X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/3f27f409dbb1f21cde182648696386708838c883..a88d4f2d2d1d5cce8f3cc1350f41d0dd52d55a31:/src/smpi/internals/smpi_shared.cpp diff --git a/src/smpi/internals/smpi_shared.cpp b/src/smpi/internals/smpi_shared.cpp index 9bf2045811..9b2a263f6e 100644 --- a/src/smpi/internals/smpi_shared.cpp +++ b/src/smpi/internals/smpi_shared.cpp @@ -1,4 +1,4 @@ -/* Copyright (c) 2007, 2009-2017. The SimGrid Team. All rights reserved. */ +/* Copyright (c) 2007-2019. The SimGrid Team. All rights reserved. */ /* This program is free software; you can redistribute it and/or modify it * under the terms of the license (GNU LGPL) which comes with this package. */ @@ -37,6 +37,8 @@ #include #include "private.hpp" +#include "xbt/config.hpp" + #include #include @@ -57,8 +59,6 @@ XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_shared, smpi, "Logging specific to SMPI (shared memory macros)"); -#define PTR_STRLEN (2 + 2 * sizeof(void*) + 1) - namespace{ /** Some location in the source code * @@ -89,7 +89,7 @@ struct shared_metadata_t { shared_data_key_type* data; }; -std::map allocs_metadata; +std::map allocs_metadata; std::map calls; #ifndef WIN32 @@ -107,36 +107,15 @@ void smpi_shared_destroy() calls.clear(); } -static size_t shm_size(int fd) { - struct stat st; - - if(fstat(fd, &st) < 0) { - xbt_die("Could not stat fd %d: %s", fd, strerror(errno)); - } - return static_cast(st.st_size); -} - #ifndef WIN32 -static void* shm_map(int fd, size_t size, shared_data_key_type* data) { - char loc[PTR_STRLEN]; +static void* shm_map(int fd, size_t size, shared_data_key_type* data) +{ + void* mem = smpi_temp_shm_mmap(fd, size); shared_metadata_t meta; - - if(size > shm_size(fd) && (ftruncate(fd, static_cast(size)) < 0)) { - xbt_die("Could not truncate fd %d to %zu: %s", fd, size, strerror(errno)); - } - - void* mem = mmap(nullptr, size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0); - if(mem == MAP_FAILED) { - xbt_die( - "Failed to map fd %d with size %zu: %s\n" - "If you are running a lot of ranks, you may be exceeding the amount of mappings allowed per process.\n" - "On Linux systems, change this value with sudo sysctl -w vm.max_map_count=newvalue (default value: 65536)\n" - "Please see http://simgrid.gforge.inria.fr/simgrid/latest/doc/html/options.html#options_virt for more info.", - fd, size, strerror(errno)); - } - snprintf(loc, PTR_STRLEN, "%p", mem); meta.size = size; meta.data = data; + meta.allocated_ptr = mem; + meta.allocated_size = size; allocs_metadata[mem] = meta; XBT_DEBUG("MMAP %zu to %p", size, mem); return mem; @@ -150,37 +129,24 @@ static void *smpi_shared_malloc_local(size_t size, const char *file, int line) auto data = res.first; if (res.second) { // The new element was inserted. - // 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; + int fd = smpi_temp_shm_get(); + 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); + XBT_DEBUG("Shared malloc %zu in %p through %d (metadata at %p)", size, mem, data->second.fd, &*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 +constexpr unsigned PAGE_SIZE = 0x1000; +constexpr unsigned HUGE_PAGE_SIZE = 1U << 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. @@ -191,14 +157,14 @@ static void *smpi_shared_malloc_local(size_t size, const char *file, int line) void* smpi_shared_malloc_partial(size_t size, size_t* shared_block_offsets, int nb_shared_blocks) { - std::string huge_page_mount_point = xbt_cfg_get_string("smpi/shared-malloc-hugepage"); + std::string huge_page_mount_point = simgrid::config::get_value("smpi/shared-malloc-hugepage"); bool use_huge_page = not huge_page_mount_point.empty(); #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")); + smpi_shared_malloc_blocksize = + static_cast(simgrid::config::get_value("smpi/shared-malloc-blocksize")); void* mem; size_t allocated_size; if(use_huge_page) { @@ -218,7 +184,7 @@ void* smpi_shared_malloc_partial(size_t size, size_t* shared_block_offsets, int "to allow big allocations.\n", size >> 20); if(use_huge_page) - mem = (void*)ALIGN_UP((uint64_t)allocated_ptr, HUGE_PAGE_SIZE); + mem = (void*)ALIGN_UP((int64_t)allocated_ptr, HUGE_PAGE_SIZE); else mem = allocated_ptr; @@ -269,41 +235,41 @@ void* smpi_shared_malloc_partial(size_t size, size_t* shared_block_offsets, int 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); + size_t start_block_offset = ALIGN_UP((int64_t)start_offset, smpi_shared_malloc_blocksize); + size_t stop_block_offset = ALIGN_DOWN((int64_t)stop_offset, smpi_shared_malloc_blocksize); for (size_t offset = start_block_offset; offset < stop_block_offset; offset += smpi_shared_malloc_blocksize) { XBT_DEBUG("\t\tglobal shared allocation, mmap block offset %zx", offset); void* pos = (void*)((unsigned long)mem + offset); 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) ? " + "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)); } - 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); + size_t low_page_start_offset = ALIGN_UP((int64_t)start_offset, PAGE_SIZE); + 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); 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); + size_t high_page_stop_offset = stop_offset == size ? size : ALIGN_DOWN((int64_t)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) ?" + "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)); } @@ -337,10 +303,28 @@ void* smpi_shared_malloc_partial(size_t size, size_t* shared_block_offsets, int return mem; } + +void *smpi_shared_malloc_intercept(size_t size, const char *file, int line) { + if( smpi_cfg_auto_shared_malloc_thresh() == 0 || size < smpi_cfg_auto_shared_malloc_thresh()) + return ::operator new(size); + else + return smpi_shared_malloc(size, file, line); +} + +void* smpi_shared_calloc_intercept(size_t num_elm, size_t elem_size, const char* file, int line){ + if( smpi_cfg_auto_shared_malloc_thresh() == 0 || elem_size*num_elm < smpi_cfg_auto_shared_malloc_thresh()){ + void* ptr = ::operator new(elem_size*num_elm); + memset(ptr, 0, elem_size*num_elm); + return ptr; + } else + return smpi_shared_malloc(elem_size*num_elm, file, line); + +} + void *smpi_shared_malloc(size_t size, const char *file, int line) { - if (size > 0 && smpi_cfg_shared_malloc == shmalloc_local) { + if (size > 0 && smpi_cfg_shared_malloc() == SharedMallocType::LOCAL) { return smpi_shared_malloc_local(size, file, line); - } else if (smpi_cfg_shared_malloc == shmalloc_global) { + } else if (smpi_cfg_shared_malloc() == SharedMallocType::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); @@ -349,11 +333,11 @@ void *smpi_shared_malloc(size_t size, const char *file, int line) { return ::operator new(size); } -int smpi_is_shared(void* ptr, std::vector> &private_blocks, size_t *offset){ +int smpi_is_shared(const 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 (smpi_cfg_shared_malloc() == SharedMallocType::LOCAL || smpi_cfg_shared_malloc() == SharedMallocType::GLOBAL) { auto low = allocs_metadata.lower_bound(ptr); if (low != allocs_metadata.end() && low->first == ptr) { private_blocks = low->second.private_blocks; @@ -375,18 +359,22 @@ int smpi_is_shared(void* ptr, std::vector> &private_bl } } -std::vector> shift_and_frame_private_blocks(const std::vector> vec, size_t offset, size_t buff_size) { - std::vector> result; - for (auto const& 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> shift_and_frame_private_blocks(const std::vector>& vec, + size_t offset, size_t buff_size) +{ + std::vector> result; + for (auto const& 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> merge_private_blocks(const std::vector>& src, + const std::vector>& dst) +{ std::vector> result; unsigned i_src = 0; unsigned i_dst = 0; @@ -413,12 +401,10 @@ std::vector> merge_private_blocks(std::vectorsecond.data->second; @@ -430,20 +416,26 @@ void smpi_shared_free(void *ptr) close(data->fd); allocs.erase(allocs.find(meta->second.data->first)); allocs_metadata.erase(ptr); - XBT_DEBUG("Shared free - with removal - of %p", ptr); + XBT_DEBUG("Shared free - Local - with removal - of %p", ptr); } else { - XBT_DEBUG("Shared free - no removal - of %p, count = %d", ptr, data->count); + XBT_DEBUG("Shared free - Local - no removal - of %p, count = %d", ptr, data->count); } - } else if (smpi_cfg_shared_malloc == shmalloc_global) { + } else if (smpi_cfg_shared_malloc() == SharedMallocType::GLOBAL) { auto meta = allocs_metadata.find(ptr); if (meta != allocs_metadata.end()){ meta->second.data->second.count--; - if(meta->second.data->second.count==0) + XBT_DEBUG("Shared free - Global - of %p", ptr); + munmap(ptr, meta->second.size); + if(meta->second.data->second.count==0){ delete meta->second.data; + allocs_metadata.erase(ptr); + } + }else{ + ::operator delete(ptr); + return; } - munmap(ptr, meta->second.size); } else { XBT_DEBUG("Classic deallocation of %p", ptr); ::operator delete(ptr);