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 "src/smpi/private.h"
21 #include "xbt/mmalloc.h"
22 #include "xbt/module.h"
24 #include "src/xbt/mmalloc/mmprivate.h"
26 #include "src/simix/smx_private.h"
28 #include <libunwind.h>
31 #include "src/mc/mc_private.h"
34 #include "src/mc/mc_hash.hpp"
35 #include "src/mc/mc_mmu.h"
36 #include "src/mc/mc_smx.h"
37 #include "src/mc/mc_snapshot.h"
38 #include "src/mc/mc_unw.h"
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,
49 "Logging specific to mc_checkpoint");
54 /************************************ Free functions **************************************/
55 /*****************************************************************************************/
57 /** @brief Restore a region from a snapshot
59 * @param region Target region
61 static void restore(mc_mem_region_t region)
63 switch(region->storage_type()) {
64 case simgrid::mc::StorageType::NoData:
66 xbt_die("Storage type not supported");
69 case simgrid::mc::StorageType::Flat:
70 mc_model_checker->process().write_bytes(region->flat_data().get(),
71 region->size(), region->permanent_address());
74 case simgrid::mc::StorageType::Chunked:
75 mc_region_restore_sparse(&mc_model_checker->process(), region);
78 case simgrid::mc::StorageType::Privatized:
79 for (auto& p : region->privatized_data())
86 RegionSnapshot privatized_region(
87 RegionType region_type, void *start_addr, void* permanent_addr,
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++)
106 data.push_back(simgrid::mc::region(region_type, start_addr,
107 privatisation_regions[i].address, size));
109 simgrid::mc::RegionSnapshot region = simgrid::mc::RegionSnapshot(
110 region_type, start_addr, permanent_addr, size);
111 region.privatized_data(std::move(data));
117 void add_region(int index, simgrid::mc::Snapshot* snapshot,
118 simgrid::mc::RegionType type,
119 simgrid::mc::ObjectInformation* object_info,
120 void *start_addr, void* permanent_addr,
123 if (type == simgrid::mc::RegionType::Data)
124 xbt_assert(object_info, "Missing object info for object.");
125 else if (type == simgrid::mc::RegionType::Heap)
126 xbt_assert(!object_info, "Unexpected object info for heap region.");
128 simgrid::mc::RegionSnapshot region;
130 const bool privatization_aware = object_info
131 && mc_model_checker->process().privatized(*object_info);
132 if (privatization_aware && MC_smpi_process_count())
133 region = simgrid::mc::privatized_region(
134 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 get_memory_regions(simgrid::mc::Process* process, simgrid::mc::Snapshot* 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 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);
157 xbt_mheap_t heap = process->get_heap();
158 void *start_heap = heap->base;
159 void *end_heap = heap->breakval;
161 add_region(n, snapshot, simgrid::mc::RegionType::Heap, nullptr,
162 start_heap, start_heap,
163 (char *) end_heap - (char *) start_heap);
164 snapshot->heap_bytes_used = mmalloc_get_bytes_used_remote(
166 process->get_malloc_info());
169 if (mc_model_checker->process().privatized() && MC_smpi_process_count())
170 // snapshot->privatization_index = smpi_loaded_page
171 mc_model_checker->process().read_variable(
172 "smpi_loaded_page", &snapshot->privatization_index,
173 sizeof(snapshot->privatization_index));
176 snapshot->privatization_index = simgrid::mc::ProcessIndexMissing;
179 #define PROT_RWX (PROT_READ | PROT_WRITE | PROT_EXEC)
180 #define PROT_RW (PROT_READ | PROT_WRITE)
181 #define PROT_RX (PROT_READ | PROT_EXEC)
183 /** \brief Fills the position of the segments (executable, read-only, read/write).
185 // TODO, use the ELF segment information for more robustness
186 void find_object_address(
187 std::vector<simgrid::xbt::VmMap> const& maps,
188 simgrid::mc::ObjectInformation* result)
190 char* name = xbt_basename(result->file_name.c_str());
192 for (size_t i = 0; i < maps.size(); ++i) {
193 simgrid::xbt::VmMap const& reg = maps[i];
194 if (maps[i].pathname.empty())
196 char* map_basename = xbt_basename(maps[i].pathname.c_str());
197 if (strcmp(name, map_basename) != 0) {
203 // This is the non-GNU_RELRO-part of the data segment:
204 if (reg.prot == PROT_RW) {
205 xbt_assert(!result->start_rw,
206 "Multiple read-write segments for %s, not supported",
207 maps[i].pathname.c_str());
208 result->start_rw = (char*) reg.start_addr;
209 result->end_rw = (char*) reg.end_addr;
211 // The next VMA might be end of the data segment:
212 if (i + 1 < maps.size()
213 && maps[i + 1].pathname.empty()
214 && maps[i + 1].prot == PROT_RW
215 && maps[i + 1].start_addr == reg.end_addr)
216 result->end_rw = (char*) maps[i + 1].end_addr;
219 // This is the text segment:
220 else if (reg.prot == PROT_RX) {
221 xbt_assert(!result->start_exec,
222 "Multiple executable segments for %s, not supported",
223 maps[i].pathname.c_str());
224 result->start_exec = (char*) reg.start_addr;
225 result->end_exec = (char*) reg.end_addr;
227 // The next VMA might be end of the data segment:
228 if (i + 1 < maps.size()
229 && maps[i + 1].pathname.empty()
230 && maps[i + 1].prot == PROT_RW
231 && maps[i + 1].start_addr == reg.end_addr) {
232 result->start_rw = (char*) maps[i + 1].start_addr;
233 result->end_rw = (char*) maps[i + 1].end_addr;
237 // This is the GNU_RELRO-part of the data segment:
238 else if (reg.prot == PROT_READ) {
239 xbt_assert(!result->start_ro,
240 "Multiple read only segments for %s, not supported",
241 maps[i].pathname.c_str());
242 result->start_ro = (char*) reg.start_addr;
243 result->end_ro = (char*) reg.end_addr;
247 result->start = result->start_rw;
248 if ((const void*) result->start_ro < result->start)
249 result->start = result->start_ro;
250 if ((const void*) result->start_exec < result->start)
251 result->start = result->start_exec;
253 result->end = result->end_rw;
254 if (result->end_ro && (const void*) result->end_ro > result->end)
255 result->end = result->end_ro;
256 if (result->end_exec && (const void*) result->end_exec > result->end)
257 result->end = result->end_exec;
259 xbt_assert(result->start_exec || result->start_rw || result->start_ro);
264 /************************************* Take Snapshot ************************************/
265 /****************************************************************************************/
267 /** \brief Checks whether the variable is in scope for a given IP.
269 * A variable may be defined only from a given value of IP.
271 * \param var Variable description
272 * \param scope Scope description
273 * \param ip Instruction pointer
274 * \return true if the variable is valid
276 static bool valid_variable(simgrid::mc::Variable* var,
277 simgrid::mc::Frame* scope,
280 // The variable is not yet valid:
281 if (scope->range.begin() + var->start_scope > (std::uint64_t) ip)
287 static void fill_local_variables_values(mc_stack_frame_t stack_frame,
288 simgrid::mc::Frame* scope,
290 std::vector<s_local_variable>& result)
292 simgrid::mc::Process* process = &mc_model_checker->process();
294 if (!scope || !scope->range.contain(stack_frame->ip))
297 for(simgrid::mc::Variable& current_variable :
300 if (!valid_variable(¤t_variable, scope, (void *) stack_frame->ip))
304 // FIXME, get rid of `region_type`
305 if ((long) stack_frame->ip > (long) process->libsimgrid_info->start_exec)
310 s_local_variable_t new_var;
311 new_var.subprogram = stack_frame->frame;
312 new_var.ip = stack_frame->ip;
313 new_var.name = current_variable.name;
314 new_var.type = current_variable.type;
315 new_var.region = region_type;
316 new_var.address = nullptr;
318 if (current_variable.address != nullptr)
319 new_var.address = current_variable.address;
320 else if (!current_variable.location_list.empty()) {
321 simgrid::dwarf::Location location =
322 simgrid::dwarf::resolve(
323 current_variable.location_list,
324 current_variable.object_info,
325 &(stack_frame->unw_cursor),
326 (void *) stack_frame->frame_base,
327 &mc_model_checker->process(), process_index);
329 if (!location.in_memory())
330 xbt_die("Cannot handle non-address variable");
331 new_var.address = location.address();
334 xbt_die("No address");
336 result.push_back(std::move(new_var));
339 // Recursive processing of nested scopes:
340 for(simgrid::mc::Frame& nested_scope : scope->scopes)
341 fill_local_variables_values(
342 stack_frame, &nested_scope, process_index, result);
345 static std::vector<s_local_variable> get_local_variables_values(
346 std::vector<s_mc_stack_frame_t>& stack_frames, int process_index)
348 std::vector<s_local_variable> variables;
349 for (s_mc_stack_frame_t& stack_frame : stack_frames)
350 fill_local_variables_values(&stack_frame, stack_frame.frame, process_index, variables);
354 static std::vector<s_mc_stack_frame_t> unwind_stack_frames(simgrid::mc::UnwindContext* stack_context)
356 simgrid::mc::Process* process = &mc_model_checker->process();
357 std::vector<s_mc_stack_frame_t> result;
359 unw_cursor_t c = stack_context->cursor();
361 // TODO, check condition check (unw_init_local==0 means end of frame)
365 s_mc_stack_frame_t stack_frame;
367 stack_frame.unw_cursor = c;
371 unw_get_reg(&c, UNW_REG_IP, &ip);
372 unw_get_reg(&c, UNW_REG_SP, &sp);
377 // TODO, use real addresses in frame_t instead of fixing it here
379 simgrid::mc::Frame* frame = process->find_function(remote(ip));
380 stack_frame.frame = frame;
383 stack_frame.frame_name = frame->name;
384 stack_frame.frame_base =
385 (unw_word_t) frame->frame_base(c);
387 stack_frame.frame_base = 0;
388 stack_frame.frame_name = std::string();
391 result.push_back(std::move(stack_frame));
393 /* Stop before context switch with maestro */
394 if (frame != nullptr &&
395 frame->name == "smx_ctx_sysv_wrapper")
398 int ret = unw_step(&c);
400 xbt_die("Unexpected end of stack.");
402 xbt_die("Error while unwinding stack");
405 if (result.empty()) {
406 XBT_INFO("unw_init_local failed");
413 static std::vector<s_mc_snapshot_stack_t> take_snapshot_stacks(simgrid::mc::Snapshot* snapshot)
415 std::vector<s_mc_snapshot_stack_t> res;
417 for (auto const& stack : mc_model_checker->process().stack_areas()) {
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(stack.context));
425 st.context.initialize(&mc_model_checker->process(), &context);
427 st.stack_frames = unwind_stack_frames(&st.context);
428 st.local_variables = get_local_variables_values(st.stack_frames, stack.process_index);
429 st.process_index = stack.process_index;
431 unw_word_t sp = st.stack_frames[0].sp;
433 res.push_back(std::move(st));
436 (char*) stack.address + stack.size - (char*) sp;
437 snapshot->stack_sizes.push_back(stack_size);
444 static void snapshot_handle_ignore(simgrid::mc::Snapshot* snapshot)
446 xbt_assert(snapshot->process());
449 for (auto const& region : mc_model_checker->process().ignored_regions()) {
450 s_mc_snapshot_ignored_data_t ignored_data;
451 ignored_data.start = (void*)region.addr;
452 ignored_data.data.resize(region.size);
453 // TODO, we should do this once per privatization segment:
454 snapshot->process()->read_bytes(
455 ignored_data.data.data(), region.size, remote(region.addr),
456 simgrid::mc::ProcessIndexDisabled);
457 snapshot->ignored_data.push_back(std::move(ignored_data));
461 for(auto const& region : mc_model_checker->process().ignored_regions())
462 snapshot->process()->clear_bytes(remote(region.addr), region.size);
466 static void snapshot_ignore_restore(simgrid::mc::Snapshot* snapshot)
468 for (auto const& ignored_data : snapshot->ignored_data)
469 snapshot->process()->write_bytes(
470 ignored_data.data.data(), ignored_data.data.size(),
471 remote(ignored_data.start));
474 static std::vector<s_fd_infos_t> get_current_fds(pid_t pid)
476 const size_t fd_dir_path_size = 20;
477 char fd_dir_path[fd_dir_path_size];
478 int res = snprintf(fd_dir_path, fd_dir_path_size,
479 "/proc/%lli/fd", (long long int) pid);
480 xbt_assert(res >= 0);
481 if ((size_t) res > fd_dir_path_size)
482 xbt_die("Unexpected buffer is too small for fd_dir_path");
484 DIR* fd_dir = opendir(fd_dir_path);
485 if (fd_dir == nullptr)
486 xbt_die("Cannot open directory '/proc/self/fd'\n");
488 std::vector<s_fd_infos_t> fds;
490 struct dirent* fd_number;
491 while ((fd_number = readdir(fd_dir))) {
493 int fd_value = xbt_str_parse_int(fd_number->d_name, "Found a non-numerical FD: %s. Freaking out!");
498 const size_t source_size = 25;
500 int res = snprintf(source, source_size, "/proc/%lli/fd/%s",
501 (long long int) pid, fd_number->d_name);
502 xbt_assert(res >= 0);
503 if ((size_t) res > source_size)
504 xbt_die("Unexpected buffer is too small for fd %s", fd_number->d_name);
506 const size_t link_size = 200;
508 res = readlink(source, link, link_size);
511 xbt_die("Could not read link for %s", source);
513 xbt_die("Buffer to small for link of %s", source);
518 if(smpi_is_privatisation_file(link))
522 // This is (probably) the DIR* we are reading:
523 // TODO, read all the file entries at once and close the DIR.*
524 if(strcmp(fd_dir_path, link) == 0)
527 // We don't handle them.
528 // It does not mean we should silently ignore them however.
529 if (strncmp(link, "pipe:", std::strlen("pipe:")) == 0 || strncmp(link, "socket:", std::strlen("socket:")) == 0)
532 // If dot_output enabled, do not handle the corresponding file
533 if (dot_output != nullptr) {
534 char* link_basename = xbt_basename(link);
535 if (strcmp(link_basename, _sg_mc_dot_output_file) == 0) {
542 // This is probably a shared memory used by lttng-ust:
543 if(strncmp("/dev/shm/ust-shm-tmp-", link, std::strlen("/dev/shm/ust-shm-tmp-"))==0)
546 // Add an entry for this FD in the snapshot:
548 fd.filename = std::string(link);
549 fd.number = fd_value;
550 fd.flags = fcntl(fd_value, F_GETFL) | fcntl(fd_value, F_GETFD) ;
551 fd.current_position = lseek(fd_value, 0, SEEK_CUR);
552 fds.push_back(std::move(fd));
559 std::shared_ptr<simgrid::mc::Snapshot> take_snapshot(int num_state)
561 XBT_DEBUG("Taking snapshot %i", num_state);
563 simgrid::mc::Process* mc_process = &mc_model_checker->process();
565 std::shared_ptr<simgrid::mc::Snapshot> snapshot = std::make_shared<simgrid::mc::Snapshot>(mc_process, num_state);
567 for (auto& p : mc_model_checker->process().actors())
568 snapshot->enabled_processes.insert(p.copy.getBuffer()->pid);
570 snapshot_handle_ignore(snapshot.get());
572 if (_sg_mc_snapshot_fds)
573 snapshot->current_fds = get_current_fds(mc_model_checker->process().pid());
575 /* Save the std heap and the writable mapped pages of libsimgrid and binary */
576 get_memory_regions(mc_process, snapshot.get());
578 snapshot->to_ignore = mc_model_checker->process().ignored_heap();
580 if (_sg_mc_max_visited_states > 0 || strcmp(_sg_mc_property_file, "")) {
581 snapshot->stacks = take_snapshot_stacks(snapshot.get());
583 snapshot->hash = simgrid::mc::hash(*snapshot);
589 snapshot_ignore_restore(snapshot.get());
594 void restore_snapshot_regions(simgrid::mc::Snapshot* snapshot)
596 for(std::unique_ptr<s_mc_mem_region_t> const& region : snapshot->snapshot_regions) {
597 // For privatized, variables we decided it was not necessary to take the snapshot:
599 restore(region.get());
603 // TODO, send a message to implement this in the MCed process
604 if(snapshot->privatization_index >= 0) {
605 // Fix the privatization mmap:
606 s_mc_restore_message message;
607 message.type = MC_MESSAGE_RESTORE;
608 message.index = snapshot->privatization_index;
609 mc_model_checker->process().getChannel().send(message);
615 void restore_snapshot_fds(simgrid::mc::Snapshot* snapshot)
617 xbt_die("FD snapshot not implemented in client/server mode.");
619 for (auto const& fd : snapshot->current_fds) {
621 int new_fd = open(fd.filename.c_str(), fd.flags);
623 xbt_die("Could not reopen the file %s fo restoring the file descriptor",
624 fd.filename.c_str());
625 if (new_fd != fd.number) {
626 dup2(new_fd, fd.number);
629 lseek(fd.number, fd.current_position, SEEK_SET);
633 void restore_snapshot(std::shared_ptr<simgrid::mc::Snapshot> snapshot)
635 XBT_DEBUG("Restore snapshot %i", snapshot->num_state);
636 restore_snapshot_regions(snapshot.get());
637 if (_sg_mc_snapshot_fds)
638 restore_snapshot_fds(snapshot.get());
639 snapshot_ignore_restore(snapshot.get());
640 mc_model_checker->process().clear_cache();