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"
37 #include "mc_hash.hpp"
39 using simgrid::mc::remote;
43 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(mc_checkpoint, mc,
44 "Logging specific to mc_checkpoint");
46 /************************************ Free functions **************************************/
47 /*****************************************************************************************/
49 /** @brief Restore a region from a snapshot
51 * @param reg Target region
53 static void MC_region_restore(mc_mem_region_t region)
55 switch(region->storage_type()) {
56 case simgrid::mc::StorageType::NoData:
58 xbt_die("Storage type not supported");
61 case simgrid::mc::StorageType::Flat:
62 mc_model_checker->process().write_bytes(region->flat_data().data(),
63 region->size(), region->permanent_address());
66 case simgrid::mc::StorageType::Chunked:
67 mc_region_restore_sparse(&mc_model_checker->process(), region);
70 case simgrid::mc::StorageType::Privatized:
71 for (auto& p : region->privatized_data())
72 MC_region_restore(&p);
82 simgrid::mc::RegionSnapshot privatized_region(
83 RegionType region_type, void *start_addr, void* permanent_addr, size_t size
86 size_t process_count = MC_smpi_process_count();
88 // Read smpi_privatisation_regions from MCed:
89 smpi_privatisation_region_t remote_smpi_privatisation_regions;
90 mc_model_checker->process().read_variable(
91 "smpi_privatisation_regions",
92 &remote_smpi_privatisation_regions, sizeof(remote_smpi_privatisation_regions));
93 s_smpi_privatisation_region_t privatisation_regions[process_count];
94 mc_model_checker->process().read_bytes(
95 &privatisation_regions, sizeof(privatisation_regions),
96 remote(remote_smpi_privatisation_regions));
98 std::vector<simgrid::mc::RegionSnapshot> data;
99 data.reserve(process_count);
100 for (size_t i = 0; i < process_count; i++)
102 simgrid::mc::region(region_type, start_addr,
103 privatisation_regions[i].address, size)
106 simgrid::mc::RegionSnapshot region = simgrid::mc::RegionSnapshot(
107 region_type, start_addr, permanent_addr, size);
108 region.privatized_data(std::move(data));
109 return std::move(region);
117 static void MC_snapshot_add_region(int index, mc_snapshot_t snapshot,
118 simgrid::mc::RegionType type,
119 simgrid::mc::ObjectInformation* object_info,
120 void *start_addr, void* permanent_addr, size_t size)
122 if (type == simgrid::mc::RegionType::Data)
123 xbt_assert(object_info, "Missing object info for object.");
124 else if (type == simgrid::mc::RegionType::Heap)
125 xbt_assert(!object_info, "Unexpected object info for heap region.");
127 const bool privatization_aware = object_info && object_info->privatized();
129 simgrid::mc::RegionSnapshot region;
130 if (privatization_aware && MC_smpi_process_count())
131 region = simgrid::mc::privatized_region(type, start_addr, permanent_addr, size);
133 region = simgrid::mc::region(type, start_addr, permanent_addr, size);
135 region.object_info(object_info);
136 snapshot->snapshot_regions[index]
137 = std::unique_ptr<simgrid::mc::RegionSnapshot>(
138 new simgrid::mc::RegionSnapshot(std::move(region)));
142 static void MC_get_memory_regions(simgrid::mc::Process* process, mc_snapshot_t snapshot)
144 const size_t n = process->object_infos.size();
145 snapshot->snapshot_regions.resize(n + 1);
147 for (auto const& object_info : process->object_infos) {
148 MC_snapshot_add_region(i, snapshot, simgrid::mc::RegionType::Data,
150 object_info->start_rw, object_info->start_rw,
151 object_info->end_rw - object_info->start_rw);
155 xbt_mheap_t heap = process->get_heap();
156 void *start_heap = heap->base;
157 void *end_heap = heap->breakval;
159 MC_snapshot_add_region(n, snapshot, simgrid::mc::RegionType::Heap, NULL,
160 start_heap, start_heap,
161 (char *) end_heap - (char *) start_heap);
162 snapshot->heap_bytes_used = mmalloc_get_bytes_used_remote(
164 process->get_malloc_info());
167 if (smpi_privatize_global_variables && MC_smpi_process_count()) {
168 // snapshot->privatization_index = smpi_loaded_page
169 mc_model_checker->process().read_variable(
170 "smpi_loaded_page", &snapshot->privatization_index,
171 sizeof(snapshot->privatization_index));
175 snapshot->privatization_index = simgrid::mc::ProcessIndexMissing;
179 /** \brief Fills the position of the segments (executable, read-only, read/write).
181 * `dl_iterate_phdr` would be more robust but would not work in cross-process.
183 void MC_find_object_address(
184 std::vector<simgrid::mc::VmMap> const& maps, simgrid::mc::ObjectInformation* result)
186 const char* file_name = xbt_strdup(result->file_name.c_str());
187 const char *name = basename(file_name);
188 for (size_t i = 0; i < maps.size(); ++i) {
189 simgrid::mc::VmMap const& reg = maps[i];
190 if (maps[i].pathname.empty()
191 || strcmp(basename(maps[i].pathname.c_str()), name)) {
193 } else if ((reg.prot & PROT_WRITE)) {
194 xbt_assert(!result->start_rw,
195 "Multiple read-write segments for %s, not supported",
196 maps[i].pathname.c_str());
197 result->start_rw = (char*) reg.start_addr;
198 result->end_rw = (char*) reg.end_addr;
199 // .bss is usually after the .data:
200 simgrid::mc::VmMap const& next = maps[i + 1];
201 if (next.pathname.empty() && (next.prot & PROT_WRITE)
202 && next.start_addr == reg.end_addr) {
203 result->end_rw = (char*) maps[i + 1].end_addr;
205 } else if ((reg.prot & PROT_READ) && (reg.prot & PROT_EXEC)) {
206 xbt_assert(!result->start_exec,
207 "Multiple executable segments for %s, not supported",
208 maps[i].pathname.c_str());
209 result->start_exec = (char*) reg.start_addr;
210 result->end_exec = (char*) reg.end_addr;
211 } else if ((reg.prot & PROT_READ) && !(reg.prot & PROT_EXEC)) {
212 xbt_assert(!result->start_ro,
213 "Multiple read only segments for %s, not supported",
214 maps[i].pathname.c_str());
215 result->start_ro = (char*) reg.start_addr;
216 result->end_ro = (char*) reg.end_addr;
220 result->start = result->start_rw;
221 if ((const void*) result->start_ro > result->start)
222 result->start = result->start_ro;
223 if ((const void*) result->start_exec > result->start)
224 result->start = result->start_exec;
226 result->end = result->end_rw;
227 if (result->end_ro && (const void*) result->end_ro < result->end)
228 result->end = result->end_ro;
229 if (result->end_exec && (const void*) result->end_exec > result->end)
230 result->end = result->end_exec;
232 xbt_assert(result->start_rw);
233 xbt_assert(result->start_exec);
236 /************************************* Take Snapshot ************************************/
237 /****************************************************************************************/
239 /** \brief Checks whether the variable is in scope for a given IP.
241 * A variable may be defined only from a given value of IP.
243 * \param var Variable description
244 * \param frame Scope description
245 * \param ip Instruction pointer
246 * \return true if the variable is valid
248 static bool mc_valid_variable(simgrid::mc::Variable* var, simgrid::mc::Frame* scope,
251 // The variable is not yet valid:
252 if ((const void *) ((const char *) scope->low_pc + var->start_scope) > ip)
258 static void mc_fill_local_variables_values(mc_stack_frame_t stack_frame,
259 simgrid::mc::Frame* scope, int process_index,
260 std::vector<s_local_variable>& result)
262 simgrid::mc::Process* process = &mc_model_checker->process();
264 void *ip = (void *) stack_frame->ip;
265 if (ip < scope->low_pc || ip >= scope->high_pc)
268 for(simgrid::mc::Variable& current_variable :
271 if (!mc_valid_variable(¤t_variable, scope, (void *) stack_frame->ip))
275 // FIXME, get rid of `region_type`
276 if ((long) stack_frame->ip > (long) process->libsimgrid_info->start_exec)
281 s_local_variable_t new_var;
282 new_var.subprogram = stack_frame->frame;
283 new_var.ip = stack_frame->ip;
284 new_var.name = current_variable.name;
285 new_var.type = current_variable.type;
286 new_var.region = region_type;
287 new_var.address = nullptr;
289 if (current_variable.address != NULL) {
290 new_var.address = current_variable.address;
291 } else if (!current_variable.location_list.empty()) {
292 s_mc_location_t location;
293 mc_dwarf_resolve_locations(
294 &location, ¤t_variable.location_list,
295 current_variable.object_info,
296 &(stack_frame->unw_cursor),
297 (void *) stack_frame->frame_base,
298 &mc_model_checker->process(), process_index);
300 switch(mc_get_location_type(&location)) {
301 case MC_LOCATION_TYPE_ADDRESS:
302 new_var.address = location.memory_location;
304 case MC_LOCATION_TYPE_REGISTER:
306 xbt_die("Cannot handle non-address variable");
310 xbt_die("No address");
313 result.push_back(std::move(new_var));
316 // Recursive processing of nested scopes:
317 for(simgrid::mc::Frame& nested_scope : scope->scopes)
318 mc_fill_local_variables_values(
319 stack_frame, &nested_scope, process_index, result);
322 static std::vector<s_local_variable> MC_get_local_variables_values(
323 std::vector<s_mc_stack_frame_t>& stack_frames, int process_index)
325 std::vector<s_local_variable> variables;
326 for (s_mc_stack_frame_t& stack_frame : stack_frames)
327 mc_fill_local_variables_values(&stack_frame, stack_frame.frame, process_index, variables);
328 return std::move(variables);
331 static void MC_stack_frame_free_voipd(void *s)
333 mc_stack_frame_t stack_frame = *(mc_stack_frame_t *) s;
337 static std::vector<s_mc_stack_frame_t> MC_unwind_stack_frames(mc_unw_context_t stack_context)
339 simgrid::mc::Process* process = &mc_model_checker->process();
340 std::vector<s_mc_stack_frame_t> result;
344 // TODO, check condition check (unw_init_local==0 means end of frame)
345 if (mc_unw_init_cursor(&c, stack_context) != 0) {
347 xbt_die("Could not initialize stack unwinding");
352 s_mc_stack_frame_t stack_frame;
354 stack_frame.unw_cursor = c;
358 unw_get_reg(&c, UNW_REG_IP, &ip);
359 unw_get_reg(&c, UNW_REG_SP, &sp);
364 // TODO, use real addresses in frame_t instead of fixing it here
366 simgrid::mc::Frame* frame = process->find_function(remote(ip));
367 stack_frame.frame = frame;
370 stack_frame.frame_name = frame->name;
371 stack_frame.frame_base =
372 (unw_word_t) mc_find_frame_base(frame, frame->object_info, &c);
374 stack_frame.frame_base = 0;
375 stack_frame.frame_name = std::string();
378 result.push_back(std::move(stack_frame));
380 /* Stop before context switch with maestro */
381 if (frame != nullptr &&
382 frame->name == "smx_ctx_sysv_wrapper")
385 int ret = unw_step(&c);
387 xbt_die("Unexpected end of stack.");
388 } else if (ret < 0) {
389 xbt_die("Error while unwinding stack");
393 if (result.empty()) {
394 XBT_INFO("unw_init_local failed");
398 return std::move(result);
401 static std::vector<s_mc_snapshot_stack_t> MC_take_snapshot_stacks(mc_snapshot_t * snapshot)
403 std::vector<s_mc_snapshot_stack_t> res;
405 unsigned int cursor = 0;
406 stack_region_t current_stack;
408 // FIXME, cross-process support (stack_areas)
409 xbt_dynar_foreach(stacks_areas, cursor, current_stack) {
410 s_mc_snapshot_stack_t st;
412 // Read the context from remote process:
413 unw_context_t context;
414 mc_model_checker->process().read_bytes(
415 &context, sizeof(context), remote(current_stack->context));
417 if (mc_unw_init_context(&st.context, &mc_model_checker->process(),
419 xbt_die("Could not initialise the libunwind context.");
421 st.stack_frames = MC_unwind_stack_frames(&st.context);
422 st.local_variables = MC_get_local_variables_values(st.stack_frames, current_stack->process_index);
423 st.process_index = current_stack->process_index;
425 unw_word_t sp = st.stack_frames[0].sp;
427 res.push_back(std::move(st));
430 (char*) current_stack->address + current_stack->size - (char*) sp;
431 (*snapshot)->stack_sizes.push_back(stack_size);
434 return std::move(res);
438 static std::vector<s_mc_heap_ignore_region_t> MC_take_snapshot_ignore()
440 std::vector<s_mc_heap_ignore_region_t> res;
442 if (mc_heap_comparison_ignore == NULL)
443 return std::move(res);
445 unsigned int cursor = 0;
446 mc_heap_ignore_region_t current_region;
448 xbt_dynar_foreach(mc_heap_comparison_ignore, cursor, current_region) {
449 s_mc_heap_ignore_region_t new_region;
450 new_region.address = current_region->address;
451 new_region.size = current_region->size;
452 new_region.block = current_region->block;
453 new_region.fragment = current_region->fragment;
454 res.push_back(std::move(new_region));
457 return std::move(res);
460 static void MC_snapshot_handle_ignore(mc_snapshot_t snapshot)
462 xbt_assert(snapshot->process);
465 for (auto const& region : mc_model_checker->process().ignored_regions()) {
466 s_mc_snapshot_ignored_data_t ignored_data;
467 ignored_data.start = (void*)region.addr;
468 ignored_data.data.resize(region.size);
469 // TODO, we should do this once per privatization segment:
470 snapshot->process->read_bytes(
471 ignored_data.data.data(), region.size, remote(region.addr),
472 simgrid::mc::ProcessIndexDisabled);
473 snapshot->ignored_data.push_back(std::move(ignored_data));
477 for(auto const& region : mc_model_checker->process().ignored_regions()) {
478 snapshot->process->clear_bytes(remote(region.addr), region.size);
483 static void MC_snapshot_ignore_restore(mc_snapshot_t snapshot)
485 for (auto const& ignored_data : snapshot->ignored_data)
486 snapshot->process->write_bytes(
487 ignored_data.data.data(), ignored_data.data.size(),
488 remote(ignored_data.start));
491 static std::vector<s_fd_infos_t> MC_get_current_fds(pid_t pid)
493 const size_t fd_dir_path_size = 20;
494 char fd_dir_path[fd_dir_path_size];
495 int res = snprintf(fd_dir_path, fd_dir_path_size,
496 "/proc/%lli/fd", (long long int) pid);
497 xbt_assert(res >= 0);
498 if ((size_t) res > fd_dir_path_size)
499 xbt_die("Unexpected buffer is too small for fd_dir_path");
501 DIR* fd_dir = opendir(fd_dir_path);
503 xbt_die("Cannot open directory '/proc/self/fd'\n");
505 std::vector<s_fd_infos_t> fds;
507 struct dirent* fd_number;
508 while ((fd_number = readdir(fd_dir))) {
510 int fd_value = atoi(fd_number->d_name);
515 const size_t source_size = 25;
517 int res = snprintf(source, source_size, "/proc/%lli/fd/%s",
518 (long long int) pid, fd_number->d_name);
519 xbt_assert(res >= 0);
520 if ((size_t) res > source_size)
521 xbt_die("Unexpected buffer is too small for fd %s", fd_number->d_name);
523 const size_t link_size = 200;
525 res = readlink(source, link, link_size);
527 xbt_die("Could not read link for %s", source);
530 xbt_die("Buffer to small for link of %s", source);
534 if(smpi_is_privatisation_file(link))
537 // This is (probably) the DIR* we are reading:
538 // TODO, read all the file entries at once and close the DIR.*
539 if(strcmp(fd_dir_path, link) == 0)
542 // We don't handle them.
543 // It does not mean we should silently ignore them however.
544 if (strncmp(link, "pipe:", 5) == 0 || strncmp(link, "socket:", 7) == 0)
547 // If dot_output enabled, do not handle the corresponding file
548 if (dot_output != NULL && strcmp(basename(link), _sg_mc_dot_output_file) == 0)
551 // This is probably a shared memory used by lttng-ust:
552 if(strncmp("/dev/shm/ust-shm-tmp-", link, 21)==0)
555 // Add an entry for this FD in the snapshot:
557 fd.filename = std::string(link);
558 fd.number = fd_value;
559 fd.flags = fcntl(fd_value, F_GETFL) | fcntl(fd_value, F_GETFD) ;
560 fd.current_position = lseek(fd_value, 0, SEEK_CUR);
561 fds.push_back(std::move(fd));
565 return std::move(fds);
568 mc_snapshot_t MC_take_snapshot(int num_state)
570 XBT_DEBUG("Taking snapshot %i", num_state);
572 simgrid::mc::Process* mc_process = &mc_model_checker->process();
574 mc_snapshot_t snapshot = new simgrid::mc::Snapshot();
576 snapshot->process = mc_process;
577 snapshot->num_state = num_state;
579 smx_process_t process;
580 MC_EACH_SIMIX_PROCESS(process,
581 snapshot->enabled_processes.insert(process->pid));
583 MC_snapshot_handle_ignore(snapshot);
585 if (_sg_mc_snapshot_fds)
586 snapshot->current_fds = MC_get_current_fds(process->pid);
588 /* Save the std heap and the writable mapped pages of libsimgrid and binary */
589 MC_get_memory_regions(mc_process, snapshot);
591 snapshot->to_ignore = MC_take_snapshot_ignore();
593 if (_sg_mc_visited > 0 || strcmp(_sg_mc_property_file, "")) {
595 MC_take_snapshot_stacks(&snapshot);
596 if (_sg_mc_hash && !snapshot->stacks.empty()) {
597 snapshot->hash = simgrid::mc::hash(*snapshot);
605 MC_snapshot_ignore_restore(snapshot);
610 void MC_restore_snapshot_regions(mc_snapshot_t snapshot)
612 for(std::unique_ptr<s_mc_mem_region_t> const& region : snapshot->snapshot_regions) {
613 // For privatized, variables we decided it was not necessary to take the snapshot:
615 MC_region_restore(region.get());
619 // TODO, send a message to implement this in the MCed process
620 if(snapshot->privatization_index >= 0) {
621 // We just rewrote the global variables.
622 // The privatisation segment SMPI thinks
623 // is mapped might be inconsistent with the segment which
624 // is really mapped in memory (kernel state).
625 // We ask politely SMPI to map the segment anyway,
626 // even if it thinks it is the current one:
627 smpi_really_switch_data_segment(snapshot->privatization_index);
633 void MC_restore_snapshot_fds(mc_snapshot_t snapshot)
635 if (mc_mode == MC_MODE_SERVER)
636 xbt_die("FD snapshot not implemented in client/server mode.");
638 for (auto const& fd : snapshot->current_fds) {
640 int new_fd = open(fd.filename.c_str(), fd.flags);
642 xbt_die("Could not reopen the file %s fo restoring the file descriptor",
643 fd.filename.c_str());
645 if (new_fd != fd.number) {
646 dup2(new_fd, fd.number);
649 lseek(fd.number, fd.current_position, SEEK_SET);
653 void MC_restore_snapshot(mc_snapshot_t snapshot)
655 XBT_DEBUG("Restore snapshot %i", snapshot->num_state);
656 MC_restore_snapshot_regions(snapshot);
657 if (_sg_mc_snapshot_fds)
658 MC_restore_snapshot_fds(snapshot);
659 MC_snapshot_ignore_restore(snapshot);
660 mc_model_checker->process().cache_flags = 0;
663 mc_snapshot_t simcall_HANDLER_mc_snapshot(smx_simcall_t simcall)
665 return MC_take_snapshot(1);