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. */
18 #include "src/internal_config.h"
19 #include "src/mc/mc_private.h"
20 #include "xbt/module.h"
21 #include <xbt/mmalloc.h>
22 #include <xbt/memory.hpp>
23 #include "src/smpi/private.h"
25 #include "src/xbt/mmalloc/mmprivate.h"
27 #include "src/simix/smx_private.h"
29 #include <libunwind.h>
32 #include "src/mc/mc_private.h"
35 #include "src/mc/mc_snapshot.h"
36 #include "src/mc/mc_mmu.h"
37 #include "src/mc/mc_unw.h"
38 #include "src/mc/mc_protocol.h"
39 #include "src/mc/mc_smx.h"
40 #include "mc_hash.hpp"
42 #include "src/mc/RegionSnapshot.hpp"
43 #include "src/mc/ObjectInformation.hpp"
44 #include "src/mc/Frame.hpp"
45 #include "src/mc/Variable.hpp"
47 using simgrid::mc::remote;
49 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(mc_checkpoint, mc,
50 "Logging specific to mc_checkpoint");
55 /************************************ Free functions **************************************/
56 /*****************************************************************************************/
58 /** @brief Restore a region from a snapshot
60 * @param reg Target region
62 static void restore(mc_mem_region_t region)
64 switch(region->storage_type()) {
65 case simgrid::mc::StorageType::NoData:
67 xbt_die("Storage type not supported");
70 case simgrid::mc::StorageType::Flat:
71 mc_model_checker->process().write_bytes(region->flat_data().get(),
72 region->size(), region->permanent_address());
75 case simgrid::mc::StorageType::Chunked:
76 mc_region_restore_sparse(&mc_model_checker->process(), region);
79 case simgrid::mc::StorageType::Privatized:
80 for (auto& p : region->privatized_data())
87 RegionSnapshot privatized_region(
88 RegionType region_type, void *start_addr, void* permanent_addr,
92 size_t process_count = MC_smpi_process_count();
94 // Read smpi_privatisation_regions from MCed:
95 smpi_privatisation_region_t remote_smpi_privatisation_regions;
96 mc_model_checker->process().read_variable(
97 "smpi_privatisation_regions",
98 &remote_smpi_privatisation_regions, sizeof(remote_smpi_privatisation_regions));
99 s_smpi_privatisation_region_t privatisation_regions[process_count];
100 mc_model_checker->process().read_bytes(
101 &privatisation_regions, sizeof(privatisation_regions),
102 remote(remote_smpi_privatisation_regions));
104 std::vector<simgrid::mc::RegionSnapshot> data;
105 data.reserve(process_count);
106 for (size_t i = 0; i < process_count; i++)
107 data.push_back(simgrid::mc::region(region_type, start_addr,
108 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));
118 void add_region(int index, simgrid::mc::Snapshot* snapshot,
119 simgrid::mc::RegionType type,
120 simgrid::mc::ObjectInformation* object_info,
121 void *start_addr, void* permanent_addr,
124 if (type == simgrid::mc::RegionType::Data)
125 xbt_assert(object_info, "Missing object info for object.");
126 else if (type == simgrid::mc::RegionType::Heap)
127 xbt_assert(!object_info, "Unexpected object info for heap region.");
129 simgrid::mc::RegionSnapshot region;
131 const bool privatization_aware = object_info
132 && mc_model_checker->process().privatized(*object_info);
133 if (privatization_aware && MC_smpi_process_count())
134 region = simgrid::mc::privatized_region(
135 type, start_addr, permanent_addr, size);
138 region = simgrid::mc::region(type, start_addr, permanent_addr, size);
140 region.object_info(object_info);
141 snapshot->snapshot_regions[index]
142 = std::unique_ptr<simgrid::mc::RegionSnapshot>(
143 new simgrid::mc::RegionSnapshot(std::move(region)));
147 static void get_memory_regions(simgrid::mc::Process* process, simgrid::mc::Snapshot* snapshot)
149 const size_t n = process->object_infos.size();
150 snapshot->snapshot_regions.resize(n + 1);
152 for (auto const& object_info : process->object_infos)
153 add_region(i++, snapshot, simgrid::mc::RegionType::Data,
155 object_info->start_rw, object_info->start_rw,
156 object_info->end_rw - object_info->start_rw);
158 xbt_mheap_t heap = process->get_heap();
159 void *start_heap = heap->base;
160 void *end_heap = heap->breakval;
162 add_region(n, snapshot, simgrid::mc::RegionType::Heap, nullptr,
163 start_heap, start_heap,
164 (char *) end_heap - (char *) start_heap);
165 snapshot->heap_bytes_used = mmalloc_get_bytes_used_remote(
167 process->get_malloc_info());
170 if (mc_model_checker->process().privatized() && MC_smpi_process_count())
171 // snapshot->privatization_index = smpi_loaded_page
172 mc_model_checker->process().read_variable(
173 "smpi_loaded_page", &snapshot->privatization_index,
174 sizeof(snapshot->privatization_index));
177 snapshot->privatization_index = simgrid::mc::ProcessIndexMissing;
180 #define PROT_RWX (PROT_READ | PROT_WRITE | PROT_EXEC)
181 #define PROT_RW (PROT_READ | PROT_WRITE)
182 #define PROT_RX (PROT_READ | PROT_EXEC)
184 /** \brief Fills the position of the segments (executable, read-only, read/write).
186 // TODO, use the ELF segment information for more robustness
187 void find_object_address(
188 std::vector<simgrid::xbt::VmMap> const& maps,
189 simgrid::mc::ObjectInformation* result)
191 char* name = xbt_basename(result->file_name.c_str());
193 for (size_t i = 0; i < maps.size(); ++i) {
194 simgrid::xbt::VmMap const& reg = maps[i];
195 if (maps[i].pathname.empty())
197 char* map_basename = xbt_basename(maps[i].pathname.c_str());
198 if (strcmp(name, map_basename) != 0) {
204 // This is the non-GNU_RELRO-part of the data segment:
205 if (reg.prot == PROT_RW) {
206 xbt_assert(!result->start_rw,
207 "Multiple read-write segments for %s, not supported",
208 maps[i].pathname.c_str());
209 result->start_rw = (char*) reg.start_addr;
210 result->end_rw = (char*) reg.end_addr;
212 // The next VMA might be end of the data segment:
213 if (i + 1 < maps.size()
214 && maps[i + 1].pathname.empty()
215 && maps[i + 1].prot == PROT_RW
216 && maps[i + 1].start_addr == reg.end_addr)
217 result->end_rw = (char*) maps[i + 1].end_addr;
220 // This is the text segment:
221 else if (reg.prot == PROT_RX) {
222 xbt_assert(!result->start_exec,
223 "Multiple executable segments for %s, not supported",
224 maps[i].pathname.c_str());
225 result->start_exec = (char*) reg.start_addr;
226 result->end_exec = (char*) reg.end_addr;
228 // The next VMA might be end of the data segment:
229 if (i + 1 < maps.size()
230 && maps[i + 1].pathname.empty()
231 && maps[i + 1].prot == PROT_RW
232 && maps[i + 1].start_addr == reg.end_addr) {
233 result->start_rw = (char*) maps[i + 1].start_addr;
234 result->end_rw = (char*) maps[i + 1].end_addr;
238 // This is the GNU_RELRO-part of the data segment:
239 else if (reg.prot == PROT_READ) {
240 xbt_assert(!result->start_ro,
241 "Multiple read only segments for %s, not supported",
242 maps[i].pathname.c_str());
243 result->start_ro = (char*) reg.start_addr;
244 result->end_ro = (char*) reg.end_addr;
248 result->start = result->start_rw;
249 if ((const void*) result->start_ro < result->start)
250 result->start = result->start_ro;
251 if ((const void*) result->start_exec < result->start)
252 result->start = result->start_exec;
254 result->end = result->end_rw;
255 if (result->end_ro && (const void*) result->end_ro > result->end)
256 result->end = result->end_ro;
257 if (result->end_exec && (const void*) result->end_exec > result->end)
258 result->end = result->end_exec;
260 xbt_assert(result->start_exec || result->start_rw || result->start_ro);
265 /************************************* Take Snapshot ************************************/
266 /****************************************************************************************/
268 /** \brief Checks whether the variable is in scope for a given IP.
270 * A variable may be defined only from a given value of IP.
272 * \param var Variable description
273 * \param frame Scope description
274 * \param ip Instruction pointer
275 * \return true if the variable is valid
277 static bool valid_variable(simgrid::mc::Variable* var,
278 simgrid::mc::Frame* scope,
281 // The variable is not yet valid:
282 if (scope->range.begin() + var->start_scope > (std::uint64_t) ip)
288 static void fill_local_variables_values(mc_stack_frame_t stack_frame,
289 simgrid::mc::Frame* scope,
291 std::vector<s_local_variable>& result)
293 simgrid::mc::Process* process = &mc_model_checker->process();
295 if (!scope->range.contain(stack_frame->ip))
298 for(simgrid::mc::Variable& current_variable :
301 if (!valid_variable(¤t_variable, scope, (void *) stack_frame->ip))
305 // FIXME, get rid of `region_type`
306 if ((long) stack_frame->ip > (long) process->libsimgrid_info->start_exec)
311 s_local_variable_t new_var;
312 new_var.subprogram = stack_frame->frame;
313 new_var.ip = stack_frame->ip;
314 new_var.name = current_variable.name;
315 new_var.type = current_variable.type;
316 new_var.region = region_type;
317 new_var.address = nullptr;
319 if (current_variable.address != nullptr)
320 new_var.address = current_variable.address;
321 else if (!current_variable.location_list.empty()) {
322 simgrid::dwarf::Location location =
323 simgrid::dwarf::resolve(
324 current_variable.location_list,
325 current_variable.object_info,
326 &(stack_frame->unw_cursor),
327 (void *) stack_frame->frame_base,
328 &mc_model_checker->process(), process_index);
330 if (!location.in_memory())
331 xbt_die("Cannot handle non-address variable");
332 new_var.address = location.address();
335 xbt_die("No address");
337 result.push_back(std::move(new_var));
340 // Recursive processing of nested scopes:
341 for(simgrid::mc::Frame& nested_scope : scope->scopes)
342 fill_local_variables_values(
343 stack_frame, &nested_scope, process_index, result);
346 static std::vector<s_local_variable> get_local_variables_values(
347 std::vector<s_mc_stack_frame_t>& stack_frames, int process_index)
349 std::vector<s_local_variable> variables;
350 for (s_mc_stack_frame_t& stack_frame : stack_frames)
351 fill_local_variables_values(&stack_frame, stack_frame.frame, process_index, variables);
355 static std::vector<s_mc_stack_frame_t> unwind_stack_frames(simgrid::mc::UnwindContext* stack_context)
357 simgrid::mc::Process* process = &mc_model_checker->process();
358 std::vector<s_mc_stack_frame_t> result;
360 unw_cursor_t c = stack_context->cursor();
362 // TODO, check condition check (unw_init_local==0 means end of frame)
366 s_mc_stack_frame_t stack_frame;
368 stack_frame.unw_cursor = c;
372 unw_get_reg(&c, UNW_REG_IP, &ip);
373 unw_get_reg(&c, UNW_REG_SP, &sp);
378 // TODO, use real addresses in frame_t instead of fixing it here
380 simgrid::mc::Frame* frame = process->find_function(remote(ip));
381 stack_frame.frame = frame;
384 stack_frame.frame_name = frame->name;
385 stack_frame.frame_base =
386 (unw_word_t) frame->frame_base(c);
388 stack_frame.frame_base = 0;
389 stack_frame.frame_name = std::string();
392 result.push_back(std::move(stack_frame));
394 /* Stop before context switch with maestro */
395 if (frame != nullptr &&
396 frame->name == "smx_ctx_sysv_wrapper")
399 int ret = unw_step(&c);
401 xbt_die("Unexpected end of stack.");
403 xbt_die("Error while unwinding stack");
406 if (result.empty()) {
407 XBT_INFO("unw_init_local failed");
414 static std::vector<s_mc_snapshot_stack_t> take_snapshot_stacks(simgrid::mc::Snapshot* snapshot)
416 std::vector<s_mc_snapshot_stack_t> res;
418 for (auto const& stack : mc_model_checker->process().stack_areas()) {
419 s_mc_snapshot_stack_t st;
421 // Read the context from remote process:
422 unw_context_t context;
423 mc_model_checker->process().read_bytes(
424 &context, sizeof(context), remote(stack.context));
426 st.context.initialize(&mc_model_checker->process(), &context);
428 st.stack_frames = unwind_stack_frames(&st.context);
429 st.local_variables = get_local_variables_values(st.stack_frames, stack.process_index);
430 st.process_index = stack.process_index;
432 unw_word_t sp = st.stack_frames[0].sp;
434 res.push_back(std::move(st));
437 (char*) stack.address + stack.size - (char*) sp;
438 snapshot->stack_sizes.push_back(stack_size);
445 static void snapshot_handle_ignore(simgrid::mc::Snapshot* snapshot)
447 xbt_assert(snapshot->process());
450 for (auto const& region : mc_model_checker->process().ignored_regions()) {
451 s_mc_snapshot_ignored_data_t ignored_data;
452 ignored_data.start = (void*)region.addr;
453 ignored_data.data.resize(region.size);
454 // TODO, we should do this once per privatization segment:
455 snapshot->process()->read_bytes(
456 ignored_data.data.data(), region.size, remote(region.addr),
457 simgrid::mc::ProcessIndexDisabled);
458 snapshot->ignored_data.push_back(std::move(ignored_data));
462 for(auto const& region : mc_model_checker->process().ignored_regions())
463 snapshot->process()->clear_bytes(remote(region.addr), region.size);
467 static void snapshot_ignore_restore(simgrid::mc::Snapshot* snapshot)
469 for (auto const& ignored_data : snapshot->ignored_data)
470 snapshot->process()->write_bytes(
471 ignored_data.data.data(), ignored_data.data.size(),
472 remote(ignored_data.start));
475 static std::vector<s_fd_infos_t> get_current_fds(pid_t pid)
477 const size_t fd_dir_path_size = 20;
478 char fd_dir_path[fd_dir_path_size];
479 int res = snprintf(fd_dir_path, fd_dir_path_size,
480 "/proc/%lli/fd", (long long int) pid);
481 xbt_assert(res >= 0);
482 if ((size_t) res > fd_dir_path_size)
483 xbt_die("Unexpected buffer is too small for fd_dir_path");
485 DIR* fd_dir = opendir(fd_dir_path);
486 if (fd_dir == nullptr)
487 xbt_die("Cannot open directory '/proc/self/fd'\n");
489 std::vector<s_fd_infos_t> fds;
491 struct dirent* fd_number;
492 while ((fd_number = readdir(fd_dir))) {
494 int fd_value = xbt_str_parse_int(fd_number->d_name, "Found a non-numerical FD: %s. Freaking out!");
499 const size_t source_size = 25;
501 int res = snprintf(source, source_size, "/proc/%lli/fd/%s",
502 (long long int) pid, fd_number->d_name);
503 xbt_assert(res >= 0);
504 if ((size_t) res > source_size)
505 xbt_die("Unexpected buffer is too small for fd %s", fd_number->d_name);
507 const size_t link_size = 200;
509 res = readlink(source, link, link_size);
512 xbt_die("Could not read link for %s", source);
514 xbt_die("Buffer to small for link of %s", source);
519 if(smpi_is_privatisation_file(link))
523 // This is (probably) the DIR* we are reading:
524 // TODO, read all the file entries at once and close the DIR.*
525 if(strcmp(fd_dir_path, link) == 0)
528 // We don't handle them.
529 // It does not mean we should silently ignore them however.
530 if (strncmp(link, "pipe:", 5) == 0 || strncmp(link, "socket:", 7) == 0)
533 // If dot_output enabled, do not handle the corresponding file
534 if (dot_output != nullptr) {
535 char* link_basename = xbt_basename(link);
536 if (strcmp(link_basename, _sg_mc_dot_output_file) == 0) {
543 // This is probably a shared memory used by lttng-ust:
544 if(strncmp("/dev/shm/ust-shm-tmp-", link, 21)==0)
547 // Add an entry for this FD in the snapshot:
549 fd.filename = std::string(link);
550 fd.number = fd_value;
551 fd.flags = fcntl(fd_value, F_GETFL) | fcntl(fd_value, F_GETFD) ;
552 fd.current_position = lseek(fd_value, 0, SEEK_CUR);
553 fds.push_back(std::move(fd));
560 std::shared_ptr<simgrid::mc::Snapshot> take_snapshot(int num_state)
562 XBT_DEBUG("Taking snapshot %i", num_state);
564 simgrid::mc::Process* mc_process = &mc_model_checker->process();
566 std::shared_ptr<simgrid::mc::Snapshot> snapshot =
567 std::make_shared<simgrid::mc::Snapshot>(mc_process);
569 snapshot->num_state = num_state;
571 for (auto& p : mc_model_checker->process().simix_processes())
572 snapshot->enabled_processes.insert(p.copy.pid);
574 snapshot_handle_ignore(snapshot.get());
576 if (_sg_mc_snapshot_fds)
577 snapshot->current_fds = get_current_fds(mc_model_checker->process().pid());
579 /* Save the std heap and the writable mapped pages of libsimgrid and binary */
580 get_memory_regions(mc_process, snapshot.get());
582 snapshot->to_ignore = mc_model_checker->process().ignored_heap();
584 if (_sg_mc_visited > 0 || strcmp(_sg_mc_property_file, "")) {
585 snapshot->stacks = take_snapshot_stacks(snapshot.get());
587 snapshot->hash = simgrid::mc::hash(*snapshot);
593 snapshot_ignore_restore(snapshot.get());
598 void restore_snapshot_regions(simgrid::mc::Snapshot* snapshot)
600 for(std::unique_ptr<s_mc_mem_region_t> const& region : snapshot->snapshot_regions) {
601 // For privatized, variables we decided it was not necessary to take the snapshot:
603 restore(region.get());
607 // TODO, send a message to implement this in the MCed process
608 if(snapshot->privatization_index >= 0) {
609 // Fix the privatization mmap:
610 s_mc_restore_message message;
611 message.type = MC_MESSAGE_RESTORE;
612 message.index = snapshot->privatization_index;
613 mc_model_checker->process().getChannel().send(message);
619 void restore_snapshot_fds(simgrid::mc::Snapshot* snapshot)
621 if (mc_mode == MC_MODE_SERVER)
622 xbt_die("FD snapshot not implemented in client/server mode.");
624 for (auto const& fd : snapshot->current_fds) {
626 int new_fd = open(fd.filename.c_str(), fd.flags);
628 xbt_die("Could not reopen the file %s fo restoring the file descriptor",
629 fd.filename.c_str());
630 if (new_fd != fd.number) {
631 dup2(new_fd, fd.number);
634 lseek(fd.number, fd.current_position, SEEK_SET);
638 void restore_snapshot(std::shared_ptr<simgrid::mc::Snapshot> snapshot)
640 XBT_DEBUG("Restore snapshot %i", snapshot->num_state);
641 restore_snapshot_regions(snapshot.get());
642 if (_sg_mc_snapshot_fds)
643 restore_snapshot_fds(snapshot.get());
644 snapshot_ignore_restore(snapshot.get());
645 mc_model_checker->process().clear_cache();