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' map ---- -.
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' map |
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"
43 #include <sys/types.h>
52 #define MAP_ANONYMOUS MAP_ANON
56 #define MAP_POPULATE 0
59 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_shared, smpi, "Logging specific to SMPI (shared memory macros)");
61 #define PTR_STRLEN (2 + 2 * sizeof(void*) + 1)
64 /** Some location in the source code
66 * This information is used by SMPI_SHARED_MALLOC to allocate some shared memory for all simulated processes.
69 class smpi_source_location : public std::string {
71 smpi_source_location(const char* filename, int line) : std::string(std::string(filename) + ":" + std::to_string(line))
76 struct shared_data_t {
81 std::unordered_map<smpi_source_location, shared_data_t, std::hash<std::string>> allocs;
82 typedef decltype(allocs)::value_type shared_data_key_type;
84 struct shared_metadata_t {
86 size_t allocated_size;
88 std::vector<std::pair<size_t, size_t>> private_blocks;
89 shared_data_key_type* data;
92 std::map<void*, shared_metadata_t> allocs_metadata;
93 std::map<std::string, void*> calls;
96 static int smpi_shared_malloc_bogusfile = -1;
97 static int smpi_shared_malloc_bogusfile_huge_page = -1;
98 static unsigned long smpi_shared_malloc_blocksize = 1UL << 20;
103 void smpi_shared_destroy()
106 allocs_metadata.clear();
110 static size_t shm_size(int fd) {
113 if(fstat(fd, &st) < 0) {
114 xbt_die("Could not stat fd %d: %s", fd, strerror(errno));
116 return static_cast<size_t>(st.st_size);
120 static void* shm_map(int fd, size_t size, shared_data_key_type* data) {
121 char loc[PTR_STRLEN];
122 shared_metadata_t meta;
124 if(size > shm_size(fd) && (ftruncate(fd, static_cast<off_t>(size)) < 0)) {
125 xbt_die("Could not truncate fd %d to %zu: %s", fd, size, strerror(errno));
128 void* mem = mmap(nullptr, size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
129 if(mem == MAP_FAILED) {
131 "Failed to map fd %d with size %zu: %s\n"
132 "If you are running a lot of ranks, you may be exceeding the amount of mappings allowed per process.\n"
133 "On Linux systems, change this value with sudo sysctl -w vm.max_map_count=newvalue (default value: 65536)\n"
134 "Please see http://simgrid.gforge.inria.fr/simgrid/latest/doc/html/options.html#options_virt for more info.",
135 fd, size, strerror(errno));
137 snprintf(loc, PTR_STRLEN, "%p", mem);
140 allocs_metadata[mem] = meta;
141 XBT_DEBUG("MMAP %zu to %p", size, mem);
145 static void *smpi_shared_malloc_local(size_t size, const char *file, int line)
148 smpi_source_location loc(file, line);
149 auto res = allocs.insert(std::make_pair(loc, shared_data_t()));
150 auto data = res.first;
152 // The insertion did not take place.
153 // Generate a shared memory name from the address of the shared_data:
154 char shmname[32]; // cannot be longer than PSHMNAMLEN = 31 on Mac OS X (shm_open raises ENAMETOOLONG otherwise)
155 snprintf(shmname, 31, "/shmalloc%p", &*data);
156 int fd = shm_open(shmname, O_RDWR | O_CREAT | O_EXCL, S_IRUSR | S_IWUSR | S_IRGRP | S_IROTH);
159 xbt_die("Please cleanup /dev/shm/%s", shmname);
161 xbt_die("An unhandled error occurred while opening %s. shm_open: %s", shmname, strerror(errno));
163 data->second.fd = fd;
164 data->second.count = 1;
165 mem = shm_map(fd, size, &*data);
166 if (shm_unlink(shmname) < 0) {
167 XBT_WARN("Could not early unlink %s. shm_unlink: %s", shmname, strerror(errno));
169 XBT_DEBUG("Mapping %s at %p through %d", shmname, mem, fd);
171 mem = shm_map(data->second.fd, size, &*data);
172 data->second.count++;
174 XBT_DEBUG("Shared malloc %zu in %p (metadata at %p)", size, mem, &*data);
178 // Align functions, from http://stackoverflow.com/questions/4840410/how-to-align-a-pointer-in-c
179 #define PAGE_SIZE 0x1000
180 #define ALIGN_UP(n, align) (((n) + (align)-1) & -(align))
181 #define ALIGN_DOWN(n, align) ((n) & -(align))
183 #define HUGE_PAGE_SIZE 1<<21
185 /* Similar to smpi_shared_malloc, but only sharing the blocks described by shared_block_offsets.
186 * This array contains the offsets (in bytes) of the block to share.
187 * Even indices are the start offsets (included), odd indices are the stop offsets (excluded).
188 * For instance, if shared_block_offsets == {27, 42}, then the elements mem[27], mem[28], ..., mem[41] are shared.
189 * The others are not.
192 void* smpi_shared_malloc_partial(size_t size, size_t* shared_block_offsets, int nb_shared_blocks)
194 char *huge_page_mount_point = xbt_cfg_get_string("smpi/shared-malloc-hugepage");
195 bool use_huge_page = huge_page_mount_point[0] != '\0';
196 #ifndef MAP_HUGETLB /* If the system header don't define that mmap flag */
197 xbt_assert(not use_huge_page,
198 "Huge pages are not available on your system, you cannot use the smpi/shared-malloc-hugepage option.");
201 smpi_shared_malloc_blocksize = static_cast<unsigned long>(xbt_cfg_get_double("smpi/shared-malloc-blocksize"));
203 size_t allocated_size;
205 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.");
206 allocated_size = size + 2*smpi_shared_malloc_blocksize;
209 xbt_assert(smpi_shared_malloc_blocksize % PAGE_SIZE == 0, "the block size of shared malloc should be a multiple of the page size.");
210 allocated_size = size;
214 /* First reserve memory area */
215 void* allocated_ptr = mmap(NULL, allocated_size, PROT_READ | PROT_WRITE, MAP_ANONYMOUS | MAP_PRIVATE, -1, 0);
217 xbt_assert(allocated_ptr != MAP_FAILED, "Failed to allocate %zuMiB of memory. Run \"sysctl vm.overcommit_memory=1\" as root "
218 "to allow big allocations.\n",
221 mem = (void*)ALIGN_UP((uint64_t)allocated_ptr, HUGE_PAGE_SIZE);
225 XBT_DEBUG("global shared allocation. Blocksize %lu", smpi_shared_malloc_blocksize);
226 /* Create a fd to a new file on disk, make it smpi_shared_malloc_blocksize big, and unlink it.
227 * It still exists in memory but not in the file system (thus it cannot be leaked). */
228 /* Create bogus file if not done already
229 * We need two different bogusfiles:
230 * smpi_shared_malloc_bogusfile_huge_page is used for calls to mmap *with* MAP_HUGETLB,
231 * smpi_shared_malloc_bogusfile is used for calls to mmap *without* MAP_HUGETLB.
232 * We cannot use a same file for the two type of calls, since the first one needs to be
233 * opened in a hugetlbfs mount point whereas the second needs to be a "classical" file. */
234 if(use_huge_page && smpi_shared_malloc_bogusfile_huge_page == -1) {
235 const char *const array[] = {huge_page_mount_point, "simgrid-shmalloc-XXXXXX", nullptr};
236 char *huge_page_filename = xbt_str_join_array(array, "/");
237 smpi_shared_malloc_bogusfile_huge_page = mkstemp(huge_page_filename);
238 XBT_DEBUG("bogusfile_huge_page: %s\n", huge_page_filename);
239 unlink(huge_page_filename);
240 xbt_free(huge_page_filename);
242 if(smpi_shared_malloc_bogusfile == -1) {
243 char *name = xbt_strdup("/tmp/simgrid-shmalloc-XXXXXX");
244 smpi_shared_malloc_bogusfile = mkstemp(name);
245 XBT_DEBUG("bogusfile : %s\n", name);
248 char* dumb = (char*)calloc(1, smpi_shared_malloc_blocksize);
249 ssize_t err = write(smpi_shared_malloc_bogusfile, dumb, smpi_shared_malloc_blocksize);
251 xbt_die("Could not write bogus file for shared malloc");
255 int mmap_base_flag = MAP_FIXED | MAP_SHARED | MAP_POPULATE;
256 int mmap_flag = mmap_base_flag;
257 int huge_fd = use_huge_page ? smpi_shared_malloc_bogusfile_huge_page : smpi_shared_malloc_bogusfile;
260 mmap_flag |= MAP_HUGETLB;
263 XBT_DEBUG("global shared allocation, begin mmap");
265 /* Map the bogus file in place of the anonymous memory */
266 for(int i_block = 0; i_block < nb_shared_blocks; i_block ++) {
267 XBT_DEBUG("\tglobal shared allocation, mmap block %d/%d", i_block+1, nb_shared_blocks);
268 size_t start_offset = shared_block_offsets[2*i_block];
269 size_t stop_offset = shared_block_offsets[2*i_block+1];
270 xbt_assert(start_offset < stop_offset, "start_offset (%zu) should be lower than stop offset (%zu)", start_offset, stop_offset);
271 xbt_assert(stop_offset <= size, "stop_offset (%zu) should be lower than size (%zu)", stop_offset, size);
272 if(i_block < nb_shared_blocks-1)
273 xbt_assert(stop_offset < shared_block_offsets[2*i_block+2],
274 "stop_offset (%zu) should be lower than its successor start offset (%zu)", stop_offset, shared_block_offsets[2*i_block+2]);
275 size_t start_block_offset = ALIGN_UP(start_offset, smpi_shared_malloc_blocksize);
276 size_t stop_block_offset = ALIGN_DOWN(stop_offset, smpi_shared_malloc_blocksize);
277 for (unsigned block_id=0, i = start_block_offset / smpi_shared_malloc_blocksize; i < stop_block_offset / smpi_shared_malloc_blocksize; block_id++, i++) {
278 XBT_DEBUG("\t\tglobal shared allocation, mmap block offset %u", block_id);
279 void* pos = (void*)((unsigned long)mem + i * smpi_shared_malloc_blocksize);
280 void* res = mmap(pos, smpi_shared_malloc_blocksize, PROT_READ | PROT_WRITE, mmap_flag,
282 xbt_assert(res == pos, "Could not map folded virtual memory (%s). Do you perhaps need to increase the "
283 "size of the mapped file using --cfg=smpi/shared-malloc-blocksize=newvalue (default 1048576) ? "
284 "You can also try using the sysctl vm.max_map_count. "
285 "If you are using huge pages, check that you have at least one huge page (/proc/sys/vm/nr_hugepages) "
286 "and that the directory you are passing is mounted correctly (mount /path/to/huge -t hugetlbfs -o rw,mode=0777).",
289 size_t low_page_start_offset = ALIGN_UP(start_offset, PAGE_SIZE);
290 size_t low_page_stop_offset = start_block_offset < ALIGN_DOWN(stop_offset, PAGE_SIZE) ? start_block_offset : ALIGN_DOWN(stop_offset, PAGE_SIZE);
291 if(low_page_start_offset < low_page_stop_offset) {
292 XBT_DEBUG("\t\tglobal shared allocation, mmap block start");
293 void* pos = (void*)((unsigned long)mem + low_page_start_offset);
294 void* res = mmap(pos, low_page_stop_offset-low_page_start_offset, PROT_READ | PROT_WRITE, mmap_base_flag, // not a full huge page
295 smpi_shared_malloc_bogusfile, 0);
296 xbt_assert(res == pos, "Could not map folded virtual memory (%s). Do you perhaps need to increase the "
297 "size of the mapped file using --cfg=smpi/shared-malloc-blocksize=newvalue (default 1048576) ?"
298 "You can also try using the sysctl vm.max_map_count",
301 if(low_page_stop_offset <= stop_block_offset) {
302 XBT_DEBUG("\t\tglobal shared allocation, mmap block stop");
303 size_t high_page_stop_offset = stop_offset == size ? size : ALIGN_DOWN(stop_offset, PAGE_SIZE);
304 if(high_page_stop_offset > stop_block_offset) {
305 void* pos = (void*)((unsigned long)mem + stop_block_offset);
306 void* res = mmap(pos, high_page_stop_offset-stop_block_offset, PROT_READ | PROT_WRITE, mmap_base_flag, // not a full huge page
307 smpi_shared_malloc_bogusfile, 0);
308 xbt_assert(res == pos, "Could not map folded virtual memory (%s). Do you perhaps need to increase the "
309 "size of the mapped file using --cfg=smpi/shared-malloc-blocksize=newvalue (default 1048576) ?"
310 "You can also try using the sysctl vm.max_map_count",
316 shared_metadata_t newmeta;
317 //register metadata for memcpy avoidance
318 shared_data_key_type* data = (shared_data_key_type*)xbt_malloc(sizeof(shared_data_key_type));
319 data->second.fd = -1;
320 data->second.count = 1;
323 newmeta.allocated_ptr = allocated_ptr;
324 newmeta.allocated_size = allocated_size;
325 if(shared_block_offsets[0] > 0) {
326 newmeta.private_blocks.push_back(std::make_pair(0, shared_block_offsets[0]));
329 for(i_block = 0; i_block < nb_shared_blocks-1; i_block ++) {
330 newmeta.private_blocks.push_back(std::make_pair(shared_block_offsets[2*i_block+1], shared_block_offsets[2*i_block+2]));
332 if(shared_block_offsets[2*i_block+1] < size) {
333 newmeta.private_blocks.push_back(std::make_pair(shared_block_offsets[2*i_block+1], size));
335 allocs_metadata[mem] = newmeta;
337 XBT_DEBUG("global shared allocation, allocated_ptr %p - %p", allocated_ptr, (void*)(((uint64_t)allocated_ptr)+allocated_size));
338 XBT_DEBUG("global shared allocation, returned_ptr %p - %p", mem, (void*)(((uint64_t)mem)+size));
343 void *smpi_shared_malloc(size_t size, const char *file, int line) {
344 if (size > 0 && smpi_cfg_shared_malloc == shmalloc_local) {
345 return smpi_shared_malloc_local(size, file, line);
346 } else if (smpi_cfg_shared_malloc == shmalloc_global) {
347 int nb_shared_blocks = 1;
348 size_t shared_block_offsets[2] = {0, size};
349 return smpi_shared_malloc_partial(size, shared_block_offsets, nb_shared_blocks);
351 XBT_DEBUG("Classic malloc %zu", size);
352 return xbt_malloc(size);
355 int smpi_is_shared(void* ptr, std::vector<std::pair<size_t, size_t>> &private_blocks, size_t *offset){
356 private_blocks.clear(); // being paranoid
357 if (allocs_metadata.empty())
359 if ( smpi_cfg_shared_malloc == shmalloc_local || smpi_cfg_shared_malloc == shmalloc_global) {
360 auto low = allocs_metadata.lower_bound(ptr);
361 if (low != allocs_metadata.end() && low->first == ptr) {
362 private_blocks = low->second.private_blocks;
366 if (low == allocs_metadata.begin())
369 if (ptr < (char*)low->first + low->second.size) {
370 xbt_assert(ptr > (char*)low->first, "Oops, there seems to be a bug in the shared memory metadata.");
371 *offset = ((uint8_t*)ptr) - ((uint8_t*) low->first);
372 private_blocks = low->second.private_blocks;
381 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) {
382 std::vector<std::pair<size_t, size_t>> result;
383 for (auto const& block : vec) {
384 auto new_block = std::make_pair(std::min(std::max((size_t)0, block.first - offset), buff_size),
385 std::min(std::max((size_t)0, block.second - offset), buff_size));
386 if (new_block.second > 0 && new_block.first < buff_size)
387 result.push_back(new_block);
392 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) {
393 std::vector<std::pair<size_t, size_t>> result;
396 while(i_src < src.size() && i_dst < dst.size()) {
397 std::pair<size_t, size_t> block;
398 if(src[i_src].second <= dst[i_dst].first) {
401 else if(dst[i_dst].second <= src[i_src].first) {
404 else { // src.second > dst.first && dst.second > src.first → the blocks are overlapping
405 block = std::make_pair(std::max(src[i_src].first, dst[i_dst].first),
406 std::min(src[i_src].second, dst[i_dst].second));
407 result.push_back(block);
408 if(src[i_src].second < dst[i_dst].second)
417 void smpi_shared_free(void *ptr)
419 if (smpi_cfg_shared_malloc == shmalloc_local) {
420 char loc[PTR_STRLEN];
421 snprintf(loc, PTR_STRLEN, "%p", ptr);
422 auto meta = allocs_metadata.find(ptr);
423 if (meta == allocs_metadata.end()) {
424 XBT_WARN("Cannot free: %p was not shared-allocated by SMPI - maybe its size was 0?", ptr);
427 shared_data_t* data = &meta->second.data->second;
428 if (munmap(meta->second.allocated_ptr, meta->second.allocated_size) < 0) {
429 XBT_WARN("Unmapping of fd %d failed: %s", data->fd, strerror(errno));
432 if (data->count <= 0) {
434 allocs.erase(allocs.find(meta->second.data->first));
435 allocs_metadata.erase(ptr);
436 XBT_DEBUG("Shared free - with removal - of %p", ptr);
438 XBT_DEBUG("Shared free - no removal - of %p, count = %d", ptr, data->count);
441 } else if (smpi_cfg_shared_malloc == shmalloc_global) {
442 auto meta = allocs_metadata.find(ptr);
443 if (meta != allocs_metadata.end()){
444 meta->second.data->second.count--;
445 if(meta->second.data->second.count==0)
446 xbt_free(meta->second.data);
449 munmap(ptr, meta->second.size);
451 XBT_DEBUG("Classic free of %p", ptr);
457 int smpi_shared_known_call(const char* func, const char* input)
459 std::string loc = std::string(func) + ":" + input;
460 return calls.find(loc) != calls.end();
463 void* smpi_shared_get_call(const char* func, const char* input) {
464 std::string loc = std::string(func) + ":" + input;
466 return calls.at(loc);
469 void* smpi_shared_set_call(const char* func, const char* input, void* data) {
470 std::string loc = std::string(func) + ":" + input;