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 {
71 smpi_source_location(const char* filename, int line)
72 : filename(xbt_strdup(filename)), filename_length(strlen(filename)), line(line)
76 /** Pointer to a static string containing the file name */
77 char* filename = nullptr;
78 int filename_length = 0;
81 bool operator==(smpi_source_location const& that) const
83 return filename_length == that.filename_length && line == that.line &&
84 std::memcmp(filename, that.filename, filename_length) == 0;
86 bool operator!=(smpi_source_location const& that) const { return not(*this == that); }
92 template <> class hash<smpi_source_location> {
94 typedef smpi_source_location argument_type;
95 typedef std::size_t result_type;
96 result_type operator()(smpi_source_location const& loc) const
98 return xbt_str_hash_ext(loc.filename, loc.filename_length) ^
99 xbt_str_hash_ext((const char*)&loc.line, sizeof(loc.line));
111 std::unordered_map<smpi_source_location, shared_data_t> allocs;
112 typedef std::unordered_map<smpi_source_location, shared_data_t>::value_type shared_data_key_type;
116 size_t allocated_size;
118 std::vector<std::pair<size_t, size_t>> private_blocks;
119 shared_data_key_type* data;
122 std::map<void*, shared_metadata_t> allocs_metadata;
123 std::map<std::string, void*> calls;
126 static int smpi_shared_malloc_bogusfile = -1;
127 static int smpi_shared_malloc_bogusfile_huge_page = -1;
128 static unsigned long smpi_shared_malloc_blocksize = 1UL << 20;
133 void smpi_shared_destroy()
136 allocs_metadata.clear();
140 static size_t shm_size(int fd) {
143 if(fstat(fd, &st) < 0) {
144 xbt_die("Could not stat fd %d: %s", fd, strerror(errno));
146 return static_cast<size_t>(st.st_size);
150 static void* shm_map(int fd, size_t size, shared_data_key_type* data) {
151 char loc[PTR_STRLEN];
152 shared_metadata_t meta;
154 if(size > shm_size(fd) && (ftruncate(fd, static_cast<off_t>(size)) < 0)) {
155 xbt_die("Could not truncate fd %d to %zu: %s", fd, size, strerror(errno));
158 void* mem = mmap(nullptr, size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
159 if(mem == MAP_FAILED) {
161 "Failed to map fd %d with size %zu: %s\n"
162 "If you are running a lot of ranks, you may be exceeding the amount of mappings allowed per process.\n"
163 "On Linux systems, change this value with sudo sysctl -w vm.max_map_count=newvalue (default value: 65536)\n"
164 "Please see http://simgrid.gforge.inria.fr/simgrid/latest/doc/html/options.html#options_virt for more info.",
165 fd, size, strerror(errno));
167 snprintf(loc, PTR_STRLEN, "%p", mem);
170 allocs_metadata[mem] = meta;
171 XBT_DEBUG("MMAP %zu to %p", size, mem);
175 static void *smpi_shared_malloc_local(size_t size, const char *file, int line)
178 smpi_source_location loc(file, line);
179 auto res = allocs.insert(std::make_pair(loc, shared_data_t()));
180 auto data = res.first;
182 // The insertion did not take place.
183 // Generate a shared memory name from the address of the shared_data:
184 char shmname[32]; // cannot be longer than PSHMNAMLEN = 31 on Mac OS X (shm_open raises ENAMETOOLONG otherwise)
185 snprintf(shmname, 31, "/shmalloc%p", &*data);
186 int fd = shm_open(shmname, O_RDWR | O_CREAT | O_EXCL, S_IRUSR | S_IWUSR | S_IRGRP | S_IROTH);
189 xbt_die("Please cleanup /dev/shm/%s", shmname);
191 xbt_die("An unhandled error occurred while opening %s. shm_open: %s", shmname, strerror(errno));
193 data->second.fd = fd;
194 data->second.count = 1;
195 mem = shm_map(fd, size, &*data);
196 if (shm_unlink(shmname) < 0) {
197 XBT_WARN("Could not early unlink %s. shm_unlink: %s", shmname, strerror(errno));
199 XBT_DEBUG("Mapping %s at %p through %d", shmname, mem, fd);
201 mem = shm_map(data->second.fd, size, &*data);
202 data->second.count++;
204 XBT_DEBUG("Shared malloc %zu in %p (metadata at %p)", size, mem, &*data);
208 // Align functions, from http://stackoverflow.com/questions/4840410/how-to-align-a-pointer-in-c
209 #define PAGE_SIZE 0x1000
210 #define ALIGN_UP(n, align) (((n) + (align)-1) & -(align))
211 #define ALIGN_DOWN(n, align) ((n) & -(align))
213 #define HUGE_PAGE_SIZE 1<<21
215 /* Similar to smpi_shared_malloc, but only sharing the blocks described by shared_block_offsets.
216 * This array contains the offsets (in bytes) of the block to share.
217 * Even indices are the start offsets (included), odd indices are the stop offsets (excluded).
218 * For instance, if shared_block_offsets == {27, 42}, then the elements mem[27], mem[28], ..., mem[41] are shared.
219 * The others are not.
222 void* smpi_shared_malloc_partial(size_t size, size_t* shared_block_offsets, int nb_shared_blocks)
224 char *huge_page_mount_point = xbt_cfg_get_string("smpi/shared-malloc-hugepage");
225 bool use_huge_page = huge_page_mount_point[0] != '\0';
226 #ifndef MAP_HUGETLB /* If the system header don't define that mmap flag */
227 xbt_assert(not use_huge_page,
228 "Huge pages are not available on your system, you cannot use the smpi/shared-malloc-hugepage option.");
231 smpi_shared_malloc_blocksize = static_cast<unsigned long>(xbt_cfg_get_double("smpi/shared-malloc-blocksize"));
233 size_t allocated_size;
235 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.");
236 allocated_size = size + 2*smpi_shared_malloc_blocksize;
239 xbt_assert(smpi_shared_malloc_blocksize % PAGE_SIZE == 0, "the block size of shared malloc should be a multiple of the page size.");
240 allocated_size = size;
244 /* First reserve memory area */
245 void* allocated_ptr = mmap(NULL, allocated_size, PROT_READ | PROT_WRITE, MAP_ANONYMOUS | MAP_PRIVATE, -1, 0);
247 xbt_assert(allocated_ptr != MAP_FAILED, "Failed to allocate %zuMiB of memory. Run \"sysctl vm.overcommit_memory=1\" as root "
248 "to allow big allocations.\n",
251 mem = (void*)ALIGN_UP((uint64_t)allocated_ptr, HUGE_PAGE_SIZE);
255 XBT_DEBUG("global shared allocation. Blocksize %lu", smpi_shared_malloc_blocksize);
256 /* Create a fd to a new file on disk, make it smpi_shared_malloc_blocksize big, and unlink it.
257 * It still exists in memory but not in the file system (thus it cannot be leaked). */
258 /* Create bogus file if not done already
259 * We need two different bogusfiles:
260 * smpi_shared_malloc_bogusfile_huge_page is used for calls to mmap *with* MAP_HUGETLB,
261 * smpi_shared_malloc_bogusfile is used for calls to mmap *without* MAP_HUGETLB.
262 * We cannot use a same file for the two type of calls, since the first one needs to be
263 * opened in a hugetlbfs mount point whereas the second needs to be a "classical" file. */
264 if(use_huge_page && smpi_shared_malloc_bogusfile_huge_page == -1) {
265 const char *const array[] = {huge_page_mount_point, "simgrid-shmalloc-XXXXXX", nullptr};
266 char *huge_page_filename = xbt_str_join_array(array, "/");
267 smpi_shared_malloc_bogusfile_huge_page = mkstemp(huge_page_filename);
268 XBT_DEBUG("bogusfile_huge_page: %s\n", huge_page_filename);
269 unlink(huge_page_filename);
270 xbt_free(huge_page_filename);
272 if(smpi_shared_malloc_bogusfile == -1) {
273 char *name = xbt_strdup("/tmp/simgrid-shmalloc-XXXXXX");
274 smpi_shared_malloc_bogusfile = mkstemp(name);
275 XBT_DEBUG("bogusfile : %s\n", name);
278 char* dumb = (char*)calloc(1, smpi_shared_malloc_blocksize);
279 ssize_t err = write(smpi_shared_malloc_bogusfile, dumb, smpi_shared_malloc_blocksize);
281 xbt_die("Could not write bogus file for shared malloc");
285 int mmap_base_flag = MAP_FIXED | MAP_SHARED | MAP_POPULATE;
286 int mmap_flag = mmap_base_flag;
287 int huge_fd = use_huge_page ? smpi_shared_malloc_bogusfile_huge_page : smpi_shared_malloc_bogusfile;
290 mmap_flag |= MAP_HUGETLB;
293 XBT_DEBUG("global shared allocation, begin mmap");
295 /* Map the bogus file in place of the anonymous memory */
296 for(int i_block = 0; i_block < nb_shared_blocks; i_block ++) {
297 XBT_DEBUG("\tglobal shared allocation, mmap block %d/%d", i_block+1, nb_shared_blocks);
298 size_t start_offset = shared_block_offsets[2*i_block];
299 size_t stop_offset = shared_block_offsets[2*i_block+1];
300 xbt_assert(start_offset < stop_offset, "start_offset (%zu) should be lower than stop offset (%zu)", start_offset, stop_offset);
301 xbt_assert(stop_offset <= size, "stop_offset (%zu) should be lower than size (%zu)", stop_offset, size);
302 if(i_block < nb_shared_blocks-1)
303 xbt_assert(stop_offset < shared_block_offsets[2*i_block+2],
304 "stop_offset (%zu) should be lower than its successor start offset (%zu)", stop_offset, shared_block_offsets[2*i_block+2]);
305 size_t start_block_offset = ALIGN_UP(start_offset, smpi_shared_malloc_blocksize);
306 size_t stop_block_offset = ALIGN_DOWN(stop_offset, smpi_shared_malloc_blocksize);
307 for (unsigned block_id=0, i = start_block_offset / smpi_shared_malloc_blocksize; i < stop_block_offset / smpi_shared_malloc_blocksize; block_id++, i++) {
308 XBT_DEBUG("\t\tglobal shared allocation, mmap block offset %u", block_id);
309 void* pos = (void*)((unsigned long)mem + i * smpi_shared_malloc_blocksize);
310 void* res = mmap(pos, smpi_shared_malloc_blocksize, PROT_READ | PROT_WRITE, mmap_flag,
312 xbt_assert(res == pos, "Could not map folded virtual memory (%s). Do you perhaps need to increase the "
313 "size of the mapped file using --cfg=smpi/shared-malloc-blocksize=newvalue (default 1048576) ? "
314 "You can also try using the sysctl vm.max_map_count. "
315 "If you are using huge pages, check that you have at least one huge page (/proc/sys/vm/nr_hugepages) "
316 "and that the directory you are passing is mounted correctly (mount /path/to/huge -t hugetlbfs -o rw,mode=0777).",
319 size_t low_page_start_offset = ALIGN_UP(start_offset, PAGE_SIZE);
320 size_t low_page_stop_offset = start_block_offset < ALIGN_DOWN(stop_offset, PAGE_SIZE) ? start_block_offset : ALIGN_DOWN(stop_offset, PAGE_SIZE);
321 if(low_page_start_offset < low_page_stop_offset) {
322 XBT_DEBUG("\t\tglobal shared allocation, mmap block start");
323 void* pos = (void*)((unsigned long)mem + low_page_start_offset);
324 void* res = mmap(pos, low_page_stop_offset-low_page_start_offset, PROT_READ | PROT_WRITE, mmap_base_flag, // not a full huge page
325 smpi_shared_malloc_bogusfile, 0);
326 xbt_assert(res == pos, "Could not map folded virtual memory (%s). Do you perhaps need to increase the "
327 "size of the mapped file using --cfg=smpi/shared-malloc-blocksize=newvalue (default 1048576) ?"
328 "You can also try using the sysctl vm.max_map_count",
331 if(low_page_stop_offset <= stop_block_offset) {
332 XBT_DEBUG("\t\tglobal shared allocation, mmap block stop");
333 size_t high_page_stop_offset = stop_offset == size ? size : ALIGN_DOWN(stop_offset, PAGE_SIZE);
334 if(high_page_stop_offset > stop_block_offset) {
335 void* pos = (void*)((unsigned long)mem + stop_block_offset);
336 void* res = mmap(pos, high_page_stop_offset-stop_block_offset, PROT_READ | PROT_WRITE, mmap_base_flag, // not a full huge page
337 smpi_shared_malloc_bogusfile, 0);
338 xbt_assert(res == pos, "Could not map folded virtual memory (%s). Do you perhaps need to increase the "
339 "size of the mapped file using --cfg=smpi/shared-malloc-blocksize=newvalue (default 1048576) ?"
340 "You can also try using the sysctl vm.max_map_count",
346 shared_metadata_t newmeta;
347 //register metadata for memcpy avoidance
348 shared_data_key_type* data = (shared_data_key_type*)xbt_malloc(sizeof(shared_data_key_type));
349 data->second.fd = -1;
350 data->second.count = 1;
353 newmeta.allocated_ptr = allocated_ptr;
354 newmeta.allocated_size = allocated_size;
355 if(shared_block_offsets[0] > 0) {
356 newmeta.private_blocks.push_back(std::make_pair(0, shared_block_offsets[0]));
359 for(i_block = 0; i_block < nb_shared_blocks-1; i_block ++) {
360 newmeta.private_blocks.push_back(std::make_pair(shared_block_offsets[2*i_block+1], shared_block_offsets[2*i_block+2]));
362 if(shared_block_offsets[2*i_block+1] < size) {
363 newmeta.private_blocks.push_back(std::make_pair(shared_block_offsets[2*i_block+1], size));
365 allocs_metadata[mem] = newmeta;
367 XBT_DEBUG("global shared allocation, allocated_ptr %p - %p", allocated_ptr, (void*)(((uint64_t)allocated_ptr)+allocated_size));
368 XBT_DEBUG("global shared allocation, returned_ptr %p - %p", mem, (void*)(((uint64_t)mem)+size));
373 void *smpi_shared_malloc(size_t size, const char *file, int line) {
374 if (size > 0 && smpi_cfg_shared_malloc == shmalloc_local) {
375 return smpi_shared_malloc_local(size, file, line);
376 } else if (smpi_cfg_shared_malloc == shmalloc_global) {
377 int nb_shared_blocks = 1;
378 size_t shared_block_offsets[2] = {0, size};
379 return smpi_shared_malloc_partial(size, shared_block_offsets, nb_shared_blocks);
381 XBT_DEBUG("Classic malloc %zu", size);
382 return xbt_malloc(size);
385 int smpi_is_shared(void* ptr, std::vector<std::pair<size_t, size_t>> &private_blocks, size_t *offset){
386 private_blocks.clear(); // being paranoid
387 if (allocs_metadata.empty())
389 if ( smpi_cfg_shared_malloc == shmalloc_local || smpi_cfg_shared_malloc == shmalloc_global) {
390 auto low = allocs_metadata.lower_bound(ptr);
391 if (low != allocs_metadata.end() && low->first == ptr) {
392 private_blocks = low->second.private_blocks;
396 if (low == allocs_metadata.begin())
399 if (ptr < (char*)low->first + low->second.size) {
400 xbt_assert(ptr > (char*)low->first, "Oops, there seems to be a bug in the shared memory metadata.");
401 *offset = ((uint8_t*)ptr) - ((uint8_t*) low->first);
402 private_blocks = low->second.private_blocks;
411 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) {
412 std::vector<std::pair<size_t, size_t>> result;
413 for (auto const& block : vec) {
414 auto new_block = std::make_pair(std::min(std::max((size_t)0, block.first-offset), buff_size),
415 std::min(std::max((size_t)0, block.second-offset), buff_size));
416 if(new_block.second > 0 && new_block.first < buff_size)
417 result.push_back(new_block);
422 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) {
423 std::vector<std::pair<size_t, size_t>> result;
426 while(i_src < src.size() && i_dst < dst.size()) {
427 std::pair<size_t, size_t> block;
428 if(src[i_src].second <= dst[i_dst].first) {
431 else if(dst[i_dst].second <= src[i_src].first) {
434 else { // src.second > dst.first && dst.second > src.first → the blocks are overlapping
435 block = std::make_pair(std::max(src[i_src].first, dst[i_dst].first),
436 std::min(src[i_src].second, dst[i_dst].second));
437 result.push_back(block);
438 if(src[i_src].second < dst[i_dst].second)
447 void smpi_shared_free(void *ptr)
449 if (smpi_cfg_shared_malloc == shmalloc_local) {
450 char loc[PTR_STRLEN];
451 snprintf(loc, PTR_STRLEN, "%p", ptr);
452 auto meta = allocs_metadata.find(ptr);
453 if (meta == allocs_metadata.end()) {
454 XBT_WARN("Cannot free: %p was not shared-allocated by SMPI - maybe its size was 0?", ptr);
457 shared_data_t* data = &meta->second.data->second;
458 if (munmap(meta->second.allocated_ptr, meta->second.allocated_size) < 0) {
459 XBT_WARN("Unmapping of fd %d failed: %s", data->fd, strerror(errno));
462 if (data->count <= 0) {
464 allocs.erase(allocs.find(meta->second.data->first));
465 allocs_metadata.erase(ptr);
466 XBT_DEBUG("Shared free - with removal - of %p", ptr);
468 XBT_DEBUG("Shared free - no removal - of %p, count = %d", ptr, data->count);
471 } else if (smpi_cfg_shared_malloc == shmalloc_global) {
472 auto meta = allocs_metadata.find(ptr);
473 if (meta != allocs_metadata.end()){
474 meta->second.data->second.count--;
475 if(meta->second.data->second.count==0)
476 xbt_free(meta->second.data);
479 munmap(ptr, meta->second.size);
481 XBT_DEBUG("Classic free of %p", ptr);
487 int smpi_shared_known_call(const char* func, const char* input)
489 std::string loc = std::string(func) + ":" + input;
490 return calls.find(loc) != calls.end();
493 void* smpi_shared_get_call(const char* func, const char* input) {
494 std::string loc = std::string(func) + ":" + input;
496 return calls.at(loc);
499 void* smpi_shared_set_call(const char* func, const char* input, void* data) {
500 std::string loc = std::string(func) + ":" + input;