1 /* Copyright (c) 2008-2017. 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. */
18 #include "src/internal_config.h"
19 #include "src/mc/mc_private.hpp"
20 #include "src/smpi/include/private.hpp"
21 #include "xbt/mmalloc.h"
22 #include "xbt/module.h"
24 #include "src/xbt/mmalloc/mmprivate.h"
26 #include "src/simix/smx_private.hpp"
28 #include <libunwind.h>
31 #include "src/mc/mc_private.hpp"
34 #include "src/mc/mc_hash.hpp"
35 #include "src/mc/mc_mmu.hpp"
36 #include "src/mc/mc_smx.hpp"
37 #include "src/mc/mc_snapshot.hpp"
38 #include "src/mc/mc_unw.hpp"
39 #include "src/mc/remote/mc_protocol.h"
41 #include "src/mc/RegionSnapshot.hpp"
42 #include "src/mc/ObjectInformation.hpp"
43 #include "src/mc/Frame.hpp"
44 #include "src/mc/Variable.hpp"
46 using simgrid::mc::remote;
48 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(mc_checkpoint, mc, "Logging specific to mc_checkpoint");
50 #define PROT_RWX (PROT_READ | PROT_WRITE | PROT_EXEC)
51 #define PROT_RW (PROT_READ | PROT_WRITE)
52 #define PROT_RX (PROT_READ | PROT_EXEC)
57 /************************************ Free functions **************************************/
58 /*****************************************************************************************/
60 /** @brief Restore a region from a snapshot
62 * @param region Target region
64 static void restore(mc_mem_region_t region)
66 switch(region->storage_type()) {
67 case simgrid::mc::StorageType::Flat:
68 mc_model_checker->process().write_bytes(region->flat_data().get(),
69 region->size(), region->permanent_address());
72 case simgrid::mc::StorageType::Chunked:
73 mc_region_restore_sparse(&mc_model_checker->process(), region);
76 case simgrid::mc::StorageType::Privatized:
77 for (auto& p : region->privatized_data())
81 default: // includes StorageType::NoData
82 xbt_die("Storage type not supported");
88 RegionSnapshot privatized_region(
89 RegionType region_type, void *start_addr, void* permanent_addr,
93 size_t process_count = MC_smpi_process_count();
95 // Read smpi_privatization_regions from MCed:
96 smpi_privatization_region_t remote_smpi_privatization_regions;
97 mc_model_checker->process().read_variable(
98 "smpi_privatization_regions",
99 &remote_smpi_privatization_regions, sizeof(remote_smpi_privatization_regions));
100 s_smpi_privatization_region_t privatization_regions[process_count];
101 mc_model_checker->process().read_bytes(
102 &privatization_regions, sizeof(privatization_regions),
103 remote(remote_smpi_privatization_regions));
105 std::vector<simgrid::mc::RegionSnapshot> data;
106 data.reserve(process_count);
107 for (size_t i = 0; i < process_count; i++)
108 data.push_back(simgrid::mc::region(region_type, start_addr,
109 privatization_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));
119 void add_region(int index, simgrid::mc::Snapshot* snapshot,
120 simgrid::mc::RegionType type,
121 simgrid::mc::ObjectInformation* object_info,
122 void *start_addr, void* permanent_addr,
125 if (type == simgrid::mc::RegionType::Data)
126 xbt_assert(object_info, "Missing object info for object.");
127 else if (type == simgrid::mc::RegionType::Heap)
128 xbt_assert(not object_info, "Unexpected object info for heap region.");
130 simgrid::mc::RegionSnapshot region;
132 const bool privatization_aware = object_info
133 && mc_model_checker->process().privatized(*object_info);
134 if (privatization_aware && MC_smpi_process_count())
135 region = simgrid::mc::privatized_region(
136 type, start_addr, permanent_addr, size);
139 region = simgrid::mc::region(type, start_addr, permanent_addr, size);
141 region.object_info(object_info);
142 snapshot->snapshot_regions[index]
143 = std::unique_ptr<simgrid::mc::RegionSnapshot>(
144 new simgrid::mc::RegionSnapshot(std::move(region)));
148 static void get_memory_regions(simgrid::mc::RemoteClient* process, simgrid::mc::Snapshot* snapshot)
150 const size_t n = process->object_infos.size();
151 snapshot->snapshot_regions.resize(n + 1);
153 for (auto const& object_info : process->object_infos)
154 add_region(i++, snapshot, simgrid::mc::RegionType::Data,
156 object_info->start_rw, object_info->start_rw,
157 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 add_region(n, snapshot, simgrid::mc::RegionType::Heap, nullptr,
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 (mc_model_checker->process().privatized() && 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));
178 snapshot->privatization_index = simgrid::mc::ProcessIndexMissing;
181 /** \brief Fills the position of the segments (executable, read-only, read/write).
183 // TODO, use the ELF segment information for more robustness
184 void find_object_address(
185 std::vector<simgrid::xbt::VmMap> const& maps,
186 simgrid::mc::ObjectInformation* result)
188 char* name = xbt_basename(result->file_name.c_str());
190 for (size_t i = 0; i < maps.size(); ++i) {
191 simgrid::xbt::VmMap const& reg = maps[i];
192 if (maps[i].pathname.empty())
194 char* map_basename = xbt_basename(maps[i].pathname.c_str());
195 if (strcmp(name, map_basename) != 0) {
201 // This is the non-GNU_RELRO-part of the data segment:
202 if (reg.prot == PROT_RW) {
203 xbt_assert(not result->start_rw, "Multiple read-write segments for %s, not supported", maps[i].pathname.c_str());
204 result->start_rw = (char*) reg.start_addr;
205 result->end_rw = (char*) reg.end_addr;
207 // The next VMA might be end of the data segment:
208 if (i + 1 < maps.size()
209 && maps[i + 1].pathname.empty()
210 && maps[i + 1].prot == PROT_RW
211 && maps[i + 1].start_addr == reg.end_addr)
212 result->end_rw = (char*) maps[i + 1].end_addr;
215 // This is the text segment:
216 else if (reg.prot == PROT_RX) {
217 xbt_assert(not result->start_exec, "Multiple executable segments for %s, not supported",
218 maps[i].pathname.c_str());
219 result->start_exec = (char*) reg.start_addr;
220 result->end_exec = (char*) reg.end_addr;
222 // The next VMA might be end of the data segment:
223 if (i + 1 < maps.size()
224 && maps[i + 1].pathname.empty()
225 && maps[i + 1].prot == PROT_RW
226 && maps[i + 1].start_addr == reg.end_addr) {
227 result->start_rw = (char*) maps[i + 1].start_addr;
228 result->end_rw = (char*) maps[i + 1].end_addr;
232 // This is the GNU_RELRO-part of the data segment:
233 else if (reg.prot == PROT_READ) {
234 xbt_assert(not result->start_ro, "Multiple read only segments for %s, not supported", maps[i].pathname.c_str());
235 result->start_ro = (char*) reg.start_addr;
236 result->end_ro = (char*) reg.end_addr;
240 result->start = result->start_rw;
241 if ((const void*) result->start_ro < result->start)
242 result->start = result->start_ro;
243 if ((const void*) result->start_exec < result->start)
244 result->start = result->start_exec;
246 result->end = result->end_rw;
247 if (result->end_ro && (const void*) result->end_ro > result->end)
248 result->end = result->end_ro;
249 if (result->end_exec && (const void*) result->end_exec > result->end)
250 result->end = result->end_exec;
252 xbt_assert(result->start_exec || result->start_rw || result->start_ro);
257 /************************************* Take Snapshot ************************************/
258 /****************************************************************************************/
260 /** \brief Checks whether the variable is in scope for a given IP.
262 * A variable may be defined only from a given value of IP.
264 * \param var Variable description
265 * \param scope Scope description
266 * \param ip Instruction pointer
267 * \return true if the variable is valid
269 static bool valid_variable(simgrid::mc::Variable* var,
270 simgrid::mc::Frame* scope,
273 // The variable is not yet valid:
274 if (scope->range.begin() + var->start_scope > (std::uint64_t) ip)
280 static void fill_local_variables_values(mc_stack_frame_t stack_frame, simgrid::mc::Frame* scope, int process_index,
281 std::vector<s_local_variable_t>& result)
283 simgrid::mc::RemoteClient* process = &mc_model_checker->process();
285 if (not scope || not scope->range.contain(stack_frame->ip))
288 for (simgrid::mc::Variable& current_variable : scope->variables) {
290 if (not valid_variable(¤t_variable, scope, (void*)stack_frame->ip))
294 // FIXME, get rid of `region_type`
295 if ((long) stack_frame->ip > (long) process->libsimgrid_info->start_exec)
300 s_local_variable_t new_var;
301 new_var.subprogram = stack_frame->frame;
302 new_var.ip = stack_frame->ip;
303 new_var.name = current_variable.name;
304 new_var.type = current_variable.type;
305 new_var.region = region_type;
306 new_var.address = nullptr;
308 if (current_variable.address != nullptr)
309 new_var.address = current_variable.address;
310 else if (not current_variable.location_list.empty()) {
311 simgrid::dwarf::Location location =
312 simgrid::dwarf::resolve(
313 current_variable.location_list,
314 current_variable.object_info,
315 &(stack_frame->unw_cursor),
316 (void *) stack_frame->frame_base,
317 &mc_model_checker->process(), process_index);
319 if (not location.in_memory())
320 xbt_die("Cannot handle non-address variable");
321 new_var.address = location.address();
324 xbt_die("No address");
326 result.push_back(std::move(new_var));
329 // Recursive processing of nested scopes:
330 for (simgrid::mc::Frame& nested_scope : scope->scopes)
331 fill_local_variables_values(
332 stack_frame, &nested_scope, process_index, result);
335 static std::vector<s_local_variable_t> get_local_variables_values(std::vector<s_mc_stack_frame_t>& stack_frames,
338 std::vector<s_local_variable_t> variables;
339 for (s_mc_stack_frame_t& stack_frame : stack_frames)
340 fill_local_variables_values(&stack_frame, stack_frame.frame, process_index, variables);
344 static std::vector<s_mc_stack_frame_t> unwind_stack_frames(simgrid::mc::UnwindContext* stack_context)
346 simgrid::mc::RemoteClient* process = &mc_model_checker->process();
347 std::vector<s_mc_stack_frame_t> result;
349 unw_cursor_t c = stack_context->cursor();
351 // TODO, check condition check (unw_init_local==0 means end of frame)
355 s_mc_stack_frame_t stack_frame;
357 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) frame->frame_base(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.");
393 xbt_die("Error while unwinding stack");
396 if (result.empty()) {
397 XBT_INFO("unw_init_local failed");
404 static std::vector<s_mc_snapshot_stack_t> take_snapshot_stacks(simgrid::mc::Snapshot* snapshot)
406 std::vector<s_mc_snapshot_stack_t> res;
408 for (auto const& stack : mc_model_checker->process().stack_areas()) {
409 s_mc_snapshot_stack_t st;
411 // Read the context from remote process:
412 unw_context_t context;
413 mc_model_checker->process().read_bytes(
414 &context, sizeof(context), remote(stack.context));
416 st.context.initialize(&mc_model_checker->process(), &context);
418 st.stack_frames = unwind_stack_frames(&st.context);
419 st.local_variables = get_local_variables_values(st.stack_frames, stack.process_index);
420 st.process_index = stack.process_index;
422 unw_word_t sp = st.stack_frames[0].sp;
424 res.push_back(std::move(st));
427 (char*) stack.address + stack.size - (char*) sp;
428 snapshot->stack_sizes.push_back(stack_size);
435 static void snapshot_handle_ignore(simgrid::mc::Snapshot* snapshot)
437 xbt_assert(snapshot->process());
440 for (auto const& region : mc_model_checker->process().ignored_regions()) {
441 s_mc_snapshot_ignored_data_t ignored_data;
442 ignored_data.start = (void*)region.addr;
443 ignored_data.data.resize(region.size);
444 // TODO, we should do this once per privatization segment:
445 snapshot->process()->read_bytes(
446 ignored_data.data.data(), region.size, remote(region.addr),
447 simgrid::mc::ProcessIndexDisabled);
448 snapshot->ignored_data.push_back(std::move(ignored_data));
452 for (auto const& region : mc_model_checker->process().ignored_regions())
453 snapshot->process()->clear_bytes(remote(region.addr), region.size);
457 static void snapshot_ignore_restore(simgrid::mc::Snapshot* snapshot)
459 for (auto const& ignored_data : snapshot->ignored_data)
460 snapshot->process()->write_bytes(
461 ignored_data.data.data(), ignored_data.data.size(),
462 remote(ignored_data.start));
465 static std::vector<s_fd_infos_t> get_current_fds(pid_t pid)
467 const size_t fd_dir_path_size = 20;
468 char fd_dir_path[fd_dir_path_size];
469 int res = snprintf(fd_dir_path, fd_dir_path_size,
470 "/proc/%lli/fd", (long long int) pid);
471 xbt_assert(res >= 0);
472 if ((size_t) res > fd_dir_path_size)
473 xbt_die("Unexpected buffer is too small for fd_dir_path");
475 DIR* fd_dir = opendir(fd_dir_path);
476 if (fd_dir == nullptr)
477 xbt_die("Cannot open directory '/proc/self/fd'\n");
479 std::vector<s_fd_infos_t> fds;
481 struct dirent* fd_number;
482 while ((fd_number = readdir(fd_dir))) {
484 int fd_value = xbt_str_parse_int(fd_number->d_name, "Found a non-numerical FD: %s. Freaking out!");
489 const size_t source_size = 25;
491 int res = snprintf(source, source_size, "/proc/%lli/fd/%s",
492 (long long int) pid, fd_number->d_name);
493 xbt_assert(res >= 0);
494 if ((size_t) res > source_size)
495 xbt_die("Unexpected buffer is too small for fd %s", fd_number->d_name);
497 const size_t link_size = 200;
499 res = readlink(source, link, link_size);
502 xbt_die("Could not read link for %s", source);
504 xbt_die("Buffer to small for link of %s", source);
509 if(smpi_is_privatization_file(link))
513 // This is (probably) the DIR* we are reading:
514 // TODO, read all the file entries at once and close the DIR.*
515 if(strcmp(fd_dir_path, link) == 0)
518 // We don't handle them.
519 // It does not mean we should silently ignore them however.
520 if (strncmp(link, "pipe:", std::strlen("pipe:")) == 0 || strncmp(link, "socket:", std::strlen("socket:")) == 0)
523 // If dot_output enabled, do not handle the corresponding file
524 if (dot_output != nullptr) {
525 char* link_basename = xbt_basename(link);
526 if (strcmp(link_basename, _sg_mc_dot_output_file) == 0) {
533 // This is probably a shared memory used by lttng-ust:
534 if(strncmp("/dev/shm/ust-shm-tmp-", link, std::strlen("/dev/shm/ust-shm-tmp-"))==0)
537 // Add an entry for this FD in the snapshot:
539 fd.filename = std::string(link);
540 fd.number = fd_value;
541 fd.flags = fcntl(fd_value, F_GETFL) | fcntl(fd_value, F_GETFD) ;
542 fd.current_position = lseek(fd_value, 0, SEEK_CUR);
543 fds.push_back(std::move(fd));
550 std::shared_ptr<simgrid::mc::Snapshot> take_snapshot(int num_state)
552 XBT_DEBUG("Taking snapshot %i", num_state);
554 simgrid::mc::RemoteClient* mc_process = &mc_model_checker->process();
556 std::shared_ptr<simgrid::mc::Snapshot> snapshot = std::make_shared<simgrid::mc::Snapshot>(mc_process, num_state);
558 for (auto const& p : mc_model_checker->process().actors())
559 snapshot->enabled_processes.insert(p.copy.getBuffer()->pid);
561 snapshot_handle_ignore(snapshot.get());
563 if (_sg_mc_snapshot_fds)
564 snapshot->current_fds = get_current_fds(mc_model_checker->process().pid());
566 /* Save the std heap and the writable mapped pages of libsimgrid and binary */
567 get_memory_regions(mc_process, snapshot.get());
569 snapshot->to_ignore = mc_model_checker->process().ignored_heap();
571 if (_sg_mc_max_visited_states > 0 || strcmp(_sg_mc_property_file, "")) {
572 snapshot->stacks = take_snapshot_stacks(snapshot.get());
574 snapshot->hash = simgrid::mc::hash(*snapshot);
580 snapshot_ignore_restore(snapshot.get());
585 void restore_snapshot_regions(simgrid::mc::Snapshot* snapshot)
587 for (std::unique_ptr<s_mc_mem_region_t> const& region : snapshot->snapshot_regions) {
588 // For privatized, variables we decided it was not necessary to take the snapshot:
590 restore(region.get());
594 if(snapshot->privatization_index >= 0) {
595 // Fix the privatization mmap:
596 s_mc_message_restore message{MC_MESSAGE_RESTORE, snapshot->privatization_index};
597 mc_model_checker->process().getChannel().send(message);
603 void restore_snapshot_fds(simgrid::mc::Snapshot* snapshot)
605 xbt_die("FD snapshot not implemented in client/server mode.");
607 for (auto const& fd : snapshot->current_fds) {
609 int new_fd = open(fd.filename.c_str(), fd.flags);
611 xbt_die("Could not reopen the file %s fo restoring the file descriptor", fd.filename.c_str());
612 if (new_fd != fd.number) {
613 dup2(new_fd, fd.number);
616 lseek(fd.number, fd.current_position, SEEK_SET);
620 void restore_snapshot(std::shared_ptr<simgrid::mc::Snapshot> snapshot)
622 XBT_DEBUG("Restore snapshot %i", snapshot->num_state);
623 restore_snapshot_regions(snapshot.get());
624 if (_sg_mc_snapshot_fds)
625 restore_snapshot_fds(snapshot.get());
626 snapshot_ignore_restore(snapshot.get());
627 mc_model_checker->process().clear_cache();