1 /* Copyright (c) 2008-2015. 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"
37 #include "mc_hash.hpp"
39 #include "mc/ObjectInformation.hpp"
40 #include "mc/Frame.hpp"
41 #include "mc/Variable.hpp"
43 using simgrid::mc::remote;
47 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(mc_checkpoint, mc,
48 "Logging specific to mc_checkpoint");
50 /************************************ Free functions **************************************/
51 /*****************************************************************************************/
53 int MC_important_snapshot(mc_snapshot_t snapshot)
55 // We need this snapshot in order to know which
56 // pages needs to be stored in the next snapshot.
57 // This field is only non-NULL when using soft-dirty
59 if (snapshot == mc_model_checker->parent_snapshot_)
65 /** @brief Restore a region from a snapshot
67 * @param reg Target region
69 static void MC_region_restore(mc_mem_region_t region)
71 switch(region->storage_type()) {
72 case simgrid::mc::StorageType::NoData:
74 xbt_die("Storage type not supported");
77 case simgrid::mc::StorageType::Flat:
78 mc_model_checker->process().write_bytes(region->flat_data(),
79 region->size(), region->permanent_address());
82 case simgrid::mc::StorageType::Chunked:
83 mc_region_restore_sparse(&mc_model_checker->process(), region);
86 case simgrid::mc::StorageType::Privatized:
87 for (auto& p : region->privatized_data())
88 MC_region_restore(&p);
99 simgrid::mc::RegionSnapshot privatized_region(
100 RegionType region_type, void *start_addr, void* permanent_addr, size_t size,
101 const simgrid::mc::RegionSnapshot* ref_region
104 size_t process_count = MC_smpi_process_count();
106 // Read smpi_privatisation_regions from MCed:
107 smpi_privatisation_region_t remote_smpi_privatisation_regions;
108 mc_model_checker->process().read_variable(
109 "smpi_privatisation_regions",
110 &remote_smpi_privatisation_regions, sizeof(remote_smpi_privatisation_regions));
111 s_smpi_privatisation_region_t privatisation_regions[process_count];
112 mc_model_checker->process().read_bytes(
113 &privatisation_regions, sizeof(privatisation_regions),
114 remote(remote_smpi_privatisation_regions));
116 std::vector<simgrid::mc::RegionSnapshot> data;
117 data.reserve(process_count);
118 for (size_t i = 0; i < process_count; i++) {
119 const simgrid::mc::RegionSnapshot* ref_privatized_region = nullptr;
120 if (ref_region && ref_region->storage_type() == StorageType::Privatized)
121 ref_privatized_region = &ref_region->privatized_data()[i];
122 data.push_back(simgrid::mc::region(region_type, start_addr,
123 privatisation_regions[i].address, size, ref_privatized_region));
126 simgrid::mc::RegionSnapshot region = simgrid::mc::RegionSnapshot(
127 region_type, start_addr, permanent_addr, size);
128 region.privatized_data(std::move(data));
129 return std::move(region);
138 static void MC_snapshot_add_region(int index, mc_snapshot_t snapshot,
139 simgrid::mc::RegionType type,
140 simgrid::mc::ObjectInformation* object_info,
141 void *start_addr, void* permanent_addr, size_t size)
143 if (type == simgrid::mc::RegionType::Data)
144 xbt_assert(object_info, "Missing object info for object.");
145 else if (type == simgrid::mc::RegionType::Heap)
146 xbt_assert(!object_info, "Unexpected object info for heap region.");
148 simgrid::mc::RegionSnapshot const* ref_region = nullptr;
149 if (mc_model_checker->parent_snapshot_)
150 ref_region = mc_model_checker->parent_snapshot_->snapshot_regions[index].get();
152 simgrid::mc::RegionSnapshot region;
154 const bool privatization_aware = object_info && object_info->privatized();
155 if (privatization_aware && MC_smpi_process_count())
156 region = simgrid::mc::privatized_region(
157 type, start_addr, permanent_addr, size, ref_region);
160 region = simgrid::mc::region(type, start_addr, permanent_addr, size, ref_region);
162 region.object_info(object_info);
163 snapshot->snapshot_regions[index]
164 = std::unique_ptr<simgrid::mc::RegionSnapshot>(
165 new simgrid::mc::RegionSnapshot(std::move(region)));
169 static void MC_get_memory_regions(simgrid::mc::Process* process, mc_snapshot_t snapshot)
171 const size_t n = process->object_infos.size();
172 snapshot->snapshot_regions.resize(n + 1);
174 for (auto const& object_info : process->object_infos) {
175 MC_snapshot_add_region(i, snapshot, simgrid::mc::RegionType::Data,
177 object_info->start_rw, object_info->start_rw,
178 object_info->end_rw - object_info->start_rw);
182 xbt_mheap_t heap = process->get_heap();
183 void *start_heap = heap->base;
184 void *end_heap = heap->breakval;
186 MC_snapshot_add_region(n, snapshot, simgrid::mc::RegionType::Heap, NULL,
187 start_heap, start_heap,
188 (char *) end_heap - (char *) start_heap);
189 snapshot->heap_bytes_used = mmalloc_get_bytes_used_remote(
191 process->get_malloc_info());
194 if (smpi_privatize_global_variables && MC_smpi_process_count()) {
195 // snapshot->privatization_index = smpi_loaded_page
196 mc_model_checker->process().read_variable(
197 "smpi_loaded_page", &snapshot->privatization_index,
198 sizeof(snapshot->privatization_index));
202 snapshot->privatization_index = simgrid::mc::ProcessIndexMissing;
206 /** \brief Fills the position of the segments (executable, read-only, read/write).
208 * `dl_iterate_phdr` would be more robust but would not work in cross-process.
210 void MC_find_object_address(
211 std::vector<simgrid::mc::VmMap> const& maps, simgrid::mc::ObjectInformation* result)
213 char* file_name = xbt_strdup(result->file_name.c_str());
214 const char *name = basename(file_name);
215 for (size_t i = 0; i < maps.size(); ++i) {
216 simgrid::mc::VmMap const& reg = maps[i];
217 if (maps[i].pathname.empty()
218 || strcmp(basename(maps[i].pathname.c_str()), name)) {
220 } else if ((reg.prot & PROT_WRITE)) {
221 xbt_assert(!result->start_rw,
222 "Multiple read-write segments for %s, not supported",
223 maps[i].pathname.c_str());
224 result->start_rw = (char*) reg.start_addr;
225 result->end_rw = (char*) reg.end_addr;
226 // .bss is usually after the .data:
227 simgrid::mc::VmMap const& next = maps[i + 1];
228 if (next.pathname.empty() && (next.prot & PROT_WRITE)
229 && next.start_addr == reg.end_addr) {
230 result->end_rw = (char*) maps[i + 1].end_addr;
232 } else if ((reg.prot & PROT_READ) && (reg.prot & PROT_EXEC)) {
233 xbt_assert(!result->start_exec,
234 "Multiple executable segments for %s, not supported",
235 maps[i].pathname.c_str());
236 result->start_exec = (char*) reg.start_addr;
237 result->end_exec = (char*) reg.end_addr;
238 } else if ((reg.prot & PROT_READ) && !(reg.prot & PROT_EXEC)) {
239 xbt_assert(!result->start_ro,
240 "Multiple read only segments for %s, not supported",
241 maps[i].pathname.c_str());
242 result->start_ro = (char*) reg.start_addr;
243 result->end_ro = (char*) reg.end_addr;
247 result->start = result->start_rw;
248 if ((const void*) result->start_ro > result->start)
249 result->start = result->start_ro;
250 if ((const void*) result->start_exec > result->start)
251 result->start = result->start_exec;
253 result->end = result->end_rw;
254 if (result->end_ro && (const void*) result->end_ro < result->end)
255 result->end = result->end_ro;
256 if (result->end_exec && (const void*) result->end_exec > result->end)
257 result->end = result->end_exec;
259 xbt_assert(result->start_rw);
260 xbt_assert(result->start_exec);
264 /************************************* Take Snapshot ************************************/
265 /****************************************************************************************/
267 /** \brief Checks whether the variable is in scope for a given IP.
269 * A variable may be defined only from a given value of IP.
271 * \param var Variable description
272 * \param frame Scope description
273 * \param ip Instruction pointer
274 * \return true if the variable is valid
276 static bool mc_valid_variable(simgrid::mc::Variable* var, simgrid::mc::Frame* scope,
279 // The variable is not yet valid:
280 if ((const void *) ((const char *) scope->low_pc + var->start_scope) > ip)
286 static void mc_fill_local_variables_values(mc_stack_frame_t stack_frame,
287 simgrid::mc::Frame* scope, int process_index,
288 std::vector<s_local_variable>& result)
290 simgrid::mc::Process* process = &mc_model_checker->process();
292 void *ip = (void *) stack_frame->ip;
293 if (ip < scope->low_pc || ip >= scope->high_pc)
296 for(simgrid::mc::Variable& current_variable :
299 if (!mc_valid_variable(¤t_variable, scope, (void *) stack_frame->ip))
303 // FIXME, get rid of `region_type`
304 if ((long) stack_frame->ip > (long) process->libsimgrid_info->start_exec)
309 s_local_variable_t new_var;
310 new_var.subprogram = stack_frame->frame;
311 new_var.ip = stack_frame->ip;
312 new_var.name = current_variable.name;
313 new_var.type = current_variable.type;
314 new_var.region = region_type;
315 new_var.address = nullptr;
317 if (current_variable.address != NULL) {
318 new_var.address = current_variable.address;
319 } else if (!current_variable.location_list.empty()) {
320 s_mc_location_t location;
321 mc_dwarf_resolve_locations(
322 &location, ¤t_variable.location_list,
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 result.push_back(std::move(new_var));
344 // Recursive processing of nested scopes:
345 for(simgrid::mc::Frame& nested_scope : scope->scopes)
346 mc_fill_local_variables_values(
347 stack_frame, &nested_scope, process_index, result);
350 static std::vector<s_local_variable> MC_get_local_variables_values(
351 std::vector<s_mc_stack_frame_t>& stack_frames, int process_index)
353 std::vector<s_local_variable> variables;
354 for (s_mc_stack_frame_t& stack_frame : stack_frames)
355 mc_fill_local_variables_values(&stack_frame, stack_frame.frame, process_index, variables);
356 return std::move(variables);
359 static void MC_stack_frame_free_voipd(void *s)
361 mc_stack_frame_t stack_frame = *(mc_stack_frame_t *) s;
365 static std::vector<s_mc_stack_frame_t> MC_unwind_stack_frames(mc_unw_context_t stack_context)
367 simgrid::mc::Process* process = &mc_model_checker->process();
368 std::vector<s_mc_stack_frame_t> result;
372 // TODO, check condition check (unw_init_local==0 means end of frame)
373 if (mc_unw_init_cursor(&c, stack_context) != 0) {
375 xbt_die("Could not initialize stack unwinding");
380 s_mc_stack_frame_t stack_frame;
382 stack_frame.unw_cursor = c;
386 unw_get_reg(&c, UNW_REG_IP, &ip);
387 unw_get_reg(&c, UNW_REG_SP, &sp);
392 // TODO, use real addresses in frame_t instead of fixing it here
394 simgrid::mc::Frame* frame = process->find_function(remote(ip));
395 stack_frame.frame = frame;
398 stack_frame.frame_name = frame->name;
399 stack_frame.frame_base =
400 (unw_word_t) mc_find_frame_base(frame, frame->object_info, &c);
402 stack_frame.frame_base = 0;
403 stack_frame.frame_name = std::string();
406 result.push_back(std::move(stack_frame));
408 /* Stop before context switch with maestro */
409 if (frame != nullptr &&
410 frame->name == "smx_ctx_sysv_wrapper")
413 int ret = unw_step(&c);
415 xbt_die("Unexpected end of stack.");
416 } else if (ret < 0) {
417 xbt_die("Error while unwinding stack");
421 if (result.empty()) {
422 XBT_INFO("unw_init_local failed");
426 return std::move(result);
429 static std::vector<s_mc_snapshot_stack_t> MC_take_snapshot_stacks(mc_snapshot_t * snapshot)
431 std::vector<s_mc_snapshot_stack_t> res;
433 unsigned int cursor = 0;
434 stack_region_t current_stack;
436 // FIXME, cross-process support (stack_areas)
437 xbt_dynar_foreach(stacks_areas, cursor, current_stack) {
438 s_mc_snapshot_stack_t st;
440 // Read the context from remote process:
441 unw_context_t context;
442 mc_model_checker->process().read_bytes(
443 &context, sizeof(context), remote(current_stack->context));
445 if (mc_unw_init_context(&st.context, &mc_model_checker->process(),
447 xbt_die("Could not initialise the libunwind context.");
449 st.stack_frames = MC_unwind_stack_frames(&st.context);
450 st.local_variables = MC_get_local_variables_values(st.stack_frames, current_stack->process_index);
451 st.process_index = current_stack->process_index;
453 unw_word_t sp = st.stack_frames[0].sp;
455 res.push_back(std::move(st));
458 (char*) current_stack->address + current_stack->size - (char*) sp;
459 (*snapshot)->stack_sizes.push_back(stack_size);
462 return std::move(res);
466 static std::vector<s_mc_heap_ignore_region_t> MC_take_snapshot_ignore()
468 std::vector<s_mc_heap_ignore_region_t> res;
470 if (mc_heap_comparison_ignore == NULL)
471 return std::move(res);
473 unsigned int cursor = 0;
474 mc_heap_ignore_region_t current_region;
476 xbt_dynar_foreach(mc_heap_comparison_ignore, cursor, current_region) {
477 s_mc_heap_ignore_region_t new_region;
478 new_region.address = current_region->address;
479 new_region.size = current_region->size;
480 new_region.block = current_region->block;
481 new_region.fragment = current_region->fragment;
482 res.push_back(std::move(new_region));
485 return std::move(res);
488 static void MC_snapshot_handle_ignore(mc_snapshot_t snapshot)
490 xbt_assert(snapshot->process());
493 for (auto const& region : mc_model_checker->process().ignored_regions()) {
494 s_mc_snapshot_ignored_data_t ignored_data;
495 ignored_data.start = (void*)region.addr;
496 ignored_data.data.resize(region.size);
497 // TODO, we should do this once per privatization segment:
498 snapshot->process()->read_bytes(
499 ignored_data.data.data(), region.size, remote(region.addr),
500 simgrid::mc::ProcessIndexDisabled);
501 snapshot->ignored_data.push_back(std::move(ignored_data));
505 for(auto const& region : mc_model_checker->process().ignored_regions()) {
506 snapshot->process()->clear_bytes(remote(region.addr), region.size);
511 static void MC_snapshot_ignore_restore(mc_snapshot_t snapshot)
513 for (auto const& ignored_data : snapshot->ignored_data)
514 snapshot->process()->write_bytes(
515 ignored_data.data.data(), ignored_data.data.size(),
516 remote(ignored_data.start));
519 static std::vector<s_fd_infos_t> MC_get_current_fds(pid_t pid)
521 const size_t fd_dir_path_size = 20;
522 char fd_dir_path[fd_dir_path_size];
523 int res = snprintf(fd_dir_path, fd_dir_path_size,
524 "/proc/%lli/fd", (long long int) pid);
525 xbt_assert(res >= 0);
526 if ((size_t) res > fd_dir_path_size)
527 xbt_die("Unexpected buffer is too small for fd_dir_path");
529 DIR* fd_dir = opendir(fd_dir_path);
531 xbt_die("Cannot open directory '/proc/self/fd'\n");
533 std::vector<s_fd_infos_t> fds;
535 struct dirent* fd_number;
536 while ((fd_number = readdir(fd_dir))) {
538 int fd_value = atoi(fd_number->d_name);
543 const size_t source_size = 25;
545 int res = snprintf(source, source_size, "/proc/%lli/fd/%s",
546 (long long int) pid, fd_number->d_name);
547 xbt_assert(res >= 0);
548 if ((size_t) res > source_size)
549 xbt_die("Unexpected buffer is too small for fd %s", fd_number->d_name);
551 const size_t link_size = 200;
553 res = readlink(source, link, link_size);
555 xbt_die("Could not read link for %s", source);
558 xbt_die("Buffer to small for link of %s", source);
563 if(smpi_is_privatisation_file(link))
567 // This is (probably) the DIR* we are reading:
568 // TODO, read all the file entries at once and close the DIR.*
569 if(strcmp(fd_dir_path, link) == 0)
572 // We don't handle them.
573 // It does not mean we should silently ignore them however.
574 if (strncmp(link, "pipe:", 5) == 0 || strncmp(link, "socket:", 7) == 0)
577 // If dot_output enabled, do not handle the corresponding file
578 if (dot_output != NULL && strcmp(basename(link), _sg_mc_dot_output_file) == 0)
581 // This is probably a shared memory used by lttng-ust:
582 if(strncmp("/dev/shm/ust-shm-tmp-", link, 21)==0)
585 // Add an entry for this FD in the snapshot:
587 fd.filename = std::string(link);
588 fd.number = fd_value;
589 fd.flags = fcntl(fd_value, F_GETFL) | fcntl(fd_value, F_GETFD) ;
590 fd.current_position = lseek(fd_value, 0, SEEK_CUR);
591 fds.push_back(std::move(fd));
595 return std::move(fds);
598 mc_snapshot_t MC_take_snapshot(int num_state)
600 XBT_DEBUG("Taking snapshot %i", num_state);
602 simgrid::mc::Process* mc_process = &mc_model_checker->process();
604 mc_snapshot_t snapshot = new simgrid::mc::Snapshot(mc_process);
606 snapshot->num_state = num_state;
608 smx_process_t process;
609 MC_EACH_SIMIX_PROCESS(process,
610 snapshot->enabled_processes.insert(process->pid));
612 MC_snapshot_handle_ignore(snapshot);
614 if (_sg_mc_snapshot_fds)
615 snapshot->current_fds = MC_get_current_fds(process->pid);
617 const bool use_soft_dirty = _sg_mc_sparse_checkpoint && _sg_mc_soft_dirty;
619 /* Save the std heap and the writable mapped pages of libsimgrid and binary */
620 MC_get_memory_regions(mc_process, snapshot);
622 mc_process->reset_soft_dirty();
624 snapshot->to_ignore = MC_take_snapshot_ignore();
626 if (_sg_mc_visited > 0 || strcmp(_sg_mc_property_file, "")) {
628 MC_take_snapshot_stacks(&snapshot);
630 snapshot->hash = simgrid::mc::hash(*snapshot);
638 MC_snapshot_ignore_restore(snapshot);
640 mc_model_checker->parent_snapshot_ = snapshot;
645 void MC_restore_snapshot_regions(mc_snapshot_t snapshot)
647 for(std::unique_ptr<s_mc_mem_region_t> const& region : snapshot->snapshot_regions) {
648 // For privatized, variables we decided it was not necessary to take the snapshot:
650 MC_region_restore(region.get());
654 // TODO, send a message to implement this in the MCed process
655 if(snapshot->privatization_index >= 0) {
656 // We just rewrote the global variables.
657 // The privatisation segment SMPI thinks
658 // is mapped might be inconsistent with the segment which
659 // is really mapped in memory (kernel state).
660 // We ask politely SMPI to map the segment anyway,
661 // even if it thinks it is the current one:
662 smpi_really_switch_data_segment(snapshot->privatization_index);
668 void MC_restore_snapshot_fds(mc_snapshot_t snapshot)
670 if (mc_mode == MC_MODE_SERVER)
671 xbt_die("FD snapshot not implemented in client/server mode.");
673 for (auto const& fd : snapshot->current_fds) {
675 int new_fd = open(fd.filename.c_str(), fd.flags);
677 xbt_die("Could not reopen the file %s fo restoring the file descriptor",
678 fd.filename.c_str());
680 if (new_fd != fd.number) {
681 dup2(new_fd, fd.number);
684 lseek(fd.number, fd.current_position, SEEK_SET);
688 void MC_restore_snapshot(mc_snapshot_t snapshot)
690 XBT_DEBUG("Restore snapshot %i", snapshot->num_state);
691 const bool use_soft_dirty = _sg_mc_sparse_checkpoint && _sg_mc_soft_dirty;
692 MC_restore_snapshot_regions(snapshot);
693 if (_sg_mc_snapshot_fds)
694 MC_restore_snapshot_fds(snapshot);
696 mc_model_checker->process().reset_soft_dirty();
697 MC_snapshot_ignore_restore(snapshot);
698 mc_model_checker->process().cache_flags = 0;
700 mc_model_checker->parent_snapshot_ = snapshot;
703 mc_snapshot_t simcall_HANDLER_mc_snapshot(smx_simcall_t simcall)
705 return MC_take_snapshot(1);