1 /* Copyright (c) 2015-2020. 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 = realpath(simgrid::xbt::binary_name.c_str(), nullptr);
59 xbt_assert(buffer != nullptr, "Could not resolve real path of binary file '%s'", simgrid::xbt::binary_name.c_str());
60 std::string full_name = buffer;
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 (j != map.end() && j->pathname.empty() && (j->prot & PROT_RWX) == PROT_RW &&
77 (char*)j->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(), [&j](const simgrid::xbt::VmMap& m) {
80 return j->start_addr <= m.start_addr && m.start_addr < j->end_addr;
82 auto end_addr = (found == initial_vm_map.end() ? j->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))
115 * @brief Uses shm_open to get a temporary shm, and returns its file descriptor.
117 int smpi_temp_shm_get()
119 constexpr unsigned VAL_MASK = 0xffffffffUL;
120 static unsigned prev_val = VAL_MASK;
121 char shmname[32]; // cannot be longer than PSHMNAMLEN = 31 on macOS (shm_open raises ENAMETOOLONG otherwise)
124 for (unsigned i = (prev_val + 1) & VAL_MASK; i != prev_val; i = (i + 1) & VAL_MASK) {
125 snprintf(shmname, sizeof(shmname), "/smpi-buffer-%016x", i);
126 fd = shm_open(shmname, O_RDWR | O_CREAT | O_EXCL, S_IRUSR | S_IWUSR);
127 if (fd != -1 || errno != EEXIST) {
133 if (errno == EMFILE) {
134 xbt_die("Impossible to create temporary file for memory mapping: %s\n\
135 The shm_open() system call failed with the EMFILE error code (too many files). \n\n\
136 This means that you reached the system limits concerning the amount of files per process. \
137 This is not a surprise if you are trying to virtualize many processes on top of SMPI. \
138 Don't panic -- you should simply increase your system limits and try again. \n\n\
139 First, check what your limits are:\n\
140 cat /proc/sys/fs/file-max # Gives you the system-wide limit\n\
141 ulimit -Hn # Gives you the per process hard limit\n\
142 ulimit -Sn # Gives you the per process soft limit\n\
143 cat /proc/self/limits # Displays any per-process limitation (including the one given above)\n\n\
144 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. \
145 Ask the Internet about tutorials on how to increase the files limit such as: https://rtcamp.com/tutorials/linux/increase-open-files-limit/",
148 xbt_die("Impossible to create temporary file for memory mapping. shm_open: %s", strerror(errno));
150 XBT_DEBUG("Got temporary shm %s (fd = %d)", shmname, fd);
151 if (shm_unlink(shmname) < 0)
152 XBT_WARN("Could not early unlink %s. shm_unlink: %s", shmname, strerror(errno));
157 * @brief Mmap a region of size bytes from temporary shm with file descriptor fd.
159 void* smpi_temp_shm_mmap(int fd, size_t size)
162 if (fstat(fd, &st) != 0)
163 xbt_die("Could not stat fd %d: %s", fd, strerror(errno));
164 if (static_cast<off_t>(size) > st.st_size && ftruncate(fd, static_cast<off_t>(size)) != 0)
165 xbt_die("Could not truncate fd %d to %zu: %s", fd, size, strerror(errno));
166 void* mem = mmap(nullptr, size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
167 if (mem == MAP_FAILED) {
168 xbt_die("Failed to map fd %d with size %zu: %s\n"
169 "If you are running a lot of ranks, you may be exceeding the amount of mappings allowed per process.\n"
170 "On Linux systems, change this value with sudo sysctl -w vm.max_map_count=newvalue (default value: 65536)\n"
172 "https://simgrid.org/doc/latest/Configuring_SimGrid.html#configuring-the-user-code-virtualization for more "
174 fd, size, strerror(errno));
179 /** Map a given SMPI privatization segment (make an SMPI process active)
181 * When doing a state restoration, the state of the restored variables might not be consistent with the state of the
182 * virtual memory. In this case, we to change the data segment.
184 void smpi_switch_data_segment(simgrid::s4u::ActorPtr actor)
186 if (smpi_loaded_page == actor->get_pid()) // no need to switch, we've already loaded the one we want
189 if (smpi_data_exe_size == 0) // no need to switch
192 #if HAVE_PRIVATIZATION
193 // FIXME, cross-process support (mmap across process when necessary)
194 XBT_DEBUG("Switching data frame to the one of process %ld", actor->get_pid());
195 const simgrid::smpi::ActorExt* process = smpi_process_remote(actor);
196 int current = process->privatized_region()->file_descriptor;
197 const void* tmp = mmap(TOPAGE(smpi_data_exe_start), smpi_data_exe_size, PROT_RW, MAP_FIXED | MAP_SHARED, current, 0);
198 if (tmp != TOPAGE(smpi_data_exe_start))
199 xbt_die("Couldn't map the new region (errno %d): %s", errno, strerror(errno));
200 smpi_loaded_page = actor->get_pid();
205 * @brief Makes a backup of the segment in memory that stores the global variables of a process.
206 * This backup is then used to initialize the global variables for every single
207 * process that is added, regardless of the progress of the simulation.
209 void smpi_backup_global_memory_segment()
211 #if HAVE_PRIVATIZATION
212 smpi_get_executable_global_size();
213 initial_vm_map.clear();
214 initial_vm_map.shrink_to_fit();
216 XBT_DEBUG("bss+data segment found : size %d starting at %p", smpi_data_exe_size, smpi_data_exe_start);
218 if (smpi_data_exe_size == 0) { // no need to do anything as global variables don't exist
219 smpi_privatize_global_variables = SmpiPrivStrategies::NONE;
223 smpi_data_exe_copy = ::operator new(smpi_data_exe_size);
224 // Make a copy of the data segment. This clean copy is retained over the whole runtime
225 // of the simulation and can be used to initialize a dynamically added, new process.
226 asan_safe_memcpy(smpi_data_exe_copy, TOPAGE(smpi_data_exe_start), smpi_data_exe_size);
227 #else /* ! HAVE_PRIVATIZATION */
228 xbt_die("You are trying to use privatization on a system that does not support it. Don't.");
232 // Initializes the memory mapping for a single process and returns the privatization region
233 smpi_privatization_region_t smpi_init_global_memory_segment_process()
235 int file_descriptor = smpi_temp_shm_get();
237 // ask for a free region
238 void* address = smpi_temp_shm_mmap(file_descriptor, smpi_data_exe_size);
240 // initialize the values
241 asan_safe_memcpy(address, smpi_data_exe_copy, smpi_data_exe_size);
243 // store the address of the mapping for further switches
244 smpi_privatization_regions.emplace_back(s_smpi_privatization_region_t{address, file_descriptor});
246 return &smpi_privatization_regions.back();
249 void smpi_destroy_global_memory_segments(){
250 if (smpi_data_exe_size == 0) // no need to switch
252 #if HAVE_PRIVATIZATION
253 for (auto const& region : smpi_privatization_regions) {
254 if (munmap(region.address, smpi_data_exe_size) < 0)
255 XBT_WARN("Unmapping of fd %d failed: %s", region.file_descriptor, strerror(errno));
256 close(region.file_descriptor);
258 smpi_privatization_regions.clear();
259 ::operator delete(smpi_data_exe_copy);
263 static std::vector<unsigned char> sendbuffer;
264 static std::vector<unsigned char> recvbuffer;
266 //allocate a single buffer for all sends, growing it if needed
267 unsigned char* smpi_get_tmp_sendbuffer(size_t size)
269 if (not smpi_process()->replaying())
270 return new unsigned char[size];
271 // FIXME: a resize() may invalidate a previous pointer. Maybe we need to handle a queue of buffers with a reference
272 // counter. The same holds for smpi_get_tmp_recvbuffer.
273 if (sendbuffer.size() < size)
274 sendbuffer.resize(size);
275 return sendbuffer.data();
278 //allocate a single buffer for all recv
279 unsigned char* smpi_get_tmp_recvbuffer(size_t size)
281 if (not smpi_process()->replaying())
282 return new unsigned char[size];
283 if (recvbuffer.size() < size)
284 recvbuffer.resize(size);
285 return recvbuffer.data();
288 void smpi_free_tmp_buffer(const unsigned char* buf)
290 if (not smpi_process()->replaying())
294 void smpi_free_replay_tmp_buffers()
296 std::vector<unsigned char>().swap(sendbuffer);
297 std::vector<unsigned char>().swap(recvbuffer);