1 /* Copyright (c) 2007-2023. 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"
41 #include "xbt/config.hpp"
42 #include "xbt/file.hpp"
46 #include "smpi_utils.hpp"
49 #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)");
62 /** Some location in the source code
64 * This information is used by SMPI_SHARED_MALLOC to allocate some shared memory for all simulated processes.
67 class smpi_source_location : public std::string {
69 smpi_source_location() = default;
70 smpi_source_location(const char* filename, int line) : std::string(std::string(filename) + ":" + std::to_string(line))
75 struct shared_data_t {
80 std::unordered_map<smpi_source_location, shared_data_t, std::hash<std::string>> allocs;
81 using shared_data_key_type = decltype(allocs)::value_type;
83 struct shared_metadata_t {
85 size_t allocated_size;
87 std::vector<std::pair<size_t, size_t>> private_blocks;
88 shared_data_key_type* data;
91 std::map<const void*, shared_metadata_t> allocs_metadata;
92 std::map<std::string, void*, std::less<>> calls;
94 int smpi_shared_malloc_bogusfile = -1;
95 int smpi_shared_malloc_bogusfile_huge_page = -1;
96 unsigned long smpi_shared_malloc_blocksize = 1UL << 20;
99 void smpi_shared_destroy()
102 allocs_metadata.clear();
106 static void* shm_map(int fd, size_t size, shared_data_key_type* data)
108 void* mem = smpi_temp_shm_mmap(fd, size);
109 shared_metadata_t meta;
112 meta.allocated_ptr = mem;
113 meta.allocated_size = size;
114 allocs_metadata[mem] = meta;
115 XBT_DEBUG("MMAP %zu to %p", size, mem);
119 static void *smpi_shared_malloc_local(size_t size, const char *file, int line)
122 smpi_source_location loc(file, line);
123 auto [data, inserted] = allocs.try_emplace(loc);
125 // The new element was inserted.
126 int fd = smpi_temp_shm_get();
127 data->second.fd = fd;
128 data->second.count = 1;
129 mem = shm_map(fd, size, &*data);
131 mem = shm_map(data->second.fd, size, &*data);
132 data->second.count++;
134 XBT_DEBUG("Shared malloc %zu in %p through %d (metadata at %p)", size, mem, data->second.fd, &*data);
138 // Align functions, from http://stackoverflow.com/questions/4840410/how-to-align-a-pointer-in-c
139 #define ALIGN_UP(n, align) (((int64_t)(n) + (int64_t)(align) - 1) & -(int64_t)(align))
140 #define ALIGN_DOWN(n, align) ((int64_t)(n) & -(int64_t)(align))
142 constexpr unsigned PAGE_SIZE = 0x1000;
143 constexpr unsigned HUGE_PAGE_SIZE = 1U << 21;
145 /* Similar to smpi_shared_malloc, but only sharing the blocks described by shared_block_offsets.
146 * This array contains the offsets (in bytes) of the block to share.
147 * Even indices are the start offsets (included), odd indices are the stop offsets (excluded).
148 * For instance, if shared_block_offsets == {27, 42}, then the elements mem[27], mem[28], ..., mem[41] are shared.
149 * The others are not.
152 void* smpi_shared_malloc_partial(size_t size, const size_t* shared_block_offsets, int nb_shared_blocks)
154 std::string huge_page_mount_point = simgrid::config::get_value<std::string>("smpi/shared-malloc-hugepage");
155 bool use_huge_page = not huge_page_mount_point.empty();
156 #ifndef MAP_HUGETLB /* If the system header don't define that mmap flag */
157 xbt_assert(not use_huge_page,
158 "Huge pages are not available on your system, you cannot use the smpi/shared-malloc-hugepage option.");
160 smpi_shared_malloc_blocksize =
161 static_cast<unsigned long>(simgrid::config::get_value<double>("smpi/shared-malloc-blocksize"));
163 size_t allocated_size;
165 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.");
166 allocated_size = size + 2*smpi_shared_malloc_blocksize;
169 xbt_assert(smpi_shared_malloc_blocksize % PAGE_SIZE == 0, "the block size of shared malloc should be a multiple of the page size.");
170 allocated_size = size;
174 /* First reserve memory area */
175 void* allocated_ptr = mmap(nullptr, allocated_size, PROT_READ | PROT_WRITE, MAP_ANONYMOUS | MAP_PRIVATE, -1, 0);
177 xbt_assert(allocated_ptr != MAP_FAILED, "Failed to allocate %zuMiB of memory. Run \"sysctl vm.overcommit_memory=1\" as root "
178 "to allow big allocations.\n",
181 mem = (void*)ALIGN_UP(allocated_ptr, HUGE_PAGE_SIZE);
185 XBT_DEBUG("global shared allocation. Blocksize %lu", smpi_shared_malloc_blocksize);
186 /* Create a fd to a new file on disk, make it smpi_shared_malloc_blocksize big, and unlink it.
187 * It still exists in memory but not in the file system (thus it cannot be leaked). */
188 /* Create bogus file if not done already
189 * We need two different bogusfiles:
190 * smpi_shared_malloc_bogusfile_huge_page is used for calls to mmap *with* MAP_HUGETLB,
191 * smpi_shared_malloc_bogusfile is used for calls to mmap *without* MAP_HUGETLB.
192 * We cannot use a same file for the two type of calls, since the first one needs to be
193 * opened in a hugetlbfs mount point whereas the second needs to be a "classical" file. */
194 if(use_huge_page && smpi_shared_malloc_bogusfile_huge_page == -1) {
195 std::string huge_page_filename = huge_page_mount_point + "/simgrid-shmalloc-XXXXXX";
196 smpi_shared_malloc_bogusfile_huge_page = mkstemp((char*)huge_page_filename.c_str());
197 XBT_DEBUG("bogusfile_huge_page: %s\n", huge_page_filename.c_str());
198 unlink(huge_page_filename.c_str());
200 if(smpi_shared_malloc_bogusfile == -1) {
201 char name[] = "/tmp/simgrid-shmalloc-XXXXXX";
202 smpi_shared_malloc_bogusfile = mkstemp(name);
203 XBT_DEBUG("bogusfile : %s\n", name);
205 xbt_assert(ftruncate(smpi_shared_malloc_bogusfile, smpi_shared_malloc_blocksize) == 0,
206 "Could not write bogus file for shared malloc");
209 int mmap_base_flag = MAP_FIXED | MAP_SHARED | MAP_POPULATE;
210 int mmap_flag = mmap_base_flag;
211 int huge_fd = use_huge_page ? smpi_shared_malloc_bogusfile_huge_page : smpi_shared_malloc_bogusfile;
214 mmap_flag |= MAP_HUGETLB;
217 XBT_DEBUG("global shared allocation, begin mmap");
219 /* Map the bogus file in place of the anonymous memory */
220 for(int i_block = 0; i_block < nb_shared_blocks; i_block ++) {
221 XBT_DEBUG("\tglobal shared allocation, mmap block %d/%d", i_block+1, nb_shared_blocks);
222 size_t start_offset = shared_block_offsets[2*i_block];
223 size_t stop_offset = shared_block_offsets[2*i_block+1];
224 xbt_assert(start_offset < stop_offset, "start_offset (%zu) should be lower than stop offset (%zu)", start_offset, stop_offset);
225 xbt_assert(stop_offset <= size, "stop_offset (%zu) should be lower than size (%zu)", stop_offset, size);
226 if(i_block < nb_shared_blocks-1)
227 xbt_assert(stop_offset < shared_block_offsets[2*i_block+2],
228 "stop_offset (%zu) should be lower than its successor start offset (%zu)", stop_offset, shared_block_offsets[2*i_block+2]);
229 size_t start_block_offset = ALIGN_UP(start_offset, smpi_shared_malloc_blocksize);
230 size_t stop_block_offset = ALIGN_DOWN(stop_offset, smpi_shared_malloc_blocksize);
231 for (size_t offset = start_block_offset; offset < stop_block_offset; offset += smpi_shared_malloc_blocksize) {
232 XBT_DEBUG("\t\tglobal shared allocation, mmap block offset %zx", offset);
233 void* pos = static_cast<char*>(mem) + offset;
234 const void* res = mmap(pos, smpi_shared_malloc_blocksize, PROT_READ | PROT_WRITE, mmap_flag, huge_fd, 0);
235 xbt_assert(res == pos, "Could not map folded virtual memory (%s). Do you perhaps need to increase the "
236 "size of the mapped file using --cfg=smpi/shared-malloc-blocksize:newvalue (default 1048576) ? "
237 "You can also try using the sysctl vm.max_map_count. "
238 "If you are using huge pages, check that you have at least one huge page (/proc/sys/vm/nr_hugepages) "
239 "and that the directory you are passing is mounted correctly (mount /path/to/huge -t hugetlbfs -o rw,mode=0777).",
242 size_t low_page_start_offset = ALIGN_UP(start_offset, PAGE_SIZE);
243 size_t low_page_stop_offset = (int64_t)start_block_offset < ALIGN_DOWN(stop_offset, PAGE_SIZE) ? start_block_offset : ALIGN_DOWN(stop_offset, PAGE_SIZE);
244 if(low_page_start_offset < low_page_stop_offset) {
245 XBT_DEBUG("\t\tglobal shared allocation, mmap block start");
246 void* pos = static_cast<char*>(mem) + low_page_start_offset;
247 const void* res = mmap(pos, low_page_stop_offset - low_page_start_offset, PROT_READ | PROT_WRITE,
248 mmap_base_flag, // not a full huge page
249 smpi_shared_malloc_bogusfile, 0);
250 xbt_assert(res == pos, "Could not map folded virtual memory (%s). Do you perhaps need to increase the "
251 "size of the mapped file using --cfg=smpi/shared-malloc-blocksize:newvalue (default 1048576) ?"
252 "You can also try using the sysctl vm.max_map_count",
255 if(low_page_stop_offset <= stop_block_offset) {
256 XBT_DEBUG("\t\tglobal shared allocation, mmap block stop");
257 size_t high_page_stop_offset = stop_offset == size ? size : ALIGN_DOWN(stop_offset, PAGE_SIZE);
258 if(high_page_stop_offset > stop_block_offset) {
259 void* pos = static_cast<char*>(mem) + stop_block_offset;
260 const void* res = mmap(pos, high_page_stop_offset - stop_block_offset, PROT_READ | PROT_WRITE,
261 mmap_base_flag, // not a full huge page
262 smpi_shared_malloc_bogusfile, 0);
263 xbt_assert(res == pos, "Could not map folded virtual memory (%s). Do you perhaps need to increase the "
264 "size of the mapped file using --cfg=smpi/shared-malloc-blocksize:newvalue (default 1048576) ?"
265 "You can also try using the sysctl vm.max_map_count",
271 shared_metadata_t newmeta;
272 //register metadata for memcpy avoidance
273 auto* data = new shared_data_key_type;
274 data->second.fd = -1;
275 data->second.count = 1;
278 newmeta.allocated_ptr = allocated_ptr;
279 newmeta.allocated_size = allocated_size;
280 if(shared_block_offsets[0] > 0) {
281 newmeta.private_blocks.emplace_back(0, shared_block_offsets[0]);
284 for(i_block = 0; i_block < nb_shared_blocks-1; i_block ++) {
285 newmeta.private_blocks.emplace_back(shared_block_offsets[2 * i_block + 1], shared_block_offsets[2 * i_block + 2]);
287 if(shared_block_offsets[2*i_block+1] < size) {
288 newmeta.private_blocks.emplace_back(shared_block_offsets[2 * i_block + 1], size);
290 allocs_metadata[mem] = newmeta;
292 XBT_DEBUG("global shared allocation, allocated_ptr %p - %p", allocated_ptr, (void*)(((uint64_t)allocated_ptr)+allocated_size));
293 XBT_DEBUG("global shared allocation, returned_ptr %p - %p", mem, (void*)(((uint64_t)mem)+size));
298 void* smpi_shared_malloc_intercept(size_t size, const char* file, int line)
300 if( smpi_cfg_auto_shared_malloc_thresh() == 0 || size < smpi_cfg_auto_shared_malloc_thresh()){
301 void* ptr = xbt_malloc(size);
302 if(not smpi_cfg_trace_call_use_absolute_path())
303 simgrid::smpi::utils::account_malloc_size(size, simgrid::xbt::Path(file).get_base_name(), line, ptr);
305 simgrid::smpi::utils::account_malloc_size(size, file, line, ptr);
308 simgrid::smpi::utils::account_shared_size(size);
309 return smpi_shared_malloc(size, file, line);
313 void* smpi_shared_calloc_intercept(size_t num_elm, size_t elem_size, const char* file, int line)
315 size_t size = elem_size * num_elm;
316 if (smpi_cfg_auto_shared_malloc_thresh() == 0 || size < smpi_cfg_auto_shared_malloc_thresh()) {
317 void* ptr = xbt_malloc0(size);
318 if(not smpi_cfg_trace_call_use_absolute_path())
319 simgrid::smpi::utils::account_malloc_size(size, simgrid::xbt::Path(file).get_base_name(), line, ptr);
321 simgrid::smpi::utils::account_malloc_size(size, file, line, ptr);
324 simgrid::smpi::utils::account_shared_size(size);
325 return memset(smpi_shared_malloc(size, file, line), 0, size);
329 void* smpi_shared_realloc_intercept(void* data, size_t size, const char* file, int line)
332 smpi_shared_free(data);
336 return smpi_shared_malloc_intercept(size, file, line);
338 auto meta = allocs_metadata.find(data);
339 if (meta == allocs_metadata.end()) {
340 XBT_DEBUG("Classical realloc(%p, %zu)", data, size);
341 return xbt_realloc(data, size);
344 XBT_DEBUG("Shared realloc(%p, %zu) (old size: %zu)", data, size, meta->second.size);
345 void* ptr = smpi_shared_malloc_intercept(size, file, line);
347 memcpy(ptr, data, std::min(size, meta->second.size));
348 smpi_shared_free(data);
353 void* smpi_shared_malloc(size_t size, const char* file, int line)
355 if (size > 0 && smpi_cfg_shared_malloc() == SharedMallocType::LOCAL) {
356 return smpi_shared_malloc_local(size, file, line);
357 } else if (smpi_cfg_shared_malloc() == SharedMallocType::GLOBAL) {
358 int nb_shared_blocks = 1;
359 const std::array<size_t, 2> shared_block_offsets = {{0, size}};
360 return smpi_shared_malloc_partial(size, shared_block_offsets.data(), nb_shared_blocks);
362 XBT_DEBUG("Classic allocation of %zu bytes", size);
363 return xbt_malloc(size);
366 int smpi_is_shared(const void* ptr, std::vector<std::pair<size_t, size_t>> &private_blocks, size_t *offset){
367 private_blocks.clear(); // being paranoid
368 if (allocs_metadata.empty())
370 if (smpi_cfg_shared_malloc() == SharedMallocType::LOCAL || smpi_cfg_shared_malloc() == SharedMallocType::GLOBAL) {
371 auto low = allocs_metadata.lower_bound(ptr);
372 if (low != allocs_metadata.end() && low->first == ptr) {
373 private_blocks = low->second.private_blocks;
377 if (low == allocs_metadata.begin())
380 if (ptr < (char*)low->first + low->second.size) {
381 xbt_assert(ptr > (char*)low->first, "Oops, there seems to be a bug in the shared memory metadata.");
382 *offset = ((uint8_t*)ptr) - ((uint8_t*) low->first);
383 private_blocks = low->second.private_blocks;
392 std::vector<std::pair<size_t, size_t>> shift_and_frame_private_blocks(const std::vector<std::pair<size_t, size_t>>& vec,
393 size_t offset, size_t buff_size)
395 std::vector<std::pair<size_t, size_t>> result;
396 for (auto const& [block_begin, block_end] : vec) {
397 auto new_block = std::make_pair(std::clamp(block_begin - offset, (size_t)0, buff_size),
398 std::clamp(block_end - offset, (size_t)0, buff_size));
399 if (new_block.second > 0 && new_block.first < buff_size)
400 result.push_back(new_block);
405 std::vector<std::pair<size_t, size_t>> merge_private_blocks(const std::vector<std::pair<size_t, size_t>>& src,
406 const std::vector<std::pair<size_t, size_t>>& dst)
408 std::vector<std::pair<size_t, size_t>> result;
411 while(i_src < src.size() && i_dst < dst.size()) {
412 std::pair<size_t, size_t> block;
413 if(src[i_src].second <= dst[i_dst].first) {
416 else if(dst[i_dst].second <= src[i_src].first) {
419 else { // src.second > dst.first && dst.second > src.first → the blocks are overlapping
420 block = std::make_pair(std::max(src[i_src].first, dst[i_dst].first),
421 std::min(src[i_src].second, dst[i_dst].second));
422 result.push_back(block);
423 if(src[i_src].second < dst[i_dst].second)
432 void smpi_shared_free(void *ptr)
434 simgrid::smpi::utils::account_free(ptr);
435 if (smpi_cfg_shared_malloc() == SharedMallocType::LOCAL) {
436 auto meta = allocs_metadata.find(ptr);
437 if (meta == allocs_metadata.end()) {
441 shared_data_t* data = &meta->second.data->second;
442 if (munmap(meta->second.allocated_ptr, meta->second.allocated_size) < 0) {
443 XBT_WARN("Unmapping of fd %d failed: %s", data->fd, strerror(errno));
446 if (data->count <= 0) {
448 allocs.erase(allocs.find(meta->second.data->first));
449 allocs_metadata.erase(meta);
450 XBT_DEBUG("Shared free - Local - with removal - of %p", ptr);
452 XBT_DEBUG("Shared free - Local - no removal - of %p, count = %d", ptr, data->count);
455 } else if (smpi_cfg_shared_malloc() == SharedMallocType::GLOBAL) {
456 auto meta = allocs_metadata.find(ptr);
457 if (meta != allocs_metadata.end()){
458 meta->second.data->second.count--;
459 XBT_DEBUG("Shared free - Global - of %p", ptr);
460 munmap(ptr, meta->second.size);
461 if(meta->second.data->second.count==0){
462 delete meta->second.data;
463 allocs_metadata.erase(meta);
471 XBT_DEBUG("Classic deallocation of %p", ptr);
476 int smpi_shared_known_call(const char* func, const char* input)
478 std::string loc = std::string(func) + ":" + input;
479 return calls.find(loc) != calls.end();
482 void* smpi_shared_get_call(const char* func, const char* input) {
483 std::string loc = std::string(func) + ":" + input;
485 return calls.at(loc);
488 void* smpi_shared_set_call(const char* func, const char* input, void* data) {
489 std::string loc = std::string(func) + ":" + input;