1 /* Copyright (c) 2015-2019. 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 "src/smpi/include/smpi_actor.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 the portion of this region not present in the initial map.
79 auto found = std::find_if(initial_vm_map.begin(), initial_vm_map.end(), [&i](const simgrid::xbt::VmMap& m) {
80 return i->start_addr <= m.start_addr && m.start_addr < i->end_addr;
82 auto end_addr = (found == initial_vm_map.end() ? i->end_addr : found->start_addr);
83 smpi_data_exe_size = (char*)end_addr - smpi_data_exe_start;
88 xbt_die("Did not find my data segment.");
92 #if HAVE_SANITIZER_ADDRESS
93 #include <sanitizer/asan_interface.h>
94 static void* asan_safe_memcpy(void* dest, void* src, size_t n)
96 char* psrc = static_cast<char*>(src);
97 char* pdest = static_cast<char*>(dest);
98 for (size_t i = 0; i < n;) {
99 while (i < n && __asan_address_is_poisoned(psrc + i))
102 char* p = static_cast<char*>(__asan_region_is_poisoned(psrc + i, n - i));
103 size_t j = p ? (p - psrc) : n;
104 memcpy(pdest + i, psrc + i, j - i);
111 #define asan_safe_memcpy(dest, src, n) memcpy(dest, src, n)
114 /** Map a given SMPI privatization segment (make a SMPI process active) */
115 void smpi_switch_data_segment(simgrid::s4u::ActorPtr actor)
117 if (smpi_loaded_page == actor->get_pid()) // no need to switch, we've already loaded the one we want
121 smpi_really_switch_data_segment(actor);
124 /** Map a given SMPI privatization segment (make a SMPI process active) even if SMPI thinks it is already active
126 * When doing a state restoration, the state of the restored variables might not be consistent with the state of the
127 * virtual memory. In this case, we to change the data segment.
129 void smpi_really_switch_data_segment(simgrid::s4u::ActorPtr actor)
131 if (smpi_data_exe_size == 0) // no need to switch
134 #if HAVE_PRIVATIZATION
135 // FIXME, cross-process support (mmap across process when necessary)
136 XBT_DEBUG("Switching data frame to the one of process %ld", actor->get_pid());
137 simgrid::smpi::ActorExt* process = smpi_process_remote(actor);
138 int current = process->privatized_region()->file_descriptor;
139 void* tmp = mmap(TOPAGE(smpi_data_exe_start), smpi_data_exe_size, PROT_RW, MAP_FIXED | MAP_SHARED, current, 0);
140 if (tmp != TOPAGE(smpi_data_exe_start))
141 xbt_die("Couldn't map the new region (errno %d): %s", errno, strerror(errno));
142 smpi_loaded_page = actor->get_pid();
146 int smpi_is_privatization_file(char* file)
148 const std::string buffer_path("/dev/shm/my-buffer-");
149 return buffer_path.compare(0, std::string::npos, file, buffer_path.length()) == 0;
153 * @brief Makes a backup of the segment in memory that stores the global variables of a process.
154 * This backup is then used to initialize the global variables for every single
155 * process that is added, regardless of the progress of the simulation.
157 void smpi_backup_global_memory_segment()
159 #if HAVE_PRIVATIZATION
160 smpi_get_executable_global_size();
161 initial_vm_map.clear();
162 initial_vm_map.shrink_to_fit();
164 XBT_DEBUG("bss+data segment found : size %d starting at %p", smpi_data_exe_size, smpi_data_exe_start);
166 if (smpi_data_exe_size == 0) { // no need to do anything as global variables don't exist
167 smpi_privatize_global_variables = SmpiPrivStrategies::NONE;
171 smpi_data_exe_copy = ::operator new(smpi_data_exe_size);
172 // Make a copy of the data segment. This clean copy is retained over the whole runtime
173 // of the simulation and can be used to initialize a dynamically added, new process.
174 asan_safe_memcpy(smpi_data_exe_copy, TOPAGE(smpi_data_exe_start), smpi_data_exe_size);
175 #else /* ! HAVE_PRIVATIZATION */
176 xbt_die("You are trying to use privatization on a system that does not support it. Don't.");
180 // Initializes the memory mapping for a single process and returns the privatization region
181 smpi_privatization_region_t smpi_init_global_memory_segment_process()
184 void* address = nullptr;
188 constexpr unsigned VAL_MASK = 0xffffffU;
189 static unsigned prev_val = VAL_MASK;
190 for (unsigned i = (prev_val + 1) & VAL_MASK; i != prev_val; i = (i + 1) & VAL_MASK) {
191 snprintf(path, sizeof(path), "/smpi-buffer-%06x", i);
192 file_descriptor = shm_open(path, O_RDWR | O_CREAT | O_EXCL, S_IRUSR | S_IWUSR);
193 if (file_descriptor != -1 || errno != EEXIST) {
198 if (file_descriptor < 0) {
199 if (errno == EMFILE) {
200 xbt_die("Impossible to create temporary file for memory mapping: %s\n\
201 The open() system call failed with the EMFILE error code (too many files). \n\n\
202 This means that you reached the system limits concerning the amount of files per process. \
203 This is not a surprise if you are trying to virtualize many processes on top of SMPI. \
204 Don't panic -- you should simply increase your system limits and try again. \n\n\
205 First, check what your limits are:\n\
206 cat /proc/sys/fs/file-max # Gives you the system-wide limit\n\
207 ulimit -Hn # Gives you the per process hard limit\n\
208 ulimit -Sn # Gives you the per process soft limit\n\
209 cat /proc/self/limits # Displays any per-process limitation (including the one given above)\n\n\
210 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. \
211 Ask the Internet about tutorials on how to increase the files limit such as: https://rtcamp.com/tutorials/linux/increase-open-files-limit/",
214 xbt_die("Impossible to create temporary file for memory mapping: %s", strerror(errno));
217 status = ftruncate(file_descriptor, smpi_data_exe_size);
219 xbt_die("Impossible to set the size of the temporary file for memory mapping");
221 /* Ask for a free region */
222 address = mmap(nullptr, smpi_data_exe_size, PROT_RW, MAP_SHARED, file_descriptor, 0);
223 if (address == MAP_FAILED)
224 xbt_die("Couldn't find a free region for memory mapping");
226 status = shm_unlink(path);
228 xbt_die("Impossible to unlink temporary file for memory mapping");
230 // initialize the values
231 asan_safe_memcpy(address, smpi_data_exe_copy, smpi_data_exe_size);
233 // store the address of the mapping for further switches
234 smpi_privatization_regions.emplace_back(s_smpi_privatization_region_t{address, file_descriptor});
236 return &smpi_privatization_regions.back();
239 void smpi_destroy_global_memory_segments(){
240 if (smpi_data_exe_size == 0) // no need to switch
242 #if HAVE_PRIVATIZATION
243 for (auto const& region : smpi_privatization_regions) {
244 if (munmap(region.address, smpi_data_exe_size) < 0)
245 XBT_WARN("Unmapping of fd %d failed: %s", region.file_descriptor, strerror(errno));
246 close(region.file_descriptor);
248 smpi_privatization_regions.clear();
249 ::operator delete(smpi_data_exe_copy);
253 static std::vector<unsigned char> sendbuffer;
254 static std::vector<unsigned char> recvbuffer;
256 //allocate a single buffer for all sends, growing it if needed
257 unsigned char* smpi_get_tmp_sendbuffer(size_t size)
259 if (not smpi_process()->replaying())
260 return new unsigned char[size];
261 // FIXME: a resize() may invalidate a previous pointer. Maybe we need to handle a queue of buffers with a reference
262 // counter. The same holds for smpi_get_tmp_recvbuffer.
263 if (sendbuffer.size() < size)
264 sendbuffer.resize(size);
265 return sendbuffer.data();
268 //allocate a single buffer for all recv
269 unsigned char* smpi_get_tmp_recvbuffer(size_t size)
271 if (not smpi_process()->replaying())
272 return new unsigned char[size];
273 if (recvbuffer.size() < size)
274 recvbuffer.resize(size);
275 return recvbuffer.data();
278 void smpi_free_tmp_buffer(const unsigned char* buf)
280 if (not smpi_process()->replaying())
284 void smpi_free_replay_tmp_buffers()
286 std::vector<unsigned char>().swap(sendbuffer);
287 std::vector<unsigned char>().swap(recvbuffer);