1 /* Copyright (c) 2007-2022. 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 <sys/types.h>
51 #include <sys/types.h>
53 #include "smpi_utils.hpp"
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)");
65 /** Some location in the source code
67 * This information is used by SMPI_SHARED_MALLOC to allocate some shared memory for all simulated processes.
70 class smpi_source_location : public std::string {
72 smpi_source_location() = default;
73 smpi_source_location(const char* filename, int line) : std::string(std::string(filename) + ":" + std::to_string(line))
78 struct shared_data_t {
83 std::unordered_map<smpi_source_location, shared_data_t, std::hash<std::string>> allocs;
84 using shared_data_key_type = decltype(allocs)::value_type;
86 struct shared_metadata_t {
88 size_t allocated_size;
90 std::vector<std::pair<size_t, size_t>> private_blocks;
91 shared_data_key_type* data;
94 std::map<const void*, shared_metadata_t> allocs_metadata;
95 std::map<std::string, void*, std::less<>> calls;
98 int smpi_shared_malloc_bogusfile = -1;
99 int smpi_shared_malloc_bogusfile_huge_page = -1;
100 unsigned long smpi_shared_malloc_blocksize = 1UL << 20;
104 void smpi_shared_destroy()
107 allocs_metadata.clear();
112 static void* shm_map(int fd, size_t size, shared_data_key_type* data)
114 void* mem = smpi_temp_shm_mmap(fd, size);
115 shared_metadata_t meta;
118 meta.allocated_ptr = mem;
119 meta.allocated_size = size;
120 allocs_metadata[mem] = meta;
121 XBT_DEBUG("MMAP %zu to %p", size, mem);
125 static void *smpi_shared_malloc_local(size_t size, const char *file, int line)
128 smpi_source_location loc(file, line);
129 auto [data, inserted] = allocs.try_emplace(loc);
131 // The new element was inserted.
132 int fd = smpi_temp_shm_get();
133 data->second.fd = fd;
134 data->second.count = 1;
135 mem = shm_map(fd, size, &*data);
137 mem = shm_map(data->second.fd, size, &*data);
138 data->second.count++;
140 XBT_DEBUG("Shared malloc %zu in %p through %d (metadata at %p)", size, mem, data->second.fd, &*data);
144 // Align functions, from http://stackoverflow.com/questions/4840410/how-to-align-a-pointer-in-c
145 #define ALIGN_UP(n, align) (((int64_t)(n) + (int64_t)(align) - 1) & -(int64_t)(align))
146 #define ALIGN_DOWN(n, align) ((int64_t)(n) & -(int64_t)(align))
148 constexpr unsigned PAGE_SIZE = 0x1000;
149 constexpr unsigned HUGE_PAGE_SIZE = 1U << 21;
151 /* Similar to smpi_shared_malloc, but only sharing the blocks described by shared_block_offsets.
152 * This array contains the offsets (in bytes) of the block to share.
153 * Even indices are the start offsets (included), odd indices are the stop offsets (excluded).
154 * For instance, if shared_block_offsets == {27, 42}, then the elements mem[27], mem[28], ..., mem[41] are shared.
155 * The others are not.
158 void* smpi_shared_malloc_partial(size_t size, const size_t* shared_block_offsets, int nb_shared_blocks)
160 std::string huge_page_mount_point = simgrid::config::get_value<std::string>("smpi/shared-malloc-hugepage");
161 bool use_huge_page = not huge_page_mount_point.empty();
162 #ifndef MAP_HUGETLB /* If the system header don't define that mmap flag */
163 xbt_assert(not use_huge_page,
164 "Huge pages are not available on your system, you cannot use the smpi/shared-malloc-hugepage option.");
166 smpi_shared_malloc_blocksize =
167 static_cast<unsigned long>(simgrid::config::get_value<double>("smpi/shared-malloc-blocksize"));
169 size_t allocated_size;
171 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.");
172 allocated_size = size + 2*smpi_shared_malloc_blocksize;
175 xbt_assert(smpi_shared_malloc_blocksize % PAGE_SIZE == 0, "the block size of shared malloc should be a multiple of the page size.");
176 allocated_size = size;
180 /* First reserve memory area */
181 void* allocated_ptr = mmap(nullptr, allocated_size, PROT_READ | PROT_WRITE, MAP_ANONYMOUS | MAP_PRIVATE, -1, 0);
183 xbt_assert(allocated_ptr != MAP_FAILED, "Failed to allocate %zuMiB of memory. Run \"sysctl vm.overcommit_memory=1\" as root "
184 "to allow big allocations.\n",
187 mem = (void*)ALIGN_UP(allocated_ptr, HUGE_PAGE_SIZE);
191 XBT_DEBUG("global shared allocation. Blocksize %lu", smpi_shared_malloc_blocksize);
192 /* Create a fd to a new file on disk, make it smpi_shared_malloc_blocksize big, and unlink it.
193 * It still exists in memory but not in the file system (thus it cannot be leaked). */
194 /* Create bogus file if not done already
195 * We need two different bogusfiles:
196 * smpi_shared_malloc_bogusfile_huge_page is used for calls to mmap *with* MAP_HUGETLB,
197 * smpi_shared_malloc_bogusfile is used for calls to mmap *without* MAP_HUGETLB.
198 * We cannot use a same file for the two type of calls, since the first one needs to be
199 * opened in a hugetlbfs mount point whereas the second needs to be a "classical" file. */
200 if(use_huge_page && smpi_shared_malloc_bogusfile_huge_page == -1) {
201 std::string huge_page_filename = huge_page_mount_point + "/simgrid-shmalloc-XXXXXX";
202 smpi_shared_malloc_bogusfile_huge_page = mkstemp((char*)huge_page_filename.c_str());
203 XBT_DEBUG("bogusfile_huge_page: %s\n", huge_page_filename.c_str());
204 unlink(huge_page_filename.c_str());
206 if(smpi_shared_malloc_bogusfile == -1) {
207 char name[] = "/tmp/simgrid-shmalloc-XXXXXX";
208 smpi_shared_malloc_bogusfile = mkstemp(name);
209 XBT_DEBUG("bogusfile : %s\n", name);
211 xbt_assert(ftruncate(smpi_shared_malloc_bogusfile, smpi_shared_malloc_blocksize) == 0,
212 "Could not write bogus file for shared malloc");
215 int mmap_base_flag = MAP_FIXED | MAP_SHARED | MAP_POPULATE;
216 int mmap_flag = mmap_base_flag;
217 int huge_fd = use_huge_page ? smpi_shared_malloc_bogusfile_huge_page : smpi_shared_malloc_bogusfile;
220 mmap_flag |= MAP_HUGETLB;
223 XBT_DEBUG("global shared allocation, begin mmap");
225 /* Map the bogus file in place of the anonymous memory */
226 for(int i_block = 0; i_block < nb_shared_blocks; i_block ++) {
227 XBT_DEBUG("\tglobal shared allocation, mmap block %d/%d", i_block+1, nb_shared_blocks);
228 size_t start_offset = shared_block_offsets[2*i_block];
229 size_t stop_offset = shared_block_offsets[2*i_block+1];
230 xbt_assert(start_offset < stop_offset, "start_offset (%zu) should be lower than stop offset (%zu)", start_offset, stop_offset);
231 xbt_assert(stop_offset <= size, "stop_offset (%zu) should be lower than size (%zu)", stop_offset, size);
232 if(i_block < nb_shared_blocks-1)
233 xbt_assert(stop_offset < shared_block_offsets[2*i_block+2],
234 "stop_offset (%zu) should be lower than its successor start offset (%zu)", stop_offset, shared_block_offsets[2*i_block+2]);
235 size_t start_block_offset = ALIGN_UP(start_offset, smpi_shared_malloc_blocksize);
236 size_t stop_block_offset = ALIGN_DOWN(stop_offset, smpi_shared_malloc_blocksize);
237 for (size_t offset = start_block_offset; offset < stop_block_offset; offset += smpi_shared_malloc_blocksize) {
238 XBT_DEBUG("\t\tglobal shared allocation, mmap block offset %zx", offset);
239 void* pos = static_cast<char*>(mem) + offset;
240 const void* res = mmap(pos, smpi_shared_malloc_blocksize, PROT_READ | PROT_WRITE, mmap_flag, huge_fd, 0);
241 xbt_assert(res == pos, "Could not map folded virtual memory (%s). Do you perhaps need to increase the "
242 "size of the mapped file using --cfg=smpi/shared-malloc-blocksize:newvalue (default 1048576) ? "
243 "You can also try using the sysctl vm.max_map_count. "
244 "If you are using huge pages, check that you have at least one huge page (/proc/sys/vm/nr_hugepages) "
245 "and that the directory you are passing is mounted correctly (mount /path/to/huge -t hugetlbfs -o rw,mode=0777).",
248 size_t low_page_start_offset = ALIGN_UP(start_offset, PAGE_SIZE);
249 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);
250 if(low_page_start_offset < low_page_stop_offset) {
251 XBT_DEBUG("\t\tglobal shared allocation, mmap block start");
252 void* pos = static_cast<char*>(mem) + low_page_start_offset;
253 const void* res = mmap(pos, low_page_stop_offset - low_page_start_offset, PROT_READ | PROT_WRITE,
254 mmap_base_flag, // not a full huge page
255 smpi_shared_malloc_bogusfile, 0);
256 xbt_assert(res == pos, "Could not map folded virtual memory (%s). Do you perhaps need to increase the "
257 "size of the mapped file using --cfg=smpi/shared-malloc-blocksize:newvalue (default 1048576) ?"
258 "You can also try using the sysctl vm.max_map_count",
261 if(low_page_stop_offset <= stop_block_offset) {
262 XBT_DEBUG("\t\tglobal shared allocation, mmap block stop");
263 size_t high_page_stop_offset = stop_offset == size ? size : ALIGN_DOWN(stop_offset, PAGE_SIZE);
264 if(high_page_stop_offset > stop_block_offset) {
265 void* pos = static_cast<char*>(mem) + stop_block_offset;
266 const void* res = mmap(pos, high_page_stop_offset - stop_block_offset, PROT_READ | PROT_WRITE,
267 mmap_base_flag, // not a full huge page
268 smpi_shared_malloc_bogusfile, 0);
269 xbt_assert(res == pos, "Could not map folded virtual memory (%s). Do you perhaps need to increase the "
270 "size of the mapped file using --cfg=smpi/shared-malloc-blocksize:newvalue (default 1048576) ?"
271 "You can also try using the sysctl vm.max_map_count",
277 shared_metadata_t newmeta;
278 //register metadata for memcpy avoidance
279 auto* data = new shared_data_key_type;
280 data->second.fd = -1;
281 data->second.count = 1;
284 newmeta.allocated_ptr = allocated_ptr;
285 newmeta.allocated_size = allocated_size;
286 if(shared_block_offsets[0] > 0) {
287 newmeta.private_blocks.emplace_back(0, shared_block_offsets[0]);
290 for(i_block = 0; i_block < nb_shared_blocks-1; i_block ++) {
291 newmeta.private_blocks.emplace_back(shared_block_offsets[2 * i_block + 1], shared_block_offsets[2 * i_block + 2]);
293 if(shared_block_offsets[2*i_block+1] < size) {
294 newmeta.private_blocks.emplace_back(shared_block_offsets[2 * i_block + 1], size);
296 allocs_metadata[mem] = newmeta;
298 XBT_DEBUG("global shared allocation, allocated_ptr %p - %p", allocated_ptr, (void*)(((uint64_t)allocated_ptr)+allocated_size));
299 XBT_DEBUG("global shared allocation, returned_ptr %p - %p", mem, (void*)(((uint64_t)mem)+size));
304 void* smpi_shared_malloc_intercept(size_t size, const char* file, int line)
306 if( smpi_cfg_auto_shared_malloc_thresh() == 0 || size < smpi_cfg_auto_shared_malloc_thresh()){
307 void* ptr = xbt_malloc(size);
308 if(not smpi_cfg_trace_call_use_absolute_path())
309 simgrid::smpi::utils::account_malloc_size(size, simgrid::xbt::Path(file).get_base_name(), line, ptr);
311 simgrid::smpi::utils::account_malloc_size(size, file, line, ptr);
314 simgrid::smpi::utils::account_shared_size(size);
315 return smpi_shared_malloc(size, file, line);
319 void* smpi_shared_calloc_intercept(size_t num_elm, size_t elem_size, const char* file, int line)
321 size_t size = elem_size * num_elm;
322 if (smpi_cfg_auto_shared_malloc_thresh() == 0 || size < smpi_cfg_auto_shared_malloc_thresh()) {
323 void* ptr = xbt_malloc0(size);
324 if(not smpi_cfg_trace_call_use_absolute_path())
325 simgrid::smpi::utils::account_malloc_size(size, simgrid::xbt::Path(file).get_base_name(), line, ptr);
327 simgrid::smpi::utils::account_malloc_size(size, file, line, ptr);
330 simgrid::smpi::utils::account_shared_size(size);
331 return memset(smpi_shared_malloc(size, file, line), 0, size);
335 void* smpi_shared_realloc_intercept(void* data, size_t size, const char* file, int line)
338 smpi_shared_free(data);
342 return smpi_shared_malloc_intercept(size, file, line);
344 auto meta = allocs_metadata.find(data);
345 if (meta == allocs_metadata.end()) {
346 XBT_DEBUG("Classical realloc(%p, %zu)", data, size);
347 return xbt_realloc(data, size);
350 XBT_DEBUG("Shared realloc(%p, %zu) (old size: %zu)", data, size, meta->second.size);
351 void* ptr = smpi_shared_malloc_intercept(size, file, line);
353 memcpy(ptr, data, std::min(size, meta->second.size));
354 smpi_shared_free(data);
359 void* smpi_shared_malloc(size_t size, const char* file, int line)
361 if (size > 0 && smpi_cfg_shared_malloc() == SharedMallocType::LOCAL) {
362 return smpi_shared_malloc_local(size, file, line);
363 } else if (smpi_cfg_shared_malloc() == SharedMallocType::GLOBAL) {
364 int nb_shared_blocks = 1;
365 const std::array<size_t, 2> shared_block_offsets = {{0, size}};
366 return smpi_shared_malloc_partial(size, shared_block_offsets.data(), nb_shared_blocks);
368 XBT_DEBUG("Classic allocation of %zu bytes", size);
369 return xbt_malloc(size);
372 int smpi_is_shared(const void* ptr, std::vector<std::pair<size_t, size_t>> &private_blocks, size_t *offset){
373 private_blocks.clear(); // being paranoid
374 if (allocs_metadata.empty())
376 if (smpi_cfg_shared_malloc() == SharedMallocType::LOCAL || smpi_cfg_shared_malloc() == SharedMallocType::GLOBAL) {
377 auto low = allocs_metadata.lower_bound(ptr);
378 if (low != allocs_metadata.end() && low->first == ptr) {
379 private_blocks = low->second.private_blocks;
383 if (low == allocs_metadata.begin())
386 if (ptr < (char*)low->first + low->second.size) {
387 xbt_assert(ptr > (char*)low->first, "Oops, there seems to be a bug in the shared memory metadata.");
388 *offset = ((uint8_t*)ptr) - ((uint8_t*) low->first);
389 private_blocks = low->second.private_blocks;
398 std::vector<std::pair<size_t, size_t>> shift_and_frame_private_blocks(const std::vector<std::pair<size_t, size_t>>& vec,
399 size_t offset, size_t buff_size)
401 std::vector<std::pair<size_t, size_t>> result;
402 for (auto const& [block_begin, block_end] : vec) {
403 auto new_block = std::make_pair(std::clamp(block_begin - offset, (size_t)0, buff_size),
404 std::clamp(block_end - offset, (size_t)0, buff_size));
405 if (new_block.second > 0 && new_block.first < buff_size)
406 result.push_back(new_block);
411 std::vector<std::pair<size_t, size_t>> merge_private_blocks(const std::vector<std::pair<size_t, size_t>>& src,
412 const std::vector<std::pair<size_t, size_t>>& dst)
414 std::vector<std::pair<size_t, size_t>> result;
417 while(i_src < src.size() && i_dst < dst.size()) {
418 std::pair<size_t, size_t> block;
419 if(src[i_src].second <= dst[i_dst].first) {
422 else if(dst[i_dst].second <= src[i_src].first) {
425 else { // src.second > dst.first && dst.second > src.first → the blocks are overlapping
426 block = std::make_pair(std::max(src[i_src].first, dst[i_dst].first),
427 std::min(src[i_src].second, dst[i_dst].second));
428 result.push_back(block);
429 if(src[i_src].second < dst[i_dst].second)
438 void smpi_shared_free(void *ptr)
440 simgrid::smpi::utils::account_free(ptr);
441 if (smpi_cfg_shared_malloc() == SharedMallocType::LOCAL) {
442 auto meta = allocs_metadata.find(ptr);
443 if (meta == allocs_metadata.end()) {
447 shared_data_t* data = &meta->second.data->second;
448 if (munmap(meta->second.allocated_ptr, meta->second.allocated_size) < 0) {
449 XBT_WARN("Unmapping of fd %d failed: %s", data->fd, strerror(errno));
452 if (data->count <= 0) {
454 allocs.erase(allocs.find(meta->second.data->first));
455 allocs_metadata.erase(meta);
456 XBT_DEBUG("Shared free - Local - with removal - of %p", ptr);
458 XBT_DEBUG("Shared free - Local - no removal - of %p, count = %d", ptr, data->count);
461 } else if (smpi_cfg_shared_malloc() == SharedMallocType::GLOBAL) {
462 auto meta = allocs_metadata.find(ptr);
463 if (meta != allocs_metadata.end()){
464 meta->second.data->second.count--;
465 XBT_DEBUG("Shared free - Global - of %p", ptr);
466 munmap(ptr, meta->second.size);
467 if(meta->second.data->second.count==0){
468 delete meta->second.data;
469 allocs_metadata.erase(meta);
477 XBT_DEBUG("Classic deallocation of %p", ptr);
483 int smpi_shared_known_call(const char* func, const char* input)
485 std::string loc = std::string(func) + ":" + input;
486 return calls.find(loc) != calls.end();
489 void* smpi_shared_get_call(const char* func, const char* input) {
490 std::string loc = std::string(func) + ":" + input;
492 return calls.at(loc);
495 void* smpi_shared_set_call(const char* func, const char* input, void* data) {
496 std::string loc = std::string(func) + ":" + input;