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 xbt_assert(full_name != nullptr, "Could not resolve real path of binary file '%s'", xbt_binary_name);
62 std::vector<simgrid::xbt::VmMap> map = simgrid::xbt::get_memory_map(getpid());
63 for (auto i = map.begin(); i != map.end() ; ++i) {
64 // TODO, In practice, this implementation would not detect a completely
65 // anonymous data segment. This does not happen in practice, however.
67 // File backed RW entry:
68 if (i->pathname == full_name && (i->prot & PROT_RWX) == PROT_RW) {
69 smpi_data_exe_start = (char*)i->start_addr;
70 smpi_data_exe_size = i->end_addr - i->start_addr;
71 /* Here we are making the assumption that a suitable empty region
72 following the rw- area is the end of the data segment. It would
73 be better to check with the size of the data segment. */
75 if (i != map.end() && i->pathname.empty() && (i->prot & PROT_RWX) == PROT_RW &&
76 (char*)i->start_addr == smpi_data_exe_start + smpi_data_exe_size) {
77 // Only count the portion of this region not present in the initial map.
78 auto found = std::find_if(initial_vm_map.begin(), initial_vm_map.end(), [&i](const simgrid::xbt::VmMap& m) {
79 return i->start_addr <= m.start_addr && m.start_addr < i->end_addr;
81 auto end_addr = (found == initial_vm_map.end() ? i->end_addr : found->start_addr);
82 smpi_data_exe_size = (char*)end_addr - smpi_data_exe_start;
87 xbt_die("Did not find my data segment.");
91 #if HAVE_SANITIZER_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))
114 * @brief Uses shm_open to get a temporary shm, and returns its file descriptor.
116 int smpi_temp_shm_get()
118 constexpr unsigned VAL_MASK = 0xffffffffUL;
119 static unsigned prev_val = VAL_MASK;
120 char shmname[32]; // cannot be longer than PSHMNAMLEN = 31 on macOS (shm_open raises ENAMETOOLONG otherwise)
123 for (unsigned i = (prev_val + 1) & VAL_MASK; i != prev_val; i = (i + 1) & VAL_MASK) {
124 snprintf(shmname, sizeof(shmname), "/smpi-buffer-%016x", i);
125 fd = shm_open(shmname, O_RDWR | O_CREAT | O_EXCL, S_IRUSR | S_IWUSR);
126 if (fd != -1 || errno != EEXIST) {
132 if (errno == EMFILE) {
133 xbt_die("Impossible to create temporary file for memory mapping: %s\n\
134 The shm_open() system call failed with the EMFILE error code (too many files). \n\n\
135 This means that you reached the system limits concerning the amount of files per process. \
136 This is not a surprise if you are trying to virtualize many processes on top of SMPI. \
137 Don't panic -- you should simply increase your system limits and try again. \n\n\
138 First, check what your limits are:\n\
139 cat /proc/sys/fs/file-max # Gives you the system-wide limit\n\
140 ulimit -Hn # Gives you the per process hard limit\n\
141 ulimit -Sn # Gives you the per process soft limit\n\
142 cat /proc/self/limits # Displays any per-process limitation (including the one given above)\n\n\
143 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. \
144 Ask the Internet about tutorials on how to increase the files limit such as: https://rtcamp.com/tutorials/linux/increase-open-files-limit/",
147 xbt_die("Impossible to create temporary file for memory mapping. shm_open: %s", strerror(errno));
149 XBT_DEBUG("Got temporary shm %s (fd = %d)", shmname, fd);
150 if (shm_unlink(shmname) < 0)
151 XBT_WARN("Could not early unlink %s. shm_unlink: %s", shmname, strerror(errno));
156 * @brief Mmap a region of size bytes from temporary shm with file descriptor fd.
158 void* smpi_temp_shm_mmap(int fd, size_t size)
161 if (fstat(fd, &st) != 0)
162 xbt_die("Could not stat fd %d: %s", fd, strerror(errno));
163 if (static_cast<off_t>(size) > st.st_size && ftruncate(fd, static_cast<off_t>(size)) != 0)
164 xbt_die("Could not truncate fd %d to %zu: %s", fd, size, strerror(errno));
165 void* mem = mmap(nullptr, size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
166 if (mem == MAP_FAILED) {
167 xbt_die("Failed to map fd %d with size %zu: %s\n"
168 "If you are running a lot of ranks, you may be exceeding the amount of mappings allowed per process.\n"
169 "On Linux systems, change this value with sudo sysctl -w vm.max_map_count=newvalue (default value: 65536)\n"
171 "https://simgrid.org/doc/latest/Configuring_SimGrid.html#configuring-the-user-code-virtualization for more "
173 fd, size, strerror(errno));
178 /** Map a given SMPI privatization segment (make a SMPI process active)
180 * When doing a state restoration, the state of the restored variables might not be consistent with the state of the
181 * virtual memory. In this case, we to change the data segment.
183 void smpi_switch_data_segment(simgrid::s4u::ActorPtr actor)
185 if (smpi_loaded_page == actor->get_pid()) // no need to switch, we've already loaded the one we want
188 if (smpi_data_exe_size == 0) // no need to switch
191 #if HAVE_PRIVATIZATION
192 // FIXME, cross-process support (mmap across process when necessary)
193 XBT_DEBUG("Switching data frame to the one of process %ld", actor->get_pid());
194 simgrid::smpi::ActorExt* process = smpi_process_remote(actor);
195 int current = process->privatized_region()->file_descriptor;
196 void* tmp = mmap(TOPAGE(smpi_data_exe_start), smpi_data_exe_size, PROT_RW, MAP_FIXED | MAP_SHARED, current, 0);
197 if (tmp != TOPAGE(smpi_data_exe_start))
198 xbt_die("Couldn't map the new region (errno %d): %s", errno, strerror(errno));
199 smpi_loaded_page = actor->get_pid();
204 * @brief Makes a backup of the segment in memory that stores the global variables of a process.
205 * This backup is then used to initialize the global variables for every single
206 * process that is added, regardless of the progress of the simulation.
208 void smpi_backup_global_memory_segment()
210 #if HAVE_PRIVATIZATION
211 smpi_get_executable_global_size();
212 initial_vm_map.clear();
213 initial_vm_map.shrink_to_fit();
215 XBT_DEBUG("bss+data segment found : size %d starting at %p", smpi_data_exe_size, smpi_data_exe_start);
217 if (smpi_data_exe_size == 0) { // no need to do anything as global variables don't exist
218 smpi_privatize_global_variables = SmpiPrivStrategies::NONE;
222 smpi_data_exe_copy = ::operator new(smpi_data_exe_size);
223 // Make a copy of the data segment. This clean copy is retained over the whole runtime
224 // of the simulation and can be used to initialize a dynamically added, new process.
225 asan_safe_memcpy(smpi_data_exe_copy, TOPAGE(smpi_data_exe_start), smpi_data_exe_size);
226 #else /* ! HAVE_PRIVATIZATION */
227 xbt_die("You are trying to use privatization on a system that does not support it. Don't.");
231 // Initializes the memory mapping for a single process and returns the privatization region
232 smpi_privatization_region_t smpi_init_global_memory_segment_process()
234 int file_descriptor = smpi_temp_shm_get();
236 // ask for a free region
237 void* address = smpi_temp_shm_mmap(file_descriptor, smpi_data_exe_size);
239 // initialize the values
240 asan_safe_memcpy(address, smpi_data_exe_copy, smpi_data_exe_size);
242 // store the address of the mapping for further switches
243 smpi_privatization_regions.emplace_back(s_smpi_privatization_region_t{address, file_descriptor});
245 return &smpi_privatization_regions.back();
248 void smpi_destroy_global_memory_segments(){
249 if (smpi_data_exe_size == 0) // no need to switch
251 #if HAVE_PRIVATIZATION
252 for (auto const& region : smpi_privatization_regions) {
253 if (munmap(region.address, smpi_data_exe_size) < 0)
254 XBT_WARN("Unmapping of fd %d failed: %s", region.file_descriptor, strerror(errno));
255 close(region.file_descriptor);
257 smpi_privatization_regions.clear();
258 ::operator delete(smpi_data_exe_copy);
262 static std::vector<unsigned char> sendbuffer;
263 static std::vector<unsigned char> recvbuffer;
265 //allocate a single buffer for all sends, growing it if needed
266 unsigned char* smpi_get_tmp_sendbuffer(size_t size)
268 if (not smpi_process()->replaying())
269 return new unsigned char[size];
270 // FIXME: a resize() may invalidate a previous pointer. Maybe we need to handle a queue of buffers with a reference
271 // counter. The same holds for smpi_get_tmp_recvbuffer.
272 if (sendbuffer.size() < size)
273 sendbuffer.resize(size);
274 return sendbuffer.data();
277 //allocate a single buffer for all recv
278 unsigned char* smpi_get_tmp_recvbuffer(size_t size)
280 if (not smpi_process()->replaying())
281 return new unsigned char[size];
282 if (recvbuffer.size() < size)
283 recvbuffer.resize(size);
284 return recvbuffer.data();
287 void smpi_free_tmp_buffer(const unsigned char* buf)
289 if (not smpi_process()->replaying())
293 void smpi_free_replay_tmp_buffers()
295 std::vector<unsigned char>().swap(sendbuffer);
296 std::vector<unsigned char>().swap(recvbuffer);