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. */
16 #include <sys/types.h>
23 #include "src/internal_config.h"
24 #include "src/xbt/memory_map.hpp"
26 #include "private.hpp"
27 #include "smpi_process.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 SmpiPrivStrategies smpi_privatize_global_variables;
35 static void* smpi_data_exe_copy;
37 // Initialized by smpi_prepare_global_memory_segment().
38 static std::vector<simgrid::xbt::VmMap> initial_vm_map;
40 // We keep a copy of all the privatization regions: We can then delete everything easily by iterating over this
41 // collection and nothing can be leaked. We could also iterate over all actors but we would have to be diligent when two
42 // actors use the same privatization region (so, smart pointers would have to be used etc.)
43 // Use a std::deque so that pointers remain valid after push_back().
44 static std::deque<s_smpi_privatization_region_t> smpi_privatization_regions;
46 static constexpr int PROT_RWX = PROT_READ | PROT_WRITE | PROT_EXEC;
47 static constexpr int PROT_RW = PROT_READ | PROT_WRITE;
49 /** Take a snapshot of the process' memory map.
51 void smpi_prepare_global_memory_segment()
53 initial_vm_map = simgrid::xbt::get_memory_map(getpid());
56 static void smpi_get_executable_global_size()
58 char buffer[PATH_MAX];
59 char* full_name = realpath(xbt_binary_name, buffer);
60 if (full_name == nullptr)
61 xbt_die("Could not resolve binary file name");
63 std::vector<simgrid::xbt::VmMap> map = simgrid::xbt::get_memory_map(getpid());
64 for (auto i = map.begin(); i != map.end() ; ++i) {
65 // TODO, In practice, this implementation would not detect a completely
66 // anonymous data segment. This does not happen in practice, however.
68 // File backed RW entry:
69 if (i->pathname == full_name && (i->prot & PROT_RWX) == PROT_RW) {
70 smpi_data_exe_start = (char*)i->start_addr;
71 smpi_data_exe_size = i->end_addr - i->start_addr;
72 /* Here we are making the assumption that a suitable empty region
73 following the rw- area is the end of the data segment. It would
74 be better to check with the size of the data segment. */
76 if (i != map.end() && i->pathname.empty() && (i->prot & PROT_RWX) == PROT_RW &&
77 (char*)i->start_addr == smpi_data_exe_start + smpi_data_exe_size) {
78 // Only count this region if it was not already present in the initial map.
79 auto found = std::find_if(begin(initial_vm_map), end(initial_vm_map),
80 [&i](const simgrid::xbt::VmMap& m) { return m.start_addr == i->start_addr; });
81 if (found == end(initial_vm_map))
82 smpi_data_exe_size = (char*)i->end_addr - smpi_data_exe_start;
87 xbt_die("Did not find my data segment.");
91 #if HAVE_SANITIZE_ADDRESS
92 #include <sanitizer/asan_interface.h>
93 static void* asan_safe_memcpy(void* dest, void* src, size_t n)
95 char* psrc = static_cast<char*>(src);
96 char* pdest = static_cast<char*>(dest);
97 for (size_t i = 0; i < n;) {
98 while (i < n && __asan_address_is_poisoned(psrc + i))
101 char* p = static_cast<char*>(__asan_region_is_poisoned(psrc + i, n - i));
102 size_t j = p ? (p - psrc) : n;
103 memcpy(pdest + i, psrc + i, j - i);
110 #define asan_safe_memcpy(dest, src, n) memcpy(dest, src, n)
113 /** Map a given SMPI privatization segment (make a SMPI process active) */
114 void smpi_switch_data_segment(simgrid::s4u::ActorPtr actor)
116 if (smpi_loaded_page == actor->get_pid()) // no need to switch, we've already loaded the one we want
120 smpi_really_switch_data_segment(actor);
123 /** Map a given SMPI privatization segment (make a SMPI process active) even if SMPI thinks it is already active
125 * When doing a state restoration, the state of the restored variables might not be consistent with the state of the
126 * virtual memory. In this case, we to change the data segment.
128 void smpi_really_switch_data_segment(simgrid::s4u::ActorPtr actor)
130 if (smpi_data_exe_size == 0) // no need to switch
133 #if HAVE_PRIVATIZATION
134 // FIXME, cross-process support (mmap across process when necessary)
135 XBT_DEBUG("Switching data frame to the one of process %ld", actor->get_pid());
136 simgrid::smpi::Process* process = smpi_process_remote(actor);
137 int current = process->privatized_region()->file_descriptor;
138 void* tmp = mmap(TOPAGE(smpi_data_exe_start), smpi_data_exe_size, PROT_RW, MAP_FIXED | MAP_SHARED, current, 0);
139 if (tmp != TOPAGE(smpi_data_exe_start))
140 xbt_die("Couldn't map the new region (errno %d): %s", errno, strerror(errno));
141 smpi_loaded_page = actor->get_pid();
145 int smpi_is_privatization_file(char* file)
147 const std::string buffer_path("/dev/shm/my-buffer-");
148 return buffer_path.compare(0, std::string::npos, file, buffer_path.length()) == 0;
152 * @brief Makes a backup of the segment in memory that stores the global variables of a process.
153 * This backup is then used to initialize the global variables for every single
154 * process that is added, regardless of the progress of the simulation.
156 void smpi_backup_global_memory_segment()
158 xbt_assert(HAVE_PRIVATIZATION, "You are trying to use privatization on a system that does not support it. Don't.");
159 smpi_get_executable_global_size();
160 initial_vm_map.clear();
161 initial_vm_map.shrink_to_fit();
163 XBT_DEBUG("bss+data segment found : size %d starting at %p", smpi_data_exe_size, smpi_data_exe_start);
165 if (smpi_data_exe_size == 0) { // no need to do anything as global variables don't exist
166 smpi_privatize_global_variables = SmpiPrivStrategies::None;
170 smpi_data_exe_copy = ::operator new(smpi_data_exe_size);
171 // Make a copy of the data segment. This clean copy is retained over the whole runtime
172 // of the simulation and can be used to initialize a dynamically added, new process.
173 asan_safe_memcpy(smpi_data_exe_copy, TOPAGE(smpi_data_exe_start), smpi_data_exe_size);
176 // Initializes the memory mapping for a single process and returns the privatization region
177 smpi_privatization_region_t smpi_init_global_memory_segment_process()
180 void* address = nullptr;
185 snprintf(path, sizeof(path), "/smpi-buffer-%06x", rand() % 0xffffffU);
186 file_descriptor = shm_open(path, O_RDWR | O_CREAT | O_EXCL, S_IRUSR | S_IWUSR);
187 } while (file_descriptor == -1 && errno == EEXIST);
188 if (file_descriptor < 0) {
189 if (errno == EMFILE) {
190 xbt_die("Impossible to create temporary file for memory mapping: %s\n\
191 The open() system call failed with the EMFILE error code (too many files). \n\n\
192 This means that you reached the system limits concerning the amount of files per process. \
193 This is not a surprise if you are trying to virtualize many processes on top of SMPI. \
194 Don't panic -- you should simply increase your system limits and try again. \n\n\
195 First, check what your limits are:\n\
196 cat /proc/sys/fs/file-max # Gives you the system-wide limit\n\
197 ulimit -Hn # Gives you the per process hard limit\n\
198 ulimit -Sn # Gives you the per process soft limit\n\
199 cat /proc/self/limits # Displays any per-process limitation (including the one given above)\n\n\
200 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. \
201 Ask the Internet about tutorials on how to increase the files limit such as: https://rtcamp.com/tutorials/linux/increase-open-files-limit/",
204 xbt_die("Impossible to create temporary file for memory mapping: %s", strerror(errno));
207 status = ftruncate(file_descriptor, smpi_data_exe_size);
209 xbt_die("Impossible to set the size of the temporary file for memory mapping");
211 /* Ask for a free region */
212 address = mmap(nullptr, smpi_data_exe_size, PROT_RW, MAP_SHARED, file_descriptor, 0);
213 if (address == MAP_FAILED)
214 xbt_die("Couldn't find a free region for memory mapping");
216 status = shm_unlink(path);
218 xbt_die("Impossible to unlink temporary file for memory mapping");
220 // initialize the values
221 asan_safe_memcpy(address, smpi_data_exe_copy, smpi_data_exe_size);
223 // store the address of the mapping for further switches
224 smpi_privatization_regions.emplace_back(s_smpi_privatization_region_t{address, file_descriptor});
226 return &smpi_privatization_regions.back();
229 void smpi_destroy_global_memory_segments(){
230 if (smpi_data_exe_size == 0) // no need to switch
232 #if HAVE_PRIVATIZATION
233 for (auto const& region : smpi_privatization_regions) {
234 if (munmap(region.address, smpi_data_exe_size) < 0)
235 XBT_WARN("Unmapping of fd %d failed: %s", region.file_descriptor, strerror(errno));
236 close(region.file_descriptor);
238 smpi_privatization_regions.clear();
239 ::operator delete(smpi_data_exe_copy);
243 static int sendbuffer_size = 0;
244 static char* sendbuffer = nullptr;
245 static int recvbuffer_size = 0;
246 static char* recvbuffer = nullptr;
248 //allocate a single buffer for all sends, growing it if needed
249 void* smpi_get_tmp_sendbuffer(int size)
251 if (not smpi_process()->replaying())
252 return xbt_malloc(size);
253 if (sendbuffer_size<size){
254 sendbuffer=static_cast<char*>(xbt_realloc(sendbuffer,size));
255 sendbuffer_size=size;
260 //allocate a single buffer for all recv
261 void* smpi_get_tmp_recvbuffer(int size){
262 if (not smpi_process()->replaying())
263 return xbt_malloc(size);
264 if (recvbuffer_size<size){
265 recvbuffer=static_cast<char*>(xbt_realloc(recvbuffer,size));
266 recvbuffer_size=size;
271 void smpi_free_tmp_buffer(void* buf){
272 if (not smpi_process()->replaying())
276 void smpi_free_replay_tmp_buffers(){
277 xbt_free(sendbuffer);
278 xbt_free(recvbuffer);