1 /* Copyright (c) 2015-2017. The SimGrid Team. All rights reserved. */
3 /* This program is free software; you can redistribute it and/or modify it
4 * under the terms of the license (GNU LGPL) which comes with this package. */
18 #include <sys/types.h>
24 #include "src/internal_config.h"
25 #include "src/xbt/memory_map.hpp"
28 #include "private.hpp"
30 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_memory, smpi, "Memory layout support for SMPI");
32 int smpi_loaded_page = -1;
33 char* smpi_start_data_exe = nullptr;
34 int smpi_size_data_exe = 0;
35 int smpi_privatize_global_variables;
37 static const int PROT_RWX = (PROT_READ | PROT_WRITE | PROT_EXEC);
38 static const int PROT_RW = (PROT_READ | PROT_WRITE );
39 XBT_ATTRIB_UNUSED static const int PROT_RX = (PROT_READ | PROT_EXEC );
41 void smpi_get_executable_global_size()
43 char buffer[PATH_MAX];
44 char* full_name = realpath(xbt_binary_name, buffer);
45 if (full_name == nullptr)
46 xbt_die("Could not resolve binary file name");
48 std::vector<simgrid::xbt::VmMap> map = simgrid::xbt::get_memory_map(getpid());
49 for (auto i = map.begin(); i != map.end() ; ++i) {
50 // TODO, In practice, this implementation would not detect a completely
51 // anonymous data segment. This does not happen in practice, however.
53 // File backed RW entry:
54 if (i->pathname == full_name && (i->prot & PROT_RWX) == PROT_RW) {
55 smpi_start_data_exe = (char*) i->start_addr;
56 smpi_size_data_exe = i->end_addr - i->start_addr;
58 /* Here we are making the assumption that a suitable empty region
59 following the rw- area is the end of the data segment. It would
60 be better to check with the size of the data segment. */
61 if (i != map.end() && i->pathname.empty() && (i->prot & PROT_RWX) == PROT_RW
62 && (char*)i->start_addr == smpi_start_data_exe + smpi_size_data_exe) {
63 smpi_size_data_exe = (char*)i->end_addr - smpi_start_data_exe;
68 xbt_die("Did not find my data segment.");
72 #if HAVE_SANITIZE_ADDRESS
73 #include <sanitizer/asan_interface.h>
74 static void* asan_safe_memcpy(void* dest, void* src, size_t n)
76 char* psrc = static_cast<char*>(src);
77 char* pdest = static_cast<char*>(dest);
78 for (size_t i = 0; i < n;) {
79 while (i < n && __asan_address_is_poisoned(psrc + i))
82 char* p = static_cast<char*>(__asan_region_is_poisoned(psrc + i, n - i));
83 size_t j = p ? (p - psrc) : n;
84 memcpy(pdest + i, psrc + i, j - i);
91 #define asan_safe_memcpy(dest, src, n) memcpy(dest, src, n)
94 /** Map a given SMPI privatization segment (make a SMPI process active) */
95 void smpi_switch_data_segment(int dest) {
96 if (smpi_loaded_page == dest)//no need to switch, we've already loaded the one we want
100 smpi_really_switch_data_segment(dest);
103 /** Map a given SMPI privatization segment (make a SMPI process active) even if SMPI thinks it is already active
105 * When doing a state restoration, the state of the restored variables might not be consistent with the state of the
106 * virtual memory. In this case, we to change the data segment.
108 void smpi_really_switch_data_segment(int dest)
110 if(smpi_size_data_exe == 0)//no need to switch
113 #if HAVE_PRIVATIZATION
114 if(smpi_loaded_page==-1){//initial switch, do the copy from the real page here
115 for (int i=0; i< smpi_process_count(); i++){
116 asan_safe_memcpy(smpi_privatization_regions[i].address, TOPAGE(smpi_start_data_exe), smpi_size_data_exe);
120 // FIXME, cross-process support (mmap across process when necessary)
121 int current = smpi_privatization_regions[dest].file_descriptor;
122 XBT_DEBUG("Switching data frame to the one of process %d", dest);
124 mmap(TOPAGE(smpi_start_data_exe), smpi_size_data_exe, PROT_READ | PROT_WRITE, MAP_FIXED | MAP_SHARED, current, 0);
125 if (tmp != TOPAGE(smpi_start_data_exe))
126 xbt_die("Couldn't map the new region (errno %d): %s", errno, strerror(errno));
127 smpi_loaded_page = dest;
131 int smpi_is_privatization_file(char* file)
133 const std::string buffer_path("/dev/shm/my-buffer-");
134 return buffer_path.compare(0, std::string::npos, file, buffer_path.length()) == 0;
137 void smpi_initialize_global_memory_segments()
140 #if HAVE_PRIVATIZATION
141 smpi_get_executable_global_size();
143 XBT_DEBUG ("bss+data segment found : size %d starting at %p", smpi_size_data_exe, smpi_start_data_exe );
145 if (smpi_size_data_exe == 0){//no need to switch
146 smpi_privatize_global_variables=false;
150 smpi_privatization_regions = static_cast<smpi_privatization_region_t>(
151 xbt_malloc(smpi_process_count() * sizeof(struct s_smpi_privatization_region)));
153 for (int i=0; i< smpi_process_count(); i++){
154 // create SIMIX_process_count() mappings of this size with the same data inside
156 void* address = nullptr;
161 snprintf(path, sizeof(path), "/smpi-buffer-%06x", rand() % 0xffffffU);
162 file_descriptor = shm_open(path, O_RDWR | O_CREAT | O_EXCL, S_IRUSR | S_IWUSR);
163 } while (file_descriptor == -1 && errno == EEXIST);
164 if (file_descriptor < 0) {
165 if (errno == EMFILE) {
166 xbt_die("Impossible to create temporary file for memory mapping: %s\n\
167 The open() system call failed with the EMFILE error code (too many files). \n\n\
168 This means that you reached the system limits concerning the amount of files per process. \
169 This is not a surprise if you are trying to virtualize many processes on top of SMPI. \
170 Don't panic -- you should simply increase your system limits and try again. \n\n\
171 First, check what your limits are:\n\
172 cat /proc/sys/fs/file-max # Gives you the system-wide limit\n\
173 ulimit -Hn # Gives you the per process hard limit\n\
174 ulimit -Sn # Gives you the per process soft limit\n\
175 cat /proc/self/limits # Displays any per-process limitation (including the one given above)\n\n\
176 If one of these values is less than the amount of MPI processes that you try to run, then you got the explanation of this error. \
177 Ask the Internet about tutorials on how to increase the files limit such as: https://rtcamp.com/tutorials/linux/increase-open-files-limit/",
180 xbt_die("Impossible to create temporary file for memory mapping: %s", strerror(errno));
183 status = ftruncate(file_descriptor, smpi_size_data_exe);
185 xbt_die("Impossible to set the size of the temporary file for memory mapping");
187 /* Ask for a free region */
188 address = mmap(nullptr, smpi_size_data_exe, PROT_READ | PROT_WRITE, MAP_SHARED, file_descriptor, 0);
189 if (address == MAP_FAILED)
190 xbt_die("Couldn't find a free region for memory mapping");
192 status = shm_unlink(path);
194 xbt_die("Impossible to unlink temporary file for memory mapping");
196 // initialize the values
197 asan_safe_memcpy(address, TOPAGE(smpi_start_data_exe), smpi_size_data_exe);
199 // store the address of the mapping for further switches
200 smpi_privatization_regions[i].file_descriptor = file_descriptor;
201 smpi_privatization_regions[i].address = address;
203 #else /* ! HAVE_PRIVATIZATION */
204 smpi_privatize_global_variables = false;
205 xbt_die("You are trying to use privatization on a system that does not support it. Don't.");
210 void smpi_destroy_global_memory_segments(){
211 if (smpi_size_data_exe == 0)//no need to switch
213 #if HAVE_PRIVATIZATION
214 for (int i=0; i< smpi_process_count(); i++) {
215 if (munmap(smpi_privatization_regions[i].address, smpi_size_data_exe) < 0)
216 XBT_WARN("Unmapping of fd %d failed: %s", smpi_privatization_regions[i].file_descriptor, strerror(errno));
217 close(smpi_privatization_regions[i].file_descriptor);
219 xbt_free(smpi_privatization_regions);