-/* 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. */
#include <map>
#include <cstring>
-#include "private.h"
#include "private.hpp"
+#include "xbt/config.hpp"
+
#include <cerrno>
#include <sys/types.h>
class smpi_source_location : public std::string {
public:
+ smpi_source_location() = default;
smpi_source_location(const char* filename, int line) : std::string(std::string(filename) + ":" + std::to_string(line))
{
}
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));
+ 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 "
+ "https://simgrid.org/doc/latest/Configuring_SimGrid.html#configuring-the-user-code-virtualization for more "
+ "information.",
+ 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;
auto res = allocs.insert(std::make_pair(loc, shared_data_t()));
auto data = res.first;
if (res.second) {
- // The insertion did not take place.
+ // 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)
+ char shmname[32]; // cannot be longer than PSHMNAMLEN = 31 on macOS (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) {
}
// 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.
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';
+ std::string huge_page_mount_point = simgrid::config::get_value<std::string>("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<unsigned long>(xbt_cfg_get_double("smpi/shared-malloc-blocksize"));
+ smpi_shared_malloc_blocksize =
+ static_cast<unsigned long>(simgrid::config::get_value<double>("smpi/shared-malloc-blocksize"));
void* mem;
size_t allocated_size;
if(use_huge_page) {
"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;
* 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);
+ std::string huge_page_filename = huge_page_mount_point + "/simgrid-shmalloc-XXXXXX";
+ smpi_shared_malloc_bogusfile_huge_page = mkstemp((char*)huge_page_filename.c_str());
+ XBT_DEBUG("bogusfile_huge_page: %s\n", huge_page_filename.c_str());
+ unlink(huge_page_filename.c_str());
}
if(smpi_shared_malloc_bogusfile == -1) {
- char *name = xbt_strdup("/tmp/simgrid-shmalloc-XXXXXX");
+ char name[] = "/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);
+ 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");
- xbt_free(dumb);
+ delete[] dumb;
}
int mmap_base_flag = MAP_FIXED | MAP_SHARED | MAP_POPULATE;
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 %u", block_id);
- void* pos = (void*)((unsigned long)mem + i * 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));
}
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));
+ shared_data_key_type* data = new shared_data_key_type;
data->second.fd = -1;
data->second.count = 1;
newmeta.size = size;
return mem;
}
+
+void *smpi_shared_malloc_intercept(size_t size, const char *file, int line) {
+ if( simgrid::config::get_value<double>("smpi/auto-shared-malloc-thresh") == 0 || size < simgrid::config::get_value<double>("smpi/auto-shared-malloc-thresh"))
+ return ::operator new(size);
+ else
+ return smpi_shared_malloc(size, 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);
}
- XBT_DEBUG("Classic malloc %zu", size);
- return xbt_malloc(size);
+ XBT_DEBUG("Classic allocation of %zu bytes", size);
+ return ::operator new(size);
}
int smpi_is_shared(void* ptr, std::vector<std::pair<size_t, size_t>> &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;
}
}
-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) {
- std::vector<std::pair<size_t, size_t>> 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<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)
+{
+ std::vector<std::pair<size_t, size_t>> 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<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) {
+std::vector<std::pair<size_t, size_t>> merge_private_blocks(const std::vector<std::pair<size_t, size_t>>& src,
+ const std::vector<std::pair<size_t, size_t>>& dst)
+{
std::vector<std::pair<size_t, size_t>> result;
unsigned i_src = 0;
unsigned i_dst = 0;
void smpi_shared_free(void *ptr)
{
- if (smpi_cfg_shared_malloc == shmalloc_local) {
+ if (smpi_cfg_shared_malloc == SharedMallocType::LOCAL) {
char loc[PTR_STRLEN];
snprintf(loc, PTR_STRLEN, "%p", ptr);
auto meta = allocs_metadata.find(ptr);
if (meta == allocs_metadata.end()) {
- XBT_WARN("Cannot free: %p was not shared-allocated by SMPI - maybe its size was 0?", ptr);
+ ::operator delete(ptr);
return;
}
shared_data_t* data = &meta->second.data->second;
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_free(meta->second.data);
+ if(meta->second.data->second.count==0){
+ delete meta->second.data;
+ allocs_metadata.erase(ptr);
+ }
+ XBT_DEBUG("Shared free - Global - of %p", ptr);
+ munmap(ptr, meta->second.size);
+ }else{
+ ::operator delete(ptr);
+ return;
}
- munmap(ptr, meta->second.size);
} else {
- XBT_DEBUG("Classic free of %p", ptr);
- xbt_free(ptr);
+ XBT_DEBUG("Classic deallocation of %p", ptr);
+ ::operator delete(ptr);
}
}
#endif
