X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/76d5ed561fdd3411adbed2c3b659a1312e138284..bf71c3c42ec41053cf17711276bfd40d9b3a93c1:/src/smpi/smpi_shared.cpp diff --git a/src/smpi/smpi_shared.cpp b/src/smpi/smpi_shared.cpp index a717d39a03..cd4b82093e 100644 --- a/src/smpi/smpi_shared.cpp +++ b/src/smpi/smpi_shared.cpp @@ -33,11 +33,13 @@ * \ | | * ---- */ -#include +#include +#include #include "private.h" #include "private.hpp" #include "xbt/dict.h" +#include "xbt/ex.hpp" #include #include @@ -84,7 +86,7 @@ public: return filename_length == that.filename_length && line == that.line && std::memcmp(filename, that.filename, filename_length) == 0; } - bool operator!=(smpi_source_location const& that) const { return !(*this == that); } + bool operator!=(smpi_source_location const& that) const { return not(*this == that); } }; } @@ -114,13 +116,17 @@ typedef std::unordered_map::value_type shar typedef struct { size_t size; + size_t allocated_size; + void *allocated_ptr; + std::vector> private_blocks; shared_data_key_type* data; } shared_metadata_t; -std::unordered_map allocs_metadata; +std::map allocs_metadata; xbt_dict_t calls = nullptr; /* Allocated on first use */ #ifndef WIN32 static int smpi_shared_malloc_bogusfile = -1; +static int smpi_shared_malloc_bogusfile_huge_page = -1; static unsigned long smpi_shared_malloc_blocksize = 1UL << 20; #endif } @@ -168,116 +174,278 @@ static void* shm_map(int fd, size_t size, shared_data_key_type* data) { return mem; } -void *smpi_shared_malloc(size_t size, const char *file, int line) +static void *smpi_shared_malloc_local(size_t size, const char *file, int line) { void* mem; - if (size > 0 && smpi_cfg_shared_malloc == shmalloc_local) { - smpi_source_location loc(file, line); - auto res = allocs.insert(std::make_pair(loc, shared_data_t())); - auto data = res.first; - if (res.second) { - // The insertion did not take place. - // Generate a shared memory name from the address of the shared_data: - char shmname[32]; // cannot be longer than PSHMNAMLEN = 31 on Mac OS X (shm_open raises ENAMETOOLONG otherwise) - snprintf(shmname, 31, "/shmalloc%p", &*data); - int fd = shm_open(shmname, O_RDWR | O_CREAT | O_EXCL, S_IRUSR | S_IWUSR | S_IRGRP | S_IROTH); - if (fd < 0) { - if (errno == EEXIST) - xbt_die("Please cleanup /dev/shm/%s", shmname); - else - xbt_die("An unhandled error occurred while opening %s. shm_open: %s", shmname, strerror(errno)); - } - data->second.fd = fd; - data->second.count = 1; - mem = shm_map(fd, size, &*data); - if (shm_unlink(shmname) < 0) { - XBT_WARN("Could not early unlink %s. shm_unlink: %s", shmname, strerror(errno)); - } - XBT_DEBUG("Mapping %s at %p through %d", shmname, mem, fd); - } else { - mem = shm_map(data->second.fd, size, &*data); - data->second.count++; + smpi_source_location loc(file, line); + auto res = allocs.insert(std::make_pair(loc, shared_data_t())); + auto data = res.first; + if (res.second) { + // The insertion did not take place. + // Generate a shared memory name from the address of the shared_data: + char shmname[32]; // cannot be longer than PSHMNAMLEN = 31 on Mac OS X (shm_open raises ENAMETOOLONG otherwise) + snprintf(shmname, 31, "/shmalloc%p", &*data); + int fd = shm_open(shmname, O_RDWR | O_CREAT | O_EXCL, S_IRUSR | S_IWUSR | S_IRGRP | S_IROTH); + if (fd < 0) { + if (errno == EEXIST) + xbt_die("Please cleanup /dev/shm/%s", shmname); + else + xbt_die("An unhandled error occurred while opening %s. shm_open: %s", shmname, strerror(errno)); } - XBT_DEBUG("Shared malloc %zu in %p (metadata at %p)", size, mem, &*data); - - } else if (smpi_cfg_shared_malloc == shmalloc_global) { - /* First reserve memory area */ - mem = mmap(NULL, size, PROT_READ | PROT_WRITE, MAP_ANONYMOUS | MAP_PRIVATE, -1, 0); - - xbt_assert(mem != MAP_FAILED, "Failed to allocate %luMiB of memory. Run \"sysctl vm.overcommit_memory=1\" as root " - "to allow big allocations.\n", - (unsigned long)(size >> 20)); - - /* Create bogus file if not done already */ - if (smpi_shared_malloc_bogusfile == -1) { - /* Create a fd to a new file on disk, make it smpi_shared_malloc_blocksize big, and unlink it. - * It still exists in memory but not in the file system (thus it cannot be leaked). */ - smpi_shared_malloc_blocksize = static_cast(xbt_cfg_get_double("smpi/shared-malloc-blocksize")); - XBT_DEBUG("global shared allocation. Blocksize %lu", smpi_shared_malloc_blocksize); - char* name = xbt_strdup("/tmp/simgrid-shmalloc-XXXXXX"); - smpi_shared_malloc_bogusfile = mkstemp(name); - unlink(name); - xbt_free(name); - char* dumb = (char*)calloc(1, smpi_shared_malloc_blocksize); - ssize_t err = write(smpi_shared_malloc_bogusfile, dumb, smpi_shared_malloc_blocksize); - if(err<0) - xbt_die("Could not write bogus file for shared malloc"); - xbt_free(dumb); + data->second.fd = fd; + data->second.count = 1; + mem = shm_map(fd, size, &*data); + if (shm_unlink(shmname) < 0) { + XBT_WARN("Could not early unlink %s. shm_unlink: %s", shmname, strerror(errno)); } + XBT_DEBUG("Mapping %s at %p through %d", shmname, mem, fd); + } else { + mem = shm_map(data->second.fd, size, &*data); + data->second.count++; + } + XBT_DEBUG("Shared malloc %zu in %p (metadata at %p)", size, mem, &*data); + return mem; +} + +// Align functions, from http://stackoverflow.com/questions/4840410/how-to-align-a-pointer-in-c +#define PAGE_SIZE 0x1000 +#define ALIGN_UP(n, align) (((n) + (align)-1) & -(align)) +#define ALIGN_DOWN(n, align) ((n) & -(align)) + +#define HUGE_PAGE_SIZE 1<<21 + +/* Similar to smpi_shared_malloc, but only sharing the blocks described by shared_block_offsets. + * This array contains the offsets (in bytes) of the block to share. + * Even indices are the start offsets (included), odd indices are the stop offsets (excluded). + * For instance, if shared_block_offsets == {27, 42}, then the elements mem[27], mem[28], ..., mem[41] are shared. + * The others are not. + */ + +void* smpi_shared_malloc_partial(size_t size, size_t* shared_block_offsets, int nb_shared_blocks) +{ + char *huge_page_mount_point = xbt_cfg_get_string("smpi/shared-malloc-hugepage"); + bool use_huge_page = huge_page_mount_point[0] != '\0'; +#ifndef MAP_HUGETLB /* If the system header don't define that mmap flag */ + xbt_assert(not use_huge_page, + "Huge pages are not available on your system, you cannot use the smpi/shared-malloc-hugepage option."); + use_huge_page = 0; +#endif + smpi_shared_malloc_blocksize = static_cast(xbt_cfg_get_double("smpi/shared-malloc-blocksize")); + void* mem; + size_t allocated_size; + if(use_huge_page) { + 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."); + allocated_size = size + 2*smpi_shared_malloc_blocksize; + } + else { + xbt_assert(smpi_shared_malloc_blocksize % PAGE_SIZE == 0, "the block size of shared malloc should be a multiple of the page size."); + allocated_size = size; + } - /* Map the bogus file in place of the anonymous memory */ - unsigned int i; - for (i = 0; i < size / smpi_shared_malloc_blocksize; i++) { + + /* First reserve memory area */ + void* allocated_ptr = mmap(NULL, allocated_size, PROT_READ | PROT_WRITE, MAP_ANONYMOUS | MAP_PRIVATE, -1, 0); + + xbt_assert(allocated_ptr != MAP_FAILED, "Failed to allocate %zuMiB of memory. Run \"sysctl vm.overcommit_memory=1\" as root " + "to allow big allocations.\n", + size >> 20); + if(use_huge_page) + mem = (void*)ALIGN_UP((uint64_t)allocated_ptr, HUGE_PAGE_SIZE); + else + mem = allocated_ptr; + + XBT_DEBUG("global shared allocation. Blocksize %lu", smpi_shared_malloc_blocksize); + /* Create a fd to a new file on disk, make it smpi_shared_malloc_blocksize big, and unlink it. + * It still exists in memory but not in the file system (thus it cannot be leaked). */ + /* Create bogus file if not done already + * We need two different bogusfiles: + * smpi_shared_malloc_bogusfile_huge_page is used for calls to mmap *with* MAP_HUGETLB, + * smpi_shared_malloc_bogusfile is used for calls to mmap *without* MAP_HUGETLB. + * We cannot use a same file for the two type of calls, since the first one needs to be + * opened in a hugetlbfs mount point whereas the second needs to be a "classical" file. */ + if(use_huge_page && smpi_shared_malloc_bogusfile_huge_page == -1) { + const char *const array[] = {huge_page_mount_point, "simgrid-shmalloc-XXXXXX", nullptr}; + char *huge_page_filename = xbt_str_join_array(array, "/"); + smpi_shared_malloc_bogusfile_huge_page = mkstemp(huge_page_filename); + XBT_DEBUG("bogusfile_huge_page: %s\n", huge_page_filename); + unlink(huge_page_filename); + xbt_free(huge_page_filename); + } + if(smpi_shared_malloc_bogusfile == -1) { + char *name = xbt_strdup("/tmp/simgrid-shmalloc-XXXXXX"); + smpi_shared_malloc_bogusfile = mkstemp(name); + XBT_DEBUG("bogusfile : %s\n", name); + unlink(name); + xbt_free(name); + char* dumb = (char*)calloc(1, smpi_shared_malloc_blocksize); + ssize_t err = write(smpi_shared_malloc_bogusfile, dumb, smpi_shared_malloc_blocksize); + if(err<0) + xbt_die("Could not write bogus file for shared malloc"); + xbt_free(dumb); + } + + int mmap_base_flag = MAP_FIXED | MAP_SHARED | MAP_POPULATE; + int mmap_flag = mmap_base_flag; + int huge_fd = use_huge_page ? smpi_shared_malloc_bogusfile_huge_page : smpi_shared_malloc_bogusfile; +#ifdef MAP_HUGETLB + if(use_huge_page) + mmap_flag |= MAP_HUGETLB; +#endif + + XBT_DEBUG("global shared allocation, begin mmap"); + + /* Map the bogus file in place of the anonymous memory */ + for(int i_block = 0; i_block < nb_shared_blocks; i_block ++) { + XBT_DEBUG("\tglobal shared allocation, mmap block %d/%d", i_block+1, nb_shared_blocks); + size_t start_offset = shared_block_offsets[2*i_block]; + size_t stop_offset = shared_block_offsets[2*i_block+1]; + xbt_assert(start_offset < stop_offset, "start_offset (%zu) should be lower than stop offset (%zu)", start_offset, stop_offset); + xbt_assert(stop_offset <= size, "stop_offset (%zu) should be lower than size (%zu)", stop_offset, size); + if(i_block < nb_shared_blocks-1) + xbt_assert(stop_offset < shared_block_offsets[2*i_block+2], + "stop_offset (%zu) should be lower than its successor start offset (%zu)", stop_offset, shared_block_offsets[2*i_block+2]); + size_t start_block_offset = ALIGN_UP(start_offset, smpi_shared_malloc_blocksize); + size_t stop_block_offset = ALIGN_DOWN(stop_offset, smpi_shared_malloc_blocksize); + for (unsigned block_id=0, i = start_block_offset / smpi_shared_malloc_blocksize; i < stop_block_offset / smpi_shared_malloc_blocksize; block_id++, i++) { + XBT_DEBUG("\t\tglobal shared allocation, mmap block offset %d", block_id); void* pos = (void*)((unsigned long)mem + i * smpi_shared_malloc_blocksize); - void* res = mmap(pos, smpi_shared_malloc_blocksize, PROT_READ | PROT_WRITE, MAP_FIXED | MAP_SHARED | MAP_POPULATE, - smpi_shared_malloc_bogusfile, 0); + void* res = mmap(pos, smpi_shared_malloc_blocksize, PROT_READ | PROT_WRITE, mmap_flag, + huge_fd, 0); xbt_assert(res == pos, "Could not map folded virtual memory (%s). Do you perhaps need to increase the " - "size of the mapped file using --cfg=smpi/shared-malloc-blocksize=newvalue (default 1048576) ?", - "You can also try using the sysctl vm.max_map_count", + "size of the mapped file using --cfg=smpi/shared-malloc-blocksize=newvalue (default 1048576) ? " + "You can also try using the sysctl vm.max_map_count. " + "If you are using huge pages, check that you have at least one huge page (/proc/sys/vm/nr_hugepages) " + "and that the directory you are passing is mounted correctly (mount /path/to/huge -t hugetlbfs -o rw,mode=0777).", strerror(errno)); } - if (size % smpi_shared_malloc_blocksize) { - void* pos = (void*)((unsigned long)mem + i * smpi_shared_malloc_blocksize); - void* res = mmap(pos, size % smpi_shared_malloc_blocksize, PROT_READ | PROT_WRITE, - MAP_FIXED | MAP_SHARED | MAP_POPULATE, smpi_shared_malloc_bogusfile, 0); + size_t low_page_start_offset = ALIGN_UP(start_offset, PAGE_SIZE); + size_t low_page_stop_offset = start_block_offset < ALIGN_DOWN(stop_offset, PAGE_SIZE) ? start_block_offset : ALIGN_DOWN(stop_offset, PAGE_SIZE); + if(low_page_start_offset < low_page_stop_offset) { + XBT_DEBUG("\t\tglobal shared allocation, mmap block start"); + void* pos = (void*)((unsigned long)mem + low_page_start_offset); + void* res = mmap(pos, low_page_stop_offset-low_page_start_offset, PROT_READ | PROT_WRITE, mmap_base_flag, // not a full huge page + smpi_shared_malloc_bogusfile, 0); xbt_assert(res == pos, "Could not map folded virtual memory (%s). Do you perhaps need to increase the " - "size of the mapped file using --cfg=smpi/shared-malloc-blocksize=newvalue (default 1048576) ?", + "size of the mapped file using --cfg=smpi/shared-malloc-blocksize=newvalue (default 1048576) ?" "You can also try using the sysctl vm.max_map_count", strerror(errno)); } + if(low_page_stop_offset <= stop_block_offset) { + XBT_DEBUG("\t\tglobal shared allocation, mmap block stop"); + size_t high_page_stop_offset = stop_offset == size ? size : ALIGN_DOWN(stop_offset, PAGE_SIZE); + if(high_page_stop_offset > stop_block_offset) { + void* pos = (void*)((unsigned long)mem + stop_block_offset); + void* res = mmap(pos, high_page_stop_offset-stop_block_offset, PROT_READ | PROT_WRITE, mmap_base_flag, // not a full huge page + smpi_shared_malloc_bogusfile, 0); + xbt_assert(res == pos, "Could not map folded virtual memory (%s). Do you perhaps need to increase the " + "size of the mapped file using --cfg=smpi/shared-malloc-blocksize=newvalue (default 1048576) ?" + "You can also try using the sysctl vm.max_map_count", + strerror(errno)); + } + } + } - shared_metadata_t newmeta; - //register metadata for memcpy avoidance - shared_data_key_type* data = (shared_data_key_type*)xbt_malloc(sizeof(shared_data_key_type)); - data->second.fd = -1; - data->second.count = 1; - newmeta.size = size; - newmeta.data = data; - allocs_metadata[mem] = newmeta; - } else { - mem = xbt_malloc(size); - XBT_DEBUG("Classic malloc %zu in %p", size, mem); + shared_metadata_t newmeta; + //register metadata for memcpy avoidance + shared_data_key_type* data = (shared_data_key_type*)xbt_malloc(sizeof(shared_data_key_type)); + data->second.fd = -1; + data->second.count = 1; + newmeta.size = size; + newmeta.data = data; + newmeta.allocated_ptr = allocated_ptr; + newmeta.allocated_size = allocated_size; + if(shared_block_offsets[0] > 0) { + newmeta.private_blocks.push_back(std::make_pair(0, shared_block_offsets[0])); + } + int i_block; + for(i_block = 0; i_block < nb_shared_blocks-1; i_block ++) { + newmeta.private_blocks.push_back(std::make_pair(shared_block_offsets[2*i_block+1], shared_block_offsets[2*i_block+2])); } + if(shared_block_offsets[2*i_block+1] < size) { + newmeta.private_blocks.push_back(std::make_pair(shared_block_offsets[2*i_block+1], size)); + } + allocs_metadata[mem] = newmeta; + + XBT_DEBUG("global shared allocation, allocated_ptr %p - %p", allocated_ptr, (void*)(((uint64_t)allocated_ptr)+allocated_size)); + XBT_DEBUG("global shared allocation, returned_ptr %p - %p", mem, (void*)(((uint64_t)mem)+size)); return mem; } -int smpi_is_shared(void*ptr){ +void *smpi_shared_malloc(size_t size, const char *file, int line) { + if (size > 0 && smpi_cfg_shared_malloc == shmalloc_local) { + return smpi_shared_malloc_local(size, file, line); + } else if (smpi_cfg_shared_malloc == shmalloc_global) { + int nb_shared_blocks = 1; + size_t shared_block_offsets[2] = {0, size}; + return smpi_shared_malloc_partial(size, shared_block_offsets, nb_shared_blocks); + } + XBT_DEBUG("Classic malloc %zu", size); + return xbt_malloc(size); +} + +int smpi_is_shared(void* ptr, std::vector> &private_blocks, size_t *offset){ + private_blocks.clear(); // being paranoid if (allocs_metadata.empty()) return 0; if ( smpi_cfg_shared_malloc == shmalloc_local || smpi_cfg_shared_malloc == shmalloc_global) { - if (allocs_metadata.count(ptr) != 0) - return 1; - for(auto it : allocs_metadata){ - if (ptr >= it.first && ptr < (char*)it.first + it.second.size) - return 1; + auto low = allocs_metadata.lower_bound(ptr); + if (low->first==ptr) { + private_blocks = low->second.private_blocks; + *offset = 0; + return 1; } + if (low == allocs_metadata.begin()) return 0; + low --; + if (ptr < (char*)low->first + low->second.size) { + xbt_assert(ptr > (char*)low->first, "Oops, there seems to be a bug in the shared memory metadata."); + *offset = ((uint8_t*)ptr) - ((uint8_t*) low->first); + private_blocks = low->second.private_blocks; + return 1; + } + return 0; } else { return 0; } } +std::vector> shift_and_frame_private_blocks(const std::vector> vec, size_t offset, size_t buff_size) { + std::vector> result; + for(auto block: vec) { + auto new_block = std::make_pair(std::min(std::max((size_t)0, block.first-offset), buff_size), + std::min(std::max((size_t)0, block.second-offset), buff_size)); + if(new_block.second > 0 && new_block.first < buff_size) + result.push_back(new_block); + } + return result; +} + +std::vector> merge_private_blocks(std::vector> src, std::vector> dst) { + std::vector> result; + unsigned i_src = 0; + unsigned i_dst = 0; + while(i_src < src.size() && i_dst < dst.size()) { + std::pair block; + if(src[i_src].second <= dst[i_dst].first) { + i_src++; + } + else if(dst[i_dst].second <= src[i_src].first) { + i_dst++; + } + else { // src.second > dst.first && dst.second > src.first → the blocks are overlapping + block = std::make_pair(std::max(src[i_src].first, dst[i_dst].first), + std::min(src[i_src].second, dst[i_dst].second)); + result.push_back(block); + if(src[i_src].second < dst[i_dst].second) + i_src ++; + else + i_dst ++; + } + } + return result; +} + void smpi_shared_free(void *ptr) { if (smpi_cfg_shared_malloc == shmalloc_local) { @@ -289,7 +457,7 @@ void smpi_shared_free(void *ptr) return; } shared_data_t* data = &meta->second.data->second; - if (munmap(ptr, meta->second.size) < 0) { + if (munmap(meta->second.allocated_ptr, meta->second.allocated_size) < 0) { XBT_WARN("Unmapping of fd %d failed: %s", data->fd, strerror(errno)); } data->count--; @@ -310,7 +478,7 @@ void smpi_shared_free(void *ptr) xbt_free(meta->second.data); } - munmap(ptr, 0); // the POSIX says that I should not give 0 as a length, but it seems to work OK + munmap(ptr, meta->second.size); } else { XBT_DEBUG("Classic free of %p", ptr); xbt_free(ptr);