1 /* Copyright (c) 2015-2018. 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. */
15 #include <sys/types.h>
22 #include "src/internal_config.h"
23 #include "src/xbt/memory_map.hpp"
25 #include "private.hpp"
26 #include "smpi_process.hpp"
28 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_memory, smpi, "Memory layout support for SMPI");
30 int smpi_loaded_page = -1;
31 char* smpi_data_exe_start = nullptr;
32 int smpi_data_exe_size = 0;
33 SmpiPrivStrategies smpi_privatize_global_variables;
34 static void* smpi_data_exe_copy;
36 // We keep a copy of all the privatization regions: We can then delete everything easily by iterating over this
37 // collection and nothing can be leaked. We could also iterate over all actors but we would have to be diligent when two
38 // actors use the same privatization region (so, smart pointers would have to be used etc.)
39 // Use a std::deque so that pointers remain valid after push_back().
40 static std::deque<s_smpi_privatization_region_t> smpi_privatization_regions;
42 static const int PROT_RWX = (PROT_READ | PROT_WRITE | PROT_EXEC);
43 static const int PROT_RW = (PROT_READ | PROT_WRITE );
44 XBT_ATTRIB_UNUSED static const int PROT_RX = (PROT_READ | PROT_EXEC );
46 void smpi_get_executable_global_size()
48 char buffer[PATH_MAX];
49 char* full_name = realpath(xbt_binary_name, buffer);
50 if (full_name == nullptr)
51 xbt_die("Could not resolve binary file name");
53 std::vector<simgrid::xbt::VmMap> map = simgrid::xbt::get_memory_map(getpid());
54 for (auto i = map.begin(); i != map.end() ; ++i) {
55 // TODO, In practice, this implementation would not detect a completely
56 // anonymous data segment. This does not happen in practice, however.
58 // File backed RW entry:
59 if (i->pathname == full_name && (i->prot & PROT_RWX) == PROT_RW) {
60 smpi_data_exe_start = (char*)i->start_addr;
61 smpi_data_exe_size = i->end_addr - i->start_addr;
63 /* Here we are making the assumption that a suitable empty region
64 following the rw- area is the end of the data segment. It would
65 be better to check with the size of the data segment. */
66 if (i != map.end() && i->pathname.empty() && (i->prot & PROT_RWX) == PROT_RW &&
67 (char*)i->start_addr == smpi_data_exe_start + smpi_data_exe_size) {
68 smpi_data_exe_size = (char*)i->end_addr - smpi_data_exe_start;
73 xbt_die("Did not find my data segment.");
77 #if HAVE_SANITIZE_ADDRESS
78 #include <sanitizer/asan_interface.h>
79 static void* asan_safe_memcpy(void* dest, void* src, size_t n)
81 char* psrc = static_cast<char*>(src);
82 char* pdest = static_cast<char*>(dest);
83 for (size_t i = 0; i < n;) {
84 while (i < n && __asan_address_is_poisoned(psrc + i))
87 char* p = static_cast<char*>(__asan_region_is_poisoned(psrc + i, n - i));
88 size_t j = p ? (p - psrc) : n;
89 memcpy(pdest + i, psrc + i, j - i);
96 #define asan_safe_memcpy(dest, src, n) memcpy(dest, src, n)
99 /** Map a given SMPI privatization segment (make a SMPI process active) */
100 void smpi_switch_data_segment(simgrid::s4u::ActorPtr actor)
102 if (smpi_loaded_page == actor->get_pid()) // no need to switch, we've already loaded the one we want
106 smpi_really_switch_data_segment(actor);
109 /** Map a given SMPI privatization segment (make a SMPI process active) even if SMPI thinks it is already active
111 * When doing a state restoration, the state of the restored variables might not be consistent with the state of the
112 * virtual memory. In this case, we to change the data segment.
114 void smpi_really_switch_data_segment(simgrid::s4u::ActorPtr actor)
116 if (smpi_data_exe_size == 0) // no need to switch
119 #if HAVE_PRIVATIZATION
120 // FIXME, cross-process support (mmap across process when necessary)
121 XBT_DEBUG("Switching data frame to the one of process %ld", actor->get_pid());
122 simgrid::smpi::Process* process = smpi_process_remote(actor);
123 int current = process->privatized_region()->file_descriptor;
125 mmap(TOPAGE(smpi_data_exe_start), smpi_data_exe_size, PROT_READ | PROT_WRITE, MAP_FIXED | MAP_SHARED, current, 0);
126 if (tmp != TOPAGE(smpi_data_exe_start))
127 xbt_die("Couldn't map the new region (errno %d): %s", errno, strerror(errno));
128 smpi_loaded_page = actor->get_pid();
132 int smpi_is_privatization_file(char* file)
134 const std::string buffer_path("/dev/shm/my-buffer-");
135 return buffer_path.compare(0, std::string::npos, file, buffer_path.length()) == 0;
139 * @brief Makes a backup of the segment in memory that stores the global variables of a process.
140 * This backup is then used to initialize the global variables for every single
141 * process that is added, regardless of the progress of the simulation.
143 void smpi_backup_global_memory_segment()
145 xbt_assert(HAVE_PRIVATIZATION, "You are trying to use privatization on a system that does not support it. Don't.");
146 smpi_get_executable_global_size();
148 XBT_DEBUG("bss+data segment found : size %d starting at %p", smpi_data_exe_size, smpi_data_exe_start);
150 if (smpi_data_exe_size == 0) { // no need to do anything as global variables don't exist
151 smpi_privatize_global_variables = SmpiPrivStrategies::None;
155 smpi_data_exe_copy = ::operator new(smpi_data_exe_size);
156 // Make a copy of the data segment. This clean copy is retained over the whole runtime
157 // of the simulation and can be used to initialize a dynamically added, new process.
158 asan_safe_memcpy(smpi_data_exe_copy, TOPAGE(smpi_data_exe_start), smpi_data_exe_size);
161 // Initializes the memory mapping for a single process and returns the privatization region
162 smpi_privatization_region_t smpi_init_global_memory_segment_process()
165 void* address = nullptr;
170 snprintf(path, sizeof(path), "/smpi-buffer-%06x", rand() % 0xffffffU);
171 file_descriptor = shm_open(path, O_RDWR | O_CREAT | O_EXCL, S_IRUSR | S_IWUSR);
172 } while (file_descriptor == -1 && errno == EEXIST);
173 if (file_descriptor < 0) {
174 if (errno == EMFILE) {
175 xbt_die("Impossible to create temporary file for memory mapping: %s\n\
176 The open() system call failed with the EMFILE error code (too many files). \n\n\
177 This means that you reached the system limits concerning the amount of files per process. \
178 This is not a surprise if you are trying to virtualize many processes on top of SMPI. \
179 Don't panic -- you should simply increase your system limits and try again. \n\n\
180 First, check what your limits are:\n\
181 cat /proc/sys/fs/file-max # Gives you the system-wide limit\n\
182 ulimit -Hn # Gives you the per process hard limit\n\
183 ulimit -Sn # Gives you the per process soft limit\n\
184 cat /proc/self/limits # Displays any per-process limitation (including the one given above)\n\n\
185 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. \
186 Ask the Internet about tutorials on how to increase the files limit such as: https://rtcamp.com/tutorials/linux/increase-open-files-limit/",
189 xbt_die("Impossible to create temporary file for memory mapping: %s", strerror(errno));
192 status = ftruncate(file_descriptor, smpi_data_exe_size);
194 xbt_die("Impossible to set the size of the temporary file for memory mapping");
196 /* Ask for a free region */
197 address = mmap(nullptr, smpi_data_exe_size, PROT_READ | PROT_WRITE, MAP_SHARED, file_descriptor, 0);
198 if (address == MAP_FAILED)
199 xbt_die("Couldn't find a free region for memory mapping");
201 status = shm_unlink(path);
203 xbt_die("Impossible to unlink temporary file for memory mapping");
205 // initialize the values
206 asan_safe_memcpy(address, smpi_data_exe_copy, smpi_data_exe_size);
208 // store the address of the mapping for further switches
209 smpi_privatization_regions.emplace_back(s_smpi_privatization_region_t{address, file_descriptor});
211 return &smpi_privatization_regions.back();
214 void smpi_destroy_global_memory_segments(){
215 if (smpi_data_exe_size == 0) // no need to switch
217 #if HAVE_PRIVATIZATION
218 for (auto const& region : smpi_privatization_regions) {
219 if (munmap(region.address, smpi_data_exe_size) < 0)
220 XBT_WARN("Unmapping of fd %d failed: %s", region.file_descriptor, strerror(errno));
221 close(region.file_descriptor);
223 smpi_privatization_regions.clear();
224 ::operator delete(smpi_data_exe_copy);
228 static int sendbuffer_size = 0;
229 static char* sendbuffer = nullptr;
230 static int recvbuffer_size = 0;
231 static char* recvbuffer = nullptr;
233 //allocate a single buffer for all sends, growing it if needed
234 void* smpi_get_tmp_sendbuffer(int size)
236 if (not smpi_process()->replaying())
237 return xbt_malloc(size);
238 if (sendbuffer_size<size){
239 sendbuffer=static_cast<char*>(xbt_realloc(sendbuffer,size));
240 sendbuffer_size=size;
245 //allocate a single buffer for all recv
246 void* smpi_get_tmp_recvbuffer(int size){
247 if (not smpi_process()->replaying())
248 return xbt_malloc(size);
249 if (recvbuffer_size<size){
250 recvbuffer=static_cast<char*>(xbt_realloc(recvbuffer,size));
251 recvbuffer_size=size;
256 void smpi_free_tmp_buffer(void* buf){
257 if (not smpi_process()->replaying())
261 void smpi_free_replay_tmp_buffers(){
262 xbt_free(sendbuffer);
263 xbt_free(recvbuffer);