#include "private.h"
#include "private.hpp"
+#include "smpi/smpi_shared_malloc.hpp"
#include "xbt/dict.h"
#include <errno.h>
typedef struct {
size_t size;
+ std::vector<std::pair<size_t, size_t>> private_blocks;
shared_data_key_type* data;
} shared_metadata_t;
return mem;
}
-void *smpi_shared_malloc_local(size_t size, const char *file, int line)
+static void *smpi_shared_malloc_local(size_t size, const char *file, int line)
{
void* mem;
smpi_source_location loc(file, line);
}
// 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))
-void *smpi_shared_malloc_global__(size_t size, const char *file, int line, int *shared_block_offsets, int nb_shared_blocks) {
+void *smpi_shared_malloc_global__(size_t size, const char *file, int line, size_t *shared_block_offsets, int nb_shared_blocks) {
void *mem;
+ xbt_assert(smpi_shared_malloc_blocksize % PAGE_SIZE == 0, "The block size of shared malloc should be a multiple of the page size.");
/* First reserve memory area */
mem = mmap(NULL, size, PROT_READ | PROT_WRITE, MAP_ANONYMOUS | MAP_PRIVATE, -1, 0);
/* Map the bogus file in place of the anonymous memory */
for(int i_block = 0; i_block < nb_shared_blocks; i_block ++) {
- int start_offset = ALIGN_UP(shared_block_offsets[2*i_block], smpi_shared_malloc_blocksize);
- int stop_offset = ALIGN_DOWN(shared_block_offsets[2*i_block+1], smpi_shared_malloc_blocksize);
+ 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 (%lu) should be lower than stop offset (%lu)", start_offset, stop_offset);
+ xbt_assert(stop_offset <= size, "stop_offset (%lu) should be lower than size (%lu)", stop_offset, size);
+ if(i_block < nb_shared_blocks-1)
+ xbt_assert(stop_offset < shared_block_offsets[2*i_block+2],
+ "stop_offset (%lu) should be lower than its successor start offset (%lu)", stop_offset, shared_block_offsets[2*i_block+2]);
+// fprintf(stderr, "shared block 0x%x - 0x%x\n", start_offset, stop_offset);
+ 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);
unsigned int i;
- for (i = start_offset; i < stop_offset / smpi_shared_malloc_blocksize; i++) {
+ for (i = start_block_offset / smpi_shared_malloc_blocksize; i < stop_block_offset / smpi_shared_malloc_blocksize; i++) {
+// fprintf(stderr, "\tmmap:for 0x%x - 0x%x\n", i*smpi_shared_malloc_blocksize, smpi_shared_malloc_blocksize);
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);
"You can also try using the sysctl vm.max_map_count",
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);
+ if(low_page_start_offset < low_page_stop_offset) {
+// fprintf(stderr, "\tmmap:low 0x%x - 0x%x\n", low_page_start_offset, low_page_stop_offset-low_page_start_offset);
+ 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, MAP_FIXED | MAP_SHARED | MAP_POPULATE,
+ 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));
+ }
+ if(low_page_stop_offset <= stop_block_offset) {
+ size_t high_page_stop_offset = stop_offset == size ? size : ALIGN_DOWN(stop_offset, PAGE_SIZE);
+ if(high_page_stop_offset > stop_block_offset) {
+// fprintf(stderr, "\tmmap:high 0x%x - 0x%x\n", stop_block_offset, 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, MAP_FIXED | MAP_SHARED | MAP_POPULATE,
+ 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));
+ }
+ }
}
- if(nb_shared_blocks == 1 && shared_block_offsets[0] == 0 && shared_block_offsets[1] == size) {
- 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;
+ 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;
+ 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;
return mem;
}
* 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_global(size_t size, const char *file, int line, int *shared_block_offsets=NULL, int nb_shared_blocks=-1) {
- int tmp_shared_block_offsets[2];
+static void *smpi_shared_malloc_global(size_t size, const char *file, int line, size_t *shared_block_offsets=NULL, int nb_shared_blocks=-1) {
+ size_t tmp_shared_block_offsets[2];
if(nb_shared_blocks == -1) {
nb_shared_blocks = 1;
shared_block_offsets = tmp_shared_block_offsets;
return mem;
}
-int smpi_is_shared(void* ptr){
+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) {
auto low = allocs_metadata.lower_bound(ptr);
- if (low->first==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)
+ 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<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 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>> result;
+ unsigned i_src=0, i_dst=0;
+ while(i_src < src.size() && i_dst < dst.size()) {
+ std::pair<size_t, size_t> 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) {
