-/* Copyright (c) 2015-2017. The SimGrid Team. All rights reserved. */
+/* Copyright (c) 2015-2020. The SimGrid Team. All rights reserved. */
/* This program is free software; you can redistribute it and/or modify it
* under the terms of the license (GNU LGPL) which comes with this package. */
-#include <cstdint>
-#include <climits>
-#include <cstring>
-
-#include <vector>
-
+#include <algorithm>
#include <cerrno>
+#include <climits>
+#include <cstdint>
#include <cstdio>
#include <cstdlib>
#include <cstring>
+#include <deque>
#include <fcntl.h>
#include <sys/stat.h>
#include <sys/types.h>
+#include <vector>
#ifndef WIN32
#include <sys/mman.h>
#include "src/xbt/memory_map.hpp"
#include "private.hpp"
+#include "src/smpi/include/smpi_actor.hpp"
XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_memory, smpi, "Memory layout support for SMPI");
-int smpi_loaded_page = -1;
+int smpi_loaded_page = -1;
char* smpi_data_exe_start = nullptr;
int smpi_data_exe_size = 0;
-int smpi_privatize_global_variables;
+SmpiPrivStrategies smpi_privatize_global_variables;
+static void* smpi_data_exe_copy;
+
+// Initialized by smpi_prepare_global_memory_segment().
+static std::vector<simgrid::xbt::VmMap> initial_vm_map;
-static const int PROT_RWX = (PROT_READ | PROT_WRITE | PROT_EXEC);
-static const int PROT_RW = (PROT_READ | PROT_WRITE );
-XBT_ATTRIB_UNUSED static const int PROT_RX = (PROT_READ | PROT_EXEC );
+// We keep a copy of all the privatization regions: We can then delete everything easily by iterating over this
+// collection and nothing can be leaked. We could also iterate over all actors but we would have to be diligent when two
+// actors use the same privatization region (so, smart pointers would have to be used etc.)
+// Use a std::deque so that pointers remain valid after push_back().
+static std::deque<s_smpi_privatization_region_t> smpi_privatization_regions;
-void smpi_get_executable_global_size()
+static constexpr int PROT_RWX = PROT_READ | PROT_WRITE | PROT_EXEC;
+static constexpr int PROT_RW = PROT_READ | PROT_WRITE;
+
+/** Take a snapshot of the process' memory map.
+ */
+void smpi_prepare_global_memory_segment()
+{
+ initial_vm_map = simgrid::xbt::get_memory_map(getpid());
+}
+
+static void smpi_get_executable_global_size()
{
- char buffer[PATH_MAX];
- char* full_name = realpath(xbt_binary_name, buffer);
- if (full_name == nullptr)
- xbt_die("Could not resolve binary file name");
+ char* buffer = realpath(simgrid::xbt::binary_name.c_str(), nullptr);
+ xbt_assert(buffer != nullptr, "Could not resolve real path of binary file '%s'", simgrid::xbt::binary_name.c_str());
+ std::string full_name = buffer;
+ free(buffer);
std::vector<simgrid::xbt::VmMap> map = simgrid::xbt::get_memory_map(getpid());
for (auto i = map.begin(); i != map.end() ; ++i) {
if (i->pathname == full_name && (i->prot & PROT_RWX) == PROT_RW) {
smpi_data_exe_start = (char*)i->start_addr;
smpi_data_exe_size = i->end_addr - i->start_addr;
- ++i;
/* Here we are making the assumption that a suitable empty region
following the rw- area is the end of the data segment. It would
be better to check with the size of the data segment. */
- if (i != map.end() && i->pathname.empty() && (i->prot & PROT_RWX) == PROT_RW &&
- (char*)i->start_addr == smpi_data_exe_start + smpi_data_exe_size) {
- smpi_data_exe_size = (char*)i->end_addr - smpi_data_exe_start;
+ auto j = i + 1;
+ if (j != map.end() && j->pathname.empty() && (j->prot & PROT_RWX) == PROT_RW &&
+ (char*)j->start_addr == smpi_data_exe_start + smpi_data_exe_size) {
+ // Only count the portion of this region not present in the initial map.
+ auto found = std::find_if(initial_vm_map.begin(), initial_vm_map.end(), [&j](const simgrid::xbt::VmMap& m) {
+ return j->start_addr <= m.start_addr && m.start_addr < j->end_addr;
+ });
+ auto end_addr = (found == initial_vm_map.end() ? j->end_addr : found->start_addr);
+ smpi_data_exe_size = (char*)end_addr - smpi_data_exe_start;
}
return;
}
}
#endif
-#if HAVE_SANITIZE_ADDRESS
+#if HAVE_SANITIZER_ADDRESS
#include <sanitizer/asan_interface.h>
static void* asan_safe_memcpy(void* dest, void* src, size_t n)
{
return dest;
}
#else
-#define asan_safe_memcpy(dest, src, n) memcpy(dest, src, n)
+#define asan_safe_memcpy(dest, src, n) memcpy((dest), (src), (n))
#endif
-/** Map a given SMPI privatization segment (make a SMPI process active) */
-void smpi_switch_data_segment(int dest) {
- if (smpi_loaded_page == dest)//no need to switch, we've already loaded the one we want
- return;
+/**
+ * @brief Uses shm_open to get a temporary shm, and returns its file descriptor.
+ */
+int smpi_temp_shm_get()
+{
+ constexpr unsigned VAL_MASK = 0xffffffffUL;
+ static unsigned prev_val = VAL_MASK;
+ char shmname[32]; // cannot be longer than PSHMNAMLEN = 31 on macOS (shm_open raises ENAMETOOLONG otherwise)
+ int fd;
+
+ for (unsigned i = (prev_val + 1) & VAL_MASK; i != prev_val; i = (i + 1) & VAL_MASK) {
+ snprintf(shmname, sizeof(shmname), "/smpi-buffer-%016x", i);
+ fd = shm_open(shmname, O_RDWR | O_CREAT | O_EXCL, S_IRUSR | S_IWUSR);
+ if (fd != -1 || errno != EEXIST) {
+ prev_val = i;
+ break;
+ }
+ }
+ if (fd < 0) {
+ if (errno == EMFILE) {
+ xbt_die("Impossible to create temporary file for memory mapping: %s\n\
+The shm_open() system call failed with the EMFILE error code (too many files). \n\n\
+This means that you reached the system limits concerning the amount of files per process. \
+This is not a surprise if you are trying to virtualize many processes on top of SMPI. \
+Don't panic -- you should simply increase your system limits and try again. \n\n\
+First, check what your limits are:\n\
+ cat /proc/sys/fs/file-max # Gives you the system-wide limit\n\
+ ulimit -Hn # Gives you the per process hard limit\n\
+ ulimit -Sn # Gives you the per process soft limit\n\
+ cat /proc/self/limits # Displays any per-process limitation (including the one given above)\n\n\
+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. \
+Ask the Internet about tutorials on how to increase the files limit such as: https://rtcamp.com/tutorials/linux/increase-open-files-limit/",
+ strerror(errno));
+ }
+ xbt_die("Impossible to create temporary file for memory mapping. shm_open: %s", strerror(errno));
+ }
+ XBT_DEBUG("Got temporary shm %s (fd = %d)", shmname, fd);
+ if (shm_unlink(shmname) < 0)
+ XBT_WARN("Could not early unlink %s. shm_unlink: %s", shmname, strerror(errno));
+ return fd;
+}
- // So the job:
- smpi_really_switch_data_segment(dest);
+/**
+ * @brief Mmap a region of size bytes from temporary shm with file descriptor fd.
+ */
+void* smpi_temp_shm_mmap(int fd, size_t size)
+{
+ struct stat st;
+ if (fstat(fd, &st) != 0)
+ xbt_die("Could not stat fd %d: %s", fd, strerror(errno));
+ if (static_cast<off_t>(size) > st.st_size && ftruncate(fd, static_cast<off_t>(size)) != 0)
+ xbt_die("Could not truncate fd %d to %zu: %s", fd, size, strerror(errno));
+ void* mem = mmap(nullptr, size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
+ if (mem == MAP_FAILED) {
+ xbt_die("Failed to map fd %d with size %zu: %s\n"
+ "If you are running a lot of ranks, you may be exceeding the amount of mappings allowed per process.\n"
+ "On Linux systems, change this value with sudo sysctl -w vm.max_map_count=newvalue (default value: 65536)\n"
+ "Please see "
+ "https://simgrid.org/doc/latest/Configuring_SimGrid.html#configuring-the-user-code-virtualization for more "
+ "information.",
+ fd, size, strerror(errno));
+ }
+ return mem;
}
-/** Map a given SMPI privatization segment (make a SMPI process active) even if SMPI thinks it is already active
+/** Map a given SMPI privatization segment (make an SMPI process active)
*
* When doing a state restoration, the state of the restored variables might not be consistent with the state of the
* virtual memory. In this case, we to change the data segment.
*/
-void smpi_really_switch_data_segment(int dest)
+void smpi_switch_data_segment(simgrid::s4u::ActorPtr actor)
{
+ if (smpi_loaded_page == actor->get_pid()) // no need to switch, we've already loaded the one we want
+ return;
+
if (smpi_data_exe_size == 0) // no need to switch
return;
#if HAVE_PRIVATIZATION
- if(smpi_loaded_page==-1){//initial switch, do the copy from the real page here
- for (int i=0; i< smpi_process_count(); i++){
- asan_safe_memcpy(smpi_privatization_regions[i].address, TOPAGE(smpi_data_exe_start), smpi_data_exe_size);
- }
- }
-
// FIXME, cross-process support (mmap across process when necessary)
- int current = smpi_privatization_regions[dest].file_descriptor;
- XBT_DEBUG("Switching data frame to the one of process %d", dest);
- void* tmp =
- mmap(TOPAGE(smpi_data_exe_start), smpi_data_exe_size, PROT_READ | PROT_WRITE, MAP_FIXED | MAP_SHARED, current, 0);
+ XBT_DEBUG("Switching data frame to the one of process %ld", actor->get_pid());
+ const simgrid::smpi::ActorExt* process = smpi_process_remote(actor);
+ int current = process->privatized_region()->file_descriptor;
+ const void* tmp = mmap(TOPAGE(smpi_data_exe_start), smpi_data_exe_size, PROT_RW, MAP_FIXED | MAP_SHARED, current, 0);
if (tmp != TOPAGE(smpi_data_exe_start))
xbt_die("Couldn't map the new region (errno %d): %s", errno, strerror(errno));
- smpi_loaded_page = dest;
+ smpi_loaded_page = actor->get_pid();
#endif
}
-int smpi_is_privatization_file(char* file)
-{
- const std::string buffer_path("/dev/shm/my-buffer-");
- return buffer_path.compare(0, std::string::npos, file, buffer_path.length()) == 0;
-}
-
-void smpi_initialize_global_memory_segments()
+/**
+ * @brief Makes a backup of the segment in memory that stores the global variables of a process.
+ * This backup is then used to initialize the global variables for every single
+ * process that is added, regardless of the progress of the simulation.
+ */
+void smpi_backup_global_memory_segment()
{
-
#if HAVE_PRIVATIZATION
smpi_get_executable_global_size();
+ initial_vm_map.clear();
+ initial_vm_map.shrink_to_fit();
XBT_DEBUG("bss+data segment found : size %d starting at %p", smpi_data_exe_size, smpi_data_exe_start);
- if (smpi_data_exe_size == 0) { // no need to switch
- smpi_privatize_global_variables=false;
+ if (smpi_data_exe_size == 0) { // no need to do anything as global variables don't exist
+ smpi_privatize_global_variables = SmpiPrivStrategies::NONE;
return;
}
- smpi_privatization_regions = static_cast<smpi_privatization_region_t>(
- xbt_malloc(smpi_process_count() * sizeof(struct s_smpi_privatization_region)));
-
- for (int i=0; i< smpi_process_count(); i++){
- // create SIMIX_process_count() mappings of this size with the same data inside
- int file_descriptor;
- void* address = nullptr;
- char path[24];
- int status;
-
- do {
- snprintf(path, sizeof(path), "/smpi-buffer-%06x", rand() % 0xffffffU);
- file_descriptor = shm_open(path, O_RDWR | O_CREAT | O_EXCL, S_IRUSR | S_IWUSR);
- } while (file_descriptor == -1 && errno == EEXIST);
- if (file_descriptor < 0) {
- if (errno == EMFILE) {
- xbt_die("Impossible to create temporary file for memory mapping: %s\n\
-The open() system call failed with the EMFILE error code (too many files). \n\n\
-This means that you reached the system limits concerning the amount of files per process. \
-This is not a surprise if you are trying to virtualize many processes on top of SMPI. \
-Don't panic -- you should simply increase your system limits and try again. \n\n\
-First, check what your limits are:\n\
- cat /proc/sys/fs/file-max # Gives you the system-wide limit\n\
- ulimit -Hn # Gives you the per process hard limit\n\
- ulimit -Sn # Gives you the per process soft limit\n\
- cat /proc/self/limits # Displays any per-process limitation (including the one given above)\n\n\
-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. \
-Ask the Internet about tutorials on how to increase the files limit such as: https://rtcamp.com/tutorials/linux/increase-open-files-limit/",
- strerror(errno));
- }
- xbt_die("Impossible to create temporary file for memory mapping: %s", strerror(errno));
- }
+ smpi_data_exe_copy = ::operator new(smpi_data_exe_size);
+ // Make a copy of the data segment. This clean copy is retained over the whole runtime
+ // of the simulation and can be used to initialize a dynamically added, new process.
+ asan_safe_memcpy(smpi_data_exe_copy, TOPAGE(smpi_data_exe_start), smpi_data_exe_size);
+#else /* ! HAVE_PRIVATIZATION */
+ xbt_die("You are trying to use privatization on a system that does not support it. Don't.");
+#endif
+}
- status = ftruncate(file_descriptor, smpi_data_exe_size);
- if (status)
- xbt_die("Impossible to set the size of the temporary file for memory mapping");
+// Initializes the memory mapping for a single process and returns the privatization region
+smpi_privatization_region_t smpi_init_global_memory_segment_process()
+{
+ int file_descriptor = smpi_temp_shm_get();
- /* Ask for a free region */
- address = mmap(nullptr, smpi_data_exe_size, PROT_READ | PROT_WRITE, MAP_SHARED, file_descriptor, 0);
- if (address == MAP_FAILED)
- xbt_die("Couldn't find a free region for memory mapping");
+ // ask for a free region
+ void* address = smpi_temp_shm_mmap(file_descriptor, smpi_data_exe_size);
- status = shm_unlink(path);
- if (status)
- xbt_die("Impossible to unlink temporary file for memory mapping");
+ // initialize the values
+ asan_safe_memcpy(address, smpi_data_exe_copy, smpi_data_exe_size);
- // initialize the values
- asan_safe_memcpy(address, TOPAGE(smpi_data_exe_start), smpi_data_exe_size);
+ // store the address of the mapping for further switches
+ smpi_privatization_regions.emplace_back(s_smpi_privatization_region_t{address, file_descriptor});
- // store the address of the mapping for further switches
- smpi_privatization_regions[i].file_descriptor = file_descriptor;
- smpi_privatization_regions[i].address = address;
- }
-#else /* ! HAVE_PRIVATIZATION */
- smpi_privatize_global_variables = false;
- xbt_die("You are trying to use privatization on a system that does not support it. Don't.");
- return;
-#endif
+ return &smpi_privatization_regions.back();
}
void smpi_destroy_global_memory_segments(){
if (smpi_data_exe_size == 0) // no need to switch
return;
#if HAVE_PRIVATIZATION
- for (int i=0; i< smpi_process_count(); i++) {
- if (munmap(smpi_privatization_regions[i].address, smpi_data_exe_size) < 0)
- XBT_WARN("Unmapping of fd %d failed: %s", smpi_privatization_regions[i].file_descriptor, strerror(errno));
- close(smpi_privatization_regions[i].file_descriptor);
+ for (auto const& region : smpi_privatization_regions) {
+ if (munmap(region.address, smpi_data_exe_size) < 0)
+ XBT_WARN("Unmapping of fd %d failed: %s", region.file_descriptor, strerror(errno));
+ close(region.file_descriptor);
}
- xbt_free(smpi_privatization_regions);
+ smpi_privatization_regions.clear();
+ ::operator delete(smpi_data_exe_copy);
#endif
}
+static std::vector<unsigned char> sendbuffer;
+static std::vector<unsigned char> recvbuffer;
+
+//allocate a single buffer for all sends, growing it if needed
+unsigned char* smpi_get_tmp_sendbuffer(size_t size)
+{
+ if (not smpi_process()->replaying())
+ return new unsigned char[size];
+ // FIXME: a resize() may invalidate a previous pointer. Maybe we need to handle a queue of buffers with a reference
+ // counter. The same holds for smpi_get_tmp_recvbuffer.
+ if (sendbuffer.size() < size)
+ sendbuffer.resize(size);
+ return sendbuffer.data();
+}
+
+//allocate a single buffer for all recv
+unsigned char* smpi_get_tmp_recvbuffer(size_t size)
+{
+ if (not smpi_process()->replaying())
+ return new unsigned char[size];
+ if (recvbuffer.size() < size)
+ recvbuffer.resize(size);
+ return recvbuffer.data();
+}
+
+void smpi_free_tmp_buffer(const unsigned char* buf)
+{
+ if (not smpi_process()->replaying())
+ delete[] buf;
+}
+
+void smpi_free_replay_tmp_buffers()
+{
+ std::vector<unsigned char>().swap(sendbuffer);
+ std::vector<unsigned char>().swap(recvbuffer);
+}