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 mc_mem_region_t mc_region_new_dense(
92 mc_region_type_t region_type,
93 void *start_addr, void* permanent_addr, size_t size)
95 mc_mem_region_t region = new simgrid::mc::RegionSnapshot();
96 region->region_type = region_type;
97 region->storage_type = MC_REGION_STORAGE_TYPE_FLAT;
98 region->start_addr = start_addr;
99 region->permanent_addr = permanent_addr;
101 region->flat_data_.resize(size);
102 mc_model_checker->process().read_bytes(region->flat_data_.data(), size,
103 remote(permanent_addr),
104 simgrid::mc::ProcessIndexDisabled);
105 XBT_DEBUG("New region : type : %d, data : %p (real addr %p), size : %zu",
106 region_type, region->flat_data_.data(), permanent_addr, size);
110 /** @brief Take a snapshot of a given region
113 * @param start_addr Address of the region in the simulated process
114 * @param permanent_addr Permanent address of this data (for privatized variables, this is the virtual address of the privatized mapping)
115 * @param size Size of the data*
117 static mc_mem_region_t MC_region_new(
118 mc_region_type_t type, void *start_addr, void* permanent_addr, size_t size)
120 if (_sg_mc_sparse_checkpoint) {
121 return mc_region_new_sparse(type, start_addr, permanent_addr, size);
123 return mc_region_new_dense(type, start_addr, permanent_addr, size);
127 /** @brief Restore a region from a snapshot
129 * @param reg Target region
131 static void MC_region_restore(mc_mem_region_t region)
133 switch(region->storage_type) {
134 case MC_REGION_STORAGE_TYPE_NONE:
136 xbt_die("Storage type not supported");
139 case MC_REGION_STORAGE_TYPE_FLAT:
140 mc_model_checker->process().write_bytes(region->flat_data_.data(), region->size,
141 remote(region->permanent_addr));
144 case MC_REGION_STORAGE_TYPE_CHUNKED:
145 mc_region_restore_sparse(&mc_model_checker->process(), region);
148 case MC_REGION_STORAGE_TYPE_PRIVATIZED:
149 for (auto const& p : region->privatized_regions_)
150 MC_region_restore(p.get());
155 static mc_mem_region_t MC_region_new_privatized(
156 mc_region_type_t region_type, void *start_addr, void* permanent_addr, size_t size
159 size_t process_count = MC_smpi_process_count();
160 mc_mem_region_t region = new simgrid::mc::RegionSnapshot();
161 region->region_type = region_type;
162 region->storage_type = MC_REGION_STORAGE_TYPE_PRIVATIZED;
163 region->start_addr = start_addr;
164 region->permanent_addr = permanent_addr;
166 region->privatized_regions_.resize(process_count);
168 // Read smpi_privatisation_regions from MCed:
169 smpi_privatisation_region_t remote_smpi_privatisation_regions;
170 mc_model_checker->process().read_variable(
171 "smpi_privatisation_regions",
172 &remote_smpi_privatisation_regions, sizeof(remote_smpi_privatisation_regions));
173 s_smpi_privatisation_region_t privatisation_regions[process_count];
174 mc_model_checker->process().read_bytes(
175 &privatisation_regions, sizeof(privatisation_regions),
176 remote(remote_smpi_privatisation_regions));
178 for (size_t i = 0; i < process_count; i++) {
179 region->privatized_regions_[i] = std::unique_ptr<simgrid::mc::RegionSnapshot>(
180 MC_region_new(region_type, start_addr,
181 privatisation_regions[i].address, size)
188 static void MC_snapshot_add_region(int index, mc_snapshot_t snapshot, mc_region_type_t type,
189 mc_object_info_t object_info,
190 void *start_addr, void* permanent_addr, size_t size)
192 if (type == MC_REGION_TYPE_DATA)
193 xbt_assert(object_info, "Missing object info for object.");
194 else if (type == MC_REGION_TYPE_HEAP)
195 xbt_assert(!object_info, "Unexpected object info for heap region.");
197 mc_mem_region_t region;
198 const bool privatization_aware = MC_object_info_is_privatized(object_info);
199 if (privatization_aware && MC_smpi_process_count())
200 region = MC_region_new_privatized(type, start_addr, permanent_addr, size);
202 region = MC_region_new(type, start_addr, permanent_addr, size);
204 region->object_info = object_info;
205 snapshot->snapshot_regions[index] = region;
209 static void MC_get_memory_regions(mc_process_t process, mc_snapshot_t snapshot)
211 const size_t n = process->object_infos_size;
212 snapshot->snapshot_regions_count = n + 1;
213 snapshot->snapshot_regions = xbt_new0(mc_mem_region_t, n + 1);
215 for (size_t i = 0; i!=n; ++i) {
216 mc_object_info_t object_info = process->object_infos[i];
217 MC_snapshot_add_region(i, snapshot, MC_REGION_TYPE_DATA, object_info,
218 object_info->start_rw, object_info->start_rw,
219 object_info->end_rw - object_info->start_rw);
222 xbt_mheap_t heap = process->get_heap();
223 void *start_heap = heap->base;
224 void *end_heap = heap->breakval;
226 MC_snapshot_add_region(n, snapshot, MC_REGION_TYPE_HEAP, NULL,
227 start_heap, start_heap,
228 (char *) end_heap - (char *) start_heap);
229 snapshot->heap_bytes_used = mmalloc_get_bytes_used_remote(
231 process->get_malloc_info());
234 if (smpi_privatize_global_variables && MC_smpi_process_count()) {
235 // snapshot->privatization_index = smpi_loaded_page
236 mc_model_checker->process().read_variable(
237 "smpi_loaded_page", &snapshot->privatization_index,
238 sizeof(snapshot->privatization_index));
242 snapshot->privatization_index = simgrid::mc::ProcessIndexMissing;
246 /** \brief Fills the position of the segments (executable, read-only, read/write).
248 * `dl_iterate_phdr` would be more robust but would not work in cross-process.
250 void MC_find_object_address(std::vector<simgrid::mc::VmMap> const& maps, mc_object_info_t result)
252 const char *name = basename(result->file_name);
253 for (size_t i = 0; i < maps.size(); ++i) {
254 simgrid::mc::VmMap const& reg = maps[i];
255 if (maps[i].pathname.empty()
256 || strcmp(basename(maps[i].pathname.c_str()), name)) {
258 } else if ((reg.prot & PROT_WRITE)) {
259 xbt_assert(!result->start_rw,
260 "Multiple read-write segments for %s, not supported",
261 maps[i].pathname.c_str());
262 result->start_rw = (char*) reg.start_addr;
263 result->end_rw = (char*) reg.end_addr;
264 // .bss is usually after the .data:
265 simgrid::mc::VmMap const& next = maps[i + 1];
266 if (next.pathname.empty() && (next.prot & PROT_WRITE)
267 && next.start_addr == reg.end_addr) {
268 result->end_rw = (char*) maps[i + 1].end_addr;
270 } else if ((reg.prot & PROT_READ) && (reg.prot & PROT_EXEC)) {
271 xbt_assert(!result->start_exec,
272 "Multiple executable segments for %s, not supported",
273 maps[i].pathname.c_str());
274 result->start_exec = (char*) reg.start_addr;
275 result->end_exec = (char*) reg.end_addr;
276 } else if ((reg.prot & PROT_READ) && !(reg.prot & PROT_EXEC)) {
277 xbt_assert(!result->start_ro,
278 "Multiple read only segments for %s, not supported",
279 maps[i].pathname.c_str());
280 result->start_ro = (char*) reg.start_addr;
281 result->end_ro = (char*) reg.end_addr;
285 result->start = result->start_rw;
286 if ((const void*) result->start_ro > result->start)
287 result->start = result->start_ro;
288 if ((const void*) result->start_exec > result->start)
289 result->start = result->start_exec;
291 result->end = result->end_rw;
292 if (result->end_ro && (const void*) result->end_ro < result->end)
293 result->end = result->end_ro;
294 if (result->end_exec && (const void*) result->end_exec > result->end)
295 result->end = result->end_exec;
297 xbt_assert(result->file_name);
298 xbt_assert(result->start_rw);
299 xbt_assert(result->start_exec);
302 /************************************* Take Snapshot ************************************/
303 /****************************************************************************************/
305 /** \brief Checks whether the variable is in scope for a given IP.
307 * A variable may be defined only from a given value of IP.
309 * \param var Variable description
310 * \param frame Scope description
311 * \param ip Instruction pointer
312 * \return true if the variable is valid
314 static bool mc_valid_variable(dw_variable_t var, dw_frame_t scope,
317 // The variable is not yet valid:
318 if ((const void *) ((const char *) scope->low_pc + var->start_scope) > ip)
324 static void mc_fill_local_variables_values(mc_stack_frame_t stack_frame,
325 dw_frame_t scope, int process_index, xbt_dynar_t result)
327 mc_process_t process = &mc_model_checker->process();
329 void *ip = (void *) stack_frame->ip;
330 if (ip < scope->low_pc || ip >= scope->high_pc)
334 dw_variable_t current_variable;
335 xbt_dynar_foreach(scope->variables, cursor, current_variable) {
337 if (!mc_valid_variable(current_variable, scope, (void *) stack_frame->ip))
341 // FIXME, get rid of `region_type`
342 if ((long) stack_frame->ip > (long) process->libsimgrid_info->start_exec)
347 local_variable_t new_var = xbt_new0(s_local_variable_t, 1);
348 new_var->subprogram = stack_frame->frame;
349 new_var->ip = stack_frame->ip;
350 new_var->name = xbt_strdup(current_variable->name);
351 new_var->type = current_variable->type;
352 new_var->region = region_type;
354 if (current_variable->address != NULL) {
355 new_var->address = current_variable->address;
356 } else if (current_variable->locations.size != 0) {
357 s_mc_location_t location;
358 mc_dwarf_resolve_locations(
359 &location, ¤t_variable->locations,
360 current_variable->object_info,
361 &(stack_frame->unw_cursor),
362 (void *) stack_frame->frame_base,
363 &mc_model_checker->process(), process_index);
365 switch(mc_get_location_type(&location)) {
366 case MC_LOCATION_TYPE_ADDRESS:
367 new_var->address = location.memory_location;
369 case MC_LOCATION_TYPE_REGISTER:
371 xbt_die("Cannot handle non-address variable");
375 xbt_die("No address");
378 xbt_dynar_push(result, &new_var);
381 // Recursive processing of nested scopes:
382 dw_frame_t nested_scope = NULL;
383 xbt_dynar_foreach(scope->scopes, cursor, nested_scope) {
384 mc_fill_local_variables_values(stack_frame, nested_scope, process_index, result);
388 static xbt_dynar_t MC_get_local_variables_values(xbt_dynar_t stack_frames, int process_index)
391 unsigned cursor1 = 0;
392 mc_stack_frame_t stack_frame;
393 xbt_dynar_t variables =
394 xbt_dynar_new(sizeof(local_variable_t), local_variable_free_voidp);
396 xbt_dynar_foreach(stack_frames, cursor1, stack_frame) {
397 mc_fill_local_variables_values(stack_frame, stack_frame->frame, process_index, variables);
403 static void MC_stack_frame_free_voipd(void *s)
405 mc_stack_frame_t stack_frame = *(mc_stack_frame_t *) s;
407 xbt_free(stack_frame->frame_name);
408 xbt_free(stack_frame);
412 static xbt_dynar_t MC_unwind_stack_frames(mc_unw_context_t stack_context)
414 mc_process_t process = &mc_model_checker->process();
416 xbt_dynar_new(sizeof(mc_stack_frame_t), MC_stack_frame_free_voipd);
420 // TODO, check condition check (unw_init_local==0 means end of frame)
421 if (mc_unw_init_cursor(&c, stack_context) != 0) {
423 xbt_die("Could not initialize stack unwinding");
428 mc_stack_frame_t stack_frame = xbt_new(s_mc_stack_frame_t, 1);
429 xbt_dynar_push(result, &stack_frame);
431 stack_frame->unw_cursor = c;
435 unw_get_reg(&c, UNW_REG_IP, &ip);
436 unw_get_reg(&c, UNW_REG_SP, &sp);
438 stack_frame->ip = ip;
439 stack_frame->sp = sp;
441 // TODO, use real addresses in frame_t instead of fixing it here
443 dw_frame_t frame = process->find_function(remote(ip));
444 stack_frame->frame = frame;
447 stack_frame->frame_name = xbt_strdup(frame->name);
448 stack_frame->frame_base =
449 (unw_word_t) mc_find_frame_base(frame, frame->object_info, &c);
451 stack_frame->frame_base = 0;
452 stack_frame->frame_name = NULL;
455 /* Stop before context switch with maestro */
456 if (frame != NULL && frame->name != NULL
457 && !strcmp(frame->name, "smx_ctx_sysv_wrapper"))
460 int ret = unw_step(&c);
462 xbt_die("Unexpected end of stack.");
463 } else if (ret < 0) {
464 xbt_die("Error while unwinding stack");
468 if (xbt_dynar_length(result) == 0) {
469 XBT_INFO("unw_init_local failed");
476 static xbt_dynar_t MC_take_snapshot_stacks(mc_snapshot_t * snapshot)
480 xbt_dynar_new(sizeof(s_mc_snapshot_stack_t),
481 MC_snapshot_stack_free_voidp);
483 unsigned int cursor = 0;
484 stack_region_t current_stack;
486 // FIXME, cross-process support (stack_areas)
487 xbt_dynar_foreach(stacks_areas, cursor, current_stack) {
488 mc_snapshot_stack_t st = xbt_new(s_mc_snapshot_stack_t, 1);
490 // Read the context from remote process:
491 unw_context_t context;
492 mc_model_checker->process().read_bytes(
493 &context, sizeof(context), remote(current_stack->context));
495 st->context = xbt_new0(s_mc_unw_context_t, 1);
496 if (mc_unw_init_context(st->context, &mc_model_checker->process(),
498 xbt_die("Could not initialise the libunwind context.");
501 st->stack_frames = MC_unwind_stack_frames(st->context);
502 st->local_variables = MC_get_local_variables_values(st->stack_frames, current_stack->process_index);
503 st->process_index = current_stack->process_index;
505 unw_word_t sp = xbt_dynar_get_as(st->stack_frames, 0, mc_stack_frame_t)->sp;
507 xbt_dynar_push(res, &st);
508 (*snapshot)->stack_sizes = (size_t*)
509 xbt_realloc((*snapshot)->stack_sizes, (cursor + 1) * sizeof(size_t));
510 (*snapshot)->stack_sizes[cursor] =
511 (char*) current_stack->address + current_stack->size - (char*) sp;
518 static xbt_dynar_t MC_take_snapshot_ignore()
521 if (mc_heap_comparison_ignore == NULL)
525 xbt_dynar_new(sizeof(mc_heap_ignore_region_t),
526 heap_ignore_region_free_voidp);
528 unsigned int cursor = 0;
529 mc_heap_ignore_region_t current_region;
531 xbt_dynar_foreach(mc_heap_comparison_ignore, cursor, current_region) {
532 mc_heap_ignore_region_t new_region = NULL;
533 new_region = xbt_new0(s_mc_heap_ignore_region_t, 1);
534 new_region->address = current_region->address;
535 new_region->size = current_region->size;
536 new_region->block = current_region->block;
537 new_region->fragment = current_region->fragment;
538 xbt_dynar_push(cpy, &new_region);
545 static void mc_free_snapshot_ignored_data_pvoid(void* data) {
546 mc_snapshot_ignored_data_t ignored_data = (mc_snapshot_ignored_data_t) data;
547 free(ignored_data->data);
550 static void MC_snapshot_handle_ignore(mc_snapshot_t snapshot)
552 xbt_assert(snapshot->process);
553 snapshot->ignored_data = xbt_dynar_new(sizeof(s_mc_snapshot_ignored_data_t), mc_free_snapshot_ignored_data_pvoid);
556 for (auto const& region : mc_model_checker->process().ignored_regions()) {
557 s_mc_snapshot_ignored_data_t ignored_data;
558 ignored_data.start = (void*)region.addr;
559 ignored_data.size = region.size;
560 ignored_data.data = malloc(region.size);
561 // TODO, we should do this once per privatization segment:
562 snapshot->process->read_bytes(
563 ignored_data.data, region.size, remote(region.addr),
564 simgrid::mc::ProcessIndexDisabled);
565 xbt_dynar_push(snapshot->ignored_data, &ignored_data);
569 for(auto const& region : mc_model_checker->process().ignored_regions()) {
570 snapshot->process->clear_bytes(remote(region.addr), region.size);
575 static void MC_snapshot_ignore_restore(mc_snapshot_t snapshot)
577 unsigned int cursor = 0;
578 s_mc_snapshot_ignored_data_t ignored_data;
579 xbt_dynar_foreach (snapshot->ignored_data, cursor, ignored_data) {
580 snapshot->process->write_bytes(ignored_data.data, ignored_data.size,
581 remote(ignored_data.start));
585 static void MC_get_current_fd(mc_snapshot_t snapshot)
587 snapshot->total_fd = 0;
589 const size_t fd_dir_path_size = 20;
590 char fd_dir_path[fd_dir_path_size];
591 int res = snprintf(fd_dir_path, fd_dir_path_size,
592 "/proc/%lli/fd", (long long int) snapshot->process->pid());
593 xbt_assert(res >= 0);
594 if ((size_t) res > fd_dir_path_size)
595 xbt_die("Unexpected buffer is too small for fd_dir_path");
597 DIR* fd_dir = opendir(fd_dir_path);
599 xbt_die("Cannot open directory '/proc/self/fd'\n");
602 struct dirent* fd_number;
603 while ((fd_number = readdir(fd_dir))) {
605 int fd_value = atoi(fd_number->d_name);
610 const size_t source_size = 25;
612 int res = snprintf(source, source_size, "/proc/%lli/fd/%s",
613 (long long int) snapshot->process->pid(), fd_number->d_name);
614 xbt_assert(res >= 0);
615 if ((size_t) res > source_size)
616 xbt_die("Unexpected buffer is too small for fd %s", fd_number->d_name);
618 const size_t link_size = 200;
620 res = readlink(source, link, link_size);
622 xbt_die("Could not read link for %s", source);
625 xbt_die("Buffer to small for link of %s", source);
629 if(smpi_is_privatisation_file(link))
632 // This is (probably) the DIR* we are reading:
633 // TODO, read all the file entries at once and close the DIR.*
634 if(strcmp(fd_dir_path, link) == 0)
637 // We don't handle them.
638 // It does not mean we should silently ignore them however.
639 if (strncmp(link, "pipe:", 5) == 0 || strncmp(link, "socket:", 7) == 0)
642 // If dot_output enabled, do not handle the corresponding file
643 if (dot_output != NULL && strcmp(basename(link), _sg_mc_dot_output_file) == 0)
646 // This is probably a shared memory used by lttng-ust:
647 if(strncmp("/dev/shm/ust-shm-tmp-", link, 21)==0)
650 // Add an entry for this FD in the snapshot:
651 fd_infos_t fd = xbt_new0(s_fd_infos_t, 1);
652 fd->filename = strdup(link);
653 fd->number = fd_value;
654 fd->flags = fcntl(fd_value, F_GETFL) | fcntl(fd_value, F_GETFD) ;
655 fd->current_position = lseek(fd_value, 0, SEEK_CUR);
656 snapshot->current_fd = (fd_infos_t*) xbt_realloc(snapshot->current_fd, (total_fd + 1) * sizeof(fd_infos_t));
657 snapshot->current_fd[total_fd] = fd;
661 snapshot->total_fd = total_fd;
665 mc_snapshot_t MC_take_snapshot(int num_state)
667 XBT_DEBUG("Taking snapshot %i", num_state);
669 mc_process_t mc_process = &mc_model_checker->process();
671 mc_snapshot_t snapshot = new simgrid::mc::Snapshot();
673 snapshot->process = mc_process;
674 snapshot->num_state = num_state;
676 snapshot->enabled_processes = xbt_dynar_new(sizeof(int), NULL);
678 smx_process_t process;
679 MC_EACH_SIMIX_PROCESS(process,
680 xbt_dynar_push_as(snapshot->enabled_processes, int, (int)process->pid));
682 MC_snapshot_handle_ignore(snapshot);
684 if (_sg_mc_snapshot_fds)
685 MC_get_current_fd(snapshot);
687 /* Save the std heap and the writable mapped pages of libsimgrid and binary */
688 MC_get_memory_regions(mc_process, snapshot);
690 snapshot->to_ignore = MC_take_snapshot_ignore();
692 if (_sg_mc_visited > 0 || strcmp(_sg_mc_property_file, "")) {
694 MC_take_snapshot_stacks(&snapshot);
695 if (_sg_mc_hash && snapshot->stacks != NULL) {
696 snapshot->hash = mc_hash_processes_state(num_state, snapshot->stacks);
704 MC_snapshot_ignore_restore(snapshot);
709 void MC_restore_snapshot_regions(mc_snapshot_t snapshot)
711 const size_t n = snapshot->snapshot_regions_count;
712 for (size_t i = 0; i < n; i++) {
713 // For privatized, variables we decided it was not necessary to take the snapshot:
714 if (snapshot->snapshot_regions[i])
715 MC_region_restore(snapshot->snapshot_regions[i]);
719 // TODO, send a message to implement this in the MCed process
720 if(snapshot->privatization_index >= 0) {
721 // We just rewrote the global variables.
722 // The privatisation segment SMPI thinks
723 // is mapped might be inconsistent with the segment which
724 // is really mapped in memory (kernel state).
725 // We ask politely SMPI to map the segment anyway,
726 // even if it thinks it is the current one:
727 smpi_really_switch_data_segment(snapshot->privatization_index);
733 void MC_restore_snapshot_fds(mc_snapshot_t snapshot)
735 if (mc_mode == MC_MODE_SERVER)
736 xbt_die("FD snapshot not implemented in client/server mode.");
739 for (int i=0; i < snapshot->total_fd; i++) {
741 new_fd = open(snapshot->current_fd[i]->filename, snapshot->current_fd[i]->flags);
743 xbt_die("Could not reopen the file %s fo restoring the file descriptor",
744 snapshot->current_fd[i]->filename);
746 if(new_fd != -1 && new_fd != snapshot->current_fd[i]->number){
747 dup2(new_fd, snapshot->current_fd[i]->number);
748 //fprintf(stderr, "%p\n", fdopen(snapshot->current_fd[i]->number, "rw"));
751 lseek(snapshot->current_fd[i]->number, snapshot->current_fd[i]->current_position, SEEK_SET);
755 void MC_restore_snapshot(mc_snapshot_t snapshot)
757 XBT_DEBUG("Restore snapshot %i", snapshot->num_state);
758 MC_restore_snapshot_regions(snapshot);
759 if (_sg_mc_snapshot_fds)
760 MC_restore_snapshot_fds(snapshot);
761 MC_snapshot_ignore_restore(snapshot);
762 mc_model_checker->process().cache_flags = 0;
765 mc_snapshot_t simcall_HANDLER_mc_snapshot(smx_simcall_t simcall)
767 return MC_take_snapshot(1);