1 /* Copyright (c) 2008-2014. The SimGrid Team.
2 * All rights reserved. */
4 /* This program is free software; you can redistribute it and/or modify it
5 * under the terms of the license (GNU LGPL) which comes with this package. */
13 #include "internal_config.h"
14 #include "mc_memory_map.h"
15 #include "mc_private.h"
16 #include "xbt/module.h"
17 #include <xbt/mmalloc.h>
18 #include "../smpi/private.h"
21 #include "xbt/mmalloc/mmprivate.h"
23 #include "../simix/smx_private.h"
25 #include <libunwind.h>
28 #include "mc_private.h"
31 #include "mc_snapshot.h"
32 #include "mc_object_info.h"
35 #include "mc_protocol.h"
38 using simgrid::mc::remote;
42 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(mc_checkpoint, mc,
43 "Logging specific to mc_checkpoint");
45 /************************************ Free functions **************************************/
46 /*****************************************************************************************/
48 static void MC_snapshot_stack_free(mc_snapshot_stack_t s)
51 xbt_dynar_free(&(s->local_variables));
52 xbt_dynar_free(&(s->stack_frames));
53 mc_unw_destroy_context(s->context);
59 static void MC_snapshot_stack_free_voidp(void *s)
61 mc_snapshot_stack_t stack = (mc_snapshot_stack_t) * (void **) s;
62 MC_snapshot_stack_free(stack);
65 static void local_variable_free(local_variable_t v)
71 static void local_variable_free_voidp(void *v)
73 local_variable_free((local_variable_t) * (void **) v);
81 RegionSnapshot::~RegionSnapshot() {}
86 /******************************* Snapshot regions ********************************/
87 /*********************************************************************************/
91 static simgrid::mc::RegionSnapshot MC_region_dense(
92 mc_region_type_t region_type,
93 void *start_addr, void* permanent_addr, size_t size)
95 std::vector<char> data(size);
96 mc_model_checker->process().read_bytes(data.data(), size,
97 remote(permanent_addr),
98 simgrid::mc::ProcessIndexDisabled);
100 simgrid::mc::RegionSnapshot region(
101 region_type, start_addr, permanent_addr, size);
102 region.flat_data(std::move(data));
104 XBT_DEBUG("New region : type : %d, data : %p (real addr %p), size : %zu",
105 region_type, region.flat_data().data(), permanent_addr, size);
106 return std::move(region);
109 /** @brief Take a snapshot of a given region
112 * @param start_addr Address of the region in the simulated process
113 * @param permanent_addr Permanent address of this data (for privatized variables, this is the virtual address of the privatized mapping)
114 * @param size Size of the data*
116 static simgrid::mc::RegionSnapshot MC_region(
117 mc_region_type_t type, void *start_addr, void* permanent_addr, size_t size)
119 if (_sg_mc_sparse_checkpoint) {
120 return MC_region_sparse(type, start_addr, permanent_addr, size);
122 return MC_region_dense(type, start_addr, permanent_addr, size);
126 /** @brief Restore a region from a snapshot
128 * @param reg Target region
130 static void MC_region_restore(mc_mem_region_t region)
132 switch(region->storage_type()) {
133 case MC_REGION_STORAGE_TYPE_NONE:
135 xbt_die("Storage type not supported");
138 case MC_REGION_STORAGE_TYPE_FLAT:
139 mc_model_checker->process().write_bytes(region->flat_data().data(),
140 region->size(), region->permanent_address());
143 case MC_REGION_STORAGE_TYPE_CHUNKED:
144 mc_region_restore_sparse(&mc_model_checker->process(), region);
147 case MC_REGION_STORAGE_TYPE_PRIVATIZED:
148 for (auto& p : region->privatized_data())
149 MC_region_restore(&p);
154 static simgrid::mc::RegionSnapshot MC_region_privatized(
155 mc_region_type_t region_type, void *start_addr, void* permanent_addr, size_t size
158 size_t process_count = MC_smpi_process_count();
160 // Read smpi_privatisation_regions from MCed:
161 smpi_privatisation_region_t remote_smpi_privatisation_regions;
162 mc_model_checker->process().read_variable(
163 "smpi_privatisation_regions",
164 &remote_smpi_privatisation_regions, sizeof(remote_smpi_privatisation_regions));
165 s_smpi_privatisation_region_t privatisation_regions[process_count];
166 mc_model_checker->process().read_bytes(
167 &privatisation_regions, sizeof(privatisation_regions),
168 remote(remote_smpi_privatisation_regions));
170 std::vector<simgrid::mc::RegionSnapshot> data;
171 data.reserve(process_count);
172 for (size_t i = 0; i < process_count; i++)
174 MC_region(region_type, start_addr,
175 privatisation_regions[i].address, size)
178 simgrid::mc::RegionSnapshot region = simgrid::mc::RegionSnapshot(
179 region_type, start_addr, permanent_addr, size);
180 region.privatized_data(std::move(data));
181 return std::move(region);
184 static void MC_snapshot_add_region(int index, mc_snapshot_t snapshot, mc_region_type_t type,
185 mc_object_info_t object_info,
186 void *start_addr, void* permanent_addr, size_t size)
188 if (type == MC_REGION_TYPE_DATA)
189 xbt_assert(object_info, "Missing object info for object.");
190 else if (type == MC_REGION_TYPE_HEAP)
191 xbt_assert(!object_info, "Unexpected object info for heap region.");
193 const bool privatization_aware = MC_object_info_is_privatized(object_info);
195 simgrid::mc::RegionSnapshot region;
196 if (privatization_aware && MC_smpi_process_count())
197 region = MC_region_privatized(type, start_addr, permanent_addr, size);
199 region = MC_region(type, start_addr, permanent_addr, size);
201 region.object_info(object_info);
202 snapshot->snapshot_regions[index]
203 = new simgrid::mc::RegionSnapshot(std::move(region));
207 static void MC_get_memory_regions(mc_process_t process, mc_snapshot_t snapshot)
209 const size_t n = process->object_infos_size;
210 snapshot->snapshot_regions_count = n + 1;
211 snapshot->snapshot_regions = xbt_new0(mc_mem_region_t, n + 1);
213 for (size_t i = 0; i!=n; ++i) {
214 mc_object_info_t object_info = process->object_infos[i];
215 MC_snapshot_add_region(i, snapshot, MC_REGION_TYPE_DATA, object_info,
216 object_info->start_rw, object_info->start_rw,
217 object_info->end_rw - object_info->start_rw);
220 xbt_mheap_t heap = process->get_heap();
221 void *start_heap = heap->base;
222 void *end_heap = heap->breakval;
224 MC_snapshot_add_region(n, snapshot, MC_REGION_TYPE_HEAP, NULL,
225 start_heap, start_heap,
226 (char *) end_heap - (char *) start_heap);
227 snapshot->heap_bytes_used = mmalloc_get_bytes_used_remote(
229 process->get_malloc_info());
232 if (smpi_privatize_global_variables && MC_smpi_process_count()) {
233 // snapshot->privatization_index = smpi_loaded_page
234 mc_model_checker->process().read_variable(
235 "smpi_loaded_page", &snapshot->privatization_index,
236 sizeof(snapshot->privatization_index));
240 snapshot->privatization_index = simgrid::mc::ProcessIndexMissing;
244 /** \brief Fills the position of the segments (executable, read-only, read/write).
246 * `dl_iterate_phdr` would be more robust but would not work in cross-process.
248 void MC_find_object_address(std::vector<simgrid::mc::VmMap> const& maps, mc_object_info_t result)
250 const char *name = basename(result->file_name);
251 for (size_t i = 0; i < maps.size(); ++i) {
252 simgrid::mc::VmMap const& reg = maps[i];
253 if (maps[i].pathname.empty()
254 || strcmp(basename(maps[i].pathname.c_str()), name)) {
256 } else if ((reg.prot & PROT_WRITE)) {
257 xbt_assert(!result->start_rw,
258 "Multiple read-write segments for %s, not supported",
259 maps[i].pathname.c_str());
260 result->start_rw = (char*) reg.start_addr;
261 result->end_rw = (char*) reg.end_addr;
262 // .bss is usually after the .data:
263 simgrid::mc::VmMap const& next = maps[i + 1];
264 if (next.pathname.empty() && (next.prot & PROT_WRITE)
265 && next.start_addr == reg.end_addr) {
266 result->end_rw = (char*) maps[i + 1].end_addr;
268 } else if ((reg.prot & PROT_READ) && (reg.prot & PROT_EXEC)) {
269 xbt_assert(!result->start_exec,
270 "Multiple executable segments for %s, not supported",
271 maps[i].pathname.c_str());
272 result->start_exec = (char*) reg.start_addr;
273 result->end_exec = (char*) reg.end_addr;
274 } else if ((reg.prot & PROT_READ) && !(reg.prot & PROT_EXEC)) {
275 xbt_assert(!result->start_ro,
276 "Multiple read only segments for %s, not supported",
277 maps[i].pathname.c_str());
278 result->start_ro = (char*) reg.start_addr;
279 result->end_ro = (char*) reg.end_addr;
283 result->start = result->start_rw;
284 if ((const void*) result->start_ro > result->start)
285 result->start = result->start_ro;
286 if ((const void*) result->start_exec > result->start)
287 result->start = result->start_exec;
289 result->end = result->end_rw;
290 if (result->end_ro && (const void*) result->end_ro < result->end)
291 result->end = result->end_ro;
292 if (result->end_exec && (const void*) result->end_exec > result->end)
293 result->end = result->end_exec;
295 xbt_assert(result->file_name);
296 xbt_assert(result->start_rw);
297 xbt_assert(result->start_exec);
300 /************************************* Take Snapshot ************************************/
301 /****************************************************************************************/
303 /** \brief Checks whether the variable is in scope for a given IP.
305 * A variable may be defined only from a given value of IP.
307 * \param var Variable description
308 * \param frame Scope description
309 * \param ip Instruction pointer
310 * \return true if the variable is valid
312 static bool mc_valid_variable(dw_variable_t var, dw_frame_t scope,
315 // The variable is not yet valid:
316 if ((const void *) ((const char *) scope->low_pc + var->start_scope) > ip)
322 static void mc_fill_local_variables_values(mc_stack_frame_t stack_frame,
323 dw_frame_t scope, int process_index, xbt_dynar_t result)
325 mc_process_t process = &mc_model_checker->process();
327 void *ip = (void *) stack_frame->ip;
328 if (ip < scope->low_pc || ip >= scope->high_pc)
332 dw_variable_t current_variable;
333 xbt_dynar_foreach(scope->variables, cursor, current_variable) {
335 if (!mc_valid_variable(current_variable, scope, (void *) stack_frame->ip))
339 // FIXME, get rid of `region_type`
340 if ((long) stack_frame->ip > (long) process->libsimgrid_info->start_exec)
345 local_variable_t new_var = xbt_new0(s_local_variable_t, 1);
346 new_var->subprogram = stack_frame->frame;
347 new_var->ip = stack_frame->ip;
348 new_var->name = xbt_strdup(current_variable->name);
349 new_var->type = current_variable->type;
350 new_var->region = region_type;
352 if (current_variable->address != NULL) {
353 new_var->address = current_variable->address;
354 } else if (current_variable->locations.size != 0) {
355 s_mc_location_t location;
356 mc_dwarf_resolve_locations(
357 &location, ¤t_variable->locations,
358 current_variable->object_info,
359 &(stack_frame->unw_cursor),
360 (void *) stack_frame->frame_base,
361 &mc_model_checker->process(), process_index);
363 switch(mc_get_location_type(&location)) {
364 case MC_LOCATION_TYPE_ADDRESS:
365 new_var->address = location.memory_location;
367 case MC_LOCATION_TYPE_REGISTER:
369 xbt_die("Cannot handle non-address variable");
373 xbt_die("No address");
376 xbt_dynar_push(result, &new_var);
379 // Recursive processing of nested scopes:
380 dw_frame_t nested_scope = NULL;
381 xbt_dynar_foreach(scope->scopes, cursor, nested_scope) {
382 mc_fill_local_variables_values(stack_frame, nested_scope, process_index, result);
386 static xbt_dynar_t MC_get_local_variables_values(xbt_dynar_t stack_frames, int process_index)
389 unsigned cursor1 = 0;
390 mc_stack_frame_t stack_frame;
391 xbt_dynar_t variables =
392 xbt_dynar_new(sizeof(local_variable_t), local_variable_free_voidp);
394 xbt_dynar_foreach(stack_frames, cursor1, stack_frame) {
395 mc_fill_local_variables_values(stack_frame, stack_frame->frame, process_index, variables);
401 static void MC_stack_frame_free_voipd(void *s)
403 mc_stack_frame_t stack_frame = *(mc_stack_frame_t *) s;
405 xbt_free(stack_frame->frame_name);
406 xbt_free(stack_frame);
410 static xbt_dynar_t MC_unwind_stack_frames(mc_unw_context_t stack_context)
412 mc_process_t process = &mc_model_checker->process();
414 xbt_dynar_new(sizeof(mc_stack_frame_t), MC_stack_frame_free_voipd);
418 // TODO, check condition check (unw_init_local==0 means end of frame)
419 if (mc_unw_init_cursor(&c, stack_context) != 0) {
421 xbt_die("Could not initialize stack unwinding");
426 mc_stack_frame_t stack_frame = xbt_new(s_mc_stack_frame_t, 1);
427 xbt_dynar_push(result, &stack_frame);
429 stack_frame->unw_cursor = c;
433 unw_get_reg(&c, UNW_REG_IP, &ip);
434 unw_get_reg(&c, UNW_REG_SP, &sp);
436 stack_frame->ip = ip;
437 stack_frame->sp = sp;
439 // TODO, use real addresses in frame_t instead of fixing it here
441 dw_frame_t frame = process->find_function(remote(ip));
442 stack_frame->frame = frame;
445 stack_frame->frame_name = xbt_strdup(frame->name);
446 stack_frame->frame_base =
447 (unw_word_t) mc_find_frame_base(frame, frame->object_info, &c);
449 stack_frame->frame_base = 0;
450 stack_frame->frame_name = NULL;
453 /* Stop before context switch with maestro */
454 if (frame != NULL && frame->name != NULL
455 && !strcmp(frame->name, "smx_ctx_sysv_wrapper"))
458 int ret = unw_step(&c);
460 xbt_die("Unexpected end of stack.");
461 } else if (ret < 0) {
462 xbt_die("Error while unwinding stack");
466 if (xbt_dynar_length(result) == 0) {
467 XBT_INFO("unw_init_local failed");
474 static xbt_dynar_t MC_take_snapshot_stacks(mc_snapshot_t * snapshot)
478 xbt_dynar_new(sizeof(s_mc_snapshot_stack_t),
479 MC_snapshot_stack_free_voidp);
481 unsigned int cursor = 0;
482 stack_region_t current_stack;
484 // FIXME, cross-process support (stack_areas)
485 xbt_dynar_foreach(stacks_areas, cursor, current_stack) {
486 mc_snapshot_stack_t st = xbt_new(s_mc_snapshot_stack_t, 1);
488 // Read the context from remote process:
489 unw_context_t context;
490 mc_model_checker->process().read_bytes(
491 &context, sizeof(context), remote(current_stack->context));
493 st->context = xbt_new0(s_mc_unw_context_t, 1);
494 if (mc_unw_init_context(st->context, &mc_model_checker->process(),
496 xbt_die("Could not initialise the libunwind context.");
499 st->stack_frames = MC_unwind_stack_frames(st->context);
500 st->local_variables = MC_get_local_variables_values(st->stack_frames, current_stack->process_index);
501 st->process_index = current_stack->process_index;
503 unw_word_t sp = xbt_dynar_get_as(st->stack_frames, 0, mc_stack_frame_t)->sp;
505 xbt_dynar_push(res, &st);
506 (*snapshot)->stack_sizes = (size_t*)
507 xbt_realloc((*snapshot)->stack_sizes, (cursor + 1) * sizeof(size_t));
508 (*snapshot)->stack_sizes[cursor] =
509 (char*) current_stack->address + current_stack->size - (char*) sp;
516 static xbt_dynar_t MC_take_snapshot_ignore()
519 if (mc_heap_comparison_ignore == NULL)
523 xbt_dynar_new(sizeof(mc_heap_ignore_region_t),
524 heap_ignore_region_free_voidp);
526 unsigned int cursor = 0;
527 mc_heap_ignore_region_t current_region;
529 xbt_dynar_foreach(mc_heap_comparison_ignore, cursor, current_region) {
530 mc_heap_ignore_region_t new_region = NULL;
531 new_region = xbt_new0(s_mc_heap_ignore_region_t, 1);
532 new_region->address = current_region->address;
533 new_region->size = current_region->size;
534 new_region->block = current_region->block;
535 new_region->fragment = current_region->fragment;
536 xbt_dynar_push(cpy, &new_region);
543 static void mc_free_snapshot_ignored_data_pvoid(void* data) {
544 mc_snapshot_ignored_data_t ignored_data = (mc_snapshot_ignored_data_t) data;
545 free(ignored_data->data);
548 static void MC_snapshot_handle_ignore(mc_snapshot_t snapshot)
550 xbt_assert(snapshot->process);
551 snapshot->ignored_data = xbt_dynar_new(sizeof(s_mc_snapshot_ignored_data_t), mc_free_snapshot_ignored_data_pvoid);
554 for (auto const& region : mc_model_checker->process().ignored_regions()) {
555 s_mc_snapshot_ignored_data_t ignored_data;
556 ignored_data.start = (void*)region.addr;
557 ignored_data.size = region.size;
558 ignored_data.data = malloc(region.size);
559 // TODO, we should do this once per privatization segment:
560 snapshot->process->read_bytes(
561 ignored_data.data, region.size, remote(region.addr),
562 simgrid::mc::ProcessIndexDisabled);
563 xbt_dynar_push(snapshot->ignored_data, &ignored_data);
567 for(auto const& region : mc_model_checker->process().ignored_regions()) {
568 snapshot->process->clear_bytes(remote(region.addr), region.size);
573 static void MC_snapshot_ignore_restore(mc_snapshot_t snapshot)
575 unsigned int cursor = 0;
576 s_mc_snapshot_ignored_data_t ignored_data;
577 xbt_dynar_foreach (snapshot->ignored_data, cursor, ignored_data) {
578 snapshot->process->write_bytes(ignored_data.data, ignored_data.size,
579 remote(ignored_data.start));
583 static void MC_get_current_fd(mc_snapshot_t snapshot)
585 snapshot->total_fd = 0;
587 const size_t fd_dir_path_size = 20;
588 char fd_dir_path[fd_dir_path_size];
589 int res = snprintf(fd_dir_path, fd_dir_path_size,
590 "/proc/%lli/fd", (long long int) snapshot->process->pid());
591 xbt_assert(res >= 0);
592 if ((size_t) res > fd_dir_path_size)
593 xbt_die("Unexpected buffer is too small for fd_dir_path");
595 DIR* fd_dir = opendir(fd_dir_path);
597 xbt_die("Cannot open directory '/proc/self/fd'\n");
600 struct dirent* fd_number;
601 while ((fd_number = readdir(fd_dir))) {
603 int fd_value = atoi(fd_number->d_name);
608 const size_t source_size = 25;
610 int res = snprintf(source, source_size, "/proc/%lli/fd/%s",
611 (long long int) snapshot->process->pid(), fd_number->d_name);
612 xbt_assert(res >= 0);
613 if ((size_t) res > source_size)
614 xbt_die("Unexpected buffer is too small for fd %s", fd_number->d_name);
616 const size_t link_size = 200;
618 res = readlink(source, link, link_size);
620 xbt_die("Could not read link for %s", source);
623 xbt_die("Buffer to small for link of %s", source);
627 if(smpi_is_privatisation_file(link))
630 // This is (probably) the DIR* we are reading:
631 // TODO, read all the file entries at once and close the DIR.*
632 if(strcmp(fd_dir_path, link) == 0)
635 // We don't handle them.
636 // It does not mean we should silently ignore them however.
637 if (strncmp(link, "pipe:", 5) == 0 || strncmp(link, "socket:", 7) == 0)
640 // If dot_output enabled, do not handle the corresponding file
641 if (dot_output != NULL && strcmp(basename(link), _sg_mc_dot_output_file) == 0)
644 // This is probably a shared memory used by lttng-ust:
645 if(strncmp("/dev/shm/ust-shm-tmp-", link, 21)==0)
648 // Add an entry for this FD in the snapshot:
649 fd_infos_t fd = xbt_new0(s_fd_infos_t, 1);
650 fd->filename = strdup(link);
651 fd->number = fd_value;
652 fd->flags = fcntl(fd_value, F_GETFL) | fcntl(fd_value, F_GETFD) ;
653 fd->current_position = lseek(fd_value, 0, SEEK_CUR);
654 snapshot->current_fd = (fd_infos_t*) xbt_realloc(snapshot->current_fd, (total_fd + 1) * sizeof(fd_infos_t));
655 snapshot->current_fd[total_fd] = fd;
659 snapshot->total_fd = total_fd;
663 mc_snapshot_t MC_take_snapshot(int num_state)
665 XBT_DEBUG("Taking snapshot %i", num_state);
667 mc_process_t mc_process = &mc_model_checker->process();
669 mc_snapshot_t snapshot = new simgrid::mc::Snapshot();
671 snapshot->process = mc_process;
672 snapshot->num_state = num_state;
674 snapshot->enabled_processes = xbt_dynar_new(sizeof(int), NULL);
676 smx_process_t process;
677 MC_EACH_SIMIX_PROCESS(process,
678 xbt_dynar_push_as(snapshot->enabled_processes, int, (int)process->pid));
680 MC_snapshot_handle_ignore(snapshot);
682 if (_sg_mc_snapshot_fds)
683 MC_get_current_fd(snapshot);
685 /* Save the std heap and the writable mapped pages of libsimgrid and binary */
686 MC_get_memory_regions(mc_process, snapshot);
688 snapshot->to_ignore = MC_take_snapshot_ignore();
690 if (_sg_mc_visited > 0 || strcmp(_sg_mc_property_file, "")) {
692 MC_take_snapshot_stacks(&snapshot);
693 if (_sg_mc_hash && snapshot->stacks != NULL) {
694 snapshot->hash = mc_hash_processes_state(num_state, snapshot->stacks);
702 MC_snapshot_ignore_restore(snapshot);
707 void MC_restore_snapshot_regions(mc_snapshot_t snapshot)
709 const size_t n = snapshot->snapshot_regions_count;
710 for (size_t i = 0; i < n; i++) {
711 // For privatized, variables we decided it was not necessary to take the snapshot:
712 if (snapshot->snapshot_regions[i])
713 MC_region_restore(snapshot->snapshot_regions[i]);
717 // TODO, send a message to implement this in the MCed process
718 if(snapshot->privatization_index >= 0) {
719 // We just rewrote the global variables.
720 // The privatisation segment SMPI thinks
721 // is mapped might be inconsistent with the segment which
722 // is really mapped in memory (kernel state).
723 // We ask politely SMPI to map the segment anyway,
724 // even if it thinks it is the current one:
725 smpi_really_switch_data_segment(snapshot->privatization_index);
731 void MC_restore_snapshot_fds(mc_snapshot_t snapshot)
733 if (mc_mode == MC_MODE_SERVER)
734 xbt_die("FD snapshot not implemented in client/server mode.");
737 for (int i=0; i < snapshot->total_fd; i++) {
739 new_fd = open(snapshot->current_fd[i]->filename, snapshot->current_fd[i]->flags);
741 xbt_die("Could not reopen the file %s fo restoring the file descriptor",
742 snapshot->current_fd[i]->filename);
744 if(new_fd != -1 && new_fd != snapshot->current_fd[i]->number){
745 dup2(new_fd, snapshot->current_fd[i]->number);
746 //fprintf(stderr, "%p\n", fdopen(snapshot->current_fd[i]->number, "rw"));
749 lseek(snapshot->current_fd[i]->number, snapshot->current_fd[i]->current_position, SEEK_SET);
753 void MC_restore_snapshot(mc_snapshot_t snapshot)
755 XBT_DEBUG("Restore snapshot %i", snapshot->num_state);
756 MC_restore_snapshot_regions(snapshot);
757 if (_sg_mc_snapshot_fds)
758 MC_restore_snapshot_fds(snapshot);
759 MC_snapshot_ignore_restore(snapshot);
760 mc_model_checker->process().cache_flags = 0;
763 mc_snapshot_t simcall_HANDLER_mc_snapshot(smx_simcall_t simcall)
765 return MC_take_snapshot(1);