, off_t offset, uintptr_t vaddr, size_t memsz) {
107 map_info* map;
108 if ( (map = (map_info*) calloc(1, sizeof(map_info))) == NULL) {
109 print_debug("can't allocate memory for map_info\n");
110 return NULL;
111 }
112
113 // initialize map
114 map->fd = fd;
115 map->offset = offset;
116 map->vaddr = vaddr;
117 map->memsz = memsz;
118 return map;
119 }
120
121 // add map info with given fd, offset, vaddr and memsz
122 static map_info* add_map_info(struct ps_prochandle* ph, int fd, off_t offset,
123 uintptr_t vaddr, size_t memsz) {
124 map_info* map;
125 if ((map = allocate_init_map(fd, offset, vaddr, memsz)) == NULL) {
126 return NULL;
127 }
128
129 // add this to map list
130 map->next = ph->core->maps;
131 ph->core->maps = map;
132 ph->core->num_maps++;
133
134 return map;
135 }
136
137 // Part of the class sharing workaround
138 static map_info* add_class_share_map_info(struct ps_prochandle* ph, off_t offset,
139 uintptr_t vaddr, size_t memsz) {
140 map_info* map;
141 if ((map = allocate_init_map(ph->core->classes_jsa_fd,
142 offset, vaddr, memsz)) == NULL) {
143 return NULL;
144 }
145
146 map->next = ph->core->class_share_maps;
147 ph->core->class_share_maps = map;
148 return map;
149 }
150
151 // Return the map_info for the given virtual address. We keep a sorted
152 // array of pointers in ph->map_array, so we can binary search.
153 static map_info* core_lookup(struct ps_prochandle *ph, uintptr_t addr) {
154 int mid, lo = 0, hi = ph->core->num_maps - 1;
155 map_info *mp;
156
157 while (hi - lo > 1) {
158 mid = (lo + hi) / 2;
159 if (addr >= ph->core->map_array[mid]->vaddr) {
160 lo = mid;
161 } else {
162 hi = mid;
163 }
164 }
165
166 if (addr < ph->core->map_array[hi]->vaddr) {
167 mp = ph->core->map_array[lo];
168 } else {
169 mp = ph->core->map_array[hi];
170 }
171
172 if (addr >= mp->vaddr && addr < mp->vaddr + mp->memsz) {
173 return (mp);
174 }
175
176
177 // Part of the class sharing workaround
178 // Unfortunately, we have no way of detecting -Xshare state.
179 // Check out the share maps atlast, if we don't find anywhere.
180 // This is done this way so to avoid reading share pages
181 // ahead of other normal maps. For eg. with -Xshare:off we don't
182 // want to prefer class sharing data to data from core.
183 mp = ph->core->class_share_maps;
184 if (mp) {
185 print_debug("can't locate map_info at 0x%lx, trying class share maps\n", addr);
186 }
187 while (mp) {
188 if (addr >= mp->vaddr && addr < mp->vaddr + mp->memsz) {
189 print_debug("located map_info at 0x%lx from class share maps\n", addr);
190 return (mp);
191 }
192 mp = mp->next;
193 }
194
195 print_debug("can't locate map_info at 0x%lx\n", addr);
196 return (NULL);
197 }
198
199 //---------------------------------------------------------------
200 // Part of the class sharing workaround:
201 //
202 // With class sharing, pages are mapped from classes.jsa file.
203 // The read-only class sharing pages are mapped as MAP_SHARED,
204 // PROT_READ pages. These pages are not dumped into core dump.
205 // With this workaround, these pages are read from classes.jsa.
206
207 static bool read_jboolean(struct ps_prochandle* ph, uintptr_t addr, jboolean* pvalue) {
208 jboolean i;
209 if (ps_pdread(ph, (psaddr_t) addr, &i, sizeof(i)) == PS_OK) {
210 *pvalue = i;
211 return true;
212 } else {
213 return false;
214 }
215 }
216
217 static bool read_pointer(struct ps_prochandle* ph, uintptr_t addr, uintptr_t* pvalue) {
218 uintptr_t uip;
219 if (ps_pdread(ph, (psaddr_t) addr, (char *)&uip, sizeof(uip)) == PS_OK) {
220 *pvalue = uip;
221 return true;
222 } else {
223 return false;
224 }
225 }
226
227 // used to read strings from debuggee
228 static bool read_string(struct ps_prochandle* ph, uintptr_t addr, char* buf, size_t size) {
229 size_t i = 0;
230 char c = ' ';
231
232 while (c != '\0') {
233 if (ps_pdread(ph, (psaddr_t) addr, &c, sizeof(char)) != PS_OK) {
234 return false;
235 }
236 if (i < size - 1) {
237 buf[i] = c;
238 } else {
239 // smaller buffer
240 return false;
241 }
242 i++; addr++;
243 }
244
245 buf[i] = '\0';
246 return true;
247 }
248
249 #define USE_SHARED_SPACES_SYM "UseSharedSpaces"
250 // mangled name of Arguments::SharedArchivePath
251 #define SHARED_ARCHIVE_PATH_SYM "_ZN9Arguments17SharedArchivePathE"
252 #define LIBJVM_NAME "/libjvm.so"
253
254 static bool init_classsharing_workaround(struct ps_prochandle* ph) {
255 lib_info* lib = ph->libs;
256 while (lib != NULL) {
257 // we are iterating over shared objects from the core dump. look for
258 // libjvm.so.
259 const char *jvm_name = 0;
260 if ((jvm_name = strstr(lib->name, LIBJVM_NAME)) != 0) {
261 char classes_jsa[PATH_MAX];
262 CDSFileMapHeaderBase header;
263 int fd = -1;
264 int m = 0;
265 size_t n = 0;
266 uintptr_t base = 0, useSharedSpacesAddr = 0;
267 uintptr_t sharedArchivePathAddrAddr = 0, sharedArchivePathAddr = 0;
268 jboolean useSharedSpaces = 0;
269 map_info* mi = 0;
270
271 memset(classes_jsa, 0, sizeof(classes_jsa));
272 jvm_name = lib->name;
273 useSharedSpacesAddr = lookup_symbol(ph, jvm_name, USE_SHARED_SPACES_SYM);
274 if (useSharedSpacesAddr == 0) {
275 print_debug("can't lookup 'UseSharedSpaces' flag\n");
276 return false;
277 }
278
279 // Hotspot vm types are not exported to build this library. So
280 // using equivalent type jboolean to read the value of
281 // UseSharedSpaces which is same as hotspot type "bool".
282 if (read_jboolean(ph, useSharedSpacesAddr, &useSharedSpaces) != true) {
283 print_debug("can't read the value of 'UseSharedSpaces' flag\n");
284 return false;
285 }
286
287 if ((int)useSharedSpaces == 0) {
288 print_debug("UseSharedSpaces is false, assuming -Xshare:off!\n");
289 return true;
290 }
291
292 sharedArchivePathAddrAddr = lookup_symbol(ph, jvm_name, SHARED_ARCHIVE_PATH_SYM);
293 if (sharedArchivePathAddrAddr == 0) {
294 print_debug("can't lookup shared archive path symbol\n");
295 return false;
296 }
297
298 if (read_pointer(ph, sharedArchivePathAddrAddr, &sharedArchivePathAddr) != true) {
299 print_debug("can't read shared archive path pointer\n");
300 return false;
301 }
302
303 if (read_string(ph, sharedArchivePathAddr, classes_jsa, sizeof(classes_jsa)) != true) {
304 print_debug("can't read shared archive path value\n");
305 return false;
306 }
307
308 print_debug("looking for %s\n", classes_jsa);
309 // open the class sharing archive file
310 fd = pathmap_open(classes_jsa);
311 if (fd < 0) {
312 print_debug("can't open %s!\n", classes_jsa);
313 ph->core->classes_jsa_fd = -1;
314 return false;
315 } else {
316 print_debug("opened %s\n", classes_jsa);
317 }
318
319 // read CDSFileMapHeaderBase from the file
320 memset(&header, 0, sizeof(CDSFileMapHeaderBase));
321 if ((n = read(fd, &header, sizeof(CDSFileMapHeaderBase)))
322 != sizeof(CDSFileMapHeaderBase)) {
323 print_debug("can't read shared archive file map header from %s\n", classes_jsa);
324 close(fd);
325 return false;
326 }
327
328 // check file magic
329 if (header._magic != CDS_ARCHIVE_MAGIC) {
330 print_debug("%s has bad shared archive file magic number 0x%x, expecting 0x%x\n",
331 classes_jsa, header._magic, CDS_ARCHIVE_MAGIC);
332 close(fd);
333 return false;
334 }
335
336 // check version
337 if (header._version != CURRENT_CDS_ARCHIVE_VERSION) {
338 print_debug("%s has wrong shared archive file version %d, expecting %d\n",
339 classes_jsa, header._version, CURRENT_CDS_ARCHIVE_VERSION);
340 close(fd);
341 return false;
342 }
343
344 ph->core->classes_jsa_fd = fd;
345 // add read-only maps from classes.jsa to the list of maps
346 for (m = 0; m < NUM_CDS_REGIONS; m++) {
347 if (header._space[m]._read_only) {
348 base = (uintptr_t) header._space[m]._addr._base;
349 // no need to worry about the fractional pages at-the-end.
350 // possible fractional pages are handled by core_read_data.
351 add_class_share_map_info(ph, (off_t) header._space[m]._file_offset,
352 base, (size_t) header._space[m]._used);
353 print_debug("added a share archive map at 0x%lx\n", base);
354 }
355 }
356 return true;
357 }
358 lib = lib->next;
359 }
360 return true;
361 }
362
363
364 //---------------------------------------------------------------------------
365 // functions to handle map_info
366
367 // Order mappings based on virtual address. We use this function as the
368 // callback for sorting the array of map_info pointers.
369 static int core_cmp_mapping(const void *lhsp, const void *rhsp)
370 {
371 const map_info *lhs = *((const map_info **)lhsp);
372 const map_info *rhs = *((const map_info **)rhsp);
373
374 if (lhs->vaddr == rhs->vaddr) {
375 return (0);
376 }
377
378 return (lhs->vaddr < rhs->vaddr ? -1 : 1);
379 }
380
381 // we sort map_info by starting virtual address so that we can do
382 // binary search to read from an address.
383 static bool sort_map_array(struct ps_prochandle* ph) {
384 size_t num_maps = ph->core->num_maps;
385 map_info* map = ph->core->maps;
386 int i = 0;
387
388 // allocate map_array
389 map_info** array;
390 if ( (array = (map_info**) malloc(sizeof(map_info*) * num_maps)) == NULL) {
391 print_debug("can't allocate memory for map array\n");
392 return false;
393 }
394
395 // add maps to array
396 while (map) {
397 array[i] = map;
398 i++;
399 map = map->next;
400 }
401
402 // sort is called twice. If this is second time, clear map array
403 if (ph->core->map_array) {
404 free(ph->core->map_array);
405 }
406
407 ph->core->map_array = array;
408 // sort the map_info array by base virtual address.
409 qsort(ph->core->map_array, ph->core->num_maps, sizeof (map_info*),
410 core_cmp_mapping);
411
412 // print map
413 if (is_debug()) {
414 int j = 0;
415 print_debug("---- sorted virtual address map ----\n");
416 for (j = 0; j < ph->core->num_maps; j++) {
417 print_debug("base = 0x%lx\tsize = %zu\n", ph->core->map_array[j]->vaddr,
418 ph->core->map_array[j]->memsz);
419 }
420 }
421
422 return true;
423 }
424
425 #ifndef MIN
426 #define MIN(x, y) (((x) < (y))? (x): (y))
427 #endif
428
429 static bool core_read_data(struct ps_prochandle* ph, uintptr_t addr, char *buf, size_t size) {
430 ssize_t resid = size;
431 int page_size=sysconf(_SC_PAGE_SIZE);
432 while (resid != 0) {
433 map_info *mp = core_lookup(ph, addr);
434 uintptr_t mapoff;
435 ssize_t len, rem;
436 off_t off;
437 int fd;
438
439 if (mp == NULL) {
440 break; /* No mapping for this address */
441 }
442
443 fd = mp->fd;
444 mapoff = addr - mp->vaddr;
445 len = MIN(resid, mp->memsz - mapoff);
446 off = mp->offset + mapoff;
447
448 if ((len = pread(fd, buf, len, off)) <= 0) {
449 break;
450 }
451
452 resid -= len;
453 addr += len;
454 buf = (char *)buf + len;
455
456 // mappings always start at page boundary. But, may end in fractional
457 // page. fill zeros for possible fractional page at the end of a mapping.
458 rem = mp->memsz % page_size;
459 if (rem > 0) {
460 rem = page_size - rem;
461 len = MIN(resid, rem);
462 resid -= len;
463 addr += len;
464 // we are not assuming 'buf' to be zero initialized.
465 memset(buf, 0, len);
466 buf += len;
467 }
468 }
469
470 if (resid) {
471 print_debug("core read failed for %d byte(s) @ 0x%lx (%d more bytes)\n",
472 size, addr, resid);
473 return false;
474 } else {
475 return true;
476 }
477 }
478
479 // null implementation for write
480 static bool core_write_data(struct ps_prochandle* ph,
481 uintptr_t addr, const char *buf , size_t size) {
482 return false;
483 }
484
485 static bool core_get_lwp_regs(struct ps_prochandle* ph, lwpid_t lwp_id,
486 struct user_regs_struct* regs) {
487 // for core we have cached the lwp regs from NOTE section
488 thread_info* thr = ph->threads;
489 while (thr) {
490 if (thr->lwp_id == lwp_id) {
491 memcpy(regs, &thr->regs, sizeof(struct user_regs_struct));
492 return true;
493 }
494 thr = thr->next;
495 }
496 return false;
497 }
498
499 static ps_prochandle_ops core_ops = {
500 .release= core_release,
501 .p_pread= core_read_data,
502 .p_pwrite= core_write_data,
503 .get_lwp_regs= core_get_lwp_regs
504 };
505
506 // read regs and create thread from NT_PRSTATUS entries from core file
507 static bool core_handle_prstatus(struct ps_prochandle* ph, const char* buf, size_t nbytes) {
508 // we have to read prstatus_t from buf
509 // assert(nbytes == sizeof(prstaus_t), "size mismatch on prstatus_t");
510 prstatus_t* prstat = (prstatus_t*) buf;
511 thread_info* newthr;
512 print_debug("got integer regset for lwp %d\n", prstat->pr_pid);
513 // we set pthread_t to -1 for core dump
514 if((newthr = add_thread_info(ph, (pthread_t) -1, prstat->pr_pid)) == NULL)
515 return false;
516
517 // copy regs
518 memcpy(&newthr->regs, prstat->pr_reg, sizeof(struct user_regs_struct));
519
520 if (is_debug()) {
521 print_debug("integer regset\n");
522 #ifdef i386
523 // print the regset
524 print_debug("\teax = 0x%x\n", newthr->regs.eax);
525 print_debug("\tebx = 0x%x\n", newthr->regs.ebx);
526 print_debug("\tecx = 0x%x\n", newthr->regs.ecx);
527 print_debug("\tedx = 0x%x\n", newthr->regs.edx);
528 print_debug("\tesp = 0x%x\n", newthr->regs.esp);
529 print_debug("\tebp = 0x%x\n", newthr->regs.ebp);
530 print_debug("\tesi = 0x%x\n", newthr->regs.esi);
531 print_debug("\tedi = 0x%x\n", newthr->regs.edi);
532 print_debug("\teip = 0x%x\n", newthr->regs.eip);
533 #endif
534
535 #if defined(amd64) || defined(x86_64)
536 // print the regset
537 print_debug("\tr15 = 0x%lx\n", newthr->regs.r15);
538 print_debug("\tr14 = 0x%lx\n", newthr->regs.r14);
539 print_debug("\tr13 = 0x%lx\n", newthr->regs.r13);
540 print_debug("\tr12 = 0x%lx\n", newthr->regs.r12);
541 print_debug("\trbp = 0x%lx\n", newthr->regs.rbp);
542 print_debug("\trbx = 0x%lx\n", newthr->regs.rbx);
543 print_debug("\tr11 = 0x%lx\n", newthr->regs.r11);
544 print_debug("\tr10 = 0x%lx\n", newthr->regs.r10);
545 print_debug("\tr9 = 0x%lx\n", newthr->regs.r9);
546 print_debug("\tr8 = 0x%lx\n", newthr->regs.r8);
547 print_debug("\trax = 0x%lx\n", newthr->regs.rax);
548 print_debug("\trcx = 0x%lx\n", newthr->regs.rcx);
549 print_debug("\trdx = 0x%lx\n", newthr->regs.rdx);
550 print_debug("\trsi = 0x%lx\n", newthr->regs.rsi);
551 print_debug("\trdi = 0x%lx\n", newthr->regs.rdi);
552 print_debug("\torig_rax = 0x%lx\n", newthr->regs.orig_rax);
553 print_debug("\trip = 0x%lx\n", newthr->regs.rip);
554 print_debug("\tcs = 0x%lx\n", newthr->regs.cs);
555 print_debug("\teflags = 0x%lx\n", newthr->regs.eflags);
556 print_debug("\trsp = 0x%lx\n", newthr->regs.rsp);
557 print_debug("\tss = 0x%lx\n", newthr->regs.ss);
558 print_debug("\tfs_base = 0x%lx\n", newthr->regs.fs_base);
559 print_debug("\tgs_base = 0x%lx\n", newthr->regs.gs_base);
560 print_debug("\tds = 0x%lx\n", newthr->regs.ds);
561 print_debug("\tes = 0x%lx\n", newthr->regs.es);
562 print_debug("\tfs = 0x%lx\n", newthr->regs.fs);
563 print_debug("\tgs = 0x%lx\n", newthr->regs.gs);
564 #endif
565 }
566
567 return true;
568 }
569
570 #define ROUNDUP(x, y) ((((x)+((y)-1))/(y))*(y))
571
572 // read NT_PRSTATUS entries from core NOTE segment
573 static bool core_handle_note(struct ps_prochandle* ph, ELF_PHDR* note_phdr) {
574 char* buf = NULL;
575 char* p = NULL;
576 size_t size = note_phdr->p_filesz;
577
578 // we are interested in just prstatus entries. we will ignore the rest.
579 // Advance the seek pointer to the start of the PT_NOTE data
580 if (lseek(ph->core->core_fd, note_phdr->p_offset, SEEK_SET) == (off_t)-1) {
581 print_debug("failed to lseek to PT_NOTE data\n");
582 return false;
583 }
584
585 // Now process the PT_NOTE structures. Each one is preceded by
586 // an Elf{32/64}_Nhdr structure describing its type and size.
587 if ( (buf = (char*) malloc(size)) == NULL) {
588 print_debug("can't allocate memory for reading core notes\n");
589 goto err;
590 }
591
592 // read notes into buffer
593 if (read(ph->core->core_fd, buf, size) != size) {
594 print_debug("failed to read notes, core file must have been truncated\n");
595 goto err;
596 }
597
598 p = buf;
599 while (p < buf + size) {
600 ELF_NHDR* notep = (ELF_NHDR*) p;
601 char* descdata = p + sizeof(ELF_NHDR) + ROUNDUP(notep->n_namesz, 4);
602 print_debug("Note header with n_type = %d and n_descsz = %u\n",
603 notep->n_type, notep->n_descsz);
604
605 if (notep->n_type == NT_PRSTATUS) {
606 if (core_handle_prstatus(ph, descdata, notep->n_descsz) != true) {
607 return false;
608 }
609 } else if (notep->n_type == NT_AUXV) {
610 // Get first segment from entry point
611 ELF_AUXV *auxv = (ELF_AUXV *)descdata;
612 while (auxv->a_type != AT_NULL) {
613 if (auxv->a_type == AT_ENTRY) {
614 // Set entry point address to address of dynamic section.
615 // We will adjust it in read_exec_segments().
616 ph->core->dynamic_addr = auxv->a_un.a_val;
617 break;
618 }
619 auxv++;
620 }
621 }
622 p = descdata + ROUNDUP(notep->n_descsz, 4);
623 }
624
625 free(buf);
626 return true;
627
628 err:
629 if (buf) free(buf);
630 return false;
631 }
632
633 // read all segments from core file
634 static bool read_core_segments(struct ps_prochandle* ph, ELF_EHDR* core_ehdr) {
635 int i = 0;
636 ELF_PHDR* phbuf = NULL;
637 ELF_PHDR* core_php = NULL;
638
639 if ((phbuf = read_program_header_table(ph->core->core_fd, core_ehdr)) == NULL)
640 return false;
641
642 /*
643 * Now iterate through the program headers in the core file.
644 * We're interested in two types of Phdrs: PT_NOTE (which
645 * contains a set of saved /proc structures), and PT_LOAD (which
646 * represents a memory mapping from the process's address space).
647 *
648 * Difference b/w Solaris PT_NOTE and Linux/BSD PT_NOTE:
649 *
650 * In Solaris there are two PT_NOTE segments the first PT_NOTE (if present)
651 * contains /proc structs in the pre-2.6 unstructured /proc format. the last
652 * PT_NOTE has data in new /proc format.
653 *
654 * In Solaris, there is only one pstatus (process status). pstatus contains
655 * integer register set among other stuff. For each LWP, we have one lwpstatus
656 * entry that has integer regset for that LWP.
657 *
658 * Linux threads are actually 'clone'd processes. To support core analysis
659 * of "multithreaded" process, Linux creates more than one pstatus (called
660 * "prstatus") entry in PT_NOTE. Each prstatus entry has integer regset for one
661 * "thread". Please refer to Linux kernel src file 'fs/binfmt_elf.c', in particular
662 * function "elf_core_dump".
663 */
664
665 for (core_php = phbuf, i = 0; i < core_ehdr->e_phnum; i++) {
666 switch (core_php->p_type) {
667 case PT_NOTE:
668 if (core_handle_note(ph, core_php) != true) {
669 goto err;
670 }
671 break;
672
673 case PT_LOAD: {
674 if (core_php->p_filesz != 0) {
675 if (add_map_info(ph, ph->core->core_fd, core_php->p_offset,
676 core_php->p_vaddr, core_php->p_filesz) == NULL) goto err;
677 }
678 break;
679 }
680 }
681
682 core_php++;
683 }
684
685 free(phbuf);
686 return true;
687 err:
688 free(phbuf);
689 return false;
690 }
691
692 // read segments of a shared object
693 static bool read_lib_segments(struct ps_prochandle* ph, int lib_fd, ELF_EHDR* lib_ehdr, uintptr_t lib_base) {
694 int i = 0;
695 ELF_PHDR* phbuf;
696 ELF_PHDR* lib_php = NULL;
697
698 int page_size = sysconf(_SC_PAGE_SIZE);
699
700 if ((phbuf = read_program_header_table(lib_fd, lib_ehdr)) == NULL) {
701 return false;
702 }
703
704 // we want to process only PT_LOAD segments that are not writable.
705 // i.e., text segments. The read/write/exec (data) segments would
706 // have been already added from core file segments.
707 for (lib_php = phbuf, i = 0; i < lib_ehdr->e_phnum; i++) {
708 if ((lib_php->p_type == PT_LOAD) && !(lib_php->p_flags & PF_W) && (lib_php->p_filesz != 0)) {
709
710 uintptr_t target_vaddr = lib_php->p_vaddr + lib_base;
711 map_info *existing_map = core_lookup(ph, target_vaddr);
712
713 if (existing_map == NULL){
714 if (add_map_info(ph, lib_fd, lib_php->p_offset,
715 target_vaddr, lib_php->p_memsz) == NULL) {
716 goto err;
717 }
718 } else {
719 // Coredump stores value of p_memsz elf field
720 // rounded up to page boundary.
721
722 if ((existing_map->memsz != page_size) &&
723 (existing_map->fd != lib_fd) &&
724 (ROUNDUP(existing_map->memsz, page_size) != ROUNDUP(lib_php->p_memsz, page_size))) {
725
726 print_debug("address conflict @ 0x%lx (existing map size = %ld, size = %ld, flags = %d)\n",
727 target_vaddr, existing_map->memsz, lib_php->p_memsz, lib_php->p_flags);
728 goto err;
729 }
730
731 /* replace PT_LOAD segment with library segment */
732 print_debug("overwrote with new address mapping (memsz %ld -> %ld)\n",
733 existing_map->memsz, ROUNDUP(lib_php->p_memsz, page_size));
734
735 existing_map->fd = lib_fd;
736 existing_map->offset = lib_php->p_offset;
737 existing_map->memsz = ROUNDUP(lib_php->p_memsz, page_size);
738 }
739 }
740
741 lib_php++;
742 }
743
744 free(phbuf);
745 return true;
746 err:
747 free(phbuf);
748 return false;
749 }
750
751 // process segments from interpreter (ld.so or ld-linux.so)
752 static bool read_interp_segments(struct ps_prochandle* ph) {
753 ELF_EHDR interp_ehdr;
754
755 if (read_elf_header(ph->core->interp_fd, &interp_ehdr) != true) {
756 print_debug("interpreter is not a valid ELF file\n");
757 return false;
758 }
759
760 if (read_lib_segments(ph, ph->core->interp_fd, &interp_ehdr, ph->core->ld_base_addr) != true) {
761 print_debug("can't read segments of interpreter\n");
762 return false;
763 }
764
765 return true;
766 }
767
768 // process segments of a a.out
769 static bool read_exec_segments(struct ps_prochandle* ph, ELF_EHDR* exec_ehdr) {
770 int i = 0;
771 ELF_PHDR* phbuf = NULL;
772 ELF_PHDR* exec_php = NULL;
773
774 if ((phbuf = read_program_header_table(ph->core->exec_fd, exec_ehdr)) == NULL) {
775 return false;
776 }
777
778 for (exec_php = phbuf, i = 0; i < exec_ehdr->e_phnum; i++) {
779 switch (exec_php->p_type) {
780
781 // add mappings for PT_LOAD segments
782 case PT_LOAD: {
783 // add only non-writable segments of non-zero filesz
784 if (!(exec_php->p_flags & PF_W) && exec_php->p_filesz != 0) {
785 if (add_map_info(ph, ph->core->exec_fd, exec_php->p_offset, exec_php->p_vaddr, exec_php->p_filesz) == NULL) goto err;
786 }
787 break;
788 }
789
790 // read the interpreter and it's segments
791 case PT_INTERP: {
792 char interp_name[BUF_SIZE + 1];
793
794 // BUF_SIZE is PATH_MAX + NAME_MAX + 1.
795 if (exec_php->p_filesz > BUF_SIZE) {
796 goto err;
797 }
798 if (pread(ph->core->exec_fd, interp_name,
799 exec_php->p_filesz, exec_php->p_offset) != exec_php->p_filesz) {
800 print_debug("Unable to read in the ELF interpreter\n");
801 goto err;
802 }
803 interp_name[exec_php->p_filesz] = '\0';
804 print_debug("ELF interpreter %s\n", interp_name);
805 // read interpreter segments as well
806 if ((ph->core->interp_fd = pathmap_open(interp_name)) < 0) {
807 print_debug("can't open runtime loader\n");
808 goto err;
809 }
810 break;
811 }
812
813 // from PT_DYNAMIC we want to read address of first link_map addr
814 case PT_DYNAMIC: {
815 if (exec_ehdr->e_type == ET_EXEC) {
816 ph->core->dynamic_addr = exec_php->p_vaddr;
817 } else { // ET_DYN
818 // dynamic_addr has entry point of executable.
819 // Thus we should substract it.
820 ph->core->dynamic_addr += exec_php->p_vaddr - exec_ehdr->e_entry;
821 }
822 print_debug("address of _DYNAMIC is 0x%lx\n", ph->core->dynamic_addr);
823 break;
824 }
825
826 } // switch
827 exec_php++;
828 } // for
829
830 free(phbuf);
831 return true;
832 err:
833 free(phbuf);
834 return false;
835 }
836
837
838 #define FIRST_LINK_MAP_OFFSET offsetof(struct r_debug, r_map)
839 #define LD_BASE_OFFSET offsetof(struct r_debug, r_ldbase)
840 #define LINK_MAP_ADDR_OFFSET offsetof(struct link_map, l_addr)
841 #define LINK_MAP_NAME_OFFSET offsetof(struct link_map, l_name)
842 #define LINK_MAP_NEXT_OFFSET offsetof(struct link_map, l_next)
843
844 // read shared library info from runtime linker's data structures.
845 // This work is done by librtlb_db in Solaris
846 static bool read_shared_lib_info(struct ps_prochandle* ph) {
847 uintptr_t addr = ph->core->dynamic_addr;
848 uintptr_t debug_base;
849 uintptr_t first_link_map_addr;
850 uintptr_t ld_base_addr;
851 uintptr_t link_map_addr;
852 uintptr_t lib_base_diff;
853 uintptr_t lib_base;
854 uintptr_t lib_name_addr;
855 char lib_name[BUF_SIZE];
856 ELF_DYN dyn;
857 ELF_EHDR elf_ehdr;
858 int lib_fd;
859
860 // _DYNAMIC has information of the form
861 // [tag] [data] [tag] [data] .....
862 // Both tag and data are pointer sized.
863 // We look for dynamic info with DT_DEBUG. This has shared object info.
864 // refer to struct r_debug in link.h
865
866 dyn.d_tag = DT_NULL;
867 while (dyn.d_tag != DT_DEBUG) {
868 if (ps_pdread(ph, (psaddr_t) addr, &dyn, sizeof(ELF_DYN)) != PS_OK) {
869 print_debug("can't read debug info from _DYNAMIC\n");
870 return false;
871 }
872 addr += sizeof(ELF_DYN);
873 }
874
875 // we have got Dyn entry with DT_DEBUG
876 debug_base = dyn.d_un.d_ptr;
877 // at debug_base we have struct r_debug. This has first link map in r_map field
878 if (ps_pdread(ph, (psaddr_t) debug_base + FIRST_LINK_MAP_OFFSET,
879 &first_link_map_addr, sizeof(uintptr_t)) != PS_OK) {
880 print_debug("can't read first link map address\n");
881 return false;
882 }
883
884 // read ld_base address from struct r_debug
885 if (ps_pdread(ph, (psaddr_t) debug_base + LD_BASE_OFFSET, &ld_base_addr,
886 sizeof(uintptr_t)) != PS_OK) {
887 print_debug("can't read ld base address\n");
888 return false;
889 }
890 ph->core->ld_base_addr = ld_base_addr;
891
892 print_debug("interpreter base address is 0x%lx\n", ld_base_addr);
893
894 // now read segments from interp (i.e ld.so or ld-linux.so or ld-elf.so)
895 if (read_interp_segments(ph) != true) {
896 return false;
897 }
898
899 // after adding interpreter (ld.so) mappings sort again
900 if (sort_map_array(ph) != true) {
901 return false;
902 }
903
904 print_debug("first link map is at 0x%lx\n", first_link_map_addr);
905
906 link_map_addr = first_link_map_addr;
907 while (link_map_addr != 0) {
908 // read library base address of the .so. Note that even though <sys/link.h> calls
909 // link_map->l_addr as "base address", this is * not * really base virtual
910 // address of the shared object. This is actually the difference b/w the virtual
911 // address mentioned in shared object and the actual virtual base where runtime
912 // linker loaded it. We use "base diff" in read_lib_segments call below.
913
914 if (ps_pdread(ph, (psaddr_t) link_map_addr + LINK_MAP_ADDR_OFFSET,
915 &lib_base_diff, sizeof(uintptr_t)) != PS_OK) {
916 print_debug("can't read shared object base address diff\n");
917 return false;
918 }
919
920 // read address of the name
921 if (ps_pdread(ph, (psaddr_t) link_map_addr + LINK_MAP_NAME_OFFSET,
922 &lib_name_addr, sizeof(uintptr_t)) != PS_OK) {
923 print_debug("can't read address of shared object name\n");
924 return false;
925 }
926
927 // read name of the shared object
928 lib_name[0] = '\0';
929 if (lib_name_addr != 0 &&
930 read_string(ph, (uintptr_t) lib_name_addr, lib_name, sizeof(lib_name)) != true) {
931 print_debug("can't read shared object name\n");
932 // don't let failure to read the name stop opening the file. If something is really wrong
933 // it will fail later.
934 }
935
936 if (lib_name[0] != '\0') {
937 // ignore empty lib names
938 lib_fd = pathmap_open(lib_name);
939
940 if (lib_fd < 0) {
941 print_debug("can't open shared object %s\n", lib_name);
942 // continue with other libraries...
943 } else {
944 if (read_elf_header(lib_fd, &elf_ehdr)) {
945 lib_base = lib_base_diff + find_base_address(lib_fd, &elf_ehdr);
946 print_debug("reading library %s @ 0x%lx [ 0x%lx ]\n",
947 lib_name, lib_base, lib_base_diff);
948 // while adding library mappings we need to use "base difference".
949 if (! read_lib_segments(ph, lib_fd, &elf_ehdr, lib_base_diff)) {
950 print_debug("can't read shared object's segments\n");
951 close(lib_fd);
952 return false;
953 }
954 add_lib_info_fd(ph, lib_name, lib_fd, lib_base);
955 // Map info is added for the library (lib_name) so
956 // we need to re-sort it before calling the p_pdread.
957 if (sort_map_array(ph) != true)
958 return false;
959 } else {
960 print_debug("can't read ELF header for shared object %s\n", lib_name);
961 close(lib_fd);
962 // continue with other libraries...
963 }
964 }
965 }
966
967 // read next link_map address
968 if (ps_pdread(ph, (psaddr_t) link_map_addr + LINK_MAP_NEXT_OFFSET,
969 &link_map_addr, sizeof(uintptr_t)) != PS_OK) {
970 print_debug("can't read next link in link_map\n");
971 return false;
972 }
973 }
974
975 return true;
976 }
977
978 // the one and only one exposed stuff from this file
979 JNIEXPORT struct ps_prochandle* JNICALL
980 Pgrab_core(const char* exec_file, const char* core_file) {
981 ELF_EHDR core_ehdr;
982 ELF_EHDR exec_ehdr;
983 ELF_EHDR lib_ehdr;
984
985 struct ps_prochandle* ph = (struct ps_prochandle*) calloc(1, sizeof(struct ps_prochandle));
986 if (ph == NULL) {
987 print_debug("can't allocate ps_prochandle\n");
988 return NULL;
989 }
990
991 if ((ph->core = (struct core_data*) calloc(1, sizeof(struct core_data))) == NULL) {
992 free(ph);
993 print_debug("can't allocate ps_prochandle\n");
994 return NULL;
995 }
996
997 // initialize ph
998 ph->ops = &core_ops;
999 ph->core->core_fd = -1;
1000 ph->core->exec_fd = -1;
1001 ph->core->interp_fd = -1;
1002
1003 // open the core file
1004 if ((ph->core->core_fd = open(core_file, O_RDONLY)) < 0) {
1005 print_debug("can't open core file\n");
1006 goto err;
1007 }
1008
1009 // read core file ELF header
1010 if (read_elf_header(ph->core->core_fd, &core_ehdr) != true || core_ehdr.e_type != ET_CORE) {
1011 print_debug("core file is not a valid ELF ET_CORE file\n");
1012 goto err;
1013 }
1014
1015 if ((ph->core->exec_fd = open(exec_file, O_RDONLY)) < 0) {
1016 print_debug("can't open executable file\n");
1017 goto err;
1018 }
1019
1020 if (read_elf_header(ph->core->exec_fd, &exec_ehdr) != true ||
1021 ((exec_ehdr.e_type != ET_EXEC) && (exec_ehdr.e_type != ET_DYN))) {
1022 print_debug("executable file is not a valid ELF file\n");
1023 goto err;
1024 }
1025
1026 // process core file segments
1027 if (read_core_segments(ph, &core_ehdr) != true) {
1028 goto err;
1029 }
1030
1031 // process exec file segments
1032 if (read_exec_segments(ph, &exec_ehdr) != true) {
1033 goto err;
1034 }
1035
1036 // exec file is also treated like a shared object for symbol search
1037 if (add_lib_info_fd(ph, exec_file, ph->core->exec_fd,
1038 (uintptr_t)0 + find_base_address(ph->core->exec_fd, &exec_ehdr)) == NULL) {
1039 goto err;
1040 }
1041
1042 // allocate and sort maps into map_array, we need to do this
1043 // here because read_shared_lib_info needs to read from debuggee
1044 // address space
1045 if (sort_map_array(ph) != true) {
1046 goto err;
1047 }
1048
1049 if (read_shared_lib_info(ph) != true) {
1050 goto err;
1051 }
1052
1053 // sort again because we have added more mappings from shared objects
1054 if (sort_map_array(ph) != true) {
1055 goto err;
1056 }
1057
1058 if (init_classsharing_workaround(ph) != true) {
1059 goto err;
1060 }
1061
1062 return ph;
1063
1064 err:
1065 Prelease(ph);
1066 return NULL;
1067 }