X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/9104957deccc59e0e804215d5db498fabfd40d29..824740eb0df1dedddb86035ff3730d87e037f356:/src/smpi/internals/smpi_shared.cpp diff --git a/src/smpi/internals/smpi_shared.cpp b/src/smpi/internals/smpi_shared.cpp index c510fc5d63..10ab0ad0ae 100644 --- a/src/smpi/internals/smpi_shared.cpp +++ b/src/smpi/internals/smpi_shared.cpp @@ -1,4 +1,4 @@ -/* Copyright (c) 2007-2020. The SimGrid Team. All rights reserved. */ +/* Copyright (c) 2007-2023. 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. */ @@ -33,22 +33,21 @@ * \ | | * ---- */ -#include +#include #include +#include #include "private.hpp" #include "xbt/config.hpp" +#include "xbt/file.hpp" #include -#include -#ifndef WIN32 +#include "smpi_utils.hpp" +#include #include -#endif -#include -#include -#include - +#include +#include #ifndef MAP_ANONYMOUS #define MAP_ANONYMOUS MAP_ANON #endif @@ -59,7 +58,7 @@ XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_shared, smpi, "Logging specific to SMPI (shared memory macros)"); -namespace{ +namespace { /** Some location in the source code * * This information is used by SMPI_SHARED_MALLOC to allocate some shared memory for all simulated processes. @@ -79,7 +78,7 @@ struct shared_data_t { }; std::unordered_map> allocs; -typedef decltype(allocs)::value_type shared_data_key_type; +using shared_data_key_type = decltype(allocs)::value_type; struct shared_metadata_t { size_t size; @@ -90,15 +89,12 @@ struct shared_metadata_t { }; std::map allocs_metadata; -std::map calls; - -#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 -} +std::map> calls; +int smpi_shared_malloc_bogusfile = -1; +int smpi_shared_malloc_bogusfile_huge_page = -1; +unsigned long smpi_shared_malloc_blocksize = 1UL << 20; +} // namespace void smpi_shared_destroy() { @@ -107,7 +103,6 @@ void smpi_shared_destroy() calls.clear(); } -#ifndef WIN32 static void* shm_map(int fd, size_t size, shared_data_key_type* data) { void* mem = smpi_temp_shm_mmap(fd, size); @@ -125,9 +120,8 @@ static void *smpi_shared_malloc_local(size_t size, const char *file, int line) { void* mem; smpi_source_location loc(file, line); - auto res = allocs.insert(std::make_pair(loc, shared_data_t())); - auto data = res.first; - if (res.second) { + auto [data, inserted] = allocs.try_emplace(loc); + if (inserted) { // The new element was inserted. int fd = smpi_temp_shm_get(); data->second.fd = fd; @@ -142,8 +136,8 @@ static void *smpi_shared_malloc_local(size_t size, const char *file, int line) } // Align functions, from http://stackoverflow.com/questions/4840410/how-to-align-a-pointer-in-c -#define ALIGN_UP(n, align) (((n) + (align)-1) & -(align)) -#define ALIGN_DOWN(n, align) ((n) & -(align)) +#define ALIGN_UP(n, align) (((int64_t)(n) + (int64_t)(align) - 1) & -(int64_t)(align)) +#define ALIGN_DOWN(n, align) ((int64_t)(n) & -(int64_t)(align)) constexpr unsigned PAGE_SIZE = 0x1000; constexpr unsigned HUGE_PAGE_SIZE = 1U << 21; @@ -155,7 +149,7 @@ constexpr unsigned HUGE_PAGE_SIZE = 1U << 21; * The others are not. */ -void* smpi_shared_malloc_partial(size_t size, size_t* shared_block_offsets, int nb_shared_blocks) +void* smpi_shared_malloc_partial(size_t size, const size_t* shared_block_offsets, int nb_shared_blocks) { std::string huge_page_mount_point = simgrid::config::get_value("smpi/shared-malloc-hugepage"); bool use_huge_page = not huge_page_mount_point.empty(); @@ -178,13 +172,13 @@ void* smpi_shared_malloc_partial(size_t size, size_t* shared_block_offsets, int /* First reserve memory area */ - void* allocated_ptr = mmap(NULL, allocated_size, PROT_READ | PROT_WRITE, MAP_ANONYMOUS | MAP_PRIVATE, -1, 0); + void* allocated_ptr = mmap(nullptr, 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((int64_t)allocated_ptr, HUGE_PAGE_SIZE); + mem = (void*)ALIGN_UP(allocated_ptr, HUGE_PAGE_SIZE); else mem = allocated_ptr; @@ -208,11 +202,8 @@ void* smpi_shared_malloc_partial(size_t size, size_t* shared_block_offsets, int smpi_shared_malloc_bogusfile = mkstemp(name); XBT_DEBUG("bogusfile : %s\n", name); unlink(name); - const char* dumb = new char[smpi_shared_malloc_blocksize](); // zero initialized - 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"); - delete[] dumb; + xbt_assert(ftruncate(smpi_shared_malloc_bogusfile, smpi_shared_malloc_blocksize) == 0, + "Could not write bogus file for shared malloc"); } int mmap_base_flag = MAP_FIXED | MAP_SHARED | MAP_POPULATE; @@ -235,11 +226,11 @@ 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((int64_t)start_offset, smpi_shared_malloc_blocksize); - size_t stop_block_offset = ALIGN_DOWN((int64_t)stop_offset, smpi_shared_malloc_blocksize); + 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 (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* pos = static_cast(mem) + offset; const 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) ? " @@ -248,11 +239,11 @@ void* smpi_shared_malloc_partial(size_t size, size_t* shared_block_offsets, int "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((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); + size_t low_page_start_offset = ALIGN_UP(start_offset, PAGE_SIZE); + 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); 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* pos = static_cast(mem) + low_page_start_offset; const 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); @@ -263,9 +254,9 @@ void* smpi_shared_malloc_partial(size_t size, size_t* shared_block_offsets, int } 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((int64_t)stop_offset, PAGE_SIZE); + 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* pos = static_cast(mem) + stop_block_offset; const 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); @@ -279,7 +270,7 @@ void* smpi_shared_malloc_partial(size_t size, size_t* shared_block_offsets, int shared_metadata_t newmeta; //register metadata for memcpy avoidance - shared_data_key_type* data = new shared_data_key_type; + auto* data = new shared_data_key_type; data->second.fd = -1; data->second.count = 1; newmeta.size = size; @@ -287,14 +278,14 @@ void* smpi_shared_malloc_partial(size_t size, size_t* shared_block_offsets, int 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])); + newmeta.private_blocks.emplace_back(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])); + newmeta.private_blocks.emplace_back(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)); + newmeta.private_blocks.emplace_back(shared_block_offsets[2 * i_block + 1], size); } allocs_metadata[mem] = newmeta; @@ -306,20 +297,57 @@ void* smpi_shared_malloc_partial(size_t size, size_t* shared_block_offsets, int 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 + if( smpi_cfg_auto_shared_malloc_thresh() == 0 || size < smpi_cfg_auto_shared_malloc_thresh()){ + void* ptr = xbt_malloc(size); + if(not smpi_cfg_trace_call_use_absolute_path()) + simgrid::smpi::utils::account_malloc_size(size, simgrid::xbt::Path(file).get_base_name(), line, ptr); + else + simgrid::smpi::utils::account_malloc_size(size, file, line, ptr); + return ptr; + } else { + simgrid::smpi::utils::account_shared_size(size); 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); + size_t size = elem_size * num_elm; + if (smpi_cfg_auto_shared_malloc_thresh() == 0 || size < smpi_cfg_auto_shared_malloc_thresh()) { + void* ptr = xbt_malloc0(size); + if(not smpi_cfg_trace_call_use_absolute_path()) + simgrid::smpi::utils::account_malloc_size(size, simgrid::xbt::Path(file).get_base_name(), line, ptr); + else + simgrid::smpi::utils::account_malloc_size(size, file, line, ptr); return ptr; - } else - return smpi_shared_malloc(elem_size*num_elm, file, line); + } else { + simgrid::smpi::utils::account_shared_size(size); + return memset(smpi_shared_malloc(size, file, line), 0, size); + } +} + +void* smpi_shared_realloc_intercept(void* data, size_t size, const char* file, int line) +{ + if (size == 0) { + smpi_shared_free(data); + return nullptr; + } + if (data == nullptr) + return smpi_shared_malloc_intercept(size, file, line); + + auto meta = allocs_metadata.find(data); + if (meta == allocs_metadata.end()) { + XBT_DEBUG("Classical realloc(%p, %zu)", data, size); + return xbt_realloc(data, size); + } + + XBT_DEBUG("Shared realloc(%p, %zu) (old size: %zu)", data, size, meta->second.size); + void* ptr = smpi_shared_malloc_intercept(size, file, line); + if (ptr != data) { + memcpy(ptr, data, std::min(size, meta->second.size)); + smpi_shared_free(data); + } + return ptr; } void* smpi_shared_malloc(size_t size, const char* file, int line) @@ -328,11 +356,11 @@ void* smpi_shared_malloc(size_t size, const char* file, int line) return smpi_shared_malloc_local(size, file, line); } 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); + const std::array shared_block_offsets = {{0, size}}; + return smpi_shared_malloc_partial(size, shared_block_offsets.data(), nb_shared_blocks); } XBT_DEBUG("Classic allocation of %zu bytes", size); - return ::operator new(size); + return xbt_malloc(size); } int smpi_is_shared(const void* ptr, std::vector> &private_blocks, size_t *offset){ @@ -365,9 +393,9 @@ std::vector> shift_and_frame_private_blocks(const std: 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)); + for (auto const& [block_begin, block_end] : vec) { + auto new_block = std::make_pair(std::clamp(block_begin - offset, (size_t)0, buff_size), + std::clamp(block_end - offset, (size_t)0, buff_size)); if (new_block.second > 0 && new_block.first < buff_size) result.push_back(new_block); } @@ -403,10 +431,11 @@ std::vector> merge_private_blocks(const std::vectorsecond.data->second; @@ -417,7 +446,7 @@ void smpi_shared_free(void *ptr) if (data->count <= 0) { close(data->fd); allocs.erase(allocs.find(meta->second.data->first)); - allocs_metadata.erase(ptr); + allocs_metadata.erase(meta); XBT_DEBUG("Shared free - Local - with removal - of %p", ptr); } else { XBT_DEBUG("Shared free - Local - no removal - of %p, count = %d", ptr, data->count); @@ -431,19 +460,18 @@ void smpi_shared_free(void *ptr) munmap(ptr, meta->second.size); if(meta->second.data->second.count==0){ delete meta->second.data; - allocs_metadata.erase(ptr); + allocs_metadata.erase(meta); } }else{ - ::operator delete(ptr); + xbt_free(ptr); return; } } else { XBT_DEBUG("Classic deallocation of %p", ptr); - ::operator delete(ptr); + xbt_free(ptr); } } -#endif int smpi_shared_known_call(const char* func, const char* input) {