1 /*
2 * Copyright (c) 2003, 2020, 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, uint32_t flags) {
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 map->flags = flags;
119 return map;
120 }
121
122 // add map info with given fd, offset, vaddr and memsz
123 static map_info* add_map_info(struct ps_prochandle* ph, int fd, off_t offset,
124 uintptr_t vaddr, size_t memsz, uint32_t flags) {
125 map_info* map;
126 if ((map = allocate_init_map(fd, offset, vaddr, memsz, flags)) == NULL) {
127 return NULL;
128 }
129
130 // add this to map list
131 map->next = ph->core->maps;
132 ph->core->maps = map;
133 ph->core->num_maps++;
134
135 return map;
136 }
137
138 // Part of the class sharing workaround
139 static map_info* add_class_share_map_info(struct ps_prochandle* ph, off_t offset,
140 uintptr_t vaddr, size_t memsz) {
141 map_info* map;
142 if ((map = allocate_init_map(ph->core->classes_jsa_fd,
143 offset, vaddr, memsz, PF_R)) == NULL) {
144 return NULL;
145 }
146
147 map->next = ph->core->class_share_maps;
148 ph->core->class_share_maps = map;
149 return map;
150 }
151
152 // Return the map_info for the given virtual address. We keep a sorted
153 // array of pointers in ph->map_array, so we can binary search.
154 static map_info* core_lookup(struct ps_prochandle *ph, uintptr_t addr) {
155 int mid, lo = 0, hi = ph->core->num_maps - 1;
156 map_info *mp;
157
158 while (hi - lo > 1) {
159 mid = (lo + hi) / 2;
160 if (addr >= ph->core->map_array[mid]->vaddr) {
161 lo = mid;
162 } else {
163 hi = mid;
164 }
165 }
166
167 if (addr < ph->core->map_array[hi]->vaddr) {
168 mp = ph->core->map_array[lo];
169 } else {
170 mp = ph->core->map_array[hi];
171 }
172
173 if (addr >= mp->vaddr && addr < mp->vaddr + mp->memsz) {
174 return (mp);
175 }
176
177
178 // Part of the class sharing workaround
179 // Unfortunately, we have no way of detecting -Xshare state.
180 // Check out the share maps atlast, if we don't find anywhere.
181 // This is done this way so to avoid reading share pages
182 // ahead of other normal maps. For eg. with -Xshare:off we don't
183 // want to prefer class sharing data to data from core.
184 mp = ph->core->class_share_maps;
185 if (mp) {
186 print_debug("can't locate map_info at 0x%lx, trying class share maps\n", addr);
187 }
188 while (mp) {
189 if (addr >= mp->vaddr && addr < mp->vaddr + mp->memsz) {
190 print_debug("located map_info at 0x%lx from class share maps\n", addr);
191 return (mp);
192 }
193 mp = mp->next;
194 }
195
196 print_debug("can't locate map_info at 0x%lx\n", addr);
197 return (NULL);
198 }
199
200 //---------------------------------------------------------------
201 // Part of the class sharing workaround:
202 //
203 // With class sharing, pages are mapped from classes.jsa file.
204 // The read-only class sharing pages are mapped as MAP_SHARED,
205 // PROT_READ pages. These pages are not dumped into core dump.
206 // With this workaround, these pages are read from classes.jsa.
207
208 static bool read_jboolean(struct ps_prochandle* ph, uintptr_t addr, jboolean* pvalue) {
209 jboolean i;
210 if (ps_pdread(ph, (psaddr_t) addr, &i, sizeof(i)) == PS_OK) {
211 *pvalue = i;
212 return true;
213 } else {
214 return false;
215 }
216 }
217
218 static bool read_pointer(struct ps_prochandle* ph, uintptr_t addr, uintptr_t* pvalue) {
219 uintptr_t uip;
220 if (ps_pdread(ph, (psaddr_t) addr, (char *)&uip, sizeof(uip)) == PS_OK) {
221 *pvalue = uip;
222 return true;
223 } else {
224 return false;
225 }
226 }
227
228 // used to read strings from debuggee
229 static bool read_string(struct ps_prochandle* ph, uintptr_t addr, char* buf, size_t size) {
230 size_t i = 0;
231 char c = ' ';
232
233 while (c != '\0') {
234 if (ps_pdread(ph, (psaddr_t) addr, &c, sizeof(char)) != PS_OK) {
235 return false;
236 }
237 if (i < size - 1) {
238 buf[i] = c;
239 } else {
240 // smaller buffer
241 return false;
242 }
243 i++; addr++;
244 }
245
246 buf[i] = '\0';
247 return true;
248 }
249
250 #define USE_SHARED_SPACES_SYM "UseSharedSpaces"
251 // mangled name of Arguments::SharedArchivePath
252 #define SHARED_ARCHIVE_PATH_SYM "_ZN9Arguments17SharedArchivePathE"
253 #define LIBJVM_NAME "/libjvm.so"
254
255 static bool init_classsharing_workaround(struct ps_prochandle* ph) {
256 lib_info* lib = ph->libs;
257 while (lib != NULL) {
258 // we are iterating over shared objects from the core dump. look for
259 // libjvm.so.
260 const char *jvm_name = 0;
261 if ((jvm_name = strstr(lib->name, LIBJVM_NAME)) != 0) {
262 char classes_jsa[PATH_MAX];
263 CDSFileMapHeaderBase header;
264 int fd = -1;
265 int m = 0;
266 size_t n = 0;
267 uintptr_t base = 0, useSharedSpacesAddr = 0;
268 uintptr_t sharedArchivePathAddrAddr = 0, sharedArchivePathAddr = 0;
269 jboolean useSharedSpaces = 0;
270 map_info* mi = 0;
271
272 memset(classes_jsa, 0, sizeof(classes_jsa));
273 jvm_name = lib->name;
274 useSharedSpacesAddr = lookup_symbol(ph, jvm_name, USE_SHARED_SPACES_SYM);
275 if (useSharedSpacesAddr == 0) {
276 print_debug("can't lookup 'UseSharedSpaces' flag\n");
277 return false;
278 }
279
280 // Hotspot vm types are not exported to build this library. So
281 // using equivalent type jboolean to read the value of
282 // UseSharedSpaces which is same as hotspot type "bool".
283 if (read_jboolean(ph, useSharedSpacesAddr, &useSharedSpaces) != true) {
284 print_debug("can't read the value of 'UseSharedSpaces' flag\n");
285 return false;
286 }
287
288 if ((int)useSharedSpaces == 0) {
289 print_debug("UseSharedSpaces is false, assuming -Xshare:off!\n");
290 return true;
291 }
292
293 sharedArchivePathAddrAddr = lookup_symbol(ph, jvm_name, SHARED_ARCHIVE_PATH_SYM);
294 if (sharedArchivePathAddrAddr == 0) {
295 print_debug("can't lookup shared archive path symbol\n");
296 return false;
297 }
298
299 if (read_pointer(ph, sharedArchivePathAddrAddr, &sharedArchivePathAddr) != true) {
300 print_debug("can't read shared archive path pointer\n");
301 return false;
302 }
303
304 if (read_string(ph, sharedArchivePathAddr, classes_jsa, sizeof(classes_jsa)) != true) {
305 print_debug("can't read shared archive path value\n");
306 return false;
307 }
308
309 print_debug("looking for %s\n", classes_jsa);
310 // open the class sharing archive file
311 fd = pathmap_open(classes_jsa);
312 if (fd < 0) {
313 print_debug("can't open %s!\n", classes_jsa);
314 ph->core->classes_jsa_fd = -1;
315 return false;
316 } else {
317 print_debug("opened %s\n", classes_jsa);
318 }
319
320 // read CDSFileMapHeaderBase from the file
321 memset(&header, 0, sizeof(CDSFileMapHeaderBase));
322 if ((n = read(fd, &header, sizeof(CDSFileMapHeaderBase)))
323 != sizeof(CDSFileMapHeaderBase)) {
324 print_debug("can't read shared archive file map header from %s\n", classes_jsa);
325 close(fd);
326 return false;
327 }
328
329 // check file magic
330 if (header._magic != CDS_ARCHIVE_MAGIC) {
331 print_debug("%s has bad shared archive file magic number 0x%x, expecting 0x%x\n",
332 classes_jsa, header._magic, CDS_ARCHIVE_MAGIC);
333 close(fd);
334 return false;
335 }
336
337 // check version
338 if (header._version != CURRENT_CDS_ARCHIVE_VERSION) {
339 print_debug("%s has wrong shared archive file version %d, expecting %d\n",
340 classes_jsa, header._version, CURRENT_CDS_ARCHIVE_VERSION);
341 close(fd);
342 return false;
343 }
344
345 ph->core->classes_jsa_fd = fd;
346 // add read-only maps from classes.jsa to the list of maps
347 for (m = 0; m < NUM_CDS_REGIONS; m++) {
348 if (header._space[m]._read_only) {
349 base = (uintptr_t) header._space[m]._addr._base;
350 // no need to worry about the fractional pages at-the-end.
351 // possible fractional pages are handled by core_read_data.
352 add_class_share_map_info(ph, (off_t) header._space[m]._file_offset,
353 base, (size_t) header._space[m]._used);
354 print_debug("added a share archive map at 0x%lx\n", base);
355 }
356 }
357 return true;
358 }
359 lib = lib->next;
360 }
361 return true;
362 }
363
364
365 //---------------------------------------------------------------------------
366 // functions to handle map_info
367
368 // Order mappings based on virtual address. We use this function as the
369 // callback for sorting the array of map_info pointers.
370 static int core_cmp_mapping(const void *lhsp, const void *rhsp)
371 {
372 const map_info *lhs = *((const map_info **)lhsp);
373 const map_info *rhs = *((const map_info **)rhsp);
374
375 if (lhs->vaddr == rhs->vaddr) {
376 return (0);
377 }
378
379 return (lhs->vaddr < rhs->vaddr ? -1 : 1);
380 }
381
382 // we sort map_info by starting virtual address so that we can do
383 // binary search to read from an address.
384 static bool sort_map_array(struct ps_prochandle* ph) {
385 size_t num_maps = ph->core->num_maps;
386 map_info* map = ph->core->maps;
387 int i = 0;
388
389 // allocate map_array
390 map_info** array;
391 if ( (array = (map_info**) malloc(sizeof(map_info*) * num_maps)) == NULL) {
392 print_debug("can't allocate memory for map array\n");
393 return false;
394 }
395
396 // add maps to array
397 while (map) {
398 array[i] = map;
399 i++;
400 map = map->next;
401 }
402
403 // sort is called twice. If this is second time, clear map array
404 if (ph->core->map_array) {
405 free(ph->core->map_array);
406 }
407
408 ph->core->map_array = array;
409 // sort the map_info array by base virtual address.
410 qsort(ph->core->map_array, ph->core->num_maps, sizeof (map_info*),
411 core_cmp_mapping);
412
413 // print map
414 if (is_debug()) {
415 int j = 0;
416 print_debug("---- sorted virtual address map ----\n");
417 for (j = 0; j < ph->core->num_maps; j++) {
418 print_debug("base = 0x%lx\tsize = %zu\n", ph->core->map_array[j]->vaddr,
419 ph->core->map_array[j]->memsz);
420 }
421 }
422
423 return true;
424 }
425
426 #ifndef MIN
427 #define MIN(x, y) (((x) < (y))? (x): (y))
428 #endif
429
430 static bool core_read_data(struct ps_prochandle* ph, uintptr_t addr, char *buf, size_t size) {
431 ssize_t resid = size;
432 int page_size=sysconf(_SC_PAGE_SIZE);
433 while (resid != 0) {
434 map_info *mp = core_lookup(ph, addr);
435 uintptr_t mapoff;
436 ssize_t len, rem;
437 off_t off;
438 int fd;
439
440 if (mp == NULL) {
441 break; /* No mapping for this address */
442 }
443
444 fd = mp->fd;
445 mapoff = addr - mp->vaddr;
446 len = MIN(resid, mp->memsz - mapoff);
447 off = mp->offset + mapoff;
448
449 if ((len = pread(fd, buf, len, off)) <= 0) {
450 break;
451 }
452
453 resid -= len;
454 addr += len;
455 buf = (char *)buf + len;
456
457 // mappings always start at page boundary. But, may end in fractional
458 // page. fill zeros for possible fractional page at the end of a mapping.
459 rem = mp->memsz % page_size;
460 if (rem > 0) {
461 rem = page_size - rem;
462 len = MIN(resid, rem);
463 resid -= len;
464 addr += len;
465 // we are not assuming 'buf' to be zero initialized.
466 memset(buf, 0, len);
467 buf += len;
468 }
469 }
470
471 if (resid) {
472 print_debug("core read failed for %d byte(s) @ 0x%lx (%d more bytes)\n",
473 size, addr, resid);
474 return false;
475 } else {
476 return true;
477 }
478 }
479
480 // null implementation for write
481 static bool core_write_data(struct ps_prochandle* ph,
482 uintptr_t addr, const char *buf , size_t size) {
483 return false;
484 }
485
486 static bool core_get_lwp_regs(struct ps_prochandle* ph, lwpid_t lwp_id,
487 struct user_regs_struct* regs) {
488 // for core we have cached the lwp regs from NOTE section
489 thread_info* thr = ph->threads;
490 while (thr) {
491 if (thr->lwp_id == lwp_id) {
492 memcpy(regs, &thr->regs, sizeof(struct user_regs_struct));
493 return true;
494 }
495 thr = thr->next;
496 }
497 return false;
498 }
499
500 static ps_prochandle_ops core_ops = {
501 .release= core_release,
502 .p_pread= core_read_data,
503 .p_pwrite= core_write_data,
504 .get_lwp_regs= core_get_lwp_regs
505 };
506
507 // read regs and create thread from NT_PRSTATUS entries from core file
508 static bool core_handle_prstatus(struct ps_prochandle* ph, const char* buf, size_t nbytes) {
509 // we have to read prstatus_t from buf
510 // assert(nbytes == sizeof(prstaus_t), "size mismatch on prstatus_t");
511 prstatus_t* prstat = (prstatus_t*) buf;
512 thread_info* newthr;
513 print_debug("got integer regset for lwp %d\n", prstat->pr_pid);
514 // we set pthread_t to -1 for core dump
515 if((newthr = add_thread_info(ph, (pthread_t) -1, prstat->pr_pid)) == NULL)
516 return false;
517
518 // copy regs
519 memcpy(&newthr->regs, prstat->pr_reg, sizeof(struct user_regs_struct));
520
521 if (is_debug()) {
522 print_debug("integer regset\n");
523 #ifdef i386
524 // print the regset
525 print_debug("\teax = 0x%x\n", newthr->regs.eax);
526 print_debug("\tebx = 0x%x\n", newthr->regs.ebx);
527 print_debug("\tecx = 0x%x\n", newthr->regs.ecx);
528 print_debug("\tedx = 0x%x\n", newthr->regs.edx);
529 print_debug("\tesp = 0x%x\n", newthr->regs.esp);
530 print_debug("\tebp = 0x%x\n", newthr->regs.ebp);
531 print_debug("\tesi = 0x%x\n", newthr->regs.esi);
532 print_debug("\tedi = 0x%x\n", newthr->regs.edi);
533 print_debug("\teip = 0x%x\n", newthr->regs.eip);
534 #endif
535
536 #if defined(amd64) || defined(x86_64)
537 // print the regset
538 print_debug("\tr15 = 0x%lx\n", newthr->regs.r15);
539 print_debug("\tr14 = 0x%lx\n", newthr->regs.r14);
540 print_debug("\tr13 = 0x%lx\n", newthr->regs.r13);
541 print_debug("\tr12 = 0x%lx\n", newthr->regs.r12);
542 print_debug("\trbp = 0x%lx\n", newthr->regs.rbp);
543 print_debug("\trbx = 0x%lx\n", newthr->regs.rbx);
544 print_debug("\tr11 = 0x%lx\n", newthr->regs.r11);
545 print_debug("\tr10 = 0x%lx\n", newthr->regs.r10);
546 print_debug("\tr9 = 0x%lx\n", newthr->regs.r9);
547 print_debug("\tr8 = 0x%lx\n", newthr->regs.r8);
548 print_debug("\trax = 0x%lx\n", newthr->regs.rax);
549 print_debug("\trcx = 0x%lx\n", newthr->regs.rcx);
550 print_debug("\trdx = 0x%lx\n", newthr->regs.rdx);
551 print_debug("\trsi = 0x%lx\n", newthr->regs.rsi);
552 print_debug("\trdi = 0x%lx\n", newthr->regs.rdi);
553 print_debug("\torig_rax = 0x%lx\n", newthr->regs.orig_rax);
554 print_debug("\trip = 0x%lx\n", newthr->regs.rip);
555 print_debug("\tcs = 0x%lx\n", newthr->regs.cs);
556 print_debug("\teflags = 0x%lx\n", newthr->regs.eflags);
557 print_debug("\trsp = 0x%lx\n", newthr->regs.rsp);
558 print_debug("\tss = 0x%lx\n", newthr->regs.ss);
559 print_debug("\tfs_base = 0x%lx\n", newthr->regs.fs_base);
560 print_debug("\tgs_base = 0x%lx\n", newthr->regs.gs_base);
561 print_debug("\tds = 0x%lx\n", newthr->regs.ds);
562 print_debug("\tes = 0x%lx\n", newthr->regs.es);
563 print_debug("\tfs = 0x%lx\n", newthr->regs.fs);
564 print_debug("\tgs = 0x%lx\n", newthr->regs.gs);
565 #endif
566 }
567
568 return true;
569 }
570
571 #define ROUNDUP(x, y) ((((x)+((y)-1))/(y))*(y))
572
573 // read NT_PRSTATUS entries from core NOTE segment
574 static bool core_handle_note(struct ps_prochandle* ph, ELF_PHDR* note_phdr) {
575 char* buf = NULL;
576 char* p = NULL;
577 size_t size = note_phdr->p_filesz;
578
579 // we are interested in just prstatus entries. we will ignore the rest.
580 // Advance the seek pointer to the start of the PT_NOTE data
581 if (lseek(ph->core->core_fd, note_phdr->p_offset, SEEK_SET) == (off_t)-1) {
582 print_debug("failed to lseek to PT_NOTE data\n");
583 return false;
584 }
585
586 // Now process the PT_NOTE structures. Each one is preceded by
587 // an Elf{32/64}_Nhdr structure describing its type and size.
588 if ( (buf = (char*) malloc(size)) == NULL) {
589 print_debug("can't allocate memory for reading core notes\n");
590 goto err;
591 }
592
593 // read notes into buffer
594 if (read(ph->core->core_fd, buf, size) != size) {
595 print_debug("failed to read notes, core file must have been truncated\n");
596 goto err;
597 }
598
599 p = buf;
600 while (p < buf + size) {
601 ELF_NHDR* notep = (ELF_NHDR*) p;
602 char* descdata = p + sizeof(ELF_NHDR) + ROUNDUP(notep->n_namesz, 4);
603 print_debug("Note header with n_type = %d and n_descsz = %u\n",
604 notep->n_type, notep->n_descsz);
605
606 if (notep->n_type == NT_PRSTATUS) {
607 if (core_handle_prstatus(ph, descdata, notep->n_descsz) != true) {
608 return false;
609 }
610 } else if (notep->n_type == NT_AUXV) {
611 // Get first segment from entry point
612 ELF_AUXV *auxv = (ELF_AUXV *)descdata;
613 while (auxv->a_type != AT_NULL) {
614 if (auxv->a_type == AT_ENTRY) {
615 // Set entry point address to address of dynamic section.
616 // We will adjust it in read_exec_segments().
617 ph->core->dynamic_addr = auxv->a_un.a_val;
618 break;
619 }
620 auxv++;
621 }
622 }
623 p = descdata + ROUNDUP(notep->n_descsz, 4);
624 }
625
626 free(buf);
627 return true;
628
629 err:
630 if (buf) free(buf);
631 return false;
632 }
633
634 // read all segments from core file
635 static bool read_core_segments(struct ps_prochandle* ph, ELF_EHDR* core_ehdr) {
636 int i = 0;
637 ELF_PHDR* phbuf = NULL;
638 ELF_PHDR* core_php = NULL;
639
640 if ((phbuf = read_program_header_table(ph->core->core_fd, core_ehdr)) == NULL)
641 return false;
642
643 /*
644 * Now iterate through the program headers in the core file.
645 * We're interested in two types of Phdrs: PT_NOTE (which
646 * contains a set of saved /proc structures), and PT_LOAD (which
647 * represents a memory mapping from the process's address space).
648 *
649 * Difference b/w Solaris PT_NOTE and Linux/BSD PT_NOTE:
650 *
651 * In Solaris there are two PT_NOTE segments the first PT_NOTE (if present)
652 * contains /proc structs in the pre-2.6 unstructured /proc format. the last
653 * PT_NOTE has data in new /proc format.
654 *
655 * In Solaris, there is only one pstatus (process status). pstatus contains
656 * integer register set among other stuff. For each LWP, we have one lwpstatus
657 * entry that has integer regset for that LWP.
658 *
659 * Linux threads are actually 'clone'd processes. To support core analysis
660 * of "multithreaded" process, Linux creates more than one pstatus (called
661 * "prstatus") entry in PT_NOTE. Each prstatus entry has integer regset for one
662 * "thread". Please refer to Linux kernel src file 'fs/binfmt_elf.c', in particular
663 * function "elf_core_dump".
664 */
665
666 for (core_php = phbuf, i = 0; i < core_ehdr->e_phnum; i++) {
667 switch (core_php->p_type) {
668 case PT_NOTE:
669 if (core_handle_note(ph, core_php) != true) {
670 goto err;
671 }
672 break;
673
674 case PT_LOAD: {
675 if (core_php->p_filesz != 0) {
676 if (add_map_info(ph, ph->core->core_fd, core_php->p_offset,
677 core_php->p_vaddr, core_php->p_filesz, core_php->p_flags) == NULL) goto err;
678 }
679 break;
680 }
681 }
682
683 core_php++;
684 }
685
686 free(phbuf);
687 return true;
688 err:
689 free(phbuf);
690 return false;
691 }
692
693 // read segments of a shared object
694 static bool read_lib_segments(struct ps_prochandle* ph, int lib_fd, ELF_EHDR* lib_ehdr, uintptr_t lib_base) {
695 int i = 0;
696 ELF_PHDR* phbuf;
697 ELF_PHDR* lib_php = NULL;
698
699 int page_size = sysconf(_SC_PAGE_SIZE);
700
701 if ((phbuf = read_program_header_table(lib_fd, lib_ehdr)) == NULL) {
702 return false;
703 }
704
705 // we want to process only PT_LOAD segments that are not writable.
706 // i.e., text segments. The read/write/exec (data) segments would
707 // have been already added from core file segments.
708 for (lib_php = phbuf, i = 0; i < lib_ehdr->e_phnum; i++) {
709 if ((lib_php->p_type == PT_LOAD) && !(lib_php->p_flags & PF_W) && (lib_php->p_filesz != 0)) {
710
711 uintptr_t target_vaddr = lib_php->p_vaddr + lib_base;
712 map_info *existing_map = core_lookup(ph, target_vaddr);
713
714 if (existing_map == NULL){
715 if (add_map_info(ph, lib_fd, lib_php->p_offset,
716 target_vaddr, lib_php->p_memsz, lib_php->p_flags) == NULL) {
717 goto err;
718 }
719 } else if (lib_php->p_flags != existing_map->flags) {
720 // Access flags for this memory region are different between the library
721 // and coredump. It might be caused by mprotect() call at runtime.
722 // We should respect the coredump.
723 continue;
724 } else {
725 // Read only segments in ELF should not be any different from PT_LOAD segments
726 // in the coredump.
727 // Also the first page of the ELF header might be included
728 // in the coredump (See JDK-7133122).
729 // Thus we need to replace the PT_LOAD segment with the library version.
730 //
731 // Coredump stores value of p_memsz elf field
732 // rounded up to page boundary.
733
734 if ((existing_map->memsz != page_size) &&
735 (existing_map->fd != lib_fd) &&
736 (ROUNDUP(existing_map->memsz, page_size) != ROUNDUP(lib_php->p_memsz, page_size))) {
737
738 print_debug("address conflict @ 0x%lx (existing map size = %ld, size = %ld, flags = %d)\n",
739 target_vaddr, existing_map->memsz, lib_php->p_memsz, lib_php->p_flags);
740 goto err;
741 }
742
743 /* replace PT_LOAD segment with library segment */
744 print_debug("overwrote with new address mapping (memsz %ld -> %ld)\n",
745 existing_map->memsz, ROUNDUP(lib_php->p_memsz, page_size));
746
747 existing_map->fd = lib_fd;
748 existing_map->offset = lib_php->p_offset;
749 existing_map->memsz = ROUNDUP(lib_php->p_memsz, page_size);
750 }
751 }
752
753 lib_php++;
754 }
755
756 free(phbuf);
757 return true;
758 err:
759 free(phbuf);
760 return false;
761 }
762
763 // process segments from interpreter (ld.so or ld-linux.so)
764 static bool read_interp_segments(struct ps_prochandle* ph) {
765 ELF_EHDR interp_ehdr;
766
767 if (read_elf_header(ph->core->interp_fd, &interp_ehdr) != true) {
768 print_debug("interpreter is not a valid ELF file\n");
769 return false;
770 }
771
772 if (read_lib_segments(ph, ph->core->interp_fd, &interp_ehdr, ph->core->ld_base_addr) != true) {
773 print_debug("can't read segments of interpreter\n");
774 return false;
775 }
776
777 return true;
778 }
779
780 // process segments of a a.out
781 static bool read_exec_segments(struct ps_prochandle* ph, ELF_EHDR* exec_ehdr) {
782 int i = 0;
783 ELF_PHDR* phbuf = NULL;
784 ELF_PHDR* exec_php = NULL;
785
786 if ((phbuf = read_program_header_table(ph->core->exec_fd, exec_ehdr)) == NULL) {
787 return false;
788 }
789
790 for (exec_php = phbuf, i = 0; i < exec_ehdr->e_phnum; i++) {
791 switch (exec_php->p_type) {
792
793 // add mappings for PT_LOAD segments
794 case PT_LOAD: {
795 // add only non-writable segments of non-zero filesz
796 if (!(exec_php->p_flags & PF_W) && exec_php->p_filesz != 0) {
797 if (add_map_info(ph, ph->core->exec_fd, exec_php->p_offset, exec_php->p_vaddr, exec_php->p_filesz, exec_php->p_flags) == NULL) goto err;
798 }
799 break;
800 }
801
802 // read the interpreter and it's segments
803 case PT_INTERP: {
804 char interp_name[BUF_SIZE + 1];
805
806 // BUF_SIZE is PATH_MAX + NAME_MAX + 1.
807 if (exec_php->p_filesz > BUF_SIZE) {
808 goto err;
809 }
810 if (pread(ph->core->exec_fd, interp_name,
811 exec_php->p_filesz, exec_php->p_offset) != exec_php->p_filesz) {
812 print_debug("Unable to read in the ELF interpreter\n");
813 goto err;
814 }
815 interp_name[exec_php->p_filesz] = '\0';
816 print_debug("ELF interpreter %s\n", interp_name);
817 // read interpreter segments as well
818 if ((ph->core->interp_fd = pathmap_open(interp_name)) < 0) {
819 print_debug("can't open runtime loader\n");
820 goto err;
821 }
822 break;
823 }
824
825 // from PT_DYNAMIC we want to read address of first link_map addr
826 case PT_DYNAMIC: {
827 if (exec_ehdr->e_type == ET_EXEC) {
828 ph->core->dynamic_addr = exec_php->p_vaddr;
829 } else { // ET_DYN
830 // dynamic_addr has entry point of executable.
831 // Thus we should substract it.
832 ph->core->dynamic_addr += exec_php->p_vaddr - exec_ehdr->e_entry;
833 }
834 print_debug("address of _DYNAMIC is 0x%lx\n", ph->core->dynamic_addr);
835 break;
836 }
837
838 } // switch
839 exec_php++;
840 } // for
841
842 free(phbuf);
843 return true;
844 err:
845 free(phbuf);
846 return false;
847 }
848
849
850 #define FIRST_LINK_MAP_OFFSET offsetof(struct r_debug, r_map)
851 #define LD_BASE_OFFSET offsetof(struct r_debug, r_ldbase)
852 #define LINK_MAP_ADDR_OFFSET offsetof(struct link_map, l_addr)
853 #define LINK_MAP_NAME_OFFSET offsetof(struct link_map, l_name)
854 #define LINK_MAP_LD_OFFSET offsetof(struct link_map, l_ld)
855 #define LINK_MAP_NEXT_OFFSET offsetof(struct link_map, l_next)
856
857 #define INVALID_LOAD_ADDRESS -1L
858 #define ZERO_LOAD_ADDRESS 0x0L
859
860 // Calculate the load address of shared library
861 // on prelink-enabled environment.
862 //
863 // In case of GDB, it would be calculated by offset of link_map.l_ld
864 // and the address of .dynamic section.
865 // See GDB implementation: lm_addr_check @ solib-svr4.c
866 static uintptr_t calc_prelinked_load_address(struct ps_prochandle* ph, int lib_fd, ELF_EHDR* elf_ehdr, uintptr_t link_map_addr) {
867 ELF_PHDR *phbuf;
868 uintptr_t lib_ld;
869 uintptr_t lib_dyn_addr = 0L;
870 uintptr_t load_addr;
871 int i;
872
873 phbuf = read_program_header_table(lib_fd, elf_ehdr);
874 if (phbuf == NULL) {
875 print_debug("can't read program header of shared object\n");
876 return INVALID_LOAD_ADDRESS;
877 }
878
879 // Get the address of .dynamic section from shared library.
880 for (i = 0; i < elf_ehdr->e_phnum; i++) {
881 if (phbuf[i].p_type == PT_DYNAMIC) {
882 lib_dyn_addr = phbuf[i].p_vaddr;
883 break;
884 }
885 }
886
887 free(phbuf);
888
889 if (ps_pdread(ph, (psaddr_t)link_map_addr + LINK_MAP_LD_OFFSET,
890 &lib_ld, sizeof(uintptr_t)) != PS_OK) {
891 print_debug("can't read address of dynamic section in shared object\n");
892 return INVALID_LOAD_ADDRESS;
893 }
894
895 // Return the load address which is calculated by the address of .dynamic
896 // and link_map.l_ld .
897 load_addr = lib_ld - lib_dyn_addr;
898 print_debug("lib_ld = 0x%lx, lib_dyn_addr = 0x%lx -> lib_base_diff = 0x%lx\n", lib_ld, lib_dyn_addr, load_addr);
899 return load_addr;
900 }
901
902 // read shared library info from runtime linker's data structures.
903 // This work is done by librtlb_db in Solaris
904 static bool read_shared_lib_info(struct ps_prochandle* ph) {
905 uintptr_t addr = ph->core->dynamic_addr;
906 uintptr_t debug_base;
907 uintptr_t first_link_map_addr;
908 uintptr_t ld_base_addr;
909 uintptr_t link_map_addr;
910 uintptr_t lib_base_diff;
911 uintptr_t lib_base;
912 uintptr_t lib_name_addr;
913 char lib_name[BUF_SIZE];
914 ELF_DYN dyn;
915 ELF_EHDR elf_ehdr;
916 int lib_fd;
917
918 // _DYNAMIC has information of the form
919 // [tag] [data] [tag] [data] .....
920 // Both tag and data are pointer sized.
921 // We look for dynamic info with DT_DEBUG. This has shared object info.
922 // refer to struct r_debug in link.h
923
924 dyn.d_tag = DT_NULL;
925 while (dyn.d_tag != DT_DEBUG) {
926 if (ps_pdread(ph, (psaddr_t) addr, &dyn, sizeof(ELF_DYN)) != PS_OK) {
927 print_debug("can't read debug info from _DYNAMIC\n");
928 return false;
929 }
930 addr += sizeof(ELF_DYN);
931 }
932
933 // we have got Dyn entry with DT_DEBUG
934 debug_base = dyn.d_un.d_ptr;
935 // at debug_base we have struct r_debug. This has first link map in r_map field
936 if (ps_pdread(ph, (psaddr_t) debug_base + FIRST_LINK_MAP_OFFSET,
937 &first_link_map_addr, sizeof(uintptr_t)) != PS_OK) {
938 print_debug("can't read first link map address\n");
939 return false;
940 }
941
942 // read ld_base address from struct r_debug
943 if (ps_pdread(ph, (psaddr_t) debug_base + LD_BASE_OFFSET, &ld_base_addr,
944 sizeof(uintptr_t)) != PS_OK) {
945 print_debug("can't read ld base address\n");
946 return false;
947 }
948 ph->core->ld_base_addr = ld_base_addr;
949
950 print_debug("interpreter base address is 0x%lx\n", ld_base_addr);
951
952 // now read segments from interp (i.e ld.so or ld-linux.so or ld-elf.so)
953 if (read_interp_segments(ph) != true) {
954 return false;
955 }
956
957 // after adding interpreter (ld.so) mappings sort again
958 if (sort_map_array(ph) != true) {
959 return false;
960 }
961
962 print_debug("first link map is at 0x%lx\n", first_link_map_addr);
963
964 link_map_addr = first_link_map_addr;
965 while (link_map_addr != 0) {
966 // read library base address of the .so. Note that even though <sys/link.h> calls
967 // link_map->l_addr as "base address", this is * not * really base virtual
968 // address of the shared object. This is actually the difference b/w the virtual
969 // address mentioned in shared object and the actual virtual base where runtime
970 // linker loaded it. We use "base diff" in read_lib_segments call below.
971
972 if (ps_pdread(ph, (psaddr_t) link_map_addr + LINK_MAP_ADDR_OFFSET,
973 &lib_base_diff, sizeof(uintptr_t)) != PS_OK) {
974 print_debug("can't read shared object base address diff\n");
975 return false;
976 }
977
978 // read address of the name
979 if (ps_pdread(ph, (psaddr_t) link_map_addr + LINK_MAP_NAME_OFFSET,
980 &lib_name_addr, sizeof(uintptr_t)) != PS_OK) {
981 print_debug("can't read address of shared object name\n");
982 return false;
983 }
984
985 // read name of the shared object
986 lib_name[0] = '\0';
987 if (lib_name_addr != 0 &&
988 read_string(ph, (uintptr_t) lib_name_addr, lib_name, sizeof(lib_name)) != true) {
989 print_debug("can't read shared object name\n");
990 // don't let failure to read the name stop opening the file. If something is really wrong
991 // it will fail later.
992 }
993
994 if (lib_name[0] != '\0') {
995 // ignore empty lib names
996 lib_fd = pathmap_open(lib_name);
997
998 if (lib_fd < 0) {
999 print_debug("can't open shared object %s\n", lib_name);
1000 // continue with other libraries...
1001 } else {
1002 if (read_elf_header(lib_fd, &elf_ehdr)) {
1003 if (lib_base_diff == ZERO_LOAD_ADDRESS ) {
1004 lib_base_diff = calc_prelinked_load_address(ph, lib_fd, &elf_ehdr, link_map_addr);
1005 if (lib_base_diff == INVALID_LOAD_ADDRESS) {
1006 close(lib_fd);
1007 return false;
1008 }
1009 }
1010
1011 lib_base = lib_base_diff + find_base_address(lib_fd, &elf_ehdr);
1012 print_debug("reading library %s @ 0x%lx [ 0x%lx ]\n",
1013 lib_name, lib_base, lib_base_diff);
1014 // while adding library mappings we need to use "base difference".
1015 if (! read_lib_segments(ph, lib_fd, &elf_ehdr, lib_base_diff)) {
1016 print_debug("can't read shared object's segments\n");
1017 close(lib_fd);
1018 return false;
1019 }
1020 add_lib_info_fd(ph, lib_name, lib_fd, lib_base);
1021 // Map info is added for the library (lib_name) so
1022 // we need to re-sort it before calling the p_pdread.
1023 if (sort_map_array(ph) != true)
1024 return false;
1025 } else {
1026 print_debug("can't read ELF header for shared object %s\n", lib_name);
1027 close(lib_fd);
1028 // continue with other libraries...
1029 }
1030 }
1031 }
1032
1033 // read next link_map address
1034 if (ps_pdread(ph, (psaddr_t) link_map_addr + LINK_MAP_NEXT_OFFSET,
1035 &link_map_addr, sizeof(uintptr_t)) != PS_OK) {
1036 print_debug("can't read next link in link_map\n");
1037 return false;
1038 }
1039 }
1040
1041 return true;
1042 }
1043
1044 // the one and only one exposed stuff from this file
1045 JNIEXPORT struct ps_prochandle* JNICALL
1046 Pgrab_core(const char* exec_file, const char* core_file) {
1047 ELF_EHDR core_ehdr;
1048 ELF_EHDR exec_ehdr;
1049 ELF_EHDR lib_ehdr;
1050
1051 struct ps_prochandle* ph = (struct ps_prochandle*) calloc(1, sizeof(struct ps_prochandle));
1052 if (ph == NULL) {
1053 print_debug("can't allocate ps_prochandle\n");
1054 return NULL;
1055 }
1056
1057 if ((ph->core = (struct core_data*) calloc(1, sizeof(struct core_data))) == NULL) {
1058 free(ph);
1059 print_debug("can't allocate ps_prochandle\n");
1060 return NULL;
1061 }
1062
1063 // initialize ph
1064 ph->ops = &core_ops;
1065 ph->core->core_fd = -1;
1066 ph->core->exec_fd = -1;
1067 ph->core->interp_fd = -1;
1068
1069 // open the core file
1070 if ((ph->core->core_fd = open(core_file, O_RDONLY)) < 0) {
1071 print_debug("can't open core file\n");
1072 goto err;
1073 }
1074
1075 // read core file ELF header
1076 if (read_elf_header(ph->core->core_fd, &core_ehdr) != true || core_ehdr.e_type != ET_CORE) {
1077 print_debug("core file is not a valid ELF ET_CORE file\n");
1078 goto err;
1079 }
1080
1081 if ((ph->core->exec_fd = open(exec_file, O_RDONLY)) < 0) {
1082 print_debug("can't open executable file\n");
1083 goto err;
1084 }
1085
1086 if (read_elf_header(ph->core->exec_fd, &exec_ehdr) != true ||
1087 ((exec_ehdr.e_type != ET_EXEC) && (exec_ehdr.e_type != ET_DYN))) {
1088 print_debug("executable file is not a valid ELF file\n");
1089 goto err;
1090 }
1091
1092 // process core file segments
1093 if (read_core_segments(ph, &core_ehdr) != true) {
1094 goto err;
1095 }
1096
1097 // process exec file segments
1098 if (read_exec_segments(ph, &exec_ehdr) != true) {
1099 goto err;
1100 }
1101
1102 // exec file is also treated like a shared object for symbol search
1103 if (add_lib_info_fd(ph, exec_file, ph->core->exec_fd,
1104 (uintptr_t)0 + find_base_address(ph->core->exec_fd, &exec_ehdr)) == NULL) {
1105 goto err;
1106 }
1107
1108 // allocate and sort maps into map_array, we need to do this
1109 // here because read_shared_lib_info needs to read from debuggee
1110 // address space
1111 if (sort_map_array(ph) != true) {
1112 goto err;
1113 }
1114
1115 if (read_shared_lib_info(ph) != true) {
1116 goto err;
1117 }
1118
1119 // sort again because we have added more mappings from shared objects
1120 if (sort_map_array(ph) != true) {
1121 goto err;
1122 }
1123
1124 if (init_classsharing_workaround(ph) != true) {
1125 goto err;
1126 }
1127
1128 return ph;
1129
1130 err:
1131 Prelease(ph);
1132 return NULL;
1133 }