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 /** @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);
86 simgrid::mc::RegionSnapshot privatized_region(
87 RegionType region_type, void *start_addr, void* permanent_addr, size_t size
90 size_t process_count = MC_smpi_process_count();
92 // Read smpi_privatisation_regions from MCed:
93 smpi_privatisation_region_t remote_smpi_privatisation_regions;
94 mc_model_checker->process().read_variable(
95 "smpi_privatisation_regions",
96 &remote_smpi_privatisation_regions, sizeof(remote_smpi_privatisation_regions));
97 s_smpi_privatisation_region_t privatisation_regions[process_count];
98 mc_model_checker->process().read_bytes(
99 &privatisation_regions, sizeof(privatisation_regions),
100 remote(remote_smpi_privatisation_regions));
102 std::vector<simgrid::mc::RegionSnapshot> data;
103 data.reserve(process_count);
104 for (size_t i = 0; i < process_count; i++)
106 simgrid::mc::region(region_type, start_addr,
107 privatisation_regions[i].address, size)
110 simgrid::mc::RegionSnapshot region = simgrid::mc::RegionSnapshot(
111 region_type, start_addr, permanent_addr, size);
112 region.privatized_data(std::move(data));
113 return std::move(region);
121 static void MC_snapshot_add_region(int index, mc_snapshot_t snapshot,
122 simgrid::mc::RegionType type,
123 simgrid::mc::ObjectInformation* object_info,
124 void *start_addr, void* permanent_addr, size_t size)
126 if (type == simgrid::mc::RegionType::Data)
127 xbt_assert(object_info, "Missing object info for object.");
128 else if (type == simgrid::mc::RegionType::Heap)
129 xbt_assert(!object_info, "Unexpected object info for heap region.");
131 const bool privatization_aware = object_info && object_info->privatized();
133 simgrid::mc::RegionSnapshot region;
134 if (privatization_aware && MC_smpi_process_count())
135 region = simgrid::mc::privatized_region(type, start_addr, permanent_addr, size);
137 region = simgrid::mc::region(type, start_addr, permanent_addr, size);
139 region.object_info(object_info);
140 snapshot->snapshot_regions[index]
141 = std::unique_ptr<simgrid::mc::RegionSnapshot>(
142 new simgrid::mc::RegionSnapshot(std::move(region)));
146 static void MC_get_memory_regions(simgrid::mc::Process* process, mc_snapshot_t snapshot)
148 const size_t n = process->object_infos.size();
149 snapshot->snapshot_regions.resize(n + 1);
151 for (auto const& object_info : process->object_infos) {
152 MC_snapshot_add_region(i, snapshot, simgrid::mc::RegionType::Data,
154 object_info->start_rw, object_info->start_rw,
155 object_info->end_rw - object_info->start_rw);
159 xbt_mheap_t heap = process->get_heap();
160 void *start_heap = heap->base;
161 void *end_heap = heap->breakval;
163 MC_snapshot_add_region(n, snapshot, simgrid::mc::RegionType::Heap, NULL,
164 start_heap, start_heap,
165 (char *) end_heap - (char *) start_heap);
166 snapshot->heap_bytes_used = mmalloc_get_bytes_used_remote(
168 process->get_malloc_info());
171 if (smpi_privatize_global_variables && MC_smpi_process_count()) {
172 // snapshot->privatization_index = smpi_loaded_page
173 mc_model_checker->process().read_variable(
174 "smpi_loaded_page", &snapshot->privatization_index,
175 sizeof(snapshot->privatization_index));
179 snapshot->privatization_index = simgrid::mc::ProcessIndexMissing;
183 /** \brief Fills the position of the segments (executable, read-only, read/write).
185 * `dl_iterate_phdr` would be more robust but would not work in cross-process.
187 void MC_find_object_address(
188 std::vector<simgrid::mc::VmMap> const& maps, simgrid::mc::ObjectInformation* result)
190 const char* file_name = xbt_strdup(result->file_name.c_str());
191 const char *name = basename(file_name);
192 for (size_t i = 0; i < maps.size(); ++i) {
193 simgrid::mc::VmMap const& reg = maps[i];
194 if (maps[i].pathname.empty()
195 || strcmp(basename(maps[i].pathname.c_str()), name)) {
197 } else if ((reg.prot & PROT_WRITE)) {
198 xbt_assert(!result->start_rw,
199 "Multiple read-write segments for %s, not supported",
200 maps[i].pathname.c_str());
201 result->start_rw = (char*) reg.start_addr;
202 result->end_rw = (char*) reg.end_addr;
203 // .bss is usually after the .data:
204 simgrid::mc::VmMap const& next = maps[i + 1];
205 if (next.pathname.empty() && (next.prot & PROT_WRITE)
206 && next.start_addr == reg.end_addr) {
207 result->end_rw = (char*) maps[i + 1].end_addr;
209 } else if ((reg.prot & PROT_READ) && (reg.prot & PROT_EXEC)) {
210 xbt_assert(!result->start_exec,
211 "Multiple executable segments for %s, not supported",
212 maps[i].pathname.c_str());
213 result->start_exec = (char*) reg.start_addr;
214 result->end_exec = (char*) reg.end_addr;
215 } else if ((reg.prot & PROT_READ) && !(reg.prot & PROT_EXEC)) {
216 xbt_assert(!result->start_ro,
217 "Multiple read only segments for %s, not supported",
218 maps[i].pathname.c_str());
219 result->start_ro = (char*) reg.start_addr;
220 result->end_ro = (char*) reg.end_addr;
224 result->start = result->start_rw;
225 if ((const void*) result->start_ro > result->start)
226 result->start = result->start_ro;
227 if ((const void*) result->start_exec > result->start)
228 result->start = result->start_exec;
230 result->end = result->end_rw;
231 if (result->end_ro && (const void*) result->end_ro < result->end)
232 result->end = result->end_ro;
233 if (result->end_exec && (const void*) result->end_exec > result->end)
234 result->end = result->end_exec;
236 xbt_assert(result->start_rw);
237 xbt_assert(result->start_exec);
240 /************************************* Take Snapshot ************************************/
241 /****************************************************************************************/
243 /** \brief Checks whether the variable is in scope for a given IP.
245 * A variable may be defined only from a given value of IP.
247 * \param var Variable description
248 * \param frame Scope description
249 * \param ip Instruction pointer
250 * \return true if the variable is valid
252 static bool mc_valid_variable(simgrid::mc::Variable* var, simgrid::mc::Frame* scope,
255 // The variable is not yet valid:
256 if ((const void *) ((const char *) scope->low_pc + var->start_scope) > ip)
262 static void mc_fill_local_variables_values(mc_stack_frame_t stack_frame,
263 simgrid::mc::Frame* scope, int process_index,
264 std::vector<s_local_variable>& result)
266 simgrid::mc::Process* process = &mc_model_checker->process();
268 void *ip = (void *) stack_frame->ip;
269 if (ip < scope->low_pc || ip >= scope->high_pc)
272 for(simgrid::mc::Variable& current_variable :
275 if (!mc_valid_variable(¤t_variable, scope, (void *) stack_frame->ip))
279 // FIXME, get rid of `region_type`
280 if ((long) stack_frame->ip > (long) process->libsimgrid_info->start_exec)
285 s_local_variable_t new_var;
286 new_var.subprogram = stack_frame->frame;
287 new_var.ip = stack_frame->ip;
288 new_var.name = current_variable.name;
289 new_var.type = current_variable.type;
290 new_var.region = region_type;
291 new_var.address = nullptr;
293 if (current_variable.address != NULL) {
294 new_var.address = current_variable.address;
295 } else if (!current_variable.location_list.empty()) {
296 s_mc_location_t location;
297 mc_dwarf_resolve_locations(
298 &location, ¤t_variable.location_list,
299 current_variable.object_info,
300 &(stack_frame->unw_cursor),
301 (void *) stack_frame->frame_base,
302 &mc_model_checker->process(), process_index);
304 switch(mc_get_location_type(&location)) {
305 case MC_LOCATION_TYPE_ADDRESS:
306 new_var.address = location.memory_location;
308 case MC_LOCATION_TYPE_REGISTER:
310 xbt_die("Cannot handle non-address variable");
314 xbt_die("No address");
317 result.push_back(std::move(new_var));
320 // Recursive processing of nested scopes:
321 for(simgrid::mc::Frame& nested_scope : scope->scopes)
322 mc_fill_local_variables_values(
323 stack_frame, &nested_scope, process_index, result);
326 static std::vector<s_local_variable> MC_get_local_variables_values(
327 std::vector<s_mc_stack_frame_t>& stack_frames, int process_index)
329 std::vector<s_local_variable> variables;
330 for (s_mc_stack_frame_t& stack_frame : stack_frames)
331 mc_fill_local_variables_values(&stack_frame, stack_frame.frame, process_index, variables);
332 return std::move(variables);
335 static void MC_stack_frame_free_voipd(void *s)
337 mc_stack_frame_t stack_frame = *(mc_stack_frame_t *) s;
341 static std::vector<s_mc_stack_frame_t> MC_unwind_stack_frames(mc_unw_context_t stack_context)
343 simgrid::mc::Process* process = &mc_model_checker->process();
344 std::vector<s_mc_stack_frame_t> result;
348 // TODO, check condition check (unw_init_local==0 means end of frame)
349 if (mc_unw_init_cursor(&c, stack_context) != 0) {
351 xbt_die("Could not initialize stack unwinding");
356 s_mc_stack_frame_t stack_frame;
358 stack_frame.unw_cursor = c;
362 unw_get_reg(&c, UNW_REG_IP, &ip);
363 unw_get_reg(&c, UNW_REG_SP, &sp);
368 // TODO, use real addresses in frame_t instead of fixing it here
370 simgrid::mc::Frame* frame = process->find_function(remote(ip));
371 stack_frame.frame = frame;
374 stack_frame.frame_name = frame->name;
375 stack_frame.frame_base =
376 (unw_word_t) mc_find_frame_base(frame, frame->object_info, &c);
378 stack_frame.frame_base = 0;
379 stack_frame.frame_name = std::string();
382 result.push_back(std::move(stack_frame));
384 /* Stop before context switch with maestro */
385 if (frame != nullptr &&
386 frame->name == "smx_ctx_sysv_wrapper")
389 int ret = unw_step(&c);
391 xbt_die("Unexpected end of stack.");
392 } else if (ret < 0) {
393 xbt_die("Error while unwinding stack");
397 if (result.empty()) {
398 XBT_INFO("unw_init_local failed");
402 return std::move(result);
405 static std::vector<s_mc_snapshot_stack_t> MC_take_snapshot_stacks(mc_snapshot_t * snapshot)
407 std::vector<s_mc_snapshot_stack_t> res;
409 unsigned int cursor = 0;
410 stack_region_t current_stack;
412 // FIXME, cross-process support (stack_areas)
413 xbt_dynar_foreach(stacks_areas, cursor, current_stack) {
414 s_mc_snapshot_stack_t st;
416 // Read the context from remote process:
417 unw_context_t context;
418 mc_model_checker->process().read_bytes(
419 &context, sizeof(context), remote(current_stack->context));
421 if (mc_unw_init_context(&st.context, &mc_model_checker->process(),
423 xbt_die("Could not initialise the libunwind context.");
425 st.stack_frames = MC_unwind_stack_frames(&st.context);
426 st.local_variables = MC_get_local_variables_values(st.stack_frames, current_stack->process_index);
427 st.process_index = current_stack->process_index;
429 unw_word_t sp = st.stack_frames[0].sp;
431 res.push_back(std::move(st));
434 (char*) current_stack->address + current_stack->size - (char*) sp;
435 (*snapshot)->stack_sizes.push_back(stack_size);
438 return std::move(res);
442 static std::vector<s_mc_heap_ignore_region_t> MC_take_snapshot_ignore()
444 std::vector<s_mc_heap_ignore_region_t> res;
446 if (mc_heap_comparison_ignore == NULL)
447 return std::move(res);
449 unsigned int cursor = 0;
450 mc_heap_ignore_region_t current_region;
452 xbt_dynar_foreach(mc_heap_comparison_ignore, cursor, current_region) {
453 s_mc_heap_ignore_region_t new_region;
454 new_region.address = current_region->address;
455 new_region.size = current_region->size;
456 new_region.block = current_region->block;
457 new_region.fragment = current_region->fragment;
458 res.push_back(std::move(new_region));
461 return std::move(res);
464 static void MC_snapshot_handle_ignore(mc_snapshot_t snapshot)
466 xbt_assert(snapshot->process);
469 for (auto const& region : mc_model_checker->process().ignored_regions()) {
470 s_mc_snapshot_ignored_data_t ignored_data;
471 ignored_data.start = (void*)region.addr;
472 ignored_data.data.resize(region.size);
473 // TODO, we should do this once per privatization segment:
474 snapshot->process->read_bytes(
475 ignored_data.data.data(), region.size, remote(region.addr),
476 simgrid::mc::ProcessIndexDisabled);
477 snapshot->ignored_data.push_back(std::move(ignored_data));
481 for(auto const& region : mc_model_checker->process().ignored_regions()) {
482 snapshot->process->clear_bytes(remote(region.addr), region.size);
487 static void MC_snapshot_ignore_restore(mc_snapshot_t snapshot)
489 for (auto const& ignored_data : snapshot->ignored_data)
490 snapshot->process->write_bytes(
491 ignored_data.data.data(), ignored_data.data.size(),
492 remote(ignored_data.start));
495 static std::vector<s_fd_infos_t> MC_get_current_fds(pid_t pid)
497 const size_t fd_dir_path_size = 20;
498 char fd_dir_path[fd_dir_path_size];
499 int res = snprintf(fd_dir_path, fd_dir_path_size,
500 "/proc/%lli/fd", (long long int) pid);
501 xbt_assert(res >= 0);
502 if ((size_t) res > fd_dir_path_size)
503 xbt_die("Unexpected buffer is too small for fd_dir_path");
505 DIR* fd_dir = opendir(fd_dir_path);
507 xbt_die("Cannot open directory '/proc/self/fd'\n");
509 std::vector<s_fd_infos_t> fds;
511 struct dirent* fd_number;
512 while ((fd_number = readdir(fd_dir))) {
514 int fd_value = atoi(fd_number->d_name);
519 const size_t source_size = 25;
521 int res = snprintf(source, source_size, "/proc/%lli/fd/%s",
522 (long long int) pid, fd_number->d_name);
523 xbt_assert(res >= 0);
524 if ((size_t) res > source_size)
525 xbt_die("Unexpected buffer is too small for fd %s", fd_number->d_name);
527 const size_t link_size = 200;
529 res = readlink(source, link, link_size);
531 xbt_die("Could not read link for %s", source);
534 xbt_die("Buffer to small for link of %s", source);
538 if(smpi_is_privatisation_file(link))
541 // This is (probably) the DIR* we are reading:
542 // TODO, read all the file entries at once and close the DIR.*
543 if(strcmp(fd_dir_path, link) == 0)
546 // We don't handle them.
547 // It does not mean we should silently ignore them however.
548 if (strncmp(link, "pipe:", 5) == 0 || strncmp(link, "socket:", 7) == 0)
551 // If dot_output enabled, do not handle the corresponding file
552 if (dot_output != NULL && strcmp(basename(link), _sg_mc_dot_output_file) == 0)
555 // This is probably a shared memory used by lttng-ust:
556 if(strncmp("/dev/shm/ust-shm-tmp-", link, 21)==0)
559 // Add an entry for this FD in the snapshot:
561 fd.filename = std::string(link);
562 fd.number = fd_value;
563 fd.flags = fcntl(fd_value, F_GETFL) | fcntl(fd_value, F_GETFD) ;
564 fd.current_position = lseek(fd_value, 0, SEEK_CUR);
565 fds.push_back(std::move(fd));
569 return std::move(fds);
572 mc_snapshot_t MC_take_snapshot(int num_state)
574 XBT_DEBUG("Taking snapshot %i", num_state);
576 simgrid::mc::Process* mc_process = &mc_model_checker->process();
578 mc_snapshot_t snapshot = new simgrid::mc::Snapshot();
580 snapshot->process = mc_process;
581 snapshot->num_state = num_state;
583 smx_process_t process;
584 MC_EACH_SIMIX_PROCESS(process,
585 snapshot->enabled_processes.insert(process->pid));
587 MC_snapshot_handle_ignore(snapshot);
589 if (_sg_mc_snapshot_fds)
590 snapshot->current_fds = MC_get_current_fds(process->pid);
592 /* Save the std heap and the writable mapped pages of libsimgrid and binary */
593 MC_get_memory_regions(mc_process, snapshot);
595 snapshot->to_ignore = MC_take_snapshot_ignore();
597 if (_sg_mc_visited > 0 || strcmp(_sg_mc_property_file, "")) {
599 MC_take_snapshot_stacks(&snapshot);
600 if (_sg_mc_hash && !snapshot->stacks.empty()) {
601 snapshot->hash = simgrid::mc::hash(*snapshot);
609 MC_snapshot_ignore_restore(snapshot);
614 void MC_restore_snapshot_regions(mc_snapshot_t snapshot)
616 for(std::unique_ptr<s_mc_mem_region_t> const& region : snapshot->snapshot_regions) {
617 // For privatized, variables we decided it was not necessary to take the snapshot:
619 MC_region_restore(region.get());
623 // TODO, send a message to implement this in the MCed process
624 if(snapshot->privatization_index >= 0) {
625 // We just rewrote the global variables.
626 // The privatisation segment SMPI thinks
627 // is mapped might be inconsistent with the segment which
628 // is really mapped in memory (kernel state).
629 // We ask politely SMPI to map the segment anyway,
630 // even if it thinks it is the current one:
631 smpi_really_switch_data_segment(snapshot->privatization_index);
637 void MC_restore_snapshot_fds(mc_snapshot_t snapshot)
639 if (mc_mode == MC_MODE_SERVER)
640 xbt_die("FD snapshot not implemented in client/server mode.");
642 for (auto const& fd : snapshot->current_fds) {
644 int new_fd = open(fd.filename.c_str(), fd.flags);
646 xbt_die("Could not reopen the file %s fo restoring the file descriptor",
647 fd.filename.c_str());
649 if (new_fd != fd.number) {
650 dup2(new_fd, fd.number);
653 lseek(fd.number, fd.current_position, SEEK_SET);
657 void MC_restore_snapshot(mc_snapshot_t snapshot)
659 XBT_DEBUG("Restore snapshot %i", snapshot->num_state);
660 MC_restore_snapshot_regions(snapshot);
661 if (_sg_mc_snapshot_fds)
662 MC_restore_snapshot_fds(snapshot);
663 MC_snapshot_ignore_restore(snapshot);
664 mc_model_checker->process().cache_flags = 0;
667 mc_snapshot_t simcall_HANDLER_mc_snapshot(smx_simcall_t simcall)
669 return MC_take_snapshot(1);