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"
27 #include "private.hpp"
29 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_memory, smpi, "Memory layout support for SMPI");
31 int smpi_loaded_page = -1;
32 char* smpi_data_exe_start = nullptr;
33 int smpi_data_exe_size = 0;
34 int smpi_privatize_global_variables;
36 static const int PROT_RWX = (PROT_READ | PROT_WRITE | PROT_EXEC);
37 static const int PROT_RW = (PROT_READ | PROT_WRITE );
38 XBT_ATTRIB_UNUSED static const int PROT_RX = (PROT_READ | PROT_EXEC );
40 void smpi_get_executable_global_size()
42 char buffer[PATH_MAX];
43 char* full_name = realpath(xbt_binary_name, buffer);
44 if (full_name == nullptr)
45 xbt_die("Could not resolve binary file name");
47 std::vector<simgrid::xbt::VmMap> map = simgrid::xbt::get_memory_map(getpid());
48 for (auto i = map.begin(); i != map.end() ; ++i) {
49 // TODO, In practice, this implementation would not detect a completely
50 // anonymous data segment. This does not happen in practice, however.
52 // File backed RW entry:
53 if (i->pathname == full_name && (i->prot & PROT_RWX) == PROT_RW) {
54 smpi_data_exe_start = (char*)i->start_addr;
55 smpi_data_exe_size = i->end_addr - i->start_addr;
57 /* Here we are making the assumption that a suitable empty region
58 following the rw- area is the end of the data segment. It would
59 be better to check with the size of the data segment. */
60 if (i != map.end() && i->pathname.empty() && (i->prot & PROT_RWX) == PROT_RW &&
61 (char*)i->start_addr == smpi_data_exe_start + smpi_data_exe_size) {
62 smpi_data_exe_size = (char*)i->end_addr - smpi_data_exe_start;
67 xbt_die("Did not find my data segment.");
71 #if HAVE_SANITIZE_ADDRESS
72 #include <sanitizer/asan_interface.h>
73 static void* asan_safe_memcpy(void* dest, void* src, size_t n)
75 char* psrc = static_cast<char*>(src);
76 char* pdest = static_cast<char*>(dest);
77 for (size_t i = 0; i < n;) {
78 while (i < n && __asan_address_is_poisoned(psrc + i))
81 char* p = static_cast<char*>(__asan_region_is_poisoned(psrc + i, n - i));
82 size_t j = p ? (p - psrc) : n;
83 memcpy(pdest + i, psrc + i, j - i);
90 #define asan_safe_memcpy(dest, src, n) memcpy(dest, src, n)
93 /** Map a given SMPI privatization segment (make a SMPI process active) */
94 void smpi_switch_data_segment(int dest) {
95 if (smpi_loaded_page == dest)//no need to switch, we've already loaded the one we want
99 smpi_really_switch_data_segment(dest);
102 /** Map a given SMPI privatization segment (make a SMPI process active) even if SMPI thinks it is already active
104 * When doing a state restoration, the state of the restored variables might not be consistent with the state of the
105 * virtual memory. In this case, we to change the data segment.
107 void smpi_really_switch_data_segment(int dest)
109 if (smpi_data_exe_size == 0) // no need to switch
112 #if HAVE_PRIVATIZATION
113 if(smpi_loaded_page==-1){//initial switch, do the copy from the real page here
114 for (int i=0; i< smpi_process_count(); i++){
115 asan_safe_memcpy(smpi_privatization_regions[i].address, TOPAGE(smpi_data_exe_start), smpi_data_exe_size);
119 // FIXME, cross-process support (mmap across process when necessary)
120 int current = smpi_privatization_regions[dest].file_descriptor;
121 XBT_DEBUG("Switching data frame to the one of process %d", dest);
123 mmap(TOPAGE(smpi_data_exe_start), smpi_data_exe_size, PROT_READ | PROT_WRITE, MAP_FIXED | MAP_SHARED, current, 0);
124 if (tmp != TOPAGE(smpi_data_exe_start))
125 xbt_die("Couldn't map the new region (errno %d): %s", errno, strerror(errno));
126 smpi_loaded_page = dest;
130 int smpi_is_privatization_file(char* file)
132 const std::string buffer_path("/dev/shm/my-buffer-");
133 return buffer_path.compare(0, std::string::npos, file, buffer_path.length()) == 0;
136 void smpi_initialize_global_memory_segments()
139 #if HAVE_PRIVATIZATION
140 smpi_get_executable_global_size();
142 XBT_DEBUG("bss+data segment found : size %d starting at %p", smpi_data_exe_size, smpi_data_exe_start);
144 if (smpi_data_exe_size == 0) { // no need to switch
145 smpi_privatize_global_variables=false;
149 smpi_privatization_regions = static_cast<smpi_privatization_region_t>(
150 xbt_malloc(smpi_process_count() * sizeof(struct s_smpi_privatization_region)));
152 for (int i=0; i< smpi_process_count(); i++){
153 // create SIMIX_process_count() mappings of this size with the same data inside
155 void* address = nullptr;
160 snprintf(path, sizeof(path), "/smpi-buffer-%06x", rand() % 0xffffffU);
161 file_descriptor = shm_open(path, O_RDWR | O_CREAT | O_EXCL, S_IRUSR | S_IWUSR);
162 } while (file_descriptor == -1 && errno == EEXIST);
163 if (file_descriptor < 0) {
164 if (errno == EMFILE) {
165 xbt_die("Impossible to create temporary file for memory mapping: %s\n\
166 The open() system call failed with the EMFILE error code (too many files). \n\n\
167 This means that you reached the system limits concerning the amount of files per process. \
168 This is not a surprise if you are trying to virtualize many processes on top of SMPI. \
169 Don't panic -- you should simply increase your system limits and try again. \n\n\
170 First, check what your limits are:\n\
171 cat /proc/sys/fs/file-max # Gives you the system-wide limit\n\
172 ulimit -Hn # Gives you the per process hard limit\n\
173 ulimit -Sn # Gives you the per process soft limit\n\
174 cat /proc/self/limits # Displays any per-process limitation (including the one given above)\n\n\
175 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. \
176 Ask the Internet about tutorials on how to increase the files limit such as: https://rtcamp.com/tutorials/linux/increase-open-files-limit/",
179 xbt_die("Impossible to create temporary file for memory mapping: %s", strerror(errno));
182 status = ftruncate(file_descriptor, smpi_data_exe_size);
184 xbt_die("Impossible to set the size of the temporary file for memory mapping");
186 /* Ask for a free region */
187 address = mmap(nullptr, smpi_data_exe_size, PROT_READ | PROT_WRITE, MAP_SHARED, file_descriptor, 0);
188 if (address == MAP_FAILED)
189 xbt_die("Couldn't find a free region for memory mapping");
191 status = shm_unlink(path);
193 xbt_die("Impossible to unlink temporary file for memory mapping");
195 // initialize the values
196 asan_safe_memcpy(address, TOPAGE(smpi_data_exe_start), smpi_data_exe_size);
198 // store the address of the mapping for further switches
199 smpi_privatization_regions[i].file_descriptor = file_descriptor;
200 smpi_privatization_regions[i].address = address;
202 #else /* ! HAVE_PRIVATIZATION */
203 smpi_privatize_global_variables = false;
204 xbt_die("You are trying to use privatization on a system that does not support it. Don't.");
209 void smpi_destroy_global_memory_segments(){
210 if (smpi_data_exe_size == 0) // no need to switch
212 #if HAVE_PRIVATIZATION
213 for (int i=0; i< smpi_process_count(); i++) {
214 if (munmap(smpi_privatization_regions[i].address, smpi_data_exe_size) < 0)
215 XBT_WARN("Unmapping of fd %d failed: %s", smpi_privatization_regions[i].file_descriptor, strerror(errno));
216 close(smpi_privatization_regions[i].file_descriptor);
218 xbt_free(smpi_privatization_regions);