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 std::vector<char> data(size);
96 mc_model_checker->process().read_bytes(data.data(), size,
97 remote(permanent_addr),
98 simgrid::mc::ProcessIndexDisabled);
100 mc_mem_region_t region = new simgrid::mc::RegionSnapshot(
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);
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 mc_mem_region_t MC_region_new(
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_new_sparse(type, start_addr, permanent_addr, size);
122 return mc_region_new_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 const& p : region->privatized_data())
149 MC_region_restore(p.get());
154 static mc_mem_region_t MC_region_new_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<std::unique_ptr<simgrid::mc::RegionSnapshot>> data;
171 data.reserve(process_count);
172 for (size_t i = 0; i < process_count; i++)
173 data.push_back(std::unique_ptr<simgrid::mc::RegionSnapshot>(
174 MC_region_new(region_type, start_addr,
175 privatisation_regions[i].address, size)
178 mc_mem_region_t region = new simgrid::mc::RegionSnapshot(
179 region_type, start_addr, permanent_addr, size);
180 region->privatized_data(std::move(data));
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 mc_mem_region_t region;
194 const bool privatization_aware = MC_object_info_is_privatized(object_info);
195 if (privatization_aware && MC_smpi_process_count())
196 region = MC_region_new_privatized(type, start_addr, permanent_addr, size);
198 region = MC_region_new(type, start_addr, permanent_addr, size);
200 region->object_info(object_info);
201 snapshot->snapshot_regions[index] = region;
205 static void MC_get_memory_regions(mc_process_t process, mc_snapshot_t snapshot)
207 const size_t n = process->object_infos_size;
208 snapshot->snapshot_regions_count = n + 1;
209 snapshot->snapshot_regions = xbt_new0(mc_mem_region_t, n + 1);
211 for (size_t i = 0; i!=n; ++i) {
212 mc_object_info_t object_info = process->object_infos[i];
213 MC_snapshot_add_region(i, snapshot, MC_REGION_TYPE_DATA, object_info,
214 object_info->start_rw, object_info->start_rw,
215 object_info->end_rw - object_info->start_rw);
218 xbt_mheap_t heap = process->get_heap();
219 void *start_heap = heap->base;
220 void *end_heap = heap->breakval;
222 MC_snapshot_add_region(n, snapshot, MC_REGION_TYPE_HEAP, NULL,
223 start_heap, start_heap,
224 (char *) end_heap - (char *) start_heap);
225 snapshot->heap_bytes_used = mmalloc_get_bytes_used_remote(
227 process->get_malloc_info());
230 if (smpi_privatize_global_variables && MC_smpi_process_count()) {
231 // snapshot->privatization_index = smpi_loaded_page
232 mc_model_checker->process().read_variable(
233 "smpi_loaded_page", &snapshot->privatization_index,
234 sizeof(snapshot->privatization_index));
238 snapshot->privatization_index = simgrid::mc::ProcessIndexMissing;
242 /** \brief Fills the position of the segments (executable, read-only, read/write).
244 * `dl_iterate_phdr` would be more robust but would not work in cross-process.
246 void MC_find_object_address(std::vector<simgrid::mc::VmMap> const& maps, mc_object_info_t result)
248 const char *name = basename(result->file_name);
249 for (size_t i = 0; i < maps.size(); ++i) {
250 simgrid::mc::VmMap const& reg = maps[i];
251 if (maps[i].pathname.empty()
252 || strcmp(basename(maps[i].pathname.c_str()), name)) {
254 } else if ((reg.prot & PROT_WRITE)) {
255 xbt_assert(!result->start_rw,
256 "Multiple read-write segments for %s, not supported",
257 maps[i].pathname.c_str());
258 result->start_rw = (char*) reg.start_addr;
259 result->end_rw = (char*) reg.end_addr;
260 // .bss is usually after the .data:
261 simgrid::mc::VmMap const& next = maps[i + 1];
262 if (next.pathname.empty() && (next.prot & PROT_WRITE)
263 && next.start_addr == reg.end_addr) {
264 result->end_rw = (char*) maps[i + 1].end_addr;
266 } else if ((reg.prot & PROT_READ) && (reg.prot & PROT_EXEC)) {
267 xbt_assert(!result->start_exec,
268 "Multiple executable segments for %s, not supported",
269 maps[i].pathname.c_str());
270 result->start_exec = (char*) reg.start_addr;
271 result->end_exec = (char*) reg.end_addr;
272 } else if ((reg.prot & PROT_READ) && !(reg.prot & PROT_EXEC)) {
273 xbt_assert(!result->start_ro,
274 "Multiple read only segments for %s, not supported",
275 maps[i].pathname.c_str());
276 result->start_ro = (char*) reg.start_addr;
277 result->end_ro = (char*) reg.end_addr;
281 result->start = result->start_rw;
282 if ((const void*) result->start_ro > result->start)
283 result->start = result->start_ro;
284 if ((const void*) result->start_exec > result->start)
285 result->start = result->start_exec;
287 result->end = result->end_rw;
288 if (result->end_ro && (const void*) result->end_ro < result->end)
289 result->end = result->end_ro;
290 if (result->end_exec && (const void*) result->end_exec > result->end)
291 result->end = result->end_exec;
293 xbt_assert(result->file_name);
294 xbt_assert(result->start_rw);
295 xbt_assert(result->start_exec);
298 /************************************* Take Snapshot ************************************/
299 /****************************************************************************************/
301 /** \brief Checks whether the variable is in scope for a given IP.
303 * A variable may be defined only from a given value of IP.
305 * \param var Variable description
306 * \param frame Scope description
307 * \param ip Instruction pointer
308 * \return true if the variable is valid
310 static bool mc_valid_variable(dw_variable_t var, dw_frame_t scope,
313 // The variable is not yet valid:
314 if ((const void *) ((const char *) scope->low_pc + var->start_scope) > ip)
320 static void mc_fill_local_variables_values(mc_stack_frame_t stack_frame,
321 dw_frame_t scope, int process_index, xbt_dynar_t result)
323 mc_process_t process = &mc_model_checker->process();
325 void *ip = (void *) stack_frame->ip;
326 if (ip < scope->low_pc || ip >= scope->high_pc)
330 dw_variable_t current_variable;
331 xbt_dynar_foreach(scope->variables, cursor, current_variable) {
333 if (!mc_valid_variable(current_variable, scope, (void *) stack_frame->ip))
337 // FIXME, get rid of `region_type`
338 if ((long) stack_frame->ip > (long) process->libsimgrid_info->start_exec)
343 local_variable_t new_var = xbt_new0(s_local_variable_t, 1);
344 new_var->subprogram = stack_frame->frame;
345 new_var->ip = stack_frame->ip;
346 new_var->name = xbt_strdup(current_variable->name);
347 new_var->type = current_variable->type;
348 new_var->region = region_type;
350 if (current_variable->address != NULL) {
351 new_var->address = current_variable->address;
352 } else if (current_variable->locations.size != 0) {
353 s_mc_location_t location;
354 mc_dwarf_resolve_locations(
355 &location, ¤t_variable->locations,
356 current_variable->object_info,
357 &(stack_frame->unw_cursor),
358 (void *) stack_frame->frame_base,
359 &mc_model_checker->process(), process_index);
361 switch(mc_get_location_type(&location)) {
362 case MC_LOCATION_TYPE_ADDRESS:
363 new_var->address = location.memory_location;
365 case MC_LOCATION_TYPE_REGISTER:
367 xbt_die("Cannot handle non-address variable");
371 xbt_die("No address");
374 xbt_dynar_push(result, &new_var);
377 // Recursive processing of nested scopes:
378 dw_frame_t nested_scope = NULL;
379 xbt_dynar_foreach(scope->scopes, cursor, nested_scope) {
380 mc_fill_local_variables_values(stack_frame, nested_scope, process_index, result);
384 static xbt_dynar_t MC_get_local_variables_values(xbt_dynar_t stack_frames, int process_index)
387 unsigned cursor1 = 0;
388 mc_stack_frame_t stack_frame;
389 xbt_dynar_t variables =
390 xbt_dynar_new(sizeof(local_variable_t), local_variable_free_voidp);
392 xbt_dynar_foreach(stack_frames, cursor1, stack_frame) {
393 mc_fill_local_variables_values(stack_frame, stack_frame->frame, process_index, variables);
399 static void MC_stack_frame_free_voipd(void *s)
401 mc_stack_frame_t stack_frame = *(mc_stack_frame_t *) s;
403 xbt_free(stack_frame->frame_name);
404 xbt_free(stack_frame);
408 static xbt_dynar_t MC_unwind_stack_frames(mc_unw_context_t stack_context)
410 mc_process_t process = &mc_model_checker->process();
412 xbt_dynar_new(sizeof(mc_stack_frame_t), MC_stack_frame_free_voipd);
416 // TODO, check condition check (unw_init_local==0 means end of frame)
417 if (mc_unw_init_cursor(&c, stack_context) != 0) {
419 xbt_die("Could not initialize stack unwinding");
424 mc_stack_frame_t stack_frame = xbt_new(s_mc_stack_frame_t, 1);
425 xbt_dynar_push(result, &stack_frame);
427 stack_frame->unw_cursor = c;
431 unw_get_reg(&c, UNW_REG_IP, &ip);
432 unw_get_reg(&c, UNW_REG_SP, &sp);
434 stack_frame->ip = ip;
435 stack_frame->sp = sp;
437 // TODO, use real addresses in frame_t instead of fixing it here
439 dw_frame_t frame = process->find_function(remote(ip));
440 stack_frame->frame = frame;
443 stack_frame->frame_name = xbt_strdup(frame->name);
444 stack_frame->frame_base =
445 (unw_word_t) mc_find_frame_base(frame, frame->object_info, &c);
447 stack_frame->frame_base = 0;
448 stack_frame->frame_name = NULL;
451 /* Stop before context switch with maestro */
452 if (frame != NULL && frame->name != NULL
453 && !strcmp(frame->name, "smx_ctx_sysv_wrapper"))
456 int ret = unw_step(&c);
458 xbt_die("Unexpected end of stack.");
459 } else if (ret < 0) {
460 xbt_die("Error while unwinding stack");
464 if (xbt_dynar_length(result) == 0) {
465 XBT_INFO("unw_init_local failed");
472 static xbt_dynar_t MC_take_snapshot_stacks(mc_snapshot_t * snapshot)
476 xbt_dynar_new(sizeof(s_mc_snapshot_stack_t),
477 MC_snapshot_stack_free_voidp);
479 unsigned int cursor = 0;
480 stack_region_t current_stack;
482 // FIXME, cross-process support (stack_areas)
483 xbt_dynar_foreach(stacks_areas, cursor, current_stack) {
484 mc_snapshot_stack_t st = xbt_new(s_mc_snapshot_stack_t, 1);
486 // Read the context from remote process:
487 unw_context_t context;
488 mc_model_checker->process().read_bytes(
489 &context, sizeof(context), remote(current_stack->context));
491 st->context = xbt_new0(s_mc_unw_context_t, 1);
492 if (mc_unw_init_context(st->context, &mc_model_checker->process(),
494 xbt_die("Could not initialise the libunwind context.");
497 st->stack_frames = MC_unwind_stack_frames(st->context);
498 st->local_variables = MC_get_local_variables_values(st->stack_frames, current_stack->process_index);
499 st->process_index = current_stack->process_index;
501 unw_word_t sp = xbt_dynar_get_as(st->stack_frames, 0, mc_stack_frame_t)->sp;
503 xbt_dynar_push(res, &st);
504 (*snapshot)->stack_sizes = (size_t*)
505 xbt_realloc((*snapshot)->stack_sizes, (cursor + 1) * sizeof(size_t));
506 (*snapshot)->stack_sizes[cursor] =
507 (char*) current_stack->address + current_stack->size - (char*) sp;
514 static xbt_dynar_t MC_take_snapshot_ignore()
517 if (mc_heap_comparison_ignore == NULL)
521 xbt_dynar_new(sizeof(mc_heap_ignore_region_t),
522 heap_ignore_region_free_voidp);
524 unsigned int cursor = 0;
525 mc_heap_ignore_region_t current_region;
527 xbt_dynar_foreach(mc_heap_comparison_ignore, cursor, current_region) {
528 mc_heap_ignore_region_t new_region = NULL;
529 new_region = xbt_new0(s_mc_heap_ignore_region_t, 1);
530 new_region->address = current_region->address;
531 new_region->size = current_region->size;
532 new_region->block = current_region->block;
533 new_region->fragment = current_region->fragment;
534 xbt_dynar_push(cpy, &new_region);
541 static void mc_free_snapshot_ignored_data_pvoid(void* data) {
542 mc_snapshot_ignored_data_t ignored_data = (mc_snapshot_ignored_data_t) data;
543 free(ignored_data->data);
546 static void MC_snapshot_handle_ignore(mc_snapshot_t snapshot)
548 xbt_assert(snapshot->process);
549 snapshot->ignored_data = xbt_dynar_new(sizeof(s_mc_snapshot_ignored_data_t), mc_free_snapshot_ignored_data_pvoid);
552 for (auto const& region : mc_model_checker->process().ignored_regions()) {
553 s_mc_snapshot_ignored_data_t ignored_data;
554 ignored_data.start = (void*)region.addr;
555 ignored_data.size = region.size;
556 ignored_data.data = malloc(region.size);
557 // TODO, we should do this once per privatization segment:
558 snapshot->process->read_bytes(
559 ignored_data.data, region.size, remote(region.addr),
560 simgrid::mc::ProcessIndexDisabled);
561 xbt_dynar_push(snapshot->ignored_data, &ignored_data);
565 for(auto const& region : mc_model_checker->process().ignored_regions()) {
566 snapshot->process->clear_bytes(remote(region.addr), region.size);
571 static void MC_snapshot_ignore_restore(mc_snapshot_t snapshot)
573 unsigned int cursor = 0;
574 s_mc_snapshot_ignored_data_t ignored_data;
575 xbt_dynar_foreach (snapshot->ignored_data, cursor, ignored_data) {
576 snapshot->process->write_bytes(ignored_data.data, ignored_data.size,
577 remote(ignored_data.start));
581 static void MC_get_current_fd(mc_snapshot_t snapshot)
583 snapshot->total_fd = 0;
585 const size_t fd_dir_path_size = 20;
586 char fd_dir_path[fd_dir_path_size];
587 int res = snprintf(fd_dir_path, fd_dir_path_size,
588 "/proc/%lli/fd", (long long int) snapshot->process->pid());
589 xbt_assert(res >= 0);
590 if ((size_t) res > fd_dir_path_size)
591 xbt_die("Unexpected buffer is too small for fd_dir_path");
593 DIR* fd_dir = opendir(fd_dir_path);
595 xbt_die("Cannot open directory '/proc/self/fd'\n");
598 struct dirent* fd_number;
599 while ((fd_number = readdir(fd_dir))) {
601 int fd_value = atoi(fd_number->d_name);
606 const size_t source_size = 25;
608 int res = snprintf(source, source_size, "/proc/%lli/fd/%s",
609 (long long int) snapshot->process->pid(), fd_number->d_name);
610 xbt_assert(res >= 0);
611 if ((size_t) res > source_size)
612 xbt_die("Unexpected buffer is too small for fd %s", fd_number->d_name);
614 const size_t link_size = 200;
616 res = readlink(source, link, link_size);
618 xbt_die("Could not read link for %s", source);
621 xbt_die("Buffer to small for link of %s", source);
625 if(smpi_is_privatisation_file(link))
628 // This is (probably) the DIR* we are reading:
629 // TODO, read all the file entries at once and close the DIR.*
630 if(strcmp(fd_dir_path, link) == 0)
633 // We don't handle them.
634 // It does not mean we should silently ignore them however.
635 if (strncmp(link, "pipe:", 5) == 0 || strncmp(link, "socket:", 7) == 0)
638 // If dot_output enabled, do not handle the corresponding file
639 if (dot_output != NULL && strcmp(basename(link), _sg_mc_dot_output_file) == 0)
642 // This is probably a shared memory used by lttng-ust:
643 if(strncmp("/dev/shm/ust-shm-tmp-", link, 21)==0)
646 // Add an entry for this FD in the snapshot:
647 fd_infos_t fd = xbt_new0(s_fd_infos_t, 1);
648 fd->filename = strdup(link);
649 fd->number = fd_value;
650 fd->flags = fcntl(fd_value, F_GETFL) | fcntl(fd_value, F_GETFD) ;
651 fd->current_position = lseek(fd_value, 0, SEEK_CUR);
652 snapshot->current_fd = (fd_infos_t*) xbt_realloc(snapshot->current_fd, (total_fd + 1) * sizeof(fd_infos_t));
653 snapshot->current_fd[total_fd] = fd;
657 snapshot->total_fd = total_fd;
661 mc_snapshot_t MC_take_snapshot(int num_state)
663 XBT_DEBUG("Taking snapshot %i", num_state);
665 mc_process_t mc_process = &mc_model_checker->process();
667 mc_snapshot_t snapshot = new simgrid::mc::Snapshot();
669 snapshot->process = mc_process;
670 snapshot->num_state = num_state;
672 snapshot->enabled_processes = xbt_dynar_new(sizeof(int), NULL);
674 smx_process_t process;
675 MC_EACH_SIMIX_PROCESS(process,
676 xbt_dynar_push_as(snapshot->enabled_processes, int, (int)process->pid));
678 MC_snapshot_handle_ignore(snapshot);
680 if (_sg_mc_snapshot_fds)
681 MC_get_current_fd(snapshot);
683 /* Save the std heap and the writable mapped pages of libsimgrid and binary */
684 MC_get_memory_regions(mc_process, snapshot);
686 snapshot->to_ignore = MC_take_snapshot_ignore();
688 if (_sg_mc_visited > 0 || strcmp(_sg_mc_property_file, "")) {
690 MC_take_snapshot_stacks(&snapshot);
691 if (_sg_mc_hash && snapshot->stacks != NULL) {
692 snapshot->hash = mc_hash_processes_state(num_state, snapshot->stacks);
700 MC_snapshot_ignore_restore(snapshot);
705 void MC_restore_snapshot_regions(mc_snapshot_t snapshot)
707 const size_t n = snapshot->snapshot_regions_count;
708 for (size_t i = 0; i < n; i++) {
709 // For privatized, variables we decided it was not necessary to take the snapshot:
710 if (snapshot->snapshot_regions[i])
711 MC_region_restore(snapshot->snapshot_regions[i]);
715 // TODO, send a message to implement this in the MCed process
716 if(snapshot->privatization_index >= 0) {
717 // We just rewrote the global variables.
718 // The privatisation segment SMPI thinks
719 // is mapped might be inconsistent with the segment which
720 // is really mapped in memory (kernel state).
721 // We ask politely SMPI to map the segment anyway,
722 // even if it thinks it is the current one:
723 smpi_really_switch_data_segment(snapshot->privatization_index);
729 void MC_restore_snapshot_fds(mc_snapshot_t snapshot)
731 if (mc_mode == MC_MODE_SERVER)
732 xbt_die("FD snapshot not implemented in client/server mode.");
735 for (int i=0; i < snapshot->total_fd; i++) {
737 new_fd = open(snapshot->current_fd[i]->filename, snapshot->current_fd[i]->flags);
739 xbt_die("Could not reopen the file %s fo restoring the file descriptor",
740 snapshot->current_fd[i]->filename);
742 if(new_fd != -1 && new_fd != snapshot->current_fd[i]->number){
743 dup2(new_fd, snapshot->current_fd[i]->number);
744 //fprintf(stderr, "%p\n", fdopen(snapshot->current_fd[i]->number, "rw"));
747 lseek(snapshot->current_fd[i]->number, snapshot->current_fd[i]->current_position, SEEK_SET);
751 void MC_restore_snapshot(mc_snapshot_t snapshot)
753 XBT_DEBUG("Restore snapshot %i", snapshot->num_state);
754 MC_restore_snapshot_regions(snapshot);
755 if (_sg_mc_snapshot_fds)
756 MC_restore_snapshot_fds(snapshot);
757 MC_snapshot_ignore_restore(snapshot);
758 mc_model_checker->process().cache_flags = 0;
761 mc_snapshot_t simcall_HANDLER_mc_snapshot(smx_simcall_t simcall)
763 return MC_take_snapshot(1);