1 /* Copyright (c) 2007-2018. 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 * ----------------- | | |
39 #include "private.hpp"
40 #include "xbt/config.hpp"
44 #include <sys/types.h>
53 #define MAP_ANONYMOUS MAP_ANON
57 #define MAP_POPULATE 0
60 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_shared, smpi, "Logging specific to SMPI (shared memory macros)");
62 #define PTR_STRLEN (2 + 2 * sizeof(void*) + 1)
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 typedef decltype(allocs)::value_type shared_data_key_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<void*, shared_metadata_t> allocs_metadata;
95 std::map<std::string, void*> calls;
98 static int smpi_shared_malloc_bogusfile = -1;
99 static int smpi_shared_malloc_bogusfile_huge_page = -1;
100 static unsigned long smpi_shared_malloc_blocksize = 1UL << 20;
105 void smpi_shared_destroy()
108 allocs_metadata.clear();
112 static size_t shm_size(int fd) {
115 if(fstat(fd, &st) < 0) {
116 xbt_die("Could not stat fd %d: %s", fd, strerror(errno));
118 return static_cast<size_t>(st.st_size);
122 static void* shm_map(int fd, size_t size, shared_data_key_type* data) {
123 char loc[PTR_STRLEN];
124 shared_metadata_t meta;
126 if(size > shm_size(fd) && (ftruncate(fd, static_cast<off_t>(size)) < 0)) {
127 xbt_die("Could not truncate fd %d to %zu: %s", fd, size, strerror(errno));
130 void* mem = mmap(nullptr, size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
131 if(mem == MAP_FAILED) {
133 "Failed to map fd %d with size %zu: %s\n"
134 "If you are running a lot of ranks, you may be exceeding the amount of mappings allowed per process.\n"
135 "On Linux systems, change this value with sudo sysctl -w vm.max_map_count=newvalue (default value: 65536)\n"
136 "Please see http://simgrid.gforge.inria.fr/simgrid/latest/doc/html/options.html#options_virt for more info.",
137 fd, size, strerror(errno));
139 snprintf(loc, PTR_STRLEN, "%p", mem);
142 allocs_metadata[mem] = meta;
143 XBT_DEBUG("MMAP %zu to %p", size, mem);
147 static void *smpi_shared_malloc_local(size_t size, const char *file, int line)
150 smpi_source_location loc(file, line);
151 auto res = allocs.insert(std::make_pair(loc, shared_data_t()));
152 auto data = res.first;
154 // The new element was inserted.
155 // Generate a shared memory name from the address of the shared_data:
156 char shmname[32]; // cannot be longer than PSHMNAMLEN = 31 on Mac OS X (shm_open raises ENAMETOOLONG otherwise)
157 snprintf(shmname, 31, "/shmalloc%p", &*data);
158 int fd = shm_open(shmname, O_RDWR | O_CREAT | O_EXCL, S_IRUSR | S_IWUSR | S_IRGRP | S_IROTH);
161 xbt_die("Please cleanup /dev/shm/%s", shmname);
163 xbt_die("An unhandled error occurred while opening %s. shm_open: %s", shmname, strerror(errno));
165 data->second.fd = fd;
166 data->second.count = 1;
167 mem = shm_map(fd, size, &*data);
168 if (shm_unlink(shmname) < 0) {
169 XBT_WARN("Could not early unlink %s. shm_unlink: %s", shmname, strerror(errno));
171 XBT_DEBUG("Mapping %s at %p through %d", shmname, mem, fd);
173 mem = shm_map(data->second.fd, size, &*data);
174 data->second.count++;
176 XBT_DEBUG("Shared malloc %zu in %p (metadata at %p)", size, mem, &*data);
180 // Align functions, from http://stackoverflow.com/questions/4840410/how-to-align-a-pointer-in-c
181 #define PAGE_SIZE 0x1000
182 #define ALIGN_UP(n, align) (((n) + (align)-1) & -(align))
183 #define ALIGN_DOWN(n, align) ((n) & -(align))
185 #define HUGE_PAGE_SIZE 1<<21
187 /* Similar to smpi_shared_malloc, but only sharing the blocks described by shared_block_offsets.
188 * This array contains the offsets (in bytes) of the block to share.
189 * Even indices are the start offsets (included), odd indices are the stop offsets (excluded).
190 * For instance, if shared_block_offsets == {27, 42}, then the elements mem[27], mem[28], ..., mem[41] are shared.
191 * The others are not.
194 void* smpi_shared_malloc_partial(size_t size, size_t* shared_block_offsets, int nb_shared_blocks)
196 std::string huge_page_mount_point = simgrid::config::get_value<std::string>("smpi/shared-malloc-hugepage");
197 bool use_huge_page = not huge_page_mount_point.empty();
198 #ifndef MAP_HUGETLB /* If the system header don't define that mmap flag */
199 xbt_assert(not use_huge_page,
200 "Huge pages are not available on your system, you cannot use the smpi/shared-malloc-hugepage option.");
202 smpi_shared_malloc_blocksize =
203 static_cast<unsigned long>(simgrid::config::get_value<double>("smpi/shared-malloc-blocksize"));
205 size_t allocated_size;
207 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.");
208 allocated_size = size + 2*smpi_shared_malloc_blocksize;
211 xbt_assert(smpi_shared_malloc_blocksize % PAGE_SIZE == 0, "the block size of shared malloc should be a multiple of the page size.");
212 allocated_size = size;
216 /* First reserve memory area */
217 void* allocated_ptr = mmap(NULL, allocated_size, PROT_READ | PROT_WRITE, MAP_ANONYMOUS | MAP_PRIVATE, -1, 0);
219 xbt_assert(allocated_ptr != MAP_FAILED, "Failed to allocate %zuMiB of memory. Run \"sysctl vm.overcommit_memory=1\" as root "
220 "to allow big allocations.\n",
223 mem = (void*)ALIGN_UP((int64_t)allocated_ptr, HUGE_PAGE_SIZE);
227 XBT_DEBUG("global shared allocation. Blocksize %lu", smpi_shared_malloc_blocksize);
228 /* Create a fd to a new file on disk, make it smpi_shared_malloc_blocksize big, and unlink it.
229 * It still exists in memory but not in the file system (thus it cannot be leaked). */
230 /* Create bogus file if not done already
231 * We need two different bogusfiles:
232 * smpi_shared_malloc_bogusfile_huge_page is used for calls to mmap *with* MAP_HUGETLB,
233 * smpi_shared_malloc_bogusfile is used for calls to mmap *without* MAP_HUGETLB.
234 * We cannot use a same file for the two type of calls, since the first one needs to be
235 * opened in a hugetlbfs mount point whereas the second needs to be a "classical" file. */
236 if(use_huge_page && smpi_shared_malloc_bogusfile_huge_page == -1) {
237 std::string huge_page_filename = huge_page_mount_point + "/simgrid-shmalloc-XXXXXX";
238 smpi_shared_malloc_bogusfile_huge_page = mkstemp((char*)huge_page_filename.c_str());
239 XBT_DEBUG("bogusfile_huge_page: %s\n", huge_page_filename.c_str());
240 unlink(huge_page_filename.c_str());
242 if(smpi_shared_malloc_bogusfile == -1) {
243 char name[] = "/tmp/simgrid-shmalloc-XXXXXX";
244 smpi_shared_malloc_bogusfile = mkstemp(name);
245 XBT_DEBUG("bogusfile : %s\n", name);
247 char* dumb = new char[smpi_shared_malloc_blocksize](); // zero initialized
248 ssize_t err = write(smpi_shared_malloc_bogusfile, dumb, smpi_shared_malloc_blocksize);
250 xbt_die("Could not write bogus file for shared malloc");
254 int mmap_base_flag = MAP_FIXED | MAP_SHARED | MAP_POPULATE;
255 int mmap_flag = mmap_base_flag;
256 int huge_fd = use_huge_page ? smpi_shared_malloc_bogusfile_huge_page : smpi_shared_malloc_bogusfile;
259 mmap_flag |= MAP_HUGETLB;
262 XBT_DEBUG("global shared allocation, begin mmap");
264 /* Map the bogus file in place of the anonymous memory */
265 for(int i_block = 0; i_block < nb_shared_blocks; i_block ++) {
266 XBT_DEBUG("\tglobal shared allocation, mmap block %d/%d", i_block+1, nb_shared_blocks);
267 size_t start_offset = shared_block_offsets[2*i_block];
268 size_t stop_offset = shared_block_offsets[2*i_block+1];
269 xbt_assert(start_offset < stop_offset, "start_offset (%zu) should be lower than stop offset (%zu)", start_offset, stop_offset);
270 xbt_assert(stop_offset <= size, "stop_offset (%zu) should be lower than size (%zu)", stop_offset, size);
271 if(i_block < nb_shared_blocks-1)
272 xbt_assert(stop_offset < shared_block_offsets[2*i_block+2],
273 "stop_offset (%zu) should be lower than its successor start offset (%zu)", stop_offset, shared_block_offsets[2*i_block+2]);
274 size_t start_block_offset = ALIGN_UP((int64_t)start_offset, smpi_shared_malloc_blocksize);
275 size_t stop_block_offset = ALIGN_DOWN((int64_t)stop_offset, smpi_shared_malloc_blocksize);
276 for (size_t offset = start_block_offset; offset < stop_block_offset; offset += smpi_shared_malloc_blocksize) {
277 XBT_DEBUG("\t\tglobal shared allocation, mmap block offset %zx", offset);
278 void* pos = (void*)((unsigned long)mem + offset);
279 void* res = mmap(pos, smpi_shared_malloc_blocksize, PROT_READ | PROT_WRITE, mmap_flag,
281 xbt_assert(res == pos, "Could not map folded virtual memory (%s). Do you perhaps need to increase the "
282 "size of the mapped file using --cfg=smpi/shared-malloc-blocksize=newvalue (default 1048576) ? "
283 "You can also try using the sysctl vm.max_map_count. "
284 "If you are using huge pages, check that you have at least one huge page (/proc/sys/vm/nr_hugepages) "
285 "and that the directory you are passing is mounted correctly (mount /path/to/huge -t hugetlbfs -o rw,mode=0777).",
288 size_t low_page_start_offset = ALIGN_UP((int64_t)start_offset, PAGE_SIZE);
289 size_t low_page_stop_offset = (int64_t)start_block_offset < ALIGN_DOWN((int64_t)stop_offset, PAGE_SIZE) ? start_block_offset : ALIGN_DOWN((int64_t)stop_offset, (int64_t)PAGE_SIZE);
290 if(low_page_start_offset < low_page_stop_offset) {
291 XBT_DEBUG("\t\tglobal shared allocation, mmap block start");
292 void* pos = (void*)((unsigned long)mem + low_page_start_offset);
293 void* res = mmap(pos, low_page_stop_offset-low_page_start_offset, PROT_READ | PROT_WRITE, mmap_base_flag, // not a full huge page
294 smpi_shared_malloc_bogusfile, 0);
295 xbt_assert(res == pos, "Could not map folded virtual memory (%s). Do you perhaps need to increase the "
296 "size of the mapped file using --cfg=smpi/shared-malloc-blocksize=newvalue (default 1048576) ?"
297 "You can also try using the sysctl vm.max_map_count",
300 if(low_page_stop_offset <= stop_block_offset) {
301 XBT_DEBUG("\t\tglobal shared allocation, mmap block stop");
302 size_t high_page_stop_offset = stop_offset == size ? size : ALIGN_DOWN((int64_t)stop_offset, PAGE_SIZE);
303 if(high_page_stop_offset > stop_block_offset) {
304 void* pos = (void*)((unsigned long)mem + stop_block_offset);
305 void* res = mmap(pos, high_page_stop_offset-stop_block_offset, PROT_READ | PROT_WRITE, mmap_base_flag, // not a full huge page
306 smpi_shared_malloc_bogusfile, 0);
307 xbt_assert(res == pos, "Could not map folded virtual memory (%s). Do you perhaps need to increase the "
308 "size of the mapped file using --cfg=smpi/shared-malloc-blocksize=newvalue (default 1048576) ?"
309 "You can also try using the sysctl vm.max_map_count",
315 shared_metadata_t newmeta;
316 //register metadata for memcpy avoidance
317 shared_data_key_type* data = new shared_data_key_type;
318 data->second.fd = -1;
319 data->second.count = 1;
322 newmeta.allocated_ptr = allocated_ptr;
323 newmeta.allocated_size = allocated_size;
324 if(shared_block_offsets[0] > 0) {
325 newmeta.private_blocks.push_back(std::make_pair(0, shared_block_offsets[0]));
328 for(i_block = 0; i_block < nb_shared_blocks-1; i_block ++) {
329 newmeta.private_blocks.push_back(std::make_pair(shared_block_offsets[2*i_block+1], shared_block_offsets[2*i_block+2]));
331 if(shared_block_offsets[2*i_block+1] < size) {
332 newmeta.private_blocks.push_back(std::make_pair(shared_block_offsets[2*i_block+1], size));
334 allocs_metadata[mem] = newmeta;
336 XBT_DEBUG("global shared allocation, allocated_ptr %p - %p", allocated_ptr, (void*)(((uint64_t)allocated_ptr)+allocated_size));
337 XBT_DEBUG("global shared allocation, returned_ptr %p - %p", mem, (void*)(((uint64_t)mem)+size));
342 void *smpi_shared_malloc(size_t size, const char *file, int line) {
343 if (size > 0 && smpi_cfg_shared_malloc == SharedMallocType::LOCAL) {
344 return smpi_shared_malloc_local(size, file, line);
345 } else if (smpi_cfg_shared_malloc == SharedMallocType::GLOBAL) {
346 int nb_shared_blocks = 1;
347 size_t shared_block_offsets[2] = {0, size};
348 return smpi_shared_malloc_partial(size, shared_block_offsets, nb_shared_blocks);
350 XBT_DEBUG("Classic allocation of %zu bytes", size);
351 return ::operator new(size);
354 int smpi_is_shared(void* ptr, std::vector<std::pair<size_t, size_t>> &private_blocks, size_t *offset){
355 private_blocks.clear(); // being paranoid
356 if (allocs_metadata.empty())
358 if (smpi_cfg_shared_malloc == SharedMallocType::LOCAL || smpi_cfg_shared_malloc == SharedMallocType::GLOBAL) {
359 auto low = allocs_metadata.lower_bound(ptr);
360 if (low != allocs_metadata.end() && low->first == ptr) {
361 private_blocks = low->second.private_blocks;
365 if (low == allocs_metadata.begin())
368 if (ptr < (char*)low->first + low->second.size) {
369 xbt_assert(ptr > (char*)low->first, "Oops, there seems to be a bug in the shared memory metadata.");
370 *offset = ((uint8_t*)ptr) - ((uint8_t*) low->first);
371 private_blocks = low->second.private_blocks;
380 std::vector<std::pair<size_t, size_t>> shift_and_frame_private_blocks(const std::vector<std::pair<size_t, size_t>>& vec,
381 size_t offset, size_t buff_size)
383 std::vector<std::pair<size_t, size_t>> result;
384 for (auto const& block : vec) {
385 auto new_block = std::make_pair(std::min(std::max((size_t)0, block.first - offset), buff_size),
386 std::min(std::max((size_t)0, block.second - offset), buff_size));
387 if (new_block.second > 0 && new_block.first < buff_size)
388 result.push_back(new_block);
393 std::vector<std::pair<size_t, size_t>> merge_private_blocks(const std::vector<std::pair<size_t, size_t>>& src,
394 const std::vector<std::pair<size_t, size_t>>& dst)
396 std::vector<std::pair<size_t, size_t>> result;
399 while(i_src < src.size() && i_dst < dst.size()) {
400 std::pair<size_t, size_t> block;
401 if(src[i_src].second <= dst[i_dst].first) {
404 else if(dst[i_dst].second <= src[i_src].first) {
407 else { // src.second > dst.first && dst.second > src.first → the blocks are overlapping
408 block = std::make_pair(std::max(src[i_src].first, dst[i_dst].first),
409 std::min(src[i_src].second, dst[i_dst].second));
410 result.push_back(block);
411 if(src[i_src].second < dst[i_dst].second)
420 void smpi_shared_free(void *ptr)
422 if (smpi_cfg_shared_malloc == SharedMallocType::LOCAL) {
423 char loc[PTR_STRLEN];
424 snprintf(loc, PTR_STRLEN, "%p", ptr);
425 auto meta = allocs_metadata.find(ptr);
426 if (meta == allocs_metadata.end()) {
427 XBT_WARN("Cannot free: %p was not shared-allocated by SMPI - maybe its size was 0?", ptr);
430 shared_data_t* data = &meta->second.data->second;
431 if (munmap(meta->second.allocated_ptr, meta->second.allocated_size) < 0) {
432 XBT_WARN("Unmapping of fd %d failed: %s", data->fd, strerror(errno));
435 if (data->count <= 0) {
437 allocs.erase(allocs.find(meta->second.data->first));
438 allocs_metadata.erase(ptr);
439 XBT_DEBUG("Shared free - with removal - of %p", ptr);
441 XBT_DEBUG("Shared free - no removal - of %p, count = %d", ptr, data->count);
444 } else if (smpi_cfg_shared_malloc == SharedMallocType::GLOBAL) {
445 auto meta = allocs_metadata.find(ptr);
446 if (meta != allocs_metadata.end()){
447 meta->second.data->second.count--;
448 if(meta->second.data->second.count==0)
449 delete meta->second.data;
452 munmap(ptr, meta->second.size);
454 XBT_DEBUG("Classic deallocation of %p", ptr);
455 ::operator delete(ptr);
460 int smpi_shared_known_call(const char* func, const char* input)
462 std::string loc = std::string(func) + ":" + input;
463 return calls.find(loc) != calls.end();
466 void* smpi_shared_get_call(const char* func, const char* input) {
467 std::string loc = std::string(func) + ":" + input;
469 return calls.at(loc);
472 void* smpi_shared_set_call(const char* func, const char* input, void* data) {
473 std::string loc = std::string(func) + ":" + input;