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 constexpr int PROT_RWX = PROT_READ | PROT_WRITE | PROT_EXEC;
43 static constexpr int PROT_RW = PROT_READ | PROT_WRITE;
45 static void smpi_get_executable_global_size()
47 char buffer[PATH_MAX];
48 char* full_name = realpath(xbt_binary_name, buffer);
49 if (full_name == nullptr)
50 xbt_die("Could not resolve binary file name");
52 std::vector<simgrid::xbt::VmMap> map = simgrid::xbt::get_memory_map(getpid());
53 for (auto i = map.begin(); i != map.end() ; ++i) {
54 // TODO, In practice, this implementation would not detect a completely
55 // anonymous data segment. This does not happen in practice, however.
57 // File backed RW entry:
58 if (i->pathname == full_name && (i->prot & PROT_RWX) == PROT_RW) {
59 smpi_data_exe_start = (char*)i->start_addr;
60 smpi_data_exe_size = i->end_addr - i->start_addr;
62 /* Here we are making the assumption that a suitable empty region
63 following the rw- area is the end of the data segment. It would
64 be better to check with the size of the data segment. */
65 if (i != map.end() && i->pathname.empty() && (i->prot & PROT_RWX) == PROT_RW &&
66 (char*)i->start_addr == smpi_data_exe_start + smpi_data_exe_size) {
67 smpi_data_exe_size = (char*)i->end_addr - smpi_data_exe_start;
72 xbt_die("Did not find my data segment.");
76 #if HAVE_SANITIZE_ADDRESS
77 #include <sanitizer/asan_interface.h>
78 static void* asan_safe_memcpy(void* dest, void* src, size_t n)
80 char* psrc = static_cast<char*>(src);
81 char* pdest = static_cast<char*>(dest);
82 for (size_t i = 0; i < n;) {
83 while (i < n && __asan_address_is_poisoned(psrc + i))
86 char* p = static_cast<char*>(__asan_region_is_poisoned(psrc + i, n - i));
87 size_t j = p ? (p - psrc) : n;
88 memcpy(pdest + i, psrc + i, j - i);
95 #define asan_safe_memcpy(dest, src, n) memcpy(dest, src, n)
98 /** Map a given SMPI privatization segment (make a SMPI process active) */
99 void smpi_switch_data_segment(simgrid::s4u::ActorPtr actor)
101 if (smpi_loaded_page == actor->get_pid()) // no need to switch, we've already loaded the one we want
105 smpi_really_switch_data_segment(actor);
108 /** Map a given SMPI privatization segment (make a SMPI process active) even if SMPI thinks it is already active
110 * When doing a state restoration, the state of the restored variables might not be consistent with the state of the
111 * virtual memory. In this case, we to change the data segment.
113 void smpi_really_switch_data_segment(simgrid::s4u::ActorPtr actor)
115 if (smpi_data_exe_size == 0) // no need to switch
118 #if HAVE_PRIVATIZATION
119 // FIXME, cross-process support (mmap across process when necessary)
120 XBT_DEBUG("Switching data frame to the one of process %ld", actor->get_pid());
121 simgrid::smpi::Process* process = smpi_process_remote(actor);
122 int current = process->privatized_region()->file_descriptor;
123 void* tmp = mmap(TOPAGE(smpi_data_exe_start), smpi_data_exe_size, PROT_RW, 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 = actor->get_pid();
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;
137 * @brief Makes a backup of the segment in memory that stores the global variables of a process.
138 * This backup is then used to initialize the global variables for every single
139 * process that is added, regardless of the progress of the simulation.
141 void smpi_backup_global_memory_segment()
143 xbt_assert(HAVE_PRIVATIZATION, "You are trying to use privatization on a system that does not support it. Don't.");
144 smpi_get_executable_global_size();
146 XBT_DEBUG("bss+data segment found : size %d starting at %p", smpi_data_exe_size, smpi_data_exe_start);
148 if (smpi_data_exe_size == 0) { // no need to do anything as global variables don't exist
149 smpi_privatize_global_variables = SmpiPrivStrategies::None;
153 smpi_data_exe_copy = ::operator new(smpi_data_exe_size);
154 // Make a copy of the data segment. This clean copy is retained over the whole runtime
155 // of the simulation and can be used to initialize a dynamically added, new process.
156 asan_safe_memcpy(smpi_data_exe_copy, TOPAGE(smpi_data_exe_start), smpi_data_exe_size);
159 // Initializes the memory mapping for a single process and returns the privatization region
160 smpi_privatization_region_t smpi_init_global_memory_segment_process()
163 void* address = nullptr;
168 snprintf(path, sizeof(path), "/smpi-buffer-%06x", rand() % 0xffffffU);
169 file_descriptor = shm_open(path, O_RDWR | O_CREAT | O_EXCL, S_IRUSR | S_IWUSR);
170 } while (file_descriptor == -1 && errno == EEXIST);
171 if (file_descriptor < 0) {
172 if (errno == EMFILE) {
173 xbt_die("Impossible to create temporary file for memory mapping: %s\n\
174 The open() system call failed with the EMFILE error code (too many files). \n\n\
175 This means that you reached the system limits concerning the amount of files per process. \
176 This is not a surprise if you are trying to virtualize many processes on top of SMPI. \
177 Don't panic -- you should simply increase your system limits and try again. \n\n\
178 First, check what your limits are:\n\
179 cat /proc/sys/fs/file-max # Gives you the system-wide limit\n\
180 ulimit -Hn # Gives you the per process hard limit\n\
181 ulimit -Sn # Gives you the per process soft limit\n\
182 cat /proc/self/limits # Displays any per-process limitation (including the one given above)\n\n\
183 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. \
184 Ask the Internet about tutorials on how to increase the files limit such as: https://rtcamp.com/tutorials/linux/increase-open-files-limit/",
187 xbt_die("Impossible to create temporary file for memory mapping: %s", strerror(errno));
190 status = ftruncate(file_descriptor, smpi_data_exe_size);
192 xbt_die("Impossible to set the size of the temporary file for memory mapping");
194 /* Ask for a free region */
195 address = mmap(nullptr, smpi_data_exe_size, PROT_RW, MAP_SHARED, file_descriptor, 0);
196 if (address == MAP_FAILED)
197 xbt_die("Couldn't find a free region for memory mapping");
199 status = shm_unlink(path);
201 xbt_die("Impossible to unlink temporary file for memory mapping");
203 // initialize the values
204 asan_safe_memcpy(address, smpi_data_exe_copy, smpi_data_exe_size);
206 // store the address of the mapping for further switches
207 smpi_privatization_regions.emplace_back(s_smpi_privatization_region_t{address, file_descriptor});
209 return &smpi_privatization_regions.back();
212 void smpi_destroy_global_memory_segments(){
213 if (smpi_data_exe_size == 0) // no need to switch
215 #if HAVE_PRIVATIZATION
216 for (auto const& region : smpi_privatization_regions) {
217 if (munmap(region.address, smpi_data_exe_size) < 0)
218 XBT_WARN("Unmapping of fd %d failed: %s", region.file_descriptor, strerror(errno));
219 close(region.file_descriptor);
221 smpi_privatization_regions.clear();
222 ::operator delete(smpi_data_exe_copy);
226 static int sendbuffer_size = 0;
227 static char* sendbuffer = nullptr;
228 static int recvbuffer_size = 0;
229 static char* recvbuffer = nullptr;
231 //allocate a single buffer for all sends, growing it if needed
232 void* smpi_get_tmp_sendbuffer(int size)
234 if (not smpi_process()->replaying())
235 return xbt_malloc(size);
236 if (sendbuffer_size<size){
237 sendbuffer=static_cast<char*>(xbt_realloc(sendbuffer,size));
238 sendbuffer_size=size;
243 //allocate a single buffer for all recv
244 void* smpi_get_tmp_recvbuffer(int size){
245 if (not smpi_process()->replaying())
246 return xbt_malloc(size);
247 if (recvbuffer_size<size){
248 recvbuffer=static_cast<char*>(xbt_realloc(recvbuffer,size));
249 recvbuffer_size=size;
254 void smpi_free_tmp_buffer(void* buf){
255 if (not smpi_process()->replaying())
259 void smpi_free_replay_tmp_buffers(){
260 xbt_free(sendbuffer);
261 xbt_free(recvbuffer);