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_dwarf.hpp"
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 /** @brief Restore a region from a snapshot
55 * @param reg Target region
57 static void MC_region_restore(mc_mem_region_t region)
59 switch(region->storage_type()) {
60 case simgrid::mc::StorageType::NoData:
62 xbt_die("Storage type not supported");
65 case simgrid::mc::StorageType::Flat:
66 mc_model_checker->process().write_bytes(region->flat_data().data(),
67 region->size(), region->permanent_address());
70 case simgrid::mc::StorageType::Chunked:
71 mc_region_restore_sparse(&mc_model_checker->process(), region);
74 case simgrid::mc::StorageType::Privatized:
75 for (auto& p : region->privatized_data())
76 MC_region_restore(&p);
87 simgrid::mc::RegionSnapshot privatized_region(
88 RegionType region_type, void *start_addr, void* permanent_addr, size_t size
91 size_t process_count = MC_smpi_process_count();
93 // Read smpi_privatisation_regions from MCed:
94 smpi_privatisation_region_t remote_smpi_privatisation_regions;
95 mc_model_checker->process().read_variable(
96 "smpi_privatisation_regions",
97 &remote_smpi_privatisation_regions, sizeof(remote_smpi_privatisation_regions));
98 s_smpi_privatisation_region_t privatisation_regions[process_count];
99 mc_model_checker->process().read_bytes(
100 &privatisation_regions, sizeof(privatisation_regions),
101 remote(remote_smpi_privatisation_regions));
103 std::vector<simgrid::mc::RegionSnapshot> data;
104 data.reserve(process_count);
105 for (size_t i = 0; i < process_count; i++)
107 simgrid::mc::region(region_type, start_addr,
108 privatisation_regions[i].address, size)
111 simgrid::mc::RegionSnapshot region = simgrid::mc::RegionSnapshot(
112 region_type, start_addr, permanent_addr, size);
113 region.privatized_data(std::move(data));
114 return std::move(region);
123 static void MC_snapshot_add_region(int index, mc_snapshot_t snapshot,
124 simgrid::mc::RegionType type,
125 simgrid::mc::ObjectInformation* object_info,
126 void *start_addr, void* permanent_addr, size_t size)
128 if (type == simgrid::mc::RegionType::Data)
129 xbt_assert(object_info, "Missing object info for object.");
130 else if (type == simgrid::mc::RegionType::Heap)
131 xbt_assert(!object_info, "Unexpected object info for heap region.");
133 simgrid::mc::RegionSnapshot region;
136 const bool privatization_aware = object_info && object_info->privatized();
137 if (privatization_aware && MC_smpi_process_count())
138 region = simgrid::mc::privatized_region(type, start_addr, permanent_addr, size);
141 region = simgrid::mc::region(type, start_addr, permanent_addr, size);
143 region.object_info(object_info);
144 snapshot->snapshot_regions[index]
145 = std::unique_ptr<simgrid::mc::RegionSnapshot>(
146 new simgrid::mc::RegionSnapshot(std::move(region)));
150 static void MC_get_memory_regions(simgrid::mc::Process* process, mc_snapshot_t snapshot)
152 const size_t n = process->object_infos.size();
153 snapshot->snapshot_regions.resize(n + 1);
155 for (auto const& object_info : process->object_infos) {
156 MC_snapshot_add_region(i, snapshot, simgrid::mc::RegionType::Data,
158 object_info->start_rw, object_info->start_rw,
159 object_info->end_rw - object_info->start_rw);
163 xbt_mheap_t heap = process->get_heap();
164 void *start_heap = heap->base;
165 void *end_heap = heap->breakval;
167 MC_snapshot_add_region(n, snapshot, simgrid::mc::RegionType::Heap, NULL,
168 start_heap, start_heap,
169 (char *) end_heap - (char *) start_heap);
170 snapshot->heap_bytes_used = mmalloc_get_bytes_used_remote(
172 process->get_malloc_info());
175 if (smpi_privatize_global_variables && MC_smpi_process_count()) {
176 // snapshot->privatization_index = smpi_loaded_page
177 mc_model_checker->process().read_variable(
178 "smpi_loaded_page", &snapshot->privatization_index,
179 sizeof(snapshot->privatization_index));
183 snapshot->privatization_index = simgrid::mc::ProcessIndexMissing;
187 /** \brief Fills the position of the segments (executable, read-only, read/write).
189 * `dl_iterate_phdr` would be more robust but would not work in cross-process.
191 void MC_find_object_address(
192 std::vector<simgrid::mc::VmMap> const& maps, simgrid::mc::ObjectInformation* result)
194 const char* file_name = xbt_strdup(result->file_name.c_str());
195 const char *name = basename(file_name);
196 for (size_t i = 0; i < maps.size(); ++i) {
197 simgrid::mc::VmMap const& reg = maps[i];
198 if (maps[i].pathname.empty()
199 || strcmp(basename(maps[i].pathname.c_str()), name)) {
201 } else if ((reg.prot & PROT_WRITE)) {
202 xbt_assert(!result->start_rw,
203 "Multiple read-write segments for %s, not supported",
204 maps[i].pathname.c_str());
205 result->start_rw = (char*) reg.start_addr;
206 result->end_rw = (char*) reg.end_addr;
207 // .bss is usually after the .data:
208 simgrid::mc::VmMap const& next = maps[i + 1];
209 if (next.pathname.empty() && (next.prot & PROT_WRITE)
210 && next.start_addr == reg.end_addr) {
211 result->end_rw = (char*) maps[i + 1].end_addr;
213 } else if ((reg.prot & PROT_READ) && (reg.prot & PROT_EXEC)) {
214 xbt_assert(!result->start_exec,
215 "Multiple executable segments for %s, not supported",
216 maps[i].pathname.c_str());
217 result->start_exec = (char*) reg.start_addr;
218 result->end_exec = (char*) reg.end_addr;
219 } else if ((reg.prot & PROT_READ) && !(reg.prot & PROT_EXEC)) {
220 xbt_assert(!result->start_ro,
221 "Multiple read only segments for %s, not supported",
222 maps[i].pathname.c_str());
223 result->start_ro = (char*) reg.start_addr;
224 result->end_ro = (char*) reg.end_addr;
228 result->start = result->start_rw;
229 if ((const void*) result->start_ro > result->start)
230 result->start = result->start_ro;
231 if ((const void*) result->start_exec > result->start)
232 result->start = result->start_exec;
234 result->end = result->end_rw;
235 if (result->end_ro && (const void*) result->end_ro < result->end)
236 result->end = result->end_ro;
237 if (result->end_exec && (const void*) result->end_exec > result->end)
238 result->end = result->end_exec;
240 xbt_assert(result->start_rw);
241 xbt_assert(result->start_exec);
244 /************************************* Take Snapshot ************************************/
245 /****************************************************************************************/
247 /** \brief Checks whether the variable is in scope for a given IP.
249 * A variable may be defined only from a given value of IP.
251 * \param var Variable description
252 * \param frame Scope description
253 * \param ip Instruction pointer
254 * \return true if the variable is valid
256 static bool mc_valid_variable(simgrid::mc::Variable* var, simgrid::mc::Frame* scope,
259 // The variable is not yet valid:
260 if ((const void *) ((const char *) scope->low_pc + var->start_scope) > ip)
266 static void mc_fill_local_variables_values(mc_stack_frame_t stack_frame,
267 simgrid::mc::Frame* scope, int process_index,
268 std::vector<s_local_variable>& result)
270 simgrid::mc::Process* process = &mc_model_checker->process();
272 void *ip = (void *) stack_frame->ip;
273 if (ip < scope->low_pc || ip >= scope->high_pc)
276 for(simgrid::mc::Variable& current_variable :
279 if (!mc_valid_variable(¤t_variable, scope, (void *) stack_frame->ip))
283 // FIXME, get rid of `region_type`
284 if ((long) stack_frame->ip > (long) process->libsimgrid_info->start_exec)
289 s_local_variable_t new_var;
290 new_var.subprogram = stack_frame->frame;
291 new_var.ip = stack_frame->ip;
292 new_var.name = current_variable.name;
293 new_var.type = current_variable.type;
294 new_var.region = region_type;
295 new_var.address = nullptr;
297 if (current_variable.address != NULL) {
298 new_var.address = current_variable.address;
299 } else if (!current_variable.location_list.empty()) {
300 s_mc_location_t location;
301 mc_dwarf_resolve_locations(
302 &location, ¤t_variable.location_list,
303 current_variable.object_info,
304 &(stack_frame->unw_cursor),
305 (void *) stack_frame->frame_base,
306 &mc_model_checker->process(), process_index);
308 switch(mc_get_location_type(&location)) {
309 case MC_LOCATION_TYPE_ADDRESS:
310 new_var.address = location.memory_location;
312 case MC_LOCATION_TYPE_REGISTER:
314 xbt_die("Cannot handle non-address variable");
318 xbt_die("No address");
321 result.push_back(std::move(new_var));
324 // Recursive processing of nested scopes:
325 for(simgrid::mc::Frame& nested_scope : scope->scopes)
326 mc_fill_local_variables_values(
327 stack_frame, &nested_scope, process_index, result);
330 static std::vector<s_local_variable> MC_get_local_variables_values(
331 std::vector<s_mc_stack_frame_t>& stack_frames, int process_index)
333 std::vector<s_local_variable> variables;
334 for (s_mc_stack_frame_t& stack_frame : stack_frames)
335 mc_fill_local_variables_values(&stack_frame, stack_frame.frame, process_index, variables);
336 return std::move(variables);
339 static void MC_stack_frame_free_voipd(void *s)
341 mc_stack_frame_t stack_frame = *(mc_stack_frame_t *) s;
345 static std::vector<s_mc_stack_frame_t> MC_unwind_stack_frames(mc_unw_context_t stack_context)
347 simgrid::mc::Process* process = &mc_model_checker->process();
348 std::vector<s_mc_stack_frame_t> result;
352 // TODO, check condition check (unw_init_local==0 means end of frame)
353 if (mc_unw_init_cursor(&c, stack_context) != 0) {
355 xbt_die("Could not initialize stack unwinding");
360 s_mc_stack_frame_t stack_frame;
362 stack_frame.unw_cursor = c;
366 unw_get_reg(&c, UNW_REG_IP, &ip);
367 unw_get_reg(&c, UNW_REG_SP, &sp);
372 // TODO, use real addresses in frame_t instead of fixing it here
374 simgrid::mc::Frame* frame = process->find_function(remote(ip));
375 stack_frame.frame = frame;
378 stack_frame.frame_name = frame->name;
379 stack_frame.frame_base =
380 (unw_word_t) mc_find_frame_base(frame, frame->object_info, &c);
382 stack_frame.frame_base = 0;
383 stack_frame.frame_name = std::string();
386 result.push_back(std::move(stack_frame));
388 /* Stop before context switch with maestro */
389 if (frame != nullptr &&
390 frame->name == "smx_ctx_sysv_wrapper")
393 int ret = unw_step(&c);
395 xbt_die("Unexpected end of stack.");
396 } else if (ret < 0) {
397 xbt_die("Error while unwinding stack");
401 if (result.empty()) {
402 XBT_INFO("unw_init_local failed");
406 return std::move(result);
409 static std::vector<s_mc_snapshot_stack_t> MC_take_snapshot_stacks(mc_snapshot_t * snapshot)
411 std::vector<s_mc_snapshot_stack_t> res;
413 unsigned int cursor = 0;
414 stack_region_t current_stack;
416 // FIXME, cross-process support (stack_areas)
417 xbt_dynar_foreach(stacks_areas, cursor, current_stack) {
418 s_mc_snapshot_stack_t st;
420 // Read the context from remote process:
421 unw_context_t context;
422 mc_model_checker->process().read_bytes(
423 &context, sizeof(context), remote(current_stack->context));
425 if (mc_unw_init_context(&st.context, &mc_model_checker->process(),
427 xbt_die("Could not initialise the libunwind context.");
429 st.stack_frames = MC_unwind_stack_frames(&st.context);
430 st.local_variables = MC_get_local_variables_values(st.stack_frames, current_stack->process_index);
431 st.process_index = current_stack->process_index;
433 unw_word_t sp = st.stack_frames[0].sp;
435 res.push_back(std::move(st));
438 (char*) current_stack->address + current_stack->size - (char*) sp;
439 (*snapshot)->stack_sizes.push_back(stack_size);
442 return std::move(res);
446 static std::vector<s_mc_heap_ignore_region_t> MC_take_snapshot_ignore()
448 std::vector<s_mc_heap_ignore_region_t> res;
450 if (mc_heap_comparison_ignore == NULL)
451 return std::move(res);
453 unsigned int cursor = 0;
454 mc_heap_ignore_region_t current_region;
456 xbt_dynar_foreach(mc_heap_comparison_ignore, cursor, current_region) {
457 s_mc_heap_ignore_region_t new_region;
458 new_region.address = current_region->address;
459 new_region.size = current_region->size;
460 new_region.block = current_region->block;
461 new_region.fragment = current_region->fragment;
462 res.push_back(std::move(new_region));
465 return std::move(res);
468 static void MC_snapshot_handle_ignore(mc_snapshot_t snapshot)
470 xbt_assert(snapshot->process);
473 for (auto const& region : mc_model_checker->process().ignored_regions()) {
474 s_mc_snapshot_ignored_data_t ignored_data;
475 ignored_data.start = (void*)region.addr;
476 ignored_data.data.resize(region.size);
477 // TODO, we should do this once per privatization segment:
478 snapshot->process->read_bytes(
479 ignored_data.data.data(), region.size, remote(region.addr),
480 simgrid::mc::ProcessIndexDisabled);
481 snapshot->ignored_data.push_back(std::move(ignored_data));
485 for(auto const& region : mc_model_checker->process().ignored_regions()) {
486 snapshot->process->clear_bytes(remote(region.addr), region.size);
491 static void MC_snapshot_ignore_restore(mc_snapshot_t snapshot)
493 for (auto const& ignored_data : snapshot->ignored_data)
494 snapshot->process->write_bytes(
495 ignored_data.data.data(), ignored_data.data.size(),
496 remote(ignored_data.start));
499 static std::vector<s_fd_infos_t> MC_get_current_fds(pid_t pid)
501 const size_t fd_dir_path_size = 20;
502 char fd_dir_path[fd_dir_path_size];
503 int res = snprintf(fd_dir_path, fd_dir_path_size,
504 "/proc/%lli/fd", (long long int) pid);
505 xbt_assert(res >= 0);
506 if ((size_t) res > fd_dir_path_size)
507 xbt_die("Unexpected buffer is too small for fd_dir_path");
509 DIR* fd_dir = opendir(fd_dir_path);
511 xbt_die("Cannot open directory '/proc/self/fd'\n");
513 std::vector<s_fd_infos_t> fds;
515 struct dirent* fd_number;
516 while ((fd_number = readdir(fd_dir))) {
518 int fd_value = atoi(fd_number->d_name);
523 const size_t source_size = 25;
525 int res = snprintf(source, source_size, "/proc/%lli/fd/%s",
526 (long long int) pid, fd_number->d_name);
527 xbt_assert(res >= 0);
528 if ((size_t) res > source_size)
529 xbt_die("Unexpected buffer is too small for fd %s", fd_number->d_name);
531 const size_t link_size = 200;
533 res = readlink(source, link, link_size);
535 xbt_die("Could not read link for %s", source);
538 xbt_die("Buffer to small for link of %s", source);
543 if(smpi_is_privatisation_file(link))
547 // This is (probably) the DIR* we are reading:
548 // TODO, read all the file entries at once and close the DIR.*
549 if(strcmp(fd_dir_path, link) == 0)
552 // We don't handle them.
553 // It does not mean we should silently ignore them however.
554 if (strncmp(link, "pipe:", 5) == 0 || strncmp(link, "socket:", 7) == 0)
557 // If dot_output enabled, do not handle the corresponding file
558 if (dot_output != NULL && strcmp(basename(link), _sg_mc_dot_output_file) == 0)
561 // This is probably a shared memory used by lttng-ust:
562 if(strncmp("/dev/shm/ust-shm-tmp-", link, 21)==0)
565 // Add an entry for this FD in the snapshot:
567 fd.filename = std::string(link);
568 fd.number = fd_value;
569 fd.flags = fcntl(fd_value, F_GETFL) | fcntl(fd_value, F_GETFD) ;
570 fd.current_position = lseek(fd_value, 0, SEEK_CUR);
571 fds.push_back(std::move(fd));
575 return std::move(fds);
578 mc_snapshot_t MC_take_snapshot(int num_state)
580 XBT_DEBUG("Taking snapshot %i", num_state);
582 simgrid::mc::Process* mc_process = &mc_model_checker->process();
584 mc_snapshot_t snapshot = new simgrid::mc::Snapshot();
586 snapshot->process = mc_process;
587 snapshot->num_state = num_state;
589 smx_process_t process;
590 MC_EACH_SIMIX_PROCESS(process,
591 snapshot->enabled_processes.insert(process->pid));
593 MC_snapshot_handle_ignore(snapshot);
595 if (_sg_mc_snapshot_fds)
596 snapshot->current_fds = MC_get_current_fds(process->pid);
598 /* Save the std heap and the writable mapped pages of libsimgrid and binary */
599 MC_get_memory_regions(mc_process, snapshot);
601 snapshot->to_ignore = MC_take_snapshot_ignore();
603 if (_sg_mc_visited > 0 || strcmp(_sg_mc_property_file, "")) {
605 MC_take_snapshot_stacks(&snapshot);
606 if (_sg_mc_hash && !snapshot->stacks.empty()) {
607 snapshot->hash = simgrid::mc::hash(*snapshot);
615 MC_snapshot_ignore_restore(snapshot);
620 void MC_restore_snapshot_regions(mc_snapshot_t snapshot)
622 for(std::unique_ptr<s_mc_mem_region_t> const& region : snapshot->snapshot_regions) {
623 // For privatized, variables we decided it was not necessary to take the snapshot:
625 MC_region_restore(region.get());
629 // TODO, send a message to implement this in the MCed process
630 if(snapshot->privatization_index >= 0) {
631 // We just rewrote the global variables.
632 // The privatisation segment SMPI thinks
633 // is mapped might be inconsistent with the segment which
634 // is really mapped in memory (kernel state).
635 // We ask politely SMPI to map the segment anyway,
636 // even if it thinks it is the current one:
637 smpi_really_switch_data_segment(snapshot->privatization_index);
643 void MC_restore_snapshot_fds(mc_snapshot_t snapshot)
645 if (mc_mode == MC_MODE_SERVER)
646 xbt_die("FD snapshot not implemented in client/server mode.");
648 for (auto const& fd : snapshot->current_fds) {
650 int new_fd = open(fd.filename.c_str(), fd.flags);
652 xbt_die("Could not reopen the file %s fo restoring the file descriptor",
653 fd.filename.c_str());
655 if (new_fd != fd.number) {
656 dup2(new_fd, fd.number);
659 lseek(fd.number, fd.current_position, SEEK_SET);
663 void MC_restore_snapshot(mc_snapshot_t snapshot)
665 XBT_DEBUG("Restore snapshot %i", snapshot->num_state);
666 MC_restore_snapshot_regions(snapshot);
667 if (_sg_mc_snapshot_fds)
668 MC_restore_snapshot_fds(snapshot);
669 MC_snapshot_ignore_restore(snapshot);
670 mc_model_checker->process().cache_flags = 0;
673 mc_snapshot_t simcall_HANDLER_mc_snapshot(smx_simcall_t simcall)
675 return MC_take_snapshot(1);