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>
54 #define MAP_ANONYMOUS MAP_ANON
58 #define MAP_POPULATE 0
61 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_shared, smpi, "Logging specific to SMPI (shared memory macros)");
63 #define PTR_STRLEN (2 + 2 * sizeof(void*) + 1)
66 /** Some location in the source code
68 * This information is used by SMPI_SHARED_MALLOC to allocate some shared memory for all simulated processes.
71 class smpi_source_location {
73 smpi_source_location(const char* filename, int line)
74 : filename(xbt_strdup(filename)), filename_length(strlen(filename)), line(line)
78 /** Pointer to a static string containing the file name */
79 char* filename = nullptr;
80 int filename_length = 0;
83 bool operator==(smpi_source_location const& that) const
85 return filename_length == that.filename_length && line == that.line &&
86 std::memcmp(filename, that.filename, filename_length) == 0;
88 bool operator!=(smpi_source_location const& that) const { return not(*this == that); }
94 template <> class hash<smpi_source_location> {
96 typedef smpi_source_location argument_type;
97 typedef std::size_t result_type;
98 result_type operator()(smpi_source_location const& loc) const
100 return xbt_str_hash_ext(loc.filename, loc.filename_length) ^
101 xbt_str_hash_ext((const char*)&loc.line, sizeof(loc.line));
113 std::unordered_map<smpi_source_location, shared_data_t> allocs;
114 typedef std::unordered_map<smpi_source_location, shared_data_t>::value_type shared_data_key_type;
118 size_t allocated_size;
120 std::vector<std::pair<size_t, size_t>> private_blocks;
121 shared_data_key_type* data;
124 std::map<void*, shared_metadata_t> allocs_metadata;
125 xbt_dict_t calls = nullptr; /* Allocated on first use */
127 static int smpi_shared_malloc_bogusfile = -1;
128 static int smpi_shared_malloc_bogusfile_huge_page = -1;
129 static unsigned long smpi_shared_malloc_blocksize = 1UL << 20;
134 void smpi_shared_destroy()
137 allocs_metadata.clear();
138 xbt_dict_free(&calls);
141 static size_t shm_size(int fd) {
144 if(fstat(fd, &st) < 0) {
145 xbt_die("Could not stat fd %d: %s", fd, strerror(errno));
147 return static_cast<size_t>(st.st_size);
151 static void* shm_map(int fd, size_t size, shared_data_key_type* data) {
152 char loc[PTR_STRLEN];
153 shared_metadata_t meta;
155 if(size > shm_size(fd) && (ftruncate(fd, static_cast<off_t>(size)) < 0)) {
156 xbt_die("Could not truncate fd %d to %zu: %s", fd, size, strerror(errno));
159 void* mem = mmap(nullptr, size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
160 if(mem == MAP_FAILED) {
162 "Failed to map fd %d with size %zu: %s\n"
163 "If you are running a lot of ranks, you may be exceeding the amount of mappings allowed per process.\n"
164 "On Linux systems, change this value with sudo sysctl -w vm.max_map_count=newvalue (default value: 65536)\n"
165 "Please see http://simgrid.gforge.inria.fr/simgrid/latest/doc/html/options.html#options_virt for more info.",
166 fd, size, strerror(errno));
168 snprintf(loc, PTR_STRLEN, "%p", mem);
171 allocs_metadata[mem] = meta;
172 XBT_DEBUG("MMAP %zu to %p", size, mem);
176 static void *smpi_shared_malloc_local(size_t size, const char *file, int line)
179 smpi_source_location loc(file, line);
180 auto res = allocs.insert(std::make_pair(loc, shared_data_t()));
181 auto data = res.first;
183 // The insertion did not take place.
184 // Generate a shared memory name from the address of the shared_data:
185 char shmname[32]; // cannot be longer than PSHMNAMLEN = 31 on Mac OS X (shm_open raises ENAMETOOLONG otherwise)
186 snprintf(shmname, 31, "/shmalloc%p", &*data);
187 int fd = shm_open(shmname, O_RDWR | O_CREAT | O_EXCL, S_IRUSR | S_IWUSR | S_IRGRP | S_IROTH);
190 xbt_die("Please cleanup /dev/shm/%s", shmname);
192 xbt_die("An unhandled error occurred while opening %s. shm_open: %s", shmname, strerror(errno));
194 data->second.fd = fd;
195 data->second.count = 1;
196 mem = shm_map(fd, size, &*data);
197 if (shm_unlink(shmname) < 0) {
198 XBT_WARN("Could not early unlink %s. shm_unlink: %s", shmname, strerror(errno));
200 XBT_DEBUG("Mapping %s at %p through %d", shmname, mem, fd);
202 mem = shm_map(data->second.fd, size, &*data);
203 data->second.count++;
205 XBT_DEBUG("Shared malloc %zu in %p (metadata at %p)", size, mem, &*data);
209 // Align functions, from http://stackoverflow.com/questions/4840410/how-to-align-a-pointer-in-c
210 #define PAGE_SIZE 0x1000
211 #define ALIGN_UP(n, align) (((n) + (align)-1) & -(align))
212 #define ALIGN_DOWN(n, align) ((n) & -(align))
214 #define HUGE_PAGE_SIZE 1<<21
216 /* Similar to smpi_shared_malloc, but only sharing the blocks described by shared_block_offsets.
217 * This array contains the offsets (in bytes) of the block to share.
218 * Even indices are the start offsets (included), odd indices are the stop offsets (excluded).
219 * For instance, if shared_block_offsets == {27, 42}, then the elements mem[27], mem[28], ..., mem[41] are shared.
220 * The others are not.
223 void* smpi_shared_malloc_partial(size_t size, size_t* shared_block_offsets, int nb_shared_blocks)
225 char *huge_page_mount_point = xbt_cfg_get_string("smpi/shared-malloc-hugepage");
226 bool use_huge_page = huge_page_mount_point[0] != '\0';
227 #ifndef MAP_HUGETLB /* If the system header don't define that mmap flag */
228 xbt_assert(not use_huge_page,
229 "Huge pages are not available on your system, you cannot use the smpi/shared-malloc-hugepage option.");
232 smpi_shared_malloc_blocksize = static_cast<unsigned long>(xbt_cfg_get_double("smpi/shared-malloc-blocksize"));
234 size_t allocated_size;
236 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.");
237 allocated_size = size + 2*smpi_shared_malloc_blocksize;
240 xbt_assert(smpi_shared_malloc_blocksize % PAGE_SIZE == 0, "the block size of shared malloc should be a multiple of the page size.");
241 allocated_size = size;
245 /* First reserve memory area */
246 void* allocated_ptr = mmap(NULL, allocated_size, PROT_READ | PROT_WRITE, MAP_ANONYMOUS | MAP_PRIVATE, -1, 0);
248 xbt_assert(allocated_ptr != MAP_FAILED, "Failed to allocate %zuMiB of memory. Run \"sysctl vm.overcommit_memory=1\" as root "
249 "to allow big allocations.\n",
252 mem = (void*)ALIGN_UP((uint64_t)allocated_ptr, HUGE_PAGE_SIZE);
256 XBT_DEBUG("global shared allocation. Blocksize %lu", smpi_shared_malloc_blocksize);
257 /* Create a fd to a new file on disk, make it smpi_shared_malloc_blocksize big, and unlink it.
258 * It still exists in memory but not in the file system (thus it cannot be leaked). */
259 /* Create bogus file if not done already
260 * We need two different bogusfiles:
261 * smpi_shared_malloc_bogusfile_huge_page is used for calls to mmap *with* MAP_HUGETLB,
262 * smpi_shared_malloc_bogusfile is used for calls to mmap *without* MAP_HUGETLB.
263 * We cannot use a same file for the two type of calls, since the first one needs to be
264 * opened in a hugetlbfs mount point whereas the second needs to be a "classical" file. */
265 if(use_huge_page && smpi_shared_malloc_bogusfile_huge_page == -1) {
266 const char *const array[] = {huge_page_mount_point, "simgrid-shmalloc-XXXXXX", nullptr};
267 char *huge_page_filename = xbt_str_join_array(array, "/");
268 smpi_shared_malloc_bogusfile_huge_page = mkstemp(huge_page_filename);
269 XBT_DEBUG("bogusfile_huge_page: %s\n", huge_page_filename);
270 unlink(huge_page_filename);
271 xbt_free(huge_page_filename);
273 if(smpi_shared_malloc_bogusfile == -1) {
274 char *name = xbt_strdup("/tmp/simgrid-shmalloc-XXXXXX");
275 smpi_shared_malloc_bogusfile = mkstemp(name);
276 XBT_DEBUG("bogusfile : %s\n", name);
279 char* dumb = (char*)calloc(1, smpi_shared_malloc_blocksize);
280 ssize_t err = write(smpi_shared_malloc_bogusfile, dumb, smpi_shared_malloc_blocksize);
282 xbt_die("Could not write bogus file for shared malloc");
286 int mmap_base_flag = MAP_FIXED | MAP_SHARED | MAP_POPULATE;
287 int mmap_flag = mmap_base_flag;
288 int huge_fd = use_huge_page ? smpi_shared_malloc_bogusfile_huge_page : smpi_shared_malloc_bogusfile;
291 mmap_flag |= MAP_HUGETLB;
294 XBT_DEBUG("global shared allocation, begin mmap");
296 /* Map the bogus file in place of the anonymous memory */
297 for(int i_block = 0; i_block < nb_shared_blocks; i_block ++) {
298 XBT_DEBUG("\tglobal shared allocation, mmap block %d/%d", i_block+1, nb_shared_blocks);
299 size_t start_offset = shared_block_offsets[2*i_block];
300 size_t stop_offset = shared_block_offsets[2*i_block+1];
301 xbt_assert(start_offset < stop_offset, "start_offset (%zu) should be lower than stop offset (%zu)", start_offset, stop_offset);
302 xbt_assert(stop_offset <= size, "stop_offset (%zu) should be lower than size (%zu)", stop_offset, size);
303 if(i_block < nb_shared_blocks-1)
304 xbt_assert(stop_offset < shared_block_offsets[2*i_block+2],
305 "stop_offset (%zu) should be lower than its successor start offset (%zu)", stop_offset, shared_block_offsets[2*i_block+2]);
306 size_t start_block_offset = ALIGN_UP(start_offset, smpi_shared_malloc_blocksize);
307 size_t stop_block_offset = ALIGN_DOWN(stop_offset, smpi_shared_malloc_blocksize);
308 for (unsigned block_id=0, i = start_block_offset / smpi_shared_malloc_blocksize; i < stop_block_offset / smpi_shared_malloc_blocksize; block_id++, i++) {
309 XBT_DEBUG("\t\tglobal shared allocation, mmap block offset %u", block_id);
310 void* pos = (void*)((unsigned long)mem + i * smpi_shared_malloc_blocksize);
311 void* res = mmap(pos, smpi_shared_malloc_blocksize, PROT_READ | PROT_WRITE, mmap_flag,
313 xbt_assert(res == pos, "Could not map folded virtual memory (%s). Do you perhaps need to increase the "
314 "size of the mapped file using --cfg=smpi/shared-malloc-blocksize=newvalue (default 1048576) ? "
315 "You can also try using the sysctl vm.max_map_count. "
316 "If you are using huge pages, check that you have at least one huge page (/proc/sys/vm/nr_hugepages) "
317 "and that the directory you are passing is mounted correctly (mount /path/to/huge -t hugetlbfs -o rw,mode=0777).",
320 size_t low_page_start_offset = ALIGN_UP(start_offset, PAGE_SIZE);
321 size_t low_page_stop_offset = start_block_offset < ALIGN_DOWN(stop_offset, PAGE_SIZE) ? start_block_offset : ALIGN_DOWN(stop_offset, PAGE_SIZE);
322 if(low_page_start_offset < low_page_stop_offset) {
323 XBT_DEBUG("\t\tglobal shared allocation, mmap block start");
324 void* pos = (void*)((unsigned long)mem + low_page_start_offset);
325 void* res = mmap(pos, low_page_stop_offset-low_page_start_offset, PROT_READ | PROT_WRITE, mmap_base_flag, // not a full huge page
326 smpi_shared_malloc_bogusfile, 0);
327 xbt_assert(res == pos, "Could not map folded virtual memory (%s). Do you perhaps need to increase the "
328 "size of the mapped file using --cfg=smpi/shared-malloc-blocksize=newvalue (default 1048576) ?"
329 "You can also try using the sysctl vm.max_map_count",
332 if(low_page_stop_offset <= stop_block_offset) {
333 XBT_DEBUG("\t\tglobal shared allocation, mmap block stop");
334 size_t high_page_stop_offset = stop_offset == size ? size : ALIGN_DOWN(stop_offset, PAGE_SIZE);
335 if(high_page_stop_offset > stop_block_offset) {
336 void* pos = (void*)((unsigned long)mem + stop_block_offset);
337 void* res = mmap(pos, high_page_stop_offset-stop_block_offset, PROT_READ | PROT_WRITE, mmap_base_flag, // not a full huge page
338 smpi_shared_malloc_bogusfile, 0);
339 xbt_assert(res == pos, "Could not map folded virtual memory (%s). Do you perhaps need to increase the "
340 "size of the mapped file using --cfg=smpi/shared-malloc-blocksize=newvalue (default 1048576) ?"
341 "You can also try using the sysctl vm.max_map_count",
347 shared_metadata_t newmeta;
348 //register metadata for memcpy avoidance
349 shared_data_key_type* data = (shared_data_key_type*)xbt_malloc(sizeof(shared_data_key_type));
350 data->second.fd = -1;
351 data->second.count = 1;
354 newmeta.allocated_ptr = allocated_ptr;
355 newmeta.allocated_size = allocated_size;
356 if(shared_block_offsets[0] > 0) {
357 newmeta.private_blocks.push_back(std::make_pair(0, shared_block_offsets[0]));
360 for(i_block = 0; i_block < nb_shared_blocks-1; i_block ++) {
361 newmeta.private_blocks.push_back(std::make_pair(shared_block_offsets[2*i_block+1], shared_block_offsets[2*i_block+2]));
363 if(shared_block_offsets[2*i_block+1] < size) {
364 newmeta.private_blocks.push_back(std::make_pair(shared_block_offsets[2*i_block+1], size));
366 allocs_metadata[mem] = newmeta;
368 XBT_DEBUG("global shared allocation, allocated_ptr %p - %p", allocated_ptr, (void*)(((uint64_t)allocated_ptr)+allocated_size));
369 XBT_DEBUG("global shared allocation, returned_ptr %p - %p", mem, (void*)(((uint64_t)mem)+size));
374 void *smpi_shared_malloc(size_t size, const char *file, int line) {
375 if (size > 0 && smpi_cfg_shared_malloc == shmalloc_local) {
376 return smpi_shared_malloc_local(size, file, line);
377 } else if (smpi_cfg_shared_malloc == shmalloc_global) {
378 int nb_shared_blocks = 1;
379 size_t shared_block_offsets[2] = {0, size};
380 return smpi_shared_malloc_partial(size, shared_block_offsets, nb_shared_blocks);
382 XBT_DEBUG("Classic malloc %zu", size);
383 return xbt_malloc(size);
386 int smpi_is_shared(void* ptr, std::vector<std::pair<size_t, size_t>> &private_blocks, size_t *offset){
387 private_blocks.clear(); // being paranoid
388 if (allocs_metadata.empty())
390 if ( smpi_cfg_shared_malloc == shmalloc_local || smpi_cfg_shared_malloc == shmalloc_global) {
391 auto low = allocs_metadata.lower_bound(ptr);
392 if (low->first==ptr) {
393 private_blocks = low->second.private_blocks;
397 if (low == allocs_metadata.begin())
400 if (ptr < (char*)low->first + low->second.size) {
401 xbt_assert(ptr > (char*)low->first, "Oops, there seems to be a bug in the shared memory metadata.");
402 *offset = ((uint8_t*)ptr) - ((uint8_t*) low->first);
403 private_blocks = low->second.private_blocks;
412 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) {
413 std::vector<std::pair<size_t, size_t>> result;
414 for(auto block: vec) {
415 auto new_block = std::make_pair(std::min(std::max((size_t)0, block.first-offset), buff_size),
416 std::min(std::max((size_t)0, block.second-offset), buff_size));
417 if(new_block.second > 0 && new_block.first < buff_size)
418 result.push_back(new_block);
423 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) {
424 std::vector<std::pair<size_t, size_t>> result;
427 while(i_src < src.size() && i_dst < dst.size()) {
428 std::pair<size_t, size_t> block;
429 if(src[i_src].second <= dst[i_dst].first) {
432 else if(dst[i_dst].second <= src[i_src].first) {
435 else { // src.second > dst.first && dst.second > src.first → the blocks are overlapping
436 block = std::make_pair(std::max(src[i_src].first, dst[i_dst].first),
437 std::min(src[i_src].second, dst[i_dst].second));
438 result.push_back(block);
439 if(src[i_src].second < dst[i_dst].second)
448 void smpi_shared_free(void *ptr)
450 if (smpi_cfg_shared_malloc == shmalloc_local) {
451 char loc[PTR_STRLEN];
452 snprintf(loc, PTR_STRLEN, "%p", ptr);
453 auto meta = allocs_metadata.find(ptr);
454 if (meta == allocs_metadata.end()) {
455 XBT_WARN("Cannot free: %p was not shared-allocated by SMPI - maybe its size was 0?", ptr);
458 shared_data_t* data = &meta->second.data->second;
459 if (munmap(meta->second.allocated_ptr, meta->second.allocated_size) < 0) {
460 XBT_WARN("Unmapping of fd %d failed: %s", data->fd, strerror(errno));
463 if (data->count <= 0) {
465 allocs.erase(allocs.find(meta->second.data->first));
466 allocs_metadata.erase(ptr);
467 XBT_DEBUG("Shared free - with removal - of %p", ptr);
469 XBT_DEBUG("Shared free - no removal - of %p, count = %d", ptr, data->count);
472 } else if (smpi_cfg_shared_malloc == shmalloc_global) {
473 auto meta = allocs_metadata.find(ptr);
474 if (meta != allocs_metadata.end()){
475 meta->second.data->second.count--;
476 if(meta->second.data->second.count==0)
477 xbt_free(meta->second.data);
480 munmap(ptr, meta->second.size);
482 XBT_DEBUG("Classic free of %p", ptr);
488 int smpi_shared_known_call(const char* func, const char* input)
490 char* loc = bprintf("%s:%s", func, input);
493 if (calls==nullptr) {
494 calls = xbt_dict_new_homogeneous(nullptr);
497 xbt_dict_get(calls, loc); /* Succeed or throw */
503 if (ex.category != not_found_error)
513 void* smpi_shared_get_call(const char* func, const char* input) {
514 char* loc = bprintf("%s:%s", func, input);
516 if (calls == nullptr)
517 calls = xbt_dict_new_homogeneous(nullptr);
518 void* data = xbt_dict_get(calls, loc);
523 void* smpi_shared_set_call(const char* func, const char* input, void* data) {
524 char* loc = bprintf("%s:%s", func, input);
526 if (calls == nullptr)
527 calls = xbt_dict_new_homogeneous(nullptr);
528 xbt_dict_set(calls, loc, data, nullptr);