1 /* Copyright (c) 2007, 2009-2017. The SimGrid Team. All rights reserved. */
3 /* This program is free software; you can redistribute it and/or modify it
4 * under the terms of the license (GNU LGPL) which comes with this package. */
6 /* Shared allocations are handled through shared memory segments.
7 * Associated data and metadata are used as follows:
10 * `allocs' dict ---- -.
11 * ---------- shared_data_t shared_metadata_t / | | |
12 * .->| <name> | ---> -------------------- <--. ----------------- | | | |
13 * | ---------- | fd of <name> | | | size of mmap | --| | | |
14 * | | count (2) | |-- | data | \ | | |
15 * `----------------- | <name> | | ----------------- ---- |
16 * -------------------- | ^ |
18 * | | `allocs_metadata' dict |
19 * | | ---------------------- |
20 * | `-- | <addr of mmap #1> |<-'
21 * | .-- | <addr of mmap #2> |<-.
22 * | | ---------------------- |
28 * | shared_metadata_t / | |
29 * | ----------------- | | |
30 * | | size of mmap | --| | |
32 * ----------------- | | |
40 #include "private.hpp"
45 #include <sys/types.h>
55 #define MAP_ANONYMOUS MAP_ANON
59 #define MAP_POPULATE 0
62 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_shared, smpi, "Logging specific to SMPI (shared memory macros)");
64 #define PTR_STRLEN (2 + 2 * sizeof(void*) + 1)
67 /** Some location in the source code
69 * This information is used by SMPI_SHARED_MALLOC to allocate some shared memory for all simulated processes.
72 class smpi_source_location {
74 smpi_source_location(const char* filename, int line)
75 : filename(xbt_strdup(filename)), filename_length(strlen(filename)), line(line)
79 /** Pointer to a static string containing the file name */
80 char* filename = nullptr;
81 int filename_length = 0;
84 bool operator==(smpi_source_location const& that) const
86 return filename_length == that.filename_length && line == that.line &&
87 std::memcmp(filename, that.filename, filename_length) == 0;
89 bool operator!=(smpi_source_location const& that) const { return not(*this == that); }
95 template <> class hash<smpi_source_location> {
97 typedef smpi_source_location argument_type;
98 typedef std::size_t result_type;
99 result_type operator()(smpi_source_location const& loc) const
101 return xbt_str_hash_ext(loc.filename, loc.filename_length) ^
102 xbt_str_hash_ext((const char*)&loc.line, sizeof(loc.line));
114 std::unordered_map<smpi_source_location, shared_data_t> allocs;
115 typedef std::unordered_map<smpi_source_location, shared_data_t>::value_type shared_data_key_type;
119 size_t allocated_size;
121 std::vector<std::pair<size_t, size_t>> private_blocks;
122 shared_data_key_type* data;
125 std::map<void*, shared_metadata_t> allocs_metadata;
126 xbt_dict_t calls = nullptr; /* Allocated on first use */
128 static int smpi_shared_malloc_bogusfile = -1;
129 static int smpi_shared_malloc_bogusfile_huge_page = -1;
130 static unsigned long smpi_shared_malloc_blocksize = 1UL << 20;
135 void smpi_shared_destroy()
138 allocs_metadata.clear();
139 xbt_dict_free(&calls);
142 static size_t shm_size(int fd) {
145 if(fstat(fd, &st) < 0) {
146 xbt_die("Could not stat fd %d: %s", fd, strerror(errno));
148 return static_cast<size_t>(st.st_size);
152 static void* shm_map(int fd, size_t size, shared_data_key_type* data) {
153 char loc[PTR_STRLEN];
154 shared_metadata_t meta;
156 if(size > shm_size(fd) && (ftruncate(fd, static_cast<off_t>(size)) < 0)) {
157 xbt_die("Could not truncate fd %d to %zu: %s", fd, size, strerror(errno));
160 void* mem = mmap(nullptr, size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
161 if(mem == MAP_FAILED) {
163 "Failed to map fd %d with size %zu: %s\n"
164 "If you are running a lot of ranks, you may be exceeding the amount of mappings allowed per process.\n"
165 "On Linux systems, change this value with sudo sysctl -w vm.max_map_count=newvalue (default value: 65536)\n"
166 "Please see http://simgrid.gforge.inria.fr/simgrid/latest/doc/html/options.html#options_virt for more info.",
167 fd, size, strerror(errno));
169 snprintf(loc, PTR_STRLEN, "%p", mem);
172 allocs_metadata[mem] = meta;
173 XBT_DEBUG("MMAP %zu to %p", size, mem);
177 static void *smpi_shared_malloc_local(size_t size, const char *file, int line)
180 smpi_source_location loc(file, line);
181 auto res = allocs.insert(std::make_pair(loc, shared_data_t()));
182 auto data = res.first;
184 // The insertion did not take place.
185 // Generate a shared memory name from the address of the shared_data:
186 char shmname[32]; // cannot be longer than PSHMNAMLEN = 31 on Mac OS X (shm_open raises ENAMETOOLONG otherwise)
187 snprintf(shmname, 31, "/shmalloc%p", &*data);
188 int fd = shm_open(shmname, O_RDWR | O_CREAT | O_EXCL, S_IRUSR | S_IWUSR | S_IRGRP | S_IROTH);
191 xbt_die("Please cleanup /dev/shm/%s", shmname);
193 xbt_die("An unhandled error occurred while opening %s. shm_open: %s", shmname, strerror(errno));
195 data->second.fd = fd;
196 data->second.count = 1;
197 mem = shm_map(fd, size, &*data);
198 if (shm_unlink(shmname) < 0) {
199 XBT_WARN("Could not early unlink %s. shm_unlink: %s", shmname, strerror(errno));
201 XBT_DEBUG("Mapping %s at %p through %d", shmname, mem, fd);
203 mem = shm_map(data->second.fd, size, &*data);
204 data->second.count++;
206 XBT_DEBUG("Shared malloc %zu in %p (metadata at %p)", size, mem, &*data);
210 // Align functions, from http://stackoverflow.com/questions/4840410/how-to-align-a-pointer-in-c
211 #define PAGE_SIZE 0x1000
212 #define ALIGN_UP(n, align) (((n) + (align)-1) & -(align))
213 #define ALIGN_DOWN(n, align) ((n) & -(align))
215 #define HUGE_PAGE_SIZE 1<<21
217 /* Similar to smpi_shared_malloc, but only sharing the blocks described by shared_block_offsets.
218 * This array contains the offsets (in bytes) of the block to share.
219 * Even indices are the start offsets (included), odd indices are the stop offsets (excluded).
220 * For instance, if shared_block_offsets == {27, 42}, then the elements mem[27], mem[28], ..., mem[41] are shared.
221 * The others are not.
224 void* smpi_shared_malloc_partial(size_t size, size_t* shared_block_offsets, int nb_shared_blocks)
226 char *huge_page_mount_point = xbt_cfg_get_string("smpi/shared-malloc-hugepage");
227 bool use_huge_page = huge_page_mount_point[0] != '\0';
228 #ifndef MAP_HUGETLB /* If the system header don't define that mmap flag */
229 xbt_assert(not use_huge_page,
230 "Huge pages are not available on your system, you cannot use the smpi/shared-malloc-hugepage option.");
233 smpi_shared_malloc_blocksize = static_cast<unsigned long>(xbt_cfg_get_double("smpi/shared-malloc-blocksize"));
235 size_t allocated_size;
237 xbt_assert(smpi_shared_malloc_blocksize == HUGE_PAGE_SIZE, "the block size of shared malloc should be equal to the size of a huge page.");
238 allocated_size = size + 2*smpi_shared_malloc_blocksize;
241 xbt_assert(smpi_shared_malloc_blocksize % PAGE_SIZE == 0, "the block size of shared malloc should be a multiple of the page size.");
242 allocated_size = size;
246 /* First reserve memory area */
247 void* allocated_ptr = mmap(NULL, allocated_size, PROT_READ | PROT_WRITE, MAP_ANONYMOUS | MAP_PRIVATE, -1, 0);
249 xbt_assert(allocated_ptr != MAP_FAILED, "Failed to allocate %zuMiB of memory. Run \"sysctl vm.overcommit_memory=1\" as root "
250 "to allow big allocations.\n",
253 mem = (void*)ALIGN_UP((uint64_t)allocated_ptr, HUGE_PAGE_SIZE);
257 XBT_DEBUG("global shared allocation. Blocksize %lu", smpi_shared_malloc_blocksize);
258 /* Create a fd to a new file on disk, make it smpi_shared_malloc_blocksize big, and unlink it.
259 * It still exists in memory but not in the file system (thus it cannot be leaked). */
260 /* Create bogus file if not done already
261 * We need two different bogusfiles:
262 * smpi_shared_malloc_bogusfile_huge_page is used for calls to mmap *with* MAP_HUGETLB,
263 * smpi_shared_malloc_bogusfile is used for calls to mmap *without* MAP_HUGETLB.
264 * We cannot use a same file for the two type of calls, since the first one needs to be
265 * opened in a hugetlbfs mount point whereas the second needs to be a "classical" file. */
266 if(use_huge_page && smpi_shared_malloc_bogusfile_huge_page == -1) {
267 const char *const array[] = {huge_page_mount_point, "simgrid-shmalloc-XXXXXX", nullptr};
268 char *huge_page_filename = xbt_str_join_array(array, "/");
269 smpi_shared_malloc_bogusfile_huge_page = mkstemp(huge_page_filename);
270 XBT_DEBUG("bogusfile_huge_page: %s\n", huge_page_filename);
271 unlink(huge_page_filename);
272 xbt_free(huge_page_filename);
274 if(smpi_shared_malloc_bogusfile == -1) {
275 char *name = xbt_strdup("/tmp/simgrid-shmalloc-XXXXXX");
276 smpi_shared_malloc_bogusfile = mkstemp(name);
277 XBT_DEBUG("bogusfile : %s\n", name);
280 char* dumb = (char*)calloc(1, smpi_shared_malloc_blocksize);
281 ssize_t err = write(smpi_shared_malloc_bogusfile, dumb, smpi_shared_malloc_blocksize);
283 xbt_die("Could not write bogus file for shared malloc");
287 int mmap_base_flag = MAP_FIXED | MAP_SHARED | MAP_POPULATE;
288 int mmap_flag = mmap_base_flag;
289 int huge_fd = use_huge_page ? smpi_shared_malloc_bogusfile_huge_page : smpi_shared_malloc_bogusfile;
292 mmap_flag |= MAP_HUGETLB;
295 XBT_DEBUG("global shared allocation, begin mmap");
297 /* Map the bogus file in place of the anonymous memory */
298 for(int i_block = 0; i_block < nb_shared_blocks; i_block ++) {
299 XBT_DEBUG("\tglobal shared allocation, mmap block %d/%d", i_block+1, nb_shared_blocks);
300 size_t start_offset = shared_block_offsets[2*i_block];
301 size_t stop_offset = shared_block_offsets[2*i_block+1];
302 xbt_assert(start_offset < stop_offset, "start_offset (%zu) should be lower than stop offset (%zu)", start_offset, stop_offset);
303 xbt_assert(stop_offset <= size, "stop_offset (%zu) should be lower than size (%zu)", stop_offset, size);
304 if(i_block < nb_shared_blocks-1)
305 xbt_assert(stop_offset < shared_block_offsets[2*i_block+2],
306 "stop_offset (%zu) should be lower than its successor start offset (%zu)", stop_offset, shared_block_offsets[2*i_block+2]);
307 size_t start_block_offset = ALIGN_UP(start_offset, smpi_shared_malloc_blocksize);
308 size_t stop_block_offset = ALIGN_DOWN(stop_offset, smpi_shared_malloc_blocksize);
309 for (unsigned block_id=0, i = start_block_offset / smpi_shared_malloc_blocksize; i < stop_block_offset / smpi_shared_malloc_blocksize; block_id++, i++) {
310 XBT_DEBUG("\t\tglobal shared allocation, mmap block offset %d", block_id);
311 void* pos = (void*)((unsigned long)mem + i * smpi_shared_malloc_blocksize);
312 void* res = mmap(pos, smpi_shared_malloc_blocksize, PROT_READ | PROT_WRITE, mmap_flag,
314 xbt_assert(res == pos, "Could not map folded virtual memory (%s). Do you perhaps need to increase the "
315 "size of the mapped file using --cfg=smpi/shared-malloc-blocksize=newvalue (default 1048576) ? "
316 "You can also try using the sysctl vm.max_map_count. "
317 "If you are using huge pages, check that you have at least one huge page (/proc/sys/vm/nr_hugepages) "
318 "and that the directory you are passing is mounted correctly (mount /path/to/huge -t hugetlbfs -o rw,mode=0777).",
321 size_t low_page_start_offset = ALIGN_UP(start_offset, PAGE_SIZE);
322 size_t low_page_stop_offset = start_block_offset < ALIGN_DOWN(stop_offset, PAGE_SIZE) ? start_block_offset : ALIGN_DOWN(stop_offset, PAGE_SIZE);
323 if(low_page_start_offset < low_page_stop_offset) {
324 XBT_DEBUG("\t\tglobal shared allocation, mmap block start");
325 void* pos = (void*)((unsigned long)mem + low_page_start_offset);
326 void* res = mmap(pos, low_page_stop_offset-low_page_start_offset, PROT_READ | PROT_WRITE, mmap_base_flag, // not a full huge page
327 smpi_shared_malloc_bogusfile, 0);
328 xbt_assert(res == pos, "Could not map folded virtual memory (%s). Do you perhaps need to increase the "
329 "size of the mapped file using --cfg=smpi/shared-malloc-blocksize=newvalue (default 1048576) ?"
330 "You can also try using the sysctl vm.max_map_count",
333 if(low_page_stop_offset <= stop_block_offset) {
334 XBT_DEBUG("\t\tglobal shared allocation, mmap block stop");
335 size_t high_page_stop_offset = stop_offset == size ? size : ALIGN_DOWN(stop_offset, PAGE_SIZE);
336 if(high_page_stop_offset > stop_block_offset) {
337 void* pos = (void*)((unsigned long)mem + stop_block_offset);
338 void* res = mmap(pos, high_page_stop_offset-stop_block_offset, PROT_READ | PROT_WRITE, mmap_base_flag, // not a full huge page
339 smpi_shared_malloc_bogusfile, 0);
340 xbt_assert(res == pos, "Could not map folded virtual memory (%s). Do you perhaps need to increase the "
341 "size of the mapped file using --cfg=smpi/shared-malloc-blocksize=newvalue (default 1048576) ?"
342 "You can also try using the sysctl vm.max_map_count",
348 shared_metadata_t newmeta;
349 //register metadata for memcpy avoidance
350 shared_data_key_type* data = (shared_data_key_type*)xbt_malloc(sizeof(shared_data_key_type));
351 data->second.fd = -1;
352 data->second.count = 1;
355 newmeta.allocated_ptr = allocated_ptr;
356 newmeta.allocated_size = allocated_size;
357 if(shared_block_offsets[0] > 0) {
358 newmeta.private_blocks.push_back(std::make_pair(0, shared_block_offsets[0]));
361 for(i_block = 0; i_block < nb_shared_blocks-1; i_block ++) {
362 newmeta.private_blocks.push_back(std::make_pair(shared_block_offsets[2*i_block+1], shared_block_offsets[2*i_block+2]));
364 if(shared_block_offsets[2*i_block+1] < size) {
365 newmeta.private_blocks.push_back(std::make_pair(shared_block_offsets[2*i_block+1], size));
367 allocs_metadata[mem] = newmeta;
369 XBT_DEBUG("global shared allocation, allocated_ptr %p - %p", allocated_ptr, (void*)(((uint64_t)allocated_ptr)+allocated_size));
370 XBT_DEBUG("global shared allocation, returned_ptr %p - %p", mem, (void*)(((uint64_t)mem)+size));
375 void *smpi_shared_malloc(size_t size, const char *file, int line) {
376 if (size > 0 && smpi_cfg_shared_malloc == shmalloc_local) {
377 return smpi_shared_malloc_local(size, file, line);
378 } else if (smpi_cfg_shared_malloc == shmalloc_global) {
379 int nb_shared_blocks = 1;
380 size_t shared_block_offsets[2] = {0, size};
381 return smpi_shared_malloc_partial(size, shared_block_offsets, nb_shared_blocks);
383 XBT_DEBUG("Classic malloc %zu", size);
384 return xbt_malloc(size);
387 int smpi_is_shared(void* ptr, std::vector<std::pair<size_t, size_t>> &private_blocks, size_t *offset){
388 private_blocks.clear(); // being paranoid
389 if (allocs_metadata.empty())
391 if ( smpi_cfg_shared_malloc == shmalloc_local || smpi_cfg_shared_malloc == shmalloc_global) {
392 auto low = allocs_metadata.lower_bound(ptr);
393 if (low->first==ptr) {
394 private_blocks = low->second.private_blocks;
398 if (low == allocs_metadata.begin())
401 if (ptr < (char*)low->first + low->second.size) {
402 xbt_assert(ptr > (char*)low->first, "Oops, there seems to be a bug in the shared memory metadata.");
403 *offset = ((uint8_t*)ptr) - ((uint8_t*) low->first);
404 private_blocks = low->second.private_blocks;
413 std::vector<std::pair<size_t, size_t>> shift_and_frame_private_blocks(const std::vector<std::pair<size_t, size_t>> vec, size_t offset, size_t buff_size) {
414 std::vector<std::pair<size_t, size_t>> result;
415 for(auto block: vec) {
416 auto new_block = std::make_pair(std::min(std::max((size_t)0, block.first-offset), buff_size),
417 std::min(std::max((size_t)0, block.second-offset), buff_size));
418 if(new_block.second > 0 && new_block.first < buff_size)
419 result.push_back(new_block);
424 std::vector<std::pair<size_t, size_t>> merge_private_blocks(std::vector<std::pair<size_t, size_t>> src, std::vector<std::pair<size_t, size_t>> dst) {
425 std::vector<std::pair<size_t, size_t>> result;
428 while(i_src < src.size() && i_dst < dst.size()) {
429 std::pair<size_t, size_t> block;
430 if(src[i_src].second <= dst[i_dst].first) {
433 else if(dst[i_dst].second <= src[i_src].first) {
436 else { // src.second > dst.first && dst.second > src.first → the blocks are overlapping
437 block = std::make_pair(std::max(src[i_src].first, dst[i_dst].first),
438 std::min(src[i_src].second, dst[i_dst].second));
439 result.push_back(block);
440 if(src[i_src].second < dst[i_dst].second)
449 void smpi_shared_free(void *ptr)
451 if (smpi_cfg_shared_malloc == shmalloc_local) {
452 char loc[PTR_STRLEN];
453 snprintf(loc, PTR_STRLEN, "%p", ptr);
454 auto meta = allocs_metadata.find(ptr);
455 if (meta == allocs_metadata.end()) {
456 XBT_WARN("Cannot free: %p was not shared-allocated by SMPI - maybe its size was 0?", ptr);
459 shared_data_t* data = &meta->second.data->second;
460 if (munmap(meta->second.allocated_ptr, meta->second.allocated_size) < 0) {
461 XBT_WARN("Unmapping of fd %d failed: %s", data->fd, strerror(errno));
464 if (data->count <= 0) {
466 allocs.erase(allocs.find(meta->second.data->first));
467 allocs_metadata.erase(ptr);
468 XBT_DEBUG("Shared free - with removal - of %p", ptr);
470 XBT_DEBUG("Shared free - no removal - of %p, count = %d", ptr, data->count);
473 } else if (smpi_cfg_shared_malloc == shmalloc_global) {
474 auto meta = allocs_metadata.find(ptr);
475 if (meta != allocs_metadata.end()){
476 meta->second.data->second.count--;
477 if(meta->second.data->second.count==0)
478 xbt_free(meta->second.data);
481 munmap(ptr, meta->second.size);
483 XBT_DEBUG("Classic free of %p", ptr);
489 int smpi_shared_known_call(const char* func, const char* input)
491 char* loc = bprintf("%s:%s", func, input);
494 if (calls==nullptr) {
495 calls = xbt_dict_new_homogeneous(nullptr);
498 xbt_dict_get(calls, loc); /* Succeed or throw */
504 if (ex.category != not_found_error)
514 void* smpi_shared_get_call(const char* func, const char* input) {
515 char* loc = bprintf("%s:%s", func, input);
517 if (calls == nullptr)
518 calls = xbt_dict_new_homogeneous(nullptr);
519 void* data = xbt_dict_get(calls, loc);
524 void* smpi_shared_set_call(const char* func, const char* input, void* data) {
525 char* loc = bprintf("%s:%s", func, input);
527 if (calls == nullptr)
528 calls = xbt_dict_new_homogeneous(nullptr);
529 xbt_dict_set(calls, loc, data, nullptr);