1 /* Copyright (c) 2008-2014. 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. */
16 #include "internal_config.h"
17 #include "mc_memory_map.h"
18 #include "mc_private.h"
19 #include "xbt/module.h"
20 #include <xbt/mmalloc.h>
21 #include "../smpi/private.h"
24 #include "xbt/mmalloc/mmprivate.h"
26 #include "../simix/smx_private.h"
28 #define UNW_LOCAL_ONLY
29 #include <libunwind.h>
32 #include "mc_private.h"
35 #include "mc_snapshot.h"
36 #include "mc_object_info.h"
39 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(mc_checkpoint, mc,
40 "Logging specific to mc_checkpoint");
42 /************************************ Free functions **************************************/
43 /*****************************************************************************************/
45 static void MC_snapshot_stack_free(mc_snapshot_stack_t s)
48 xbt_dynar_free(&(s->local_variables));
49 xbt_dynar_free(&(s->stack_frames));
54 static void MC_snapshot_stack_free_voidp(void *s)
56 MC_snapshot_stack_free((mc_snapshot_stack_t) * (void **) s);
59 static void local_variable_free(local_variable_t v)
65 static void local_variable_free_voidp(void *v)
67 local_variable_free((local_variable_t) * (void **) v);
70 void MC_region_destroy(mc_mem_region_t reg)
75 //munmap(reg->data, reg->size);
77 if (reg->page_numbers) {
78 mc_free_page_snapshot_region(reg->page_numbers, mc_page_count(reg->size));
83 void MC_free_snapshot(mc_snapshot_t snapshot)
86 for (i = 0; i < NB_REGIONS; i++) {
87 MC_region_destroy(snapshot->regions[i]);
90 xbt_free(snapshot->stack_sizes);
91 xbt_dynar_free(&(snapshot->stacks));
92 xbt_dynar_free(&(snapshot->to_ignore));
93 xbt_dynar_free(&snapshot->ignored_data);
95 if (snapshot->privatization_regions) {
96 size_t n = xbt_dynar_length(snapshot->enabled_processes);
97 for (i = 0; i != n; ++i) {
98 MC_region_destroy(snapshot->privatization_regions[i]);
100 xbt_free(snapshot->privatization_regions);
106 /******************************* Snapshot regions ********************************/
107 /*********************************************************************************/
109 static mc_mem_region_t mc_region_new_dense(int type, void *start_addr, void* permanent_addr, size_t size, mc_mem_region_t ref_reg)
111 mc_mem_region_t new_reg = xbt_new(s_mc_mem_region_t, 1);
112 new_reg->start_addr = start_addr;
113 new_reg->permanent_addr = permanent_addr;
114 new_reg->data = NULL;
115 new_reg->size = size;
116 new_reg->page_numbers = NULL;
117 new_reg->data = xbt_malloc(size);
118 memcpy(new_reg->data, permanent_addr, size);
119 XBT_DEBUG("New region : type : %d, data : %p (real addr %p), size : %zu",
120 type, new_reg->data, permanent_addr, size);
125 /** @brief Take a snapshot of a given region
128 * @param start_addr Address of the region in the simulated process
129 * @param permanent_addr Permanent address of this data (for privatized variables, this is the virtual address of the privatized mapping)
130 * @param size Size of the data*
131 * @param ref_reg Reference corresponding region
133 static mc_mem_region_t MC_region_new(int type, void *start_addr, void* permanent_addr, size_t size, mc_mem_region_t ref_reg)
135 if (_sg_mc_sparse_checkpoint) {
136 return mc_region_new_sparse(type, start_addr, permanent_addr, size, ref_reg);
138 return mc_region_new_dense(type, start_addr, permanent_addr, size, ref_reg);
142 /** @brief Restore a region from a snapshot
144 * If we are using per page snapshots, it is possible to use the reference
145 * region in order to do an incremental restoration of the region: the
146 * softclean pages which are shared between the two snapshots do not need
149 * @param reg Target region
150 * @param reg_reg Current region (if not NULL), used for lazy per page restoration
152 static void MC_region_restore(mc_mem_region_t reg, mc_mem_region_t ref_reg)
154 /*FIXME: check if start_addr is still mapped, if it is not, then map it
155 before copying the data */
156 if (!reg->page_numbers) {
157 memcpy(reg->permanent_addr, reg->data, reg->size);
159 mc_region_restore_sparse(reg, ref_reg);
164 static void MC_snapshot_add_region(mc_snapshot_t snapshot, int type,
165 void *start_addr, void* permanent_addr, size_t size)
168 mc_mem_region_t ref_reg =
169 mc_model_checker->parent_snapshot ? mc_model_checker->parent_snapshot->regions[type] : NULL;
170 mc_mem_region_t new_reg = MC_region_new(type, start_addr, permanent_addr, size, ref_reg);
171 snapshot->regions[type] = new_reg;
175 static void MC_get_memory_regions(mc_snapshot_t snapshot)
177 mc_process_t process = &mc_model_checker->process;
179 void *start_heap = std_heap->base;
180 void *end_heap = std_heap->breakval;
181 MC_snapshot_add_region(snapshot, 0, start_heap, start_heap,
182 (char *) end_heap - (char *) start_heap);
183 snapshot->heap_bytes_used = mmalloc_get_bytes_used(std_heap);
184 snapshot->privatization_regions = NULL;
186 MC_snapshot_add_region(snapshot, 1,
187 process->libsimgrid_info->start_rw, process->libsimgrid_info->start_rw,
188 process->libsimgrid_info->end_rw - process->libsimgrid_info->start_rw);
192 if (smpi_privatize_global_variables && smpi_process_count()) {
193 // Snapshot the global variable of the application separately for each
194 // simulated process:
195 snapshot->privatization_regions =
196 xbt_new(mc_mem_region_t, smpi_process_count());
197 for (i = 0; i < smpi_process_count(); i++) {
198 mc_mem_region_t ref_reg =
199 mc_model_checker->parent_snapshot ? mc_model_checker->parent_snapshot->privatization_regions[i] : NULL;
200 snapshot->privatization_regions[i] =
201 MC_region_new(-1, process->binary_info->start_rw, smpi_privatisation_regions[i].address, size_data_exe, ref_reg);
203 snapshot->privatization_index = smpi_loaded_page;
204 snapshot->regions[2] = NULL;
208 MC_snapshot_add_region(snapshot, 2,
209 process->binary_info->start_rw, process->binary_info->start_rw,
210 process->binary_info->end_rw - process->binary_info->start_rw);
211 snapshot->privatization_regions = NULL;
212 snapshot->privatization_index = -1;
216 /** \brief Fills the position of the segments (executable, read-only, read/write).
218 * TODO, use dl_iterate_phdr to be more robust
220 void MC_find_object_address(memory_map_t maps, mc_object_info_t result)
225 const char *name = basename(result->file_name);
226 while (i < maps->mapsize) {
227 reg = maps->regions[i];
228 if (maps->regions[i].pathname == NULL
229 || strcmp(basename(maps->regions[i].pathname), name)) {
231 } else if ((reg.prot & PROT_WRITE)) {
232 xbt_assert(!result->start_rw,
233 "Multiple read-write segments for %s, not supported",
234 maps->regions[i].pathname);
235 result->start_rw = reg.start_addr;
236 result->end_rw = reg.end_addr;
237 // .bss is usually after the .data:
238 s_map_region_t *next = &(maps->regions[i + 1]);
239 if (next->pathname == NULL && (next->prot & PROT_WRITE)
240 && next->start_addr == reg.end_addr) {
241 result->end_rw = maps->regions[i + 1].end_addr;
243 } else if ((reg.prot & PROT_READ) && (reg.prot & PROT_EXEC)) {
244 xbt_assert(!result->start_exec,
245 "Multiple executable segments for %s, not supported",
246 maps->regions[i].pathname);
247 result->start_exec = reg.start_addr;
248 result->end_exec = reg.end_addr;
249 } else if ((reg.prot & PROT_READ) && !(reg.prot & PROT_EXEC)) {
250 xbt_assert(!result->start_ro,
251 "Multiple read only segments for %s, not supported",
252 maps->regions[i].pathname);
253 result->start_ro = reg.start_addr;
254 result->end_ro = reg.end_addr;
259 xbt_assert(result->file_name);
260 xbt_assert(result->start_rw);
261 xbt_assert(result->start_exec);
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 frame Scope description
273 * \param ip Instruction pointer
274 * \return true if the variable is valid
276 static bool mc_valid_variable(dw_variable_t var, dw_frame_t scope,
279 // The variable is not yet valid:
280 if ((const void *) ((const char *) scope->low_pc + var->start_scope) > ip)
286 static void mc_fill_local_variables_values(mc_stack_frame_t stack_frame,
287 dw_frame_t scope, int process_index, xbt_dynar_t result)
289 mc_process_t process = &mc_model_checker->process;
291 void *ip = (void *) stack_frame->ip;
292 if (ip < scope->low_pc || ip >= scope->high_pc)
296 dw_variable_t current_variable;
297 xbt_dynar_foreach(scope->variables, cursor, current_variable) {
299 if (!mc_valid_variable(current_variable, scope, (void *) stack_frame->ip))
303 if ((long) stack_frame->ip > (long) process->libsimgrid_info->start_exec)
308 local_variable_t new_var = xbt_new0(s_local_variable_t, 1);
309 new_var->subprogram = stack_frame->frame;
310 new_var->ip = stack_frame->ip;
311 new_var->name = xbt_strdup(current_variable->name);
312 new_var->type = current_variable->type;
313 new_var->region = region_type;
315 if (current_variable->address != NULL) {
316 new_var->address = current_variable->address;
317 } else if (current_variable->locations.size != 0) {
318 s_mc_location_t location;
319 mc_dwarf_resolve_locations(&location, ¤t_variable->locations,
320 current_variable->object_info,
321 &(stack_frame->unw_cursor),
322 (void *) stack_frame->frame_base,
323 NULL, process_index);
325 switch(mc_get_location_type(&location)) {
326 case MC_LOCATION_TYPE_ADDRESS:
327 new_var->address = location.memory_location;
329 case MC_LOCATION_TYPE_REGISTER:
331 xbt_die("Cannot handle non-address variable");
335 xbt_die("No address");
338 xbt_dynar_push(result, &new_var);
341 // Recursive processing of nested scopes:
342 dw_frame_t nested_scope = NULL;
343 xbt_dynar_foreach(scope->scopes, cursor, nested_scope) {
344 mc_fill_local_variables_values(stack_frame, nested_scope, process_index, result);
348 static xbt_dynar_t MC_get_local_variables_values(xbt_dynar_t stack_frames, int process_index)
351 unsigned cursor1 = 0;
352 mc_stack_frame_t stack_frame;
353 xbt_dynar_t variables =
354 xbt_dynar_new(sizeof(local_variable_t), local_variable_free_voidp);
356 xbt_dynar_foreach(stack_frames, cursor1, stack_frame) {
357 mc_fill_local_variables_values(stack_frame, stack_frame->frame, process_index, variables);
363 static void MC_stack_frame_free_voipd(void *s)
365 mc_stack_frame_t stack_frame = *(mc_stack_frame_t *) s;
367 xbt_free(stack_frame->frame_name);
368 xbt_free(stack_frame);
372 static xbt_dynar_t MC_unwind_stack_frames(void *stack_context)
374 mc_process_t process = &mc_model_checker->process;
376 xbt_dynar_new(sizeof(mc_stack_frame_t), MC_stack_frame_free_voipd);
380 // TODO, check condition check (unw_init_local==0 means end of frame)
381 if (unw_init_local(&c, (unw_context_t *) stack_context) != 0) {
383 xbt_die("Could not initialize stack unwinding");
388 mc_stack_frame_t stack_frame = xbt_new(s_mc_stack_frame_t, 1);
389 xbt_dynar_push(result, &stack_frame);
391 stack_frame->unw_cursor = c;
395 unw_get_reg(&c, UNW_REG_IP, &ip);
396 unw_get_reg(&c, UNW_REG_SP, &sp);
398 stack_frame->ip = ip;
399 stack_frame->sp = sp;
401 // TODO, use real addresses in frame_t instead of fixing it here
403 dw_frame_t frame = MC_process_find_function(process, (void *) ip);
404 stack_frame->frame = frame;
407 stack_frame->frame_name = xbt_strdup(frame->name);
408 stack_frame->frame_base =
409 (unw_word_t) mc_find_frame_base(frame, frame->object_info, &c);
411 stack_frame->frame_base = 0;
412 stack_frame->frame_name = NULL;
415 /* Stop before context switch with maestro */
416 if (frame != NULL && frame->name != NULL
417 && !strcmp(frame->name, "smx_ctx_sysv_wrapper"))
420 int ret = ret = unw_step(&c);
422 xbt_die("Unexpected end of stack.");
423 } else if (ret < 0) {
424 xbt_die("Error while unwinding stack.");
428 if (xbt_dynar_length(result) == 0) {
429 XBT_INFO("unw_init_local failed");
436 static xbt_dynar_t MC_take_snapshot_stacks(mc_snapshot_t * snapshot)
440 xbt_dynar_new(sizeof(s_mc_snapshot_stack_t),
441 MC_snapshot_stack_free_voidp);
443 unsigned int cursor = 0;
444 stack_region_t current_stack;
446 xbt_dynar_foreach(stacks_areas, cursor, current_stack) {
447 mc_snapshot_stack_t st = xbt_new(s_mc_snapshot_stack_t, 1);
448 st->stack_frames = MC_unwind_stack_frames(current_stack->context);
449 st->local_variables = MC_get_local_variables_values(st->stack_frames, current_stack->process_index);
450 st->process_index = current_stack->process_index;
452 unw_word_t sp = xbt_dynar_get_as(st->stack_frames, 0, mc_stack_frame_t)->sp;
454 xbt_dynar_push(res, &st);
455 (*snapshot)->stack_sizes =
456 xbt_realloc((*snapshot)->stack_sizes, (cursor + 1) * sizeof(size_t));
457 (*snapshot)->stack_sizes[cursor] =
458 (char*) current_stack->address + current_stack->size - (char*) sp;
465 static xbt_dynar_t MC_take_snapshot_ignore()
468 if (mc_heap_comparison_ignore == NULL)
472 xbt_dynar_new(sizeof(mc_heap_ignore_region_t),
473 heap_ignore_region_free_voidp);
475 unsigned int cursor = 0;
476 mc_heap_ignore_region_t current_region;
478 xbt_dynar_foreach(mc_heap_comparison_ignore, cursor, current_region) {
479 mc_heap_ignore_region_t new_region = NULL;
480 new_region = xbt_new0(s_mc_heap_ignore_region_t, 1);
481 new_region->address = current_region->address;
482 new_region->size = current_region->size;
483 new_region->block = current_region->block;
484 new_region->fragment = current_region->fragment;
485 xbt_dynar_push(cpy, &new_region);
492 static void mc_free_snapshot_ignored_data_pvoid(void* data) {
493 mc_snapshot_ignored_data_t ignored_data = (mc_snapshot_ignored_data_t) data;
494 free(ignored_data->data);
497 static void MC_snapshot_handle_ignore(mc_snapshot_t snapshot)
499 snapshot->ignored_data = xbt_dynar_new(sizeof(s_mc_snapshot_ignored_data_t), mc_free_snapshot_ignored_data_pvoid);
502 unsigned int cursor = 0;
503 mc_checkpoint_ignore_region_t region;
504 xbt_dynar_foreach (mc_checkpoint_ignore, cursor, region) {
505 s_mc_snapshot_ignored_data_t ignored_data;
506 ignored_data.start = region->addr;
507 ignored_data.size = region->size;
508 ignored_data.data = malloc(region->size);
509 memcpy(ignored_data.data, region->addr, region->size);
510 xbt_dynar_push(snapshot->ignored_data, &ignored_data);
514 xbt_dynar_foreach (mc_checkpoint_ignore, cursor, region) {
515 memset(region->addr, 0, region->size);
520 static void MC_snapshot_ignore_restore(mc_snapshot_t snapshot)
522 unsigned int cursor = 0;
523 s_mc_snapshot_ignored_data_t ignored_data;
524 xbt_dynar_foreach (snapshot->ignored_data, cursor, ignored_data) {
525 memcpy(ignored_data.start, ignored_data.data, ignored_data.size);
529 /** @brief Can we remove this snapshot?
531 * Some snapshots cannot be removed (yet) because we need them
536 int mc_important_snapshot(mc_snapshot_t snapshot)
538 // We need this snapshot in order to know which
539 // pages needs to be stored in the next snapshot.
540 // This field is only non-NULL when using soft-dirty
542 if (snapshot == mc_model_checker->parent_snapshot)
548 static void MC_get_current_fd(mc_snapshot_t snapshot){
550 snapshot->total_fd = 0;
552 const size_t fd_dir_path_size = 20;
553 char fd_dir_path[fd_dir_path_size];
554 if (snprintf(fd_dir_path, fd_dir_path_size,
555 "/proc/%lli/fd", (long long int) getpid()) > fd_dir_path_size)
556 xbt_die("Unexpected buffer is too small for fd_dir_path");
558 DIR* fd_dir = opendir (fd_dir_path);
560 xbt_die("Cannot open directory '/proc/self/fd'\n");
563 struct dirent* fd_number;
564 while ((fd_number = readdir(fd_dir))) {
566 int fd_value = atoi(fd_number->d_name);
571 const size_t source_size = 25;
573 if (snprintf(source, source_size, "/proc/self/fd/%s", fd_number->d_name) > source_size)
574 xbt_die("Unexpected buffer is too small for fd %s", fd_number->d_name);
576 const size_t link_size = 200;
578 int res = readlink(source, link, link_size);
580 xbt_die("Could not read link for %s", source);
583 xbt_die("Buffer to small for link of %s", source);
587 if(smpi_is_privatisation_file(link))
590 // This is (probably) the DIR* we are reading:
591 // TODO, read all the file entries at once and close the DIR.*
592 if(strcmp(fd_dir_path, link) == 0)
595 // We don't handle them.
596 // It does not mean we should silently ignore them however.
597 if (strncmp(link, "pipe:", 5) == 0 || strncmp(link, "socket:", 7) == 0)
600 // This is probably a shared memory used by lttng-ust:
601 if(strncmp("/dev/shm/ust-shm-tmp-", link, 21)==0)
604 // Add an entry for this FD in the snapshot:
605 fd_infos_t fd = xbt_new0(s_fd_infos_t, 1);
606 fd->filename = strdup(link);
607 fd->number = fd_value;
608 fd->flags = fcntl(fd_value, F_GETFL) | fcntl(fd_value, F_GETFD) ;
609 fd->current_position = lseek(fd_value, 0, SEEK_CUR);
610 snapshot->current_fd = xbt_realloc(snapshot->current_fd, (total_fd + 1) * sizeof(fd_infos_t));
611 snapshot->current_fd[total_fd] = fd;
615 snapshot->total_fd = total_fd;
619 mc_snapshot_t MC_take_snapshot(int num_state)
622 mc_snapshot_t snapshot = xbt_new0(s_mc_snapshot_t, 1);
623 snapshot->enabled_processes = xbt_dynar_new(sizeof(int), NULL);
624 smx_process_t process;
625 xbt_swag_foreach(process, simix_global->process_list) {
626 xbt_dynar_push_as(snapshot->enabled_processes, int, (int)process->pid);
629 MC_snapshot_handle_ignore(snapshot);
631 MC_get_current_fd(snapshot);
633 /* Save the std heap and the writable mapped pages of libsimgrid and binary */
634 MC_get_memory_regions(snapshot);
635 if (_sg_mc_sparse_checkpoint && _sg_mc_soft_dirty) {
636 mc_softdirty_reset();
639 snapshot->to_ignore = MC_take_snapshot_ignore();
641 if (_sg_mc_visited > 0 || strcmp(_sg_mc_property_file, "")) {
643 MC_take_snapshot_stacks(&snapshot);
644 if (_sg_mc_hash && snapshot->stacks != NULL) {
645 snapshot->hash = mc_hash_processes_state(num_state, snapshot->stacks);
653 MC_snapshot_ignore_restore(snapshot);
654 if (_sg_mc_sparse_checkpoint && _sg_mc_soft_dirty) {
655 mc_model_checker->parent_snapshot = snapshot;
660 void MC_restore_snapshot(mc_snapshot_t snapshot)
662 mc_snapshot_t parent_snapshot = mc_model_checker->parent_snapshot;
666 for (i = 0; i < NB_REGIONS; i++) {
667 // For privatized, variables we decided it was not necessary to take the snapshot:
668 if (snapshot->regions[i])
669 MC_region_restore(snapshot->regions[i],
670 parent_snapshot ? parent_snapshot->regions[i] : NULL);
674 if (snapshot->privatization_regions) {
675 // Restore the global variables of the application separately for each
676 // simulated process:
677 for (i = 0; i < smpi_process_count(); i++) {
678 if (snapshot->privatization_regions[i]) {
679 MC_region_restore(snapshot->privatization_regions[i],
680 parent_snapshot ? parent_snapshot->privatization_regions[i] : NULL);
684 if(snapshot->privatization_index >= 0) {
685 // We just rewrote the global variables.
686 // The privatisation segment SMPI thinks
687 // is mapped might be inconsistent with the segment which
688 // is really mapped in memory (kernel state).
689 // We ask politely SMPI to map the segment anyway,
690 // even if it thinks it is the current one:
691 smpi_really_switch_data_segment(snapshot->privatization_index);
695 for(i=0; i < snapshot->total_fd; i++){
697 new_fd = open(snapshot->current_fd[i]->filename, snapshot->current_fd[i]->flags);
699 xbt_die("Could not reopen the file %s fo restoring the file descriptor",
700 snapshot->current_fd[i]->filename);
702 if(new_fd != -1 && new_fd != snapshot->current_fd[i]->number){
703 dup2(new_fd, snapshot->current_fd[i]->number);
704 //fprintf(stderr, "%p\n", fdopen(snapshot->current_fd[i]->number, "rw"));
707 lseek(snapshot->current_fd[i]->number, snapshot->current_fd[i]->current_position, SEEK_SET);
710 if (_sg_mc_sparse_checkpoint && _sg_mc_soft_dirty) {
711 mc_softdirty_reset();
714 MC_snapshot_ignore_restore(snapshot);
715 if (_sg_mc_sparse_checkpoint && _sg_mc_soft_dirty) {
716 mc_model_checker->parent_snapshot = snapshot;
721 mc_snapshot_t simcall_HANDLER_mc_snapshot(smx_simcall_t simcall)
723 return MC_take_snapshot(1);
726 void *MC_snapshot(void)
728 return simcall_mc_snapshot();