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);
80 /** @brief Restore a region from a snapshot
82 * @param reg Target region
84 static void MC_region_restore(mc_mem_region_t region)
86 switch(region->storage_type()) {
87 case simgrid::mc::StorageType::NoData:
89 xbt_die("Storage type not supported");
92 case simgrid::mc::StorageType::Flat:
93 mc_model_checker->process().write_bytes(region->flat_data().data(),
94 region->size(), region->permanent_address());
97 case simgrid::mc::StorageType::Chunked:
98 mc_region_restore_sparse(&mc_model_checker->process(), region);
101 case simgrid::mc::StorageType::Privatized:
102 for (auto& p : region->privatized_data())
103 MC_region_restore(&p);
113 simgrid::mc::RegionSnapshot privatized_region(
114 RegionType region_type, void *start_addr, void* permanent_addr, size_t size
117 size_t process_count = MC_smpi_process_count();
119 // Read smpi_privatisation_regions from MCed:
120 smpi_privatisation_region_t remote_smpi_privatisation_regions;
121 mc_model_checker->process().read_variable(
122 "smpi_privatisation_regions",
123 &remote_smpi_privatisation_regions, sizeof(remote_smpi_privatisation_regions));
124 s_smpi_privatisation_region_t privatisation_regions[process_count];
125 mc_model_checker->process().read_bytes(
126 &privatisation_regions, sizeof(privatisation_regions),
127 remote(remote_smpi_privatisation_regions));
129 std::vector<simgrid::mc::RegionSnapshot> data;
130 data.reserve(process_count);
131 for (size_t i = 0; i < process_count; i++)
133 simgrid::mc::region(region_type, start_addr,
134 privatisation_regions[i].address, size)
137 simgrid::mc::RegionSnapshot region = simgrid::mc::RegionSnapshot(
138 region_type, start_addr, permanent_addr, size);
139 region.privatized_data(std::move(data));
140 return std::move(region);
148 static void MC_snapshot_add_region(int index, mc_snapshot_t snapshot,
149 simgrid::mc::RegionType type,
150 mc_object_info_t object_info,
151 void *start_addr, void* permanent_addr, size_t size)
153 if (type == simgrid::mc::RegionType::Data)
154 xbt_assert(object_info, "Missing object info for object.");
155 else if (type == simgrid::mc::RegionType::Heap)
156 xbt_assert(!object_info, "Unexpected object info for heap region.");
158 const bool privatization_aware = MC_object_info_is_privatized(object_info);
160 simgrid::mc::RegionSnapshot region;
161 if (privatization_aware && MC_smpi_process_count())
162 region = simgrid::mc::privatized_region(type, start_addr, permanent_addr, size);
164 region = simgrid::mc::region(type, start_addr, permanent_addr, size);
166 region.object_info(object_info);
167 snapshot->snapshot_regions[index]
168 = new simgrid::mc::RegionSnapshot(std::move(region));
172 static void MC_get_memory_regions(mc_process_t process, mc_snapshot_t snapshot)
174 const size_t n = process->object_infos_size;
175 snapshot->snapshot_regions_count = n + 1;
176 snapshot->snapshot_regions = xbt_new0(mc_mem_region_t, n + 1);
178 for (size_t i = 0; i!=n; ++i) {
179 mc_object_info_t object_info = process->object_infos[i];
180 MC_snapshot_add_region(i, snapshot, simgrid::mc::RegionType::Data, object_info,
181 object_info->start_rw, object_info->start_rw,
182 object_info->end_rw - object_info->start_rw);
185 xbt_mheap_t heap = process->get_heap();
186 void *start_heap = heap->base;
187 void *end_heap = heap->breakval;
189 MC_snapshot_add_region(n, snapshot, simgrid::mc::RegionType::Heap, NULL,
190 start_heap, start_heap,
191 (char *) end_heap - (char *) start_heap);
192 snapshot->heap_bytes_used = mmalloc_get_bytes_used_remote(
194 process->get_malloc_info());
197 if (smpi_privatize_global_variables && MC_smpi_process_count()) {
198 // snapshot->privatization_index = smpi_loaded_page
199 mc_model_checker->process().read_variable(
200 "smpi_loaded_page", &snapshot->privatization_index,
201 sizeof(snapshot->privatization_index));
205 snapshot->privatization_index = simgrid::mc::ProcessIndexMissing;
209 /** \brief Fills the position of the segments (executable, read-only, read/write).
211 * `dl_iterate_phdr` would be more robust but would not work in cross-process.
213 void MC_find_object_address(std::vector<simgrid::mc::VmMap> const& maps, mc_object_info_t result)
215 const char *name = basename(result->file_name);
216 for (size_t i = 0; i < maps.size(); ++i) {
217 simgrid::mc::VmMap const& reg = maps[i];
218 if (maps[i].pathname.empty()
219 || strcmp(basename(maps[i].pathname.c_str()), name)) {
221 } else if ((reg.prot & PROT_WRITE)) {
222 xbt_assert(!result->start_rw,
223 "Multiple read-write segments for %s, not supported",
224 maps[i].pathname.c_str());
225 result->start_rw = (char*) reg.start_addr;
226 result->end_rw = (char*) reg.end_addr;
227 // .bss is usually after the .data:
228 simgrid::mc::VmMap const& next = maps[i + 1];
229 if (next.pathname.empty() && (next.prot & PROT_WRITE)
230 && next.start_addr == reg.end_addr) {
231 result->end_rw = (char*) maps[i + 1].end_addr;
233 } else if ((reg.prot & PROT_READ) && (reg.prot & PROT_EXEC)) {
234 xbt_assert(!result->start_exec,
235 "Multiple executable segments for %s, not supported",
236 maps[i].pathname.c_str());
237 result->start_exec = (char*) reg.start_addr;
238 result->end_exec = (char*) reg.end_addr;
239 } else if ((reg.prot & PROT_READ) && !(reg.prot & PROT_EXEC)) {
240 xbt_assert(!result->start_ro,
241 "Multiple read only segments for %s, not supported",
242 maps[i].pathname.c_str());
243 result->start_ro = (char*) reg.start_addr;
244 result->end_ro = (char*) reg.end_addr;
248 result->start = result->start_rw;
249 if ((const void*) result->start_ro > result->start)
250 result->start = result->start_ro;
251 if ((const void*) result->start_exec > result->start)
252 result->start = result->start_exec;
254 result->end = result->end_rw;
255 if (result->end_ro && (const void*) result->end_ro < result->end)
256 result->end = result->end_ro;
257 if (result->end_exec && (const void*) result->end_exec > result->end)
258 result->end = result->end_exec;
260 xbt_assert(result->file_name);
261 xbt_assert(result->start_rw);
262 xbt_assert(result->start_exec);
265 /************************************* Take Snapshot ************************************/
266 /****************************************************************************************/
268 /** \brief Checks whether the variable is in scope for a given IP.
270 * A variable may be defined only from a given value of IP.
272 * \param var Variable description
273 * \param frame Scope description
274 * \param ip Instruction pointer
275 * \return true if the variable is valid
277 static bool mc_valid_variable(dw_variable_t var, dw_frame_t scope,
280 // The variable is not yet valid:
281 if ((const void *) ((const char *) scope->low_pc + var->start_scope) > ip)
287 static void mc_fill_local_variables_values(mc_stack_frame_t stack_frame,
288 dw_frame_t scope, int process_index, xbt_dynar_t result)
290 mc_process_t process = &mc_model_checker->process();
292 void *ip = (void *) stack_frame->ip;
293 if (ip < scope->low_pc || ip >= scope->high_pc)
297 dw_variable_t current_variable;
298 xbt_dynar_foreach(scope->variables, cursor, current_variable) {
300 if (!mc_valid_variable(current_variable, scope, (void *) stack_frame->ip))
304 // FIXME, get rid of `region_type`
305 if ((long) stack_frame->ip > (long) process->libsimgrid_info->start_exec)
310 local_variable_t new_var = xbt_new0(s_local_variable_t, 1);
311 new_var->subprogram = stack_frame->frame;
312 new_var->ip = stack_frame->ip;
313 new_var->name = xbt_strdup(current_variable->name);
314 new_var->type = current_variable->type;
315 new_var->region = region_type;
317 if (current_variable->address != NULL) {
318 new_var->address = current_variable->address;
319 } else if (current_variable->locations.size != 0) {
320 s_mc_location_t location;
321 mc_dwarf_resolve_locations(
322 &location, ¤t_variable->locations,
323 current_variable->object_info,
324 &(stack_frame->unw_cursor),
325 (void *) stack_frame->frame_base,
326 &mc_model_checker->process(), process_index);
328 switch(mc_get_location_type(&location)) {
329 case MC_LOCATION_TYPE_ADDRESS:
330 new_var->address = location.memory_location;
332 case MC_LOCATION_TYPE_REGISTER:
334 xbt_die("Cannot handle non-address variable");
338 xbt_die("No address");
341 xbt_dynar_push(result, &new_var);
344 // Recursive processing of nested scopes:
345 dw_frame_t nested_scope = NULL;
346 xbt_dynar_foreach(scope->scopes, cursor, nested_scope) {
347 mc_fill_local_variables_values(stack_frame, nested_scope, process_index, result);
351 static xbt_dynar_t MC_get_local_variables_values(xbt_dynar_t stack_frames, int process_index)
354 unsigned cursor1 = 0;
355 mc_stack_frame_t stack_frame;
356 xbt_dynar_t variables =
357 xbt_dynar_new(sizeof(local_variable_t), local_variable_free_voidp);
359 xbt_dynar_foreach(stack_frames, cursor1, stack_frame) {
360 mc_fill_local_variables_values(stack_frame, stack_frame->frame, process_index, variables);
366 static void MC_stack_frame_free_voipd(void *s)
368 mc_stack_frame_t stack_frame = *(mc_stack_frame_t *) s;
370 xbt_free(stack_frame->frame_name);
371 xbt_free(stack_frame);
375 static xbt_dynar_t MC_unwind_stack_frames(mc_unw_context_t stack_context)
377 mc_process_t process = &mc_model_checker->process();
379 xbt_dynar_new(sizeof(mc_stack_frame_t), MC_stack_frame_free_voipd);
383 // TODO, check condition check (unw_init_local==0 means end of frame)
384 if (mc_unw_init_cursor(&c, stack_context) != 0) {
386 xbt_die("Could not initialize stack unwinding");
391 mc_stack_frame_t stack_frame = xbt_new(s_mc_stack_frame_t, 1);
392 xbt_dynar_push(result, &stack_frame);
394 stack_frame->unw_cursor = c;
398 unw_get_reg(&c, UNW_REG_IP, &ip);
399 unw_get_reg(&c, UNW_REG_SP, &sp);
401 stack_frame->ip = ip;
402 stack_frame->sp = sp;
404 // TODO, use real addresses in frame_t instead of fixing it here
406 dw_frame_t frame = process->find_function(remote(ip));
407 stack_frame->frame = frame;
410 stack_frame->frame_name = xbt_strdup(frame->name);
411 stack_frame->frame_base =
412 (unw_word_t) mc_find_frame_base(frame, frame->object_info, &c);
414 stack_frame->frame_base = 0;
415 stack_frame->frame_name = NULL;
418 /* Stop before context switch with maestro */
419 if (frame != NULL && frame->name != NULL
420 && !strcmp(frame->name, "smx_ctx_sysv_wrapper"))
423 int ret = unw_step(&c);
425 xbt_die("Unexpected end of stack.");
426 } else if (ret < 0) {
427 xbt_die("Error while unwinding stack");
431 if (xbt_dynar_length(result) == 0) {
432 XBT_INFO("unw_init_local failed");
439 static xbt_dynar_t MC_take_snapshot_stacks(mc_snapshot_t * snapshot)
443 xbt_dynar_new(sizeof(s_mc_snapshot_stack_t),
444 MC_snapshot_stack_free_voidp);
446 unsigned int cursor = 0;
447 stack_region_t current_stack;
449 // FIXME, cross-process support (stack_areas)
450 xbt_dynar_foreach(stacks_areas, cursor, current_stack) {
451 mc_snapshot_stack_t st = xbt_new(s_mc_snapshot_stack_t, 1);
453 // Read the context from remote process:
454 unw_context_t context;
455 mc_model_checker->process().read_bytes(
456 &context, sizeof(context), remote(current_stack->context));
458 st->context = xbt_new0(s_mc_unw_context_t, 1);
459 if (mc_unw_init_context(st->context, &mc_model_checker->process(),
461 xbt_die("Could not initialise the libunwind context.");
464 st->stack_frames = MC_unwind_stack_frames(st->context);
465 st->local_variables = MC_get_local_variables_values(st->stack_frames, current_stack->process_index);
466 st->process_index = current_stack->process_index;
468 unw_word_t sp = xbt_dynar_get_as(st->stack_frames, 0, mc_stack_frame_t)->sp;
470 xbt_dynar_push(res, &st);
472 (char*) current_stack->address + current_stack->size - (char*) sp;
473 (*snapshot)->stack_sizes.push_back(stack_size);
480 static std::vector<s_mc_heap_ignore_region_t> MC_take_snapshot_ignore()
482 std::vector<s_mc_heap_ignore_region_t> res;
484 if (mc_heap_comparison_ignore == NULL)
485 return std::move(res);
487 unsigned int cursor = 0;
488 mc_heap_ignore_region_t current_region;
490 xbt_dynar_foreach(mc_heap_comparison_ignore, cursor, current_region) {
491 s_mc_heap_ignore_region_t new_region;
492 new_region.address = current_region->address;
493 new_region.size = current_region->size;
494 new_region.block = current_region->block;
495 new_region.fragment = current_region->fragment;
496 res.push_back(std::move(new_region));
499 return std::move(res);
502 static void MC_snapshot_handle_ignore(mc_snapshot_t snapshot)
504 xbt_assert(snapshot->process);
507 for (auto const& region : mc_model_checker->process().ignored_regions()) {
508 s_mc_snapshot_ignored_data_t ignored_data;
509 ignored_data.start = (void*)region.addr;
510 ignored_data.data.resize(region.size);
511 // TODO, we should do this once per privatization segment:
512 snapshot->process->read_bytes(
513 ignored_data.data.data(), region.size, remote(region.addr),
514 simgrid::mc::ProcessIndexDisabled);
515 snapshot->ignored_data.push_back(std::move(ignored_data));
519 for(auto const& region : mc_model_checker->process().ignored_regions()) {
520 snapshot->process->clear_bytes(remote(region.addr), region.size);
525 static void MC_snapshot_ignore_restore(mc_snapshot_t snapshot)
527 for (auto const& ignored_data : snapshot->ignored_data)
528 snapshot->process->write_bytes(
529 ignored_data.data.data(), ignored_data.data.size(),
530 remote(ignored_data.start));
533 static std::vector<s_fd_infos_t> MC_get_current_fds(pid_t pid)
535 const size_t fd_dir_path_size = 20;
536 char fd_dir_path[fd_dir_path_size];
537 int res = snprintf(fd_dir_path, fd_dir_path_size,
538 "/proc/%lli/fd", (long long int) pid);
539 xbt_assert(res >= 0);
540 if ((size_t) res > fd_dir_path_size)
541 xbt_die("Unexpected buffer is too small for fd_dir_path");
543 DIR* fd_dir = opendir(fd_dir_path);
545 xbt_die("Cannot open directory '/proc/self/fd'\n");
547 std::vector<s_fd_infos_t> fds;
549 struct dirent* fd_number;
550 while ((fd_number = readdir(fd_dir))) {
552 int fd_value = atoi(fd_number->d_name);
557 const size_t source_size = 25;
559 int res = snprintf(source, source_size, "/proc/%lli/fd/%s",
560 (long long int) pid, fd_number->d_name);
561 xbt_assert(res >= 0);
562 if ((size_t) res > source_size)
563 xbt_die("Unexpected buffer is too small for fd %s", fd_number->d_name);
565 const size_t link_size = 200;
567 res = readlink(source, link, link_size);
569 xbt_die("Could not read link for %s", source);
572 xbt_die("Buffer to small for link of %s", source);
576 if(smpi_is_privatisation_file(link))
579 // This is (probably) the DIR* we are reading:
580 // TODO, read all the file entries at once and close the DIR.*
581 if(strcmp(fd_dir_path, link) == 0)
584 // We don't handle them.
585 // It does not mean we should silently ignore them however.
586 if (strncmp(link, "pipe:", 5) == 0 || strncmp(link, "socket:", 7) == 0)
589 // If dot_output enabled, do not handle the corresponding file
590 if (dot_output != NULL && strcmp(basename(link), _sg_mc_dot_output_file) == 0)
593 // This is probably a shared memory used by lttng-ust:
594 if(strncmp("/dev/shm/ust-shm-tmp-", link, 21)==0)
597 // Add an entry for this FD in the snapshot:
599 fd.filename = std::string(link);
600 fd.number = fd_value;
601 fd.flags = fcntl(fd_value, F_GETFL) | fcntl(fd_value, F_GETFD) ;
602 fd.current_position = lseek(fd_value, 0, SEEK_CUR);
603 fds.push_back(std::move(fd));
607 return std::move(fds);
610 mc_snapshot_t MC_take_snapshot(int num_state)
612 XBT_DEBUG("Taking snapshot %i", num_state);
614 mc_process_t mc_process = &mc_model_checker->process();
616 mc_snapshot_t snapshot = new simgrid::mc::Snapshot();
618 snapshot->process = mc_process;
619 snapshot->num_state = num_state;
621 smx_process_t process;
622 MC_EACH_SIMIX_PROCESS(process,
623 snapshot->enabled_processes.insert(process->pid));
625 MC_snapshot_handle_ignore(snapshot);
627 if (_sg_mc_snapshot_fds)
628 snapshot->current_fds = MC_get_current_fds(process->pid);
630 /* Save the std heap and the writable mapped pages of libsimgrid and binary */
631 MC_get_memory_regions(mc_process, snapshot);
633 snapshot->to_ignore = MC_take_snapshot_ignore();
635 if (_sg_mc_visited > 0 || strcmp(_sg_mc_property_file, "")) {
637 MC_take_snapshot_stacks(&snapshot);
638 if (_sg_mc_hash && snapshot->stacks != NULL) {
639 snapshot->hash = mc_hash_processes_state(num_state, snapshot->stacks);
647 MC_snapshot_ignore_restore(snapshot);
652 void MC_restore_snapshot_regions(mc_snapshot_t snapshot)
654 const size_t n = snapshot->snapshot_regions_count;
655 for (size_t i = 0; i < n; i++) {
656 // For privatized, variables we decided it was not necessary to take the snapshot:
657 if (snapshot->snapshot_regions[i])
658 MC_region_restore(snapshot->snapshot_regions[i]);
662 // TODO, send a message to implement this in the MCed process
663 if(snapshot->privatization_index >= 0) {
664 // We just rewrote the global variables.
665 // The privatisation segment SMPI thinks
666 // is mapped might be inconsistent with the segment which
667 // is really mapped in memory (kernel state).
668 // We ask politely SMPI to map the segment anyway,
669 // even if it thinks it is the current one:
670 smpi_really_switch_data_segment(snapshot->privatization_index);
676 void MC_restore_snapshot_fds(mc_snapshot_t snapshot)
678 if (mc_mode == MC_MODE_SERVER)
679 xbt_die("FD snapshot not implemented in client/server mode.");
681 for (auto const& fd : snapshot->current_fds) {
683 int new_fd = open(fd.filename.c_str(), fd.flags);
685 xbt_die("Could not reopen the file %s fo restoring the file descriptor",
686 fd.filename.c_str());
688 if (new_fd != fd.number) {
689 dup2(new_fd, fd.number);
692 lseek(fd.number, fd.current_position, SEEK_SET);
696 void MC_restore_snapshot(mc_snapshot_t snapshot)
698 XBT_DEBUG("Restore snapshot %i", snapshot->num_state);
699 MC_restore_snapshot_regions(snapshot);
700 if (_sg_mc_snapshot_fds)
701 MC_restore_snapshot_fds(snapshot);
702 MC_snapshot_ignore_restore(snapshot);
703 mc_model_checker->process().cache_flags = 0;
706 mc_snapshot_t simcall_HANDLER_mc_snapshot(smx_simcall_t simcall)
708 return MC_take_snapshot(1);