1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/agent/src/os/linux/ps_core.c Sat Dec 01 00:00:00 2007 +0000
1.3 @@ -0,0 +1,1011 @@
1.4 +/*
1.5 + * Copyright 2003-2007 Sun Microsystems, Inc. All Rights Reserved.
1.6 + * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
1.7 + *
1.8 + * This code is free software; you can redistribute it and/or modify it
1.9 + * under the terms of the GNU General Public License version 2 only, as
1.10 + * published by the Free Software Foundation.
1.11 + *
1.12 + * This code is distributed in the hope that it will be useful, but WITHOUT
1.13 + * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
1.14 + * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
1.15 + * version 2 for more details (a copy is included in the LICENSE file that
1.16 + * accompanied this code).
1.17 + *
1.18 + * You should have received a copy of the GNU General Public License version
1.19 + * 2 along with this work; if not, write to the Free Software Foundation,
1.20 + * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
1.21 + *
1.22 + * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
1.23 + * CA 95054 USA or visit www.sun.com if you need additional information or
1.24 + * have any questions.
1.25 + *
1.26 + */
1.27 +
1.28 +#include <jni.h>
1.29 +#include <unistd.h>
1.30 +#include <fcntl.h>
1.31 +#include <string.h>
1.32 +#include <stdlib.h>
1.33 +#include <stddef.h>
1.34 +#include <elf.h>
1.35 +#include <link.h>
1.36 +#include "libproc_impl.h"
1.37 +#include "salibelf.h"
1.38 +
1.39 +// This file has the libproc implementation to read core files.
1.40 +// For live processes, refer to ps_proc.c. Portions of this is adapted
1.41 +// /modelled after Solaris libproc.so (in particular Pcore.c)
1.42 +
1.43 +//----------------------------------------------------------------------
1.44 +// ps_prochandle cleanup helper functions
1.45 +
1.46 +// close all file descriptors
1.47 +static void close_elf_files(struct ps_prochandle* ph) {
1.48 + lib_info* lib = NULL;
1.49 +
1.50 + // close core file descriptor
1.51 + if (ph->core->core_fd >= 0)
1.52 + close(ph->core->core_fd);
1.53 +
1.54 + // close exec file descriptor
1.55 + if (ph->core->exec_fd >= 0)
1.56 + close(ph->core->exec_fd);
1.57 +
1.58 + // close interp file descriptor
1.59 + if (ph->core->interp_fd >= 0)
1.60 + close(ph->core->interp_fd);
1.61 +
1.62 + // close class share archive file
1.63 + if (ph->core->classes_jsa_fd >= 0)
1.64 + close(ph->core->classes_jsa_fd);
1.65 +
1.66 + // close all library file descriptors
1.67 + lib = ph->libs;
1.68 + while (lib) {
1.69 + int fd = lib->fd;
1.70 + if (fd >= 0 && fd != ph->core->exec_fd) close(fd);
1.71 + lib = lib->next;
1.72 + }
1.73 +}
1.74 +
1.75 +// clean all map_info stuff
1.76 +static void destroy_map_info(struct ps_prochandle* ph) {
1.77 + map_info* map = ph->core->maps;
1.78 + while (map) {
1.79 + map_info* next = map->next;
1.80 + free(map);
1.81 + map = next;
1.82 + }
1.83 +
1.84 + if (ph->core->map_array) {
1.85 + free(ph->core->map_array);
1.86 + }
1.87 +
1.88 + // Part of the class sharing workaround
1.89 + map = ph->core->class_share_maps;
1.90 + while (map) {
1.91 + map_info* next = map->next;
1.92 + free(map);
1.93 + map = next;
1.94 + }
1.95 +}
1.96 +
1.97 +// ps_prochandle operations
1.98 +static void core_release(struct ps_prochandle* ph) {
1.99 + if (ph->core) {
1.100 + close_elf_files(ph);
1.101 + destroy_map_info(ph);
1.102 + free(ph->core);
1.103 + }
1.104 +}
1.105 +
1.106 +static map_info* allocate_init_map(int fd, off_t offset, uintptr_t vaddr, size_t memsz) {
1.107 + map_info* map;
1.108 + if ( (map = (map_info*) calloc(1, sizeof(map_info))) == NULL) {
1.109 + print_debug("can't allocate memory for map_info\n");
1.110 + return NULL;
1.111 + }
1.112 +
1.113 + // initialize map
1.114 + map->fd = fd;
1.115 + map->offset = offset;
1.116 + map->vaddr = vaddr;
1.117 + map->memsz = memsz;
1.118 + return map;
1.119 +}
1.120 +
1.121 +// add map info with given fd, offset, vaddr and memsz
1.122 +static map_info* add_map_info(struct ps_prochandle* ph, int fd, off_t offset,
1.123 + uintptr_t vaddr, size_t memsz) {
1.124 + map_info* map;
1.125 + if ((map = allocate_init_map(fd, offset, vaddr, memsz)) == NULL) {
1.126 + return NULL;
1.127 + }
1.128 +
1.129 + // add this to map list
1.130 + map->next = ph->core->maps;
1.131 + ph->core->maps = map;
1.132 + ph->core->num_maps++;
1.133 +
1.134 + return map;
1.135 +}
1.136 +
1.137 +// Part of the class sharing workaround
1.138 +static map_info* add_class_share_map_info(struct ps_prochandle* ph, off_t offset,
1.139 + uintptr_t vaddr, size_t memsz) {
1.140 + map_info* map;
1.141 + if ((map = allocate_init_map(ph->core->classes_jsa_fd,
1.142 + offset, vaddr, memsz)) == NULL) {
1.143 + return NULL;
1.144 + }
1.145 +
1.146 + map->next = ph->core->class_share_maps;
1.147 + ph->core->class_share_maps = map;
1.148 +}
1.149 +
1.150 +// Return the map_info for the given virtual address. We keep a sorted
1.151 +// array of pointers in ph->map_array, so we can binary search.
1.152 +static map_info* core_lookup(struct ps_prochandle *ph, uintptr_t addr)
1.153 +{
1.154 + int mid, lo = 0, hi = ph->core->num_maps - 1;
1.155 + map_info *mp;
1.156 +
1.157 + while (hi - lo > 1) {
1.158 + mid = (lo + hi) / 2;
1.159 + if (addr >= ph->core->map_array[mid]->vaddr)
1.160 + lo = mid;
1.161 + else
1.162 + hi = mid;
1.163 + }
1.164 +
1.165 + if (addr < ph->core->map_array[hi]->vaddr)
1.166 + mp = ph->core->map_array[lo];
1.167 + else
1.168 + mp = ph->core->map_array[hi];
1.169 +
1.170 + if (addr >= mp->vaddr && addr < mp->vaddr + mp->memsz)
1.171 + return (mp);
1.172 +
1.173 +
1.174 + // Part of the class sharing workaround
1.175 + // Unfortunately, we have no way of detecting -Xshare state.
1.176 + // Check out the share maps atlast, if we don't find anywhere.
1.177 + // This is done this way so to avoid reading share pages
1.178 + // ahead of other normal maps. For eg. with -Xshare:off we don't
1.179 + // want to prefer class sharing data to data from core.
1.180 + mp = ph->core->class_share_maps;
1.181 + if (mp) {
1.182 + print_debug("can't locate map_info at 0x%lx, trying class share maps\n",
1.183 + addr);
1.184 + }
1.185 + while (mp) {
1.186 + if (addr >= mp->vaddr && addr < mp->vaddr + mp->memsz) {
1.187 + print_debug("located map_info at 0x%lx from class share maps\n",
1.188 + addr);
1.189 + return (mp);
1.190 + }
1.191 + mp = mp->next;
1.192 + }
1.193 +
1.194 + print_debug("can't locate map_info at 0x%lx\n", addr);
1.195 + return (NULL);
1.196 +}
1.197 +
1.198 +//---------------------------------------------------------------
1.199 +// Part of the class sharing workaround:
1.200 +//
1.201 +// With class sharing, pages are mapped from classes[_g].jsa file.
1.202 +// The read-only class sharing pages are mapped as MAP_SHARED,
1.203 +// PROT_READ pages. These pages are not dumped into core dump.
1.204 +// With this workaround, these pages are read from classes[_g].jsa.
1.205 +
1.206 +// FIXME: !HACK ALERT!
1.207 +// The format of sharing achive file header is needed to read shared heap
1.208 +// file mappings. For now, I am hard coding portion of FileMapHeader here.
1.209 +// Refer to filemap.hpp.
1.210 +
1.211 +// FileMapHeader describes the shared space data in the file to be
1.212 +// mapped. This structure gets written to a file. It is not a class,
1.213 +// so that the compilers don't add any compiler-private data to it.
1.214 +
1.215 +// Refer to CompactingPermGenGen::n_regions in compactingPermGenGen.hpp
1.216 +#define NUM_SHARED_MAPS 4
1.217 +
1.218 +// Refer to FileMapInfo::_current_version in filemap.hpp
1.219 +#define CURRENT_ARCHIVE_VERSION 1
1.220 +
1.221 +struct FileMapHeader {
1.222 + int _magic; // identify file type.
1.223 + int _version; // (from enum, above.)
1.224 + size_t _alignment; // how shared archive should be aligned
1.225 +
1.226 + struct space_info {
1.227 + int _file_offset; // sizeof(this) rounded to vm page size
1.228 + char* _base; // copy-on-write base address
1.229 + size_t _capacity; // for validity checking
1.230 + size_t _used; // for setting space top on read
1.231 +
1.232 + // 4991491 NOTICE These are C++ bool's in filemap.hpp and must match up with
1.233 + // the C type matching the C++ bool type on any given platform. For
1.234 + // Hotspot on Linux we assume the corresponding C type is char but
1.235 + // licensees on Linux versions may need to adjust the type of these fields.
1.236 + char _read_only; // read only space?
1.237 + char _allow_exec; // executable code in space?
1.238 +
1.239 + } _space[NUM_SHARED_MAPS]; // was _space[CompactingPermGenGen::n_regions];
1.240 +
1.241 + // Ignore the rest of the FileMapHeader. We don't need those fields here.
1.242 +};
1.243 +
1.244 +static bool read_int(struct ps_prochandle* ph, uintptr_t addr, int* pvalue) {
1.245 + int i;
1.246 + if (ps_pdread(ph, (psaddr_t) addr, &i, sizeof(i)) == PS_OK) {
1.247 + *pvalue = i;
1.248 + return true;
1.249 + } else {
1.250 + return false;
1.251 + }
1.252 +}
1.253 +
1.254 +static bool read_pointer(struct ps_prochandle* ph, uintptr_t addr, uintptr_t* pvalue) {
1.255 + uintptr_t uip;
1.256 + if (ps_pdread(ph, (psaddr_t) addr, &uip, sizeof(uip)) == PS_OK) {
1.257 + *pvalue = uip;
1.258 + return true;
1.259 + } else {
1.260 + return false;
1.261 + }
1.262 +}
1.263 +
1.264 +// used to read strings from debuggee
1.265 +static bool read_string(struct ps_prochandle* ph, uintptr_t addr, char* buf, size_t size) {
1.266 + size_t i = 0;
1.267 + char c = ' ';
1.268 +
1.269 + while (c != '\0') {
1.270 + if (ps_pdread(ph, (psaddr_t) addr, &c, sizeof(char)) != PS_OK)
1.271 + return false;
1.272 + if (i < size - 1)
1.273 + buf[i] = c;
1.274 + else // smaller buffer
1.275 + return false;
1.276 + i++; addr++;
1.277 + }
1.278 +
1.279 + buf[i] = '\0';
1.280 + return true;
1.281 +}
1.282 +
1.283 +#define USE_SHARED_SPACES_SYM "UseSharedSpaces"
1.284 +// mangled name of Arguments::SharedArchivePath
1.285 +#define SHARED_ARCHIVE_PATH_SYM "_ZN9Arguments17SharedArchivePathE"
1.286 +
1.287 +static bool init_classsharing_workaround(struct ps_prochandle* ph) {
1.288 + lib_info* lib = ph->libs;
1.289 + while (lib != NULL) {
1.290 + // we are iterating over shared objects from the core dump. look for
1.291 + // libjvm[_g].so.
1.292 + const char *jvm_name = 0;
1.293 + if ((jvm_name = strstr(lib->name, "/libjvm.so")) != 0 ||
1.294 + (jvm_name = strstr(lib->name, "/libjvm_g.so")) != 0) {
1.295 + char classes_jsa[PATH_MAX];
1.296 + struct FileMapHeader header;
1.297 + size_t n = 0;
1.298 + int fd = -1, m = 0;
1.299 + uintptr_t base = 0, useSharedSpacesAddr = 0;
1.300 + uintptr_t sharedArchivePathAddrAddr = 0, sharedArchivePathAddr = 0;
1.301 + int useSharedSpaces = 0;
1.302 + map_info* mi = 0;
1.303 +
1.304 + memset(classes_jsa, 0, sizeof(classes_jsa));
1.305 + jvm_name = lib->name;
1.306 + useSharedSpacesAddr = lookup_symbol(ph, jvm_name, USE_SHARED_SPACES_SYM);
1.307 + if (useSharedSpacesAddr == 0) {
1.308 + print_debug("can't lookup 'UseSharedSpaces' flag\n");
1.309 + return false;
1.310 + }
1.311 +
1.312 + if (read_int(ph, useSharedSpacesAddr, &useSharedSpaces) != true) {
1.313 + print_debug("can't read the value of 'UseSharedSpaces' flag\n");
1.314 + return false;
1.315 + }
1.316 +
1.317 + if (useSharedSpaces == 0) {
1.318 + print_debug("UseSharedSpaces is false, assuming -Xshare:off!\n");
1.319 + return true;
1.320 + }
1.321 +
1.322 + sharedArchivePathAddrAddr = lookup_symbol(ph, jvm_name, SHARED_ARCHIVE_PATH_SYM);
1.323 + if (sharedArchivePathAddrAddr == 0) {
1.324 + print_debug("can't lookup shared archive path symbol\n");
1.325 + return false;
1.326 + }
1.327 +
1.328 + if (read_pointer(ph, sharedArchivePathAddrAddr, &sharedArchivePathAddr) != true) {
1.329 + print_debug("can't read shared archive path pointer\n");
1.330 + return false;
1.331 + }
1.332 +
1.333 + if (read_string(ph, sharedArchivePathAddr, classes_jsa, sizeof(classes_jsa)) != true) {
1.334 + print_debug("can't read shared archive path value\n");
1.335 + return false;
1.336 + }
1.337 +
1.338 + print_debug("looking for %s\n", classes_jsa);
1.339 + // open the class sharing archive file
1.340 + fd = pathmap_open(classes_jsa);
1.341 + if (fd < 0) {
1.342 + print_debug("can't open %s!\n", classes_jsa);
1.343 + ph->core->classes_jsa_fd = -1;
1.344 + return false;
1.345 + } else {
1.346 + print_debug("opened %s\n", classes_jsa);
1.347 + }
1.348 +
1.349 + // read FileMapHeader from the file
1.350 + memset(&header, 0, sizeof(struct FileMapHeader));
1.351 + if ((n = read(fd, &header, sizeof(struct FileMapHeader)))
1.352 + != sizeof(struct FileMapHeader)) {
1.353 + print_debug("can't read shared archive file map header from %s\n", classes_jsa);
1.354 + close(fd);
1.355 + return false;
1.356 + }
1.357 +
1.358 + // check file magic
1.359 + if (header._magic != 0xf00baba2) {
1.360 + print_debug("%s has bad shared archive file magic number 0x%x, expecing 0xf00baba2\n",
1.361 + classes_jsa, header._magic);
1.362 + close(fd);
1.363 + return false;
1.364 + }
1.365 +
1.366 + // check version
1.367 + if (header._version != CURRENT_ARCHIVE_VERSION) {
1.368 + print_debug("%s has wrong shared archive file version %d, expecting %d\n",
1.369 + classes_jsa, header._version, CURRENT_ARCHIVE_VERSION);
1.370 + close(fd);
1.371 + return false;
1.372 + }
1.373 +
1.374 + ph->core->classes_jsa_fd = fd;
1.375 + // add read-only maps from classes[_g].jsa to the list of maps
1.376 + for (m = 0; m < NUM_SHARED_MAPS; m++) {
1.377 + if (header._space[m]._read_only) {
1.378 + base = (uintptr_t) header._space[m]._base;
1.379 + // no need to worry about the fractional pages at-the-end.
1.380 + // possible fractional pages are handled by core_read_data.
1.381 + add_class_share_map_info(ph, (off_t) header._space[m]._file_offset,
1.382 + base, (size_t) header._space[m]._used);
1.383 + print_debug("added a share archive map at 0x%lx\n", base);
1.384 + }
1.385 + }
1.386 + return true;
1.387 + }
1.388 + lib = lib->next;
1.389 + }
1.390 + return true;
1.391 +}
1.392 +
1.393 +
1.394 +//---------------------------------------------------------------------------
1.395 +// functions to handle map_info
1.396 +
1.397 +// Order mappings based on virtual address. We use this function as the
1.398 +// callback for sorting the array of map_info pointers.
1.399 +static int core_cmp_mapping(const void *lhsp, const void *rhsp)
1.400 +{
1.401 + const map_info *lhs = *((const map_info **)lhsp);
1.402 + const map_info *rhs = *((const map_info **)rhsp);
1.403 +
1.404 + if (lhs->vaddr == rhs->vaddr)
1.405 + return (0);
1.406 +
1.407 + return (lhs->vaddr < rhs->vaddr ? -1 : 1);
1.408 +}
1.409 +
1.410 +// we sort map_info by starting virtual address so that we can do
1.411 +// binary search to read from an address.
1.412 +static bool sort_map_array(struct ps_prochandle* ph) {
1.413 + size_t num_maps = ph->core->num_maps;
1.414 + map_info* map = ph->core->maps;
1.415 + int i = 0;
1.416 +
1.417 + // allocate map_array
1.418 + map_info** array;
1.419 + if ( (array = (map_info**) malloc(sizeof(map_info*) * num_maps)) == NULL) {
1.420 + print_debug("can't allocate memory for map array\n");
1.421 + return false;
1.422 + }
1.423 +
1.424 + // add maps to array
1.425 + while (map) {
1.426 + array[i] = map;
1.427 + i++;
1.428 + map = map->next;
1.429 + }
1.430 +
1.431 + // sort is called twice. If this is second time, clear map array
1.432 + if (ph->core->map_array) free(ph->core->map_array);
1.433 + ph->core->map_array = array;
1.434 + // sort the map_info array by base virtual address.
1.435 + qsort(ph->core->map_array, ph->core->num_maps, sizeof (map_info*),
1.436 + core_cmp_mapping);
1.437 +
1.438 + // print map
1.439 + if (is_debug()) {
1.440 + int j = 0;
1.441 + print_debug("---- sorted virtual address map ----\n");
1.442 + for (j = 0; j < ph->core->num_maps; j++) {
1.443 + print_debug("base = 0x%lx\tsize = %d\n", ph->core->map_array[j]->vaddr,
1.444 + ph->core->map_array[j]->memsz);
1.445 + }
1.446 + }
1.447 +
1.448 + return true;
1.449 +}
1.450 +
1.451 +#ifndef MIN
1.452 +#define MIN(x, y) (((x) < (y))? (x): (y))
1.453 +#endif
1.454 +
1.455 +static bool core_read_data(struct ps_prochandle* ph, uintptr_t addr, char *buf, size_t size) {
1.456 + ssize_t resid = size;
1.457 + int page_size=sysconf(_SC_PAGE_SIZE);
1.458 + while (resid != 0) {
1.459 + map_info *mp = core_lookup(ph, addr);
1.460 + uintptr_t mapoff;
1.461 + ssize_t len, rem;
1.462 + off_t off;
1.463 + int fd;
1.464 +
1.465 + if (mp == NULL)
1.466 + break; /* No mapping for this address */
1.467 +
1.468 + fd = mp->fd;
1.469 + mapoff = addr - mp->vaddr;
1.470 + len = MIN(resid, mp->memsz - mapoff);
1.471 + off = mp->offset + mapoff;
1.472 +
1.473 + if ((len = pread(fd, buf, len, off)) <= 0)
1.474 + break;
1.475 +
1.476 + resid -= len;
1.477 + addr += len;
1.478 + buf = (char *)buf + len;
1.479 +
1.480 + // mappings always start at page boundary. But, may end in fractional
1.481 + // page. fill zeros for possible fractional page at the end of a mapping.
1.482 + rem = mp->memsz % page_size;
1.483 + if (rem > 0) {
1.484 + rem = page_size - rem;
1.485 + len = MIN(resid, rem);
1.486 + resid -= len;
1.487 + addr += len;
1.488 + // we are not assuming 'buf' to be zero initialized.
1.489 + memset(buf, 0, len);
1.490 + buf += len;
1.491 + }
1.492 + }
1.493 +
1.494 + if (resid) {
1.495 + print_debug("core read failed for %d byte(s) @ 0x%lx (%d more bytes)\n",
1.496 + size, addr, resid);
1.497 + return false;
1.498 + } else {
1.499 + return true;
1.500 + }
1.501 +}
1.502 +
1.503 +// null implementation for write
1.504 +static bool core_write_data(struct ps_prochandle* ph,
1.505 + uintptr_t addr, const char *buf , size_t size) {
1.506 + return false;
1.507 +}
1.508 +
1.509 +static bool core_get_lwp_regs(struct ps_prochandle* ph, lwpid_t lwp_id,
1.510 + struct user_regs_struct* regs) {
1.511 + // for core we have cached the lwp regs from NOTE section
1.512 + thread_info* thr = ph->threads;
1.513 + while (thr) {
1.514 + if (thr->lwp_id == lwp_id) {
1.515 + memcpy(regs, &thr->regs, sizeof(struct user_regs_struct));
1.516 + return true;
1.517 + }
1.518 + thr = thr->next;
1.519 + }
1.520 + return false;
1.521 +}
1.522 +
1.523 +static ps_prochandle_ops core_ops = {
1.524 + release: core_release,
1.525 + p_pread: core_read_data,
1.526 + p_pwrite: core_write_data,
1.527 + get_lwp_regs: core_get_lwp_regs
1.528 +};
1.529 +
1.530 +// read regs and create thread from NT_PRSTATUS entries from core file
1.531 +static bool core_handle_prstatus(struct ps_prochandle* ph, const char* buf, size_t nbytes) {
1.532 + // we have to read prstatus_t from buf
1.533 + // assert(nbytes == sizeof(prstaus_t), "size mismatch on prstatus_t");
1.534 + prstatus_t* prstat = (prstatus_t*) buf;
1.535 + thread_info* newthr;
1.536 + print_debug("got integer regset for lwp %d\n", prstat->pr_pid);
1.537 + // we set pthread_t to -1 for core dump
1.538 + if((newthr = add_thread_info(ph, (pthread_t) -1, prstat->pr_pid)) == NULL)
1.539 + return false;
1.540 +
1.541 + // copy regs
1.542 + memcpy(&newthr->regs, prstat->pr_reg, sizeof(struct user_regs_struct));
1.543 +
1.544 + if (is_debug()) {
1.545 + print_debug("integer regset\n");
1.546 +#ifdef i386
1.547 + // print the regset
1.548 + print_debug("\teax = 0x%x\n", newthr->regs.eax);
1.549 + print_debug("\tebx = 0x%x\n", newthr->regs.ebx);
1.550 + print_debug("\tecx = 0x%x\n", newthr->regs.ecx);
1.551 + print_debug("\tedx = 0x%x\n", newthr->regs.edx);
1.552 + print_debug("\tesp = 0x%x\n", newthr->regs.esp);
1.553 + print_debug("\tebp = 0x%x\n", newthr->regs.ebp);
1.554 + print_debug("\tesi = 0x%x\n", newthr->regs.esi);
1.555 + print_debug("\tedi = 0x%x\n", newthr->regs.edi);
1.556 + print_debug("\teip = 0x%x\n", newthr->regs.eip);
1.557 +#endif
1.558 +
1.559 +#if defined(amd64) || defined(x86_64)
1.560 + // print the regset
1.561 + print_debug("\tr15 = 0x%lx\n", newthr->regs.r15);
1.562 + print_debug("\tr14 = 0x%lx\n", newthr->regs.r14);
1.563 + print_debug("\tr13 = 0x%lx\n", newthr->regs.r13);
1.564 + print_debug("\tr12 = 0x%lx\n", newthr->regs.r12);
1.565 + print_debug("\trbp = 0x%lx\n", newthr->regs.rbp);
1.566 + print_debug("\trbx = 0x%lx\n", newthr->regs.rbx);
1.567 + print_debug("\tr11 = 0x%lx\n", newthr->regs.r11);
1.568 + print_debug("\tr10 = 0x%lx\n", newthr->regs.r10);
1.569 + print_debug("\tr9 = 0x%lx\n", newthr->regs.r9);
1.570 + print_debug("\tr8 = 0x%lx\n", newthr->regs.r8);
1.571 + print_debug("\trax = 0x%lx\n", newthr->regs.rax);
1.572 + print_debug("\trcx = 0x%lx\n", newthr->regs.rcx);
1.573 + print_debug("\trdx = 0x%lx\n", newthr->regs.rdx);
1.574 + print_debug("\trsi = 0x%lx\n", newthr->regs.rsi);
1.575 + print_debug("\trdi = 0x%lx\n", newthr->regs.rdi);
1.576 + print_debug("\torig_rax = 0x%lx\n", newthr->regs.orig_rax);
1.577 + print_debug("\trip = 0x%lx\n", newthr->regs.rip);
1.578 + print_debug("\tcs = 0x%lx\n", newthr->regs.cs);
1.579 + print_debug("\teflags = 0x%lx\n", newthr->regs.eflags);
1.580 + print_debug("\trsp = 0x%lx\n", newthr->regs.rsp);
1.581 + print_debug("\tss = 0x%lx\n", newthr->regs.ss);
1.582 + print_debug("\tfs_base = 0x%lx\n", newthr->regs.fs_base);
1.583 + print_debug("\tgs_base = 0x%lx\n", newthr->regs.gs_base);
1.584 + print_debug("\tds = 0x%lx\n", newthr->regs.ds);
1.585 + print_debug("\tes = 0x%lx\n", newthr->regs.es);
1.586 + print_debug("\tfs = 0x%lx\n", newthr->regs.fs);
1.587 + print_debug("\tgs = 0x%lx\n", newthr->regs.gs);
1.588 +#endif
1.589 + }
1.590 +
1.591 + return true;
1.592 +}
1.593 +
1.594 +#define ROUNDUP(x, y) ((((x)+((y)-1))/(y))*(y))
1.595 +
1.596 +// read NT_PRSTATUS entries from core NOTE segment
1.597 +static bool core_handle_note(struct ps_prochandle* ph, ELF_PHDR* note_phdr) {
1.598 + char* buf = NULL;
1.599 + char* p = NULL;
1.600 + size_t size = note_phdr->p_filesz;
1.601 +
1.602 + // we are interested in just prstatus entries. we will ignore the rest.
1.603 + // Advance the seek pointer to the start of the PT_NOTE data
1.604 + if (lseek(ph->core->core_fd, note_phdr->p_offset, SEEK_SET) == (off_t)-1) {
1.605 + print_debug("failed to lseek to PT_NOTE data\n");
1.606 + return false;
1.607 + }
1.608 +
1.609 + // Now process the PT_NOTE structures. Each one is preceded by
1.610 + // an Elf{32/64}_Nhdr structure describing its type and size.
1.611 + if ( (buf = (char*) malloc(size)) == NULL) {
1.612 + print_debug("can't allocate memory for reading core notes\n");
1.613 + goto err;
1.614 + }
1.615 +
1.616 + // read notes into buffer
1.617 + if (read(ph->core->core_fd, buf, size) != size) {
1.618 + print_debug("failed to read notes, core file must have been truncated\n");
1.619 + goto err;
1.620 + }
1.621 +
1.622 + p = buf;
1.623 + while (p < buf + size) {
1.624 + ELF_NHDR* notep = (ELF_NHDR*) p;
1.625 + char* descdata = p + sizeof(ELF_NHDR) + ROUNDUP(notep->n_namesz, 4);
1.626 + print_debug("Note header with n_type = %d and n_descsz = %u\n",
1.627 + notep->n_type, notep->n_descsz);
1.628 +
1.629 + if (notep->n_type == NT_PRSTATUS) {
1.630 + if (core_handle_prstatus(ph, descdata, notep->n_descsz) != true)
1.631 + return false;
1.632 + }
1.633 + p = descdata + ROUNDUP(notep->n_descsz, 4);
1.634 + }
1.635 +
1.636 + free(buf);
1.637 + return true;
1.638 +
1.639 +err:
1.640 + if (buf) free(buf);
1.641 + return false;
1.642 +}
1.643 +
1.644 +// read all segments from core file
1.645 +static bool read_core_segments(struct ps_prochandle* ph, ELF_EHDR* core_ehdr) {
1.646 + int i = 0;
1.647 + ELF_PHDR* phbuf = NULL;
1.648 + ELF_PHDR* core_php = NULL;
1.649 +
1.650 + if ((phbuf = read_program_header_table(ph->core->core_fd, core_ehdr)) == NULL)
1.651 + return false;
1.652 +
1.653 + /*
1.654 + * Now iterate through the program headers in the core file.
1.655 + * We're interested in two types of Phdrs: PT_NOTE (which
1.656 + * contains a set of saved /proc structures), and PT_LOAD (which
1.657 + * represents a memory mapping from the process's address space).
1.658 + *
1.659 + * Difference b/w Solaris PT_NOTE and Linux PT_NOTE:
1.660 + *
1.661 + * In Solaris there are two PT_NOTE segments the first PT_NOTE (if present)
1.662 + * contains /proc structs in the pre-2.6 unstructured /proc format. the last
1.663 + * PT_NOTE has data in new /proc format.
1.664 + *
1.665 + * In Solaris, there is only one pstatus (process status). pstatus contains
1.666 + * integer register set among other stuff. For each LWP, we have one lwpstatus
1.667 + * entry that has integer regset for that LWP.
1.668 + *
1.669 + * Linux threads are actually 'clone'd processes. To support core analysis
1.670 + * of "multithreaded" process, Linux creates more than one pstatus (called
1.671 + * "prstatus") entry in PT_NOTE. Each prstatus entry has integer regset for one
1.672 + * "thread". Please refer to Linux kernel src file 'fs/binfmt_elf.c', in particular
1.673 + * function "elf_core_dump".
1.674 + */
1.675 +
1.676 + for (core_php = phbuf, i = 0; i < core_ehdr->e_phnum; i++) {
1.677 + switch (core_php->p_type) {
1.678 + case PT_NOTE:
1.679 + if (core_handle_note(ph, core_php) != true) goto err;
1.680 + break;
1.681 +
1.682 + case PT_LOAD: {
1.683 + if (core_php->p_filesz != 0) {
1.684 + if (add_map_info(ph, ph->core->core_fd, core_php->p_offset,
1.685 + core_php->p_vaddr, core_php->p_filesz) == NULL) goto err;
1.686 + }
1.687 + break;
1.688 + }
1.689 + }
1.690 +
1.691 + core_php++;
1.692 + }
1.693 +
1.694 + free(phbuf);
1.695 + return true;
1.696 +err:
1.697 + free(phbuf);
1.698 + return false;
1.699 +}
1.700 +
1.701 +// read segments of a shared object
1.702 +static bool read_lib_segments(struct ps_prochandle* ph, int lib_fd, ELF_EHDR* lib_ehdr, uintptr_t lib_base) {
1.703 + int i = 0;
1.704 + ELF_PHDR* phbuf;
1.705 + ELF_PHDR* lib_php = NULL;
1.706 +
1.707 + if ((phbuf = read_program_header_table(lib_fd, lib_ehdr)) == NULL)
1.708 + return false;
1.709 +
1.710 + // we want to process only PT_LOAD segments that are not writable.
1.711 + // i.e., text segments. The read/write/exec (data) segments would
1.712 + // have been already added from core file segments.
1.713 + for (lib_php = phbuf, i = 0; i < lib_ehdr->e_phnum; i++) {
1.714 + if ((lib_php->p_type == PT_LOAD) && !(lib_php->p_flags & PF_W) && (lib_php->p_filesz != 0)) {
1.715 + if (add_map_info(ph, lib_fd, lib_php->p_offset, lib_php->p_vaddr + lib_base, lib_php->p_filesz) == NULL)
1.716 + goto err;
1.717 + }
1.718 + lib_php++;
1.719 + }
1.720 +
1.721 + free(phbuf);
1.722 + return true;
1.723 +err:
1.724 + free(phbuf);
1.725 + return false;
1.726 +}
1.727 +
1.728 +// process segments from interpreter (ld.so or ld-linux.so)
1.729 +static bool read_interp_segments(struct ps_prochandle* ph) {
1.730 + ELF_EHDR interp_ehdr;
1.731 +
1.732 + if (read_elf_header(ph->core->interp_fd, &interp_ehdr) != true) {
1.733 + print_debug("interpreter is not a valid ELF file\n");
1.734 + return false;
1.735 + }
1.736 +
1.737 + if (read_lib_segments(ph, ph->core->interp_fd, &interp_ehdr, ph->core->ld_base_addr) != true) {
1.738 + print_debug("can't read segments of interpreter\n");
1.739 + return false;
1.740 + }
1.741 +
1.742 + return true;
1.743 +}
1.744 +
1.745 +// process segments of a a.out
1.746 +static bool read_exec_segments(struct ps_prochandle* ph, ELF_EHDR* exec_ehdr) {
1.747 + int i = 0;
1.748 + ELF_PHDR* phbuf = NULL;
1.749 + ELF_PHDR* exec_php = NULL;
1.750 +
1.751 + if ((phbuf = read_program_header_table(ph->core->exec_fd, exec_ehdr)) == NULL)
1.752 + return false;
1.753 +
1.754 + for (exec_php = phbuf, i = 0; i < exec_ehdr->e_phnum; i++) {
1.755 + switch (exec_php->p_type) {
1.756 +
1.757 + // add mappings for PT_LOAD segments
1.758 + case PT_LOAD: {
1.759 + // add only non-writable segments of non-zero filesz
1.760 + if (!(exec_php->p_flags & PF_W) && exec_php->p_filesz != 0) {
1.761 + if (add_map_info(ph, ph->core->exec_fd, exec_php->p_offset, exec_php->p_vaddr, exec_php->p_filesz) == NULL) goto err;
1.762 + }
1.763 + break;
1.764 + }
1.765 +
1.766 + // read the interpreter and it's segments
1.767 + case PT_INTERP: {
1.768 + char interp_name[BUF_SIZE];
1.769 +
1.770 + pread(ph->core->exec_fd, interp_name, MIN(exec_php->p_filesz, BUF_SIZE), exec_php->p_offset);
1.771 + print_debug("ELF interpreter %s\n", interp_name);
1.772 + // read interpreter segments as well
1.773 + if ((ph->core->interp_fd = pathmap_open(interp_name)) < 0) {
1.774 + print_debug("can't open runtime loader\n");
1.775 + goto err;
1.776 + }
1.777 + break;
1.778 + }
1.779 +
1.780 + // from PT_DYNAMIC we want to read address of first link_map addr
1.781 + case PT_DYNAMIC: {
1.782 + ph->core->dynamic_addr = exec_php->p_vaddr;
1.783 + print_debug("address of _DYNAMIC is 0x%lx\n", ph->core->dynamic_addr);
1.784 + break;
1.785 + }
1.786 +
1.787 + } // switch
1.788 + exec_php++;
1.789 + } // for
1.790 +
1.791 + free(phbuf);
1.792 + return true;
1.793 +err:
1.794 + free(phbuf);
1.795 + return false;
1.796 +}
1.797 +
1.798 +
1.799 +#define FIRST_LINK_MAP_OFFSET offsetof(struct r_debug, r_map)
1.800 +#define LD_BASE_OFFSET offsetof(struct r_debug, r_ldbase)
1.801 +#define LINK_MAP_ADDR_OFFSET offsetof(struct link_map, l_addr)
1.802 +#define LINK_MAP_NAME_OFFSET offsetof(struct link_map, l_name)
1.803 +#define LINK_MAP_NEXT_OFFSET offsetof(struct link_map, l_next)
1.804 +
1.805 +// read shared library info from runtime linker's data structures.
1.806 +// This work is done by librtlb_db in Solaris
1.807 +static bool read_shared_lib_info(struct ps_prochandle* ph) {
1.808 + uintptr_t addr = ph->core->dynamic_addr;
1.809 + uintptr_t debug_base;
1.810 + uintptr_t first_link_map_addr;
1.811 + uintptr_t ld_base_addr;
1.812 + uintptr_t link_map_addr;
1.813 + uintptr_t lib_base_diff;
1.814 + uintptr_t lib_base;
1.815 + uintptr_t lib_name_addr;
1.816 + char lib_name[BUF_SIZE];
1.817 + ELF_DYN dyn;
1.818 + ELF_EHDR elf_ehdr;
1.819 + int lib_fd;
1.820 +
1.821 + // _DYNAMIC has information of the form
1.822 + // [tag] [data] [tag] [data] .....
1.823 + // Both tag and data are pointer sized.
1.824 + // We look for dynamic info with DT_DEBUG. This has shared object info.
1.825 + // refer to struct r_debug in link.h
1.826 +
1.827 + dyn.d_tag = DT_NULL;
1.828 + while (dyn.d_tag != DT_DEBUG) {
1.829 + if (ps_pdread(ph, (psaddr_t) addr, &dyn, sizeof(ELF_DYN)) != PS_OK) {
1.830 + print_debug("can't read debug info from _DYNAMIC\n");
1.831 + return false;
1.832 + }
1.833 + addr += sizeof(ELF_DYN);
1.834 + }
1.835 +
1.836 + // we have got Dyn entry with DT_DEBUG
1.837 + debug_base = dyn.d_un.d_ptr;
1.838 + // at debug_base we have struct r_debug. This has first link map in r_map field
1.839 + if (ps_pdread(ph, (psaddr_t) debug_base + FIRST_LINK_MAP_OFFSET,
1.840 + &first_link_map_addr, sizeof(uintptr_t)) != PS_OK) {
1.841 + print_debug("can't read first link map address\n");
1.842 + return false;
1.843 + }
1.844 +
1.845 + // read ld_base address from struct r_debug
1.846 + if (ps_pdread(ph, (psaddr_t) debug_base + LD_BASE_OFFSET, &ld_base_addr,
1.847 + sizeof(uintptr_t)) != PS_OK) {
1.848 + print_debug("can't read ld base address\n");
1.849 + return false;
1.850 + }
1.851 + ph->core->ld_base_addr = ld_base_addr;
1.852 +
1.853 + print_debug("interpreter base address is 0x%lx\n", ld_base_addr);
1.854 +
1.855 + // now read segments from interp (i.e ld.so or ld-linux.so)
1.856 + if (read_interp_segments(ph) != true)
1.857 + return false;
1.858 +
1.859 + // after adding interpreter (ld.so) mappings sort again
1.860 + if (sort_map_array(ph) != true)
1.861 + return false;
1.862 +
1.863 + print_debug("first link map is at 0x%lx\n", first_link_map_addr);
1.864 +
1.865 + link_map_addr = first_link_map_addr;
1.866 + while (link_map_addr != 0) {
1.867 + // read library base address of the .so. Note that even though <sys/link.h> calls
1.868 + // link_map->l_addr as "base address", this is * not * really base virtual
1.869 + // address of the shared object. This is actually the difference b/w the virtual
1.870 + // address mentioned in shared object and the actual virtual base where runtime
1.871 + // linker loaded it. We use "base diff" in read_lib_segments call below.
1.872 +
1.873 + if (ps_pdread(ph, (psaddr_t) link_map_addr + LINK_MAP_ADDR_OFFSET,
1.874 + &lib_base_diff, sizeof(uintptr_t)) != PS_OK) {
1.875 + print_debug("can't read shared object base address diff\n");
1.876 + return false;
1.877 + }
1.878 +
1.879 + // read address of the name
1.880 + if (ps_pdread(ph, (psaddr_t) link_map_addr + LINK_MAP_NAME_OFFSET,
1.881 + &lib_name_addr, sizeof(uintptr_t)) != PS_OK) {
1.882 + print_debug("can't read address of shared object name\n");
1.883 + return false;
1.884 + }
1.885 +
1.886 + // read name of the shared object
1.887 + if (read_string(ph, (uintptr_t) lib_name_addr, lib_name, sizeof(lib_name)) != true) {
1.888 + print_debug("can't read shared object name\n");
1.889 + return false;
1.890 + }
1.891 +
1.892 + if (lib_name[0] != '\0') {
1.893 + // ignore empty lib names
1.894 + lib_fd = pathmap_open(lib_name);
1.895 +
1.896 + if (lib_fd < 0) {
1.897 + print_debug("can't open shared object %s\n", lib_name);
1.898 + // continue with other libraries...
1.899 + } else {
1.900 + if (read_elf_header(lib_fd, &elf_ehdr)) {
1.901 + lib_base = lib_base_diff + find_base_address(lib_fd, &elf_ehdr);
1.902 + print_debug("reading library %s @ 0x%lx [ 0x%lx ]\n",
1.903 + lib_name, lib_base, lib_base_diff);
1.904 + // while adding library mappings we need to use "base difference".
1.905 + if (! read_lib_segments(ph, lib_fd, &elf_ehdr, lib_base_diff)) {
1.906 + print_debug("can't read shared object's segments\n");
1.907 + close(lib_fd);
1.908 + return false;
1.909 + }
1.910 + add_lib_info_fd(ph, lib_name, lib_fd, lib_base);
1.911 + // Map info is added for the library (lib_name) so
1.912 + // we need to re-sort it before calling the p_pdread.
1.913 + if (sort_map_array(ph) != true)
1.914 + return false;
1.915 + } else {
1.916 + print_debug("can't read ELF header for shared object %s\n", lib_name);
1.917 + close(lib_fd);
1.918 + // continue with other libraries...
1.919 + }
1.920 + }
1.921 + }
1.922 +
1.923 + // read next link_map address
1.924 + if (ps_pdread(ph, (psaddr_t) link_map_addr + LINK_MAP_NEXT_OFFSET,
1.925 + &link_map_addr, sizeof(uintptr_t)) != PS_OK) {
1.926 + print_debug("can't read next link in link_map\n");
1.927 + return false;
1.928 + }
1.929 + }
1.930 +
1.931 + return true;
1.932 +}
1.933 +
1.934 +// the one and only one exposed stuff from this file
1.935 +struct ps_prochandle* Pgrab_core(const char* exec_file, const char* core_file) {
1.936 + ELF_EHDR core_ehdr;
1.937 + ELF_EHDR exec_ehdr;
1.938 + ELF_EHDR lib_ehdr;
1.939 +
1.940 + struct ps_prochandle* ph = (struct ps_prochandle*) calloc(1, sizeof(struct ps_prochandle));
1.941 + if (ph == NULL) {
1.942 + print_debug("can't allocate ps_prochandle\n");
1.943 + return NULL;
1.944 + }
1.945 +
1.946 + if ((ph->core = (struct core_data*) calloc(1, sizeof(struct core_data))) == NULL) {
1.947 + free(ph);
1.948 + print_debug("can't allocate ps_prochandle\n");
1.949 + return NULL;
1.950 + }
1.951 +
1.952 + // initialize ph
1.953 + ph->ops = &core_ops;
1.954 + ph->core->core_fd = -1;
1.955 + ph->core->exec_fd = -1;
1.956 + ph->core->interp_fd = -1;
1.957 +
1.958 + // open the core file
1.959 + if ((ph->core->core_fd = open(core_file, O_RDONLY)) < 0) {
1.960 + print_debug("can't open core file\n");
1.961 + goto err;
1.962 + }
1.963 +
1.964 + // read core file ELF header
1.965 + if (read_elf_header(ph->core->core_fd, &core_ehdr) != true || core_ehdr.e_type != ET_CORE) {
1.966 + print_debug("core file is not a valid ELF ET_CORE file\n");
1.967 + goto err;
1.968 + }
1.969 +
1.970 + if ((ph->core->exec_fd = open(exec_file, O_RDONLY)) < 0) {
1.971 + print_debug("can't open executable file\n");
1.972 + goto err;
1.973 + }
1.974 +
1.975 + if (read_elf_header(ph->core->exec_fd, &exec_ehdr) != true || exec_ehdr.e_type != ET_EXEC) {
1.976 + print_debug("executable file is not a valid ELF ET_EXEC file\n");
1.977 + goto err;
1.978 + }
1.979 +
1.980 + // process core file segments
1.981 + if (read_core_segments(ph, &core_ehdr) != true)
1.982 + goto err;
1.983 +
1.984 + // process exec file segments
1.985 + if (read_exec_segments(ph, &exec_ehdr) != true)
1.986 + goto err;
1.987 +
1.988 + // exec file is also treated like a shared object for symbol search
1.989 + if (add_lib_info_fd(ph, exec_file, ph->core->exec_fd,
1.990 + (uintptr_t)0 + find_base_address(ph->core->exec_fd, &exec_ehdr)) == NULL)
1.991 + goto err;
1.992 +
1.993 + // allocate and sort maps into map_array, we need to do this
1.994 + // here because read_shared_lib_info needs to read from debuggee
1.995 + // address space
1.996 + if (sort_map_array(ph) != true)
1.997 + goto err;
1.998 +
1.999 + if (read_shared_lib_info(ph) != true)
1.1000 + goto err;
1.1001 +
1.1002 + // sort again because we have added more mappings from shared objects
1.1003 + if (sort_map_array(ph) != true)
1.1004 + goto err;
1.1005 +
1.1006 + if (init_classsharing_workaround(ph) != true)
1.1007 + goto err;
1.1008 +
1.1009 + return ph;
1.1010 +
1.1011 +err:
1.1012 + Prelease(ph);
1.1013 + return NULL;
1.1014 +}
