X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/b7f06539e6732f81fa6ec70488b6da3e8a5e6b4b..c04b15f5346d042218b4fb8357cace04585e067b:/src/smpi/smpi_global.cpp diff --git a/src/smpi/smpi_global.cpp b/src/smpi/smpi_global.cpp index f08bea57c8..63f2a6cdbe 100644 --- a/src/smpi/smpi_global.cpp +++ b/src/smpi/smpi_global.cpp @@ -3,36 +3,31 @@ /* 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 -#include -#include -#include - #include "mc/mc.h" -#include "private.h" -#include "private.hpp" #include "simgrid/s4u/Mailbox.hpp" -#include "simgrid/sg_config.h" -#include "src/kernel/activity/SynchroComm.hpp" -#include "src/mc/mc_record.h" -#include "src/mc/mc_replay.h" +#include "simgrid/s4u/Host.hpp" #include "src/msg/msg_private.h" #include "src/simix/smx_private.h" #include "src/surf/surf_interface.hpp" -#include "surf/surf.h" -#include "xbt/replay.hpp" -#include +#include "src/smpi/SmpiHost.hpp" +#include "xbt/config.hpp" +#include "src/smpi/private.h" +#include "smpi/smpi_shared_malloc.hpp" +#include "src/smpi/smpi_coll.hpp" +#include "src/smpi/smpi_comm.hpp" +#include "src/smpi/smpi_group.hpp" +#include "src/smpi/smpi_info.hpp" +#include "src/smpi/smpi_process.hpp" +#include +#include +#include #include /* DBL_MAX */ #include -#include -#include -#include -#include -#include -#include -#include -#include + +#if HAVE_SENDFILE +#include +#endif XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_kernel, smpi, "Logging specific to SMPI (kernel)"); #include @@ -46,9 +41,6 @@ XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_kernel, smpi, "Logging specific to SMPI (ke #define RTLD_DEEPBIND 0 #endif -/* Mac OSX does not have any header file providing that definition so we have to duplicate it here. Bummers. */ -extern char** environ; /* we use it in posix_spawnp below */ - #if HAVE_PAPI #include "papi.h" const char* papi_default_config_name = "default"; @@ -78,8 +70,6 @@ static simgrid::config::Flag smpi_init_sleep( void (*smpi_comm_copy_data_callback) (smx_activity_t, void*, size_t) = &smpi_comm_copy_buffer_callback; - - int smpi_process_count() { return process_count; @@ -87,7 +77,10 @@ int smpi_process_count() simgrid::smpi::Process* smpi_process() { - simgrid::MsgActorExt* msgExt = static_cast(SIMIX_process_self()->data); + smx_actor_t me = SIMIX_process_self(); + if (me == nullptr) // This happens sometimes (eg, when linking against NS3 because it pulls openMPI...) + return nullptr; + simgrid::MsgActorExt* msgExt = static_cast(me->data); return static_cast(msgExt->data); } @@ -108,6 +101,14 @@ int smpi_process_index(){ return smpi_process()->index(); } +void * smpi_process_get_user_data(){ + return smpi_process()->get_user_data(); +} + +void smpi_process_set_user_data(void *data){ + return smpi_process()->set_user_data(data); +} + int smpi_global_size() { @@ -122,49 +123,88 @@ void smpi_comm_set_copy_data_callback(void (*callback) (smx_activity_t, void*, s smpi_comm_copy_data_callback = callback; } -void smpi_comm_copy_buffer_callback(smx_activity_t synchro, void *buff, size_t buff_size) +static void print(std::vector> vec) { + std::fprintf(stderr, "{"); + for (auto elt : vec) { + std::fprintf(stderr, "(0x%zx, 0x%zx),", elt.first, elt.second); + } + std::fprintf(stderr, "}\n"); +} +static void memcpy_private(void* dest, const void* src, std::vector>& private_blocks) { + for(auto block : private_blocks) { + memcpy((uint8_t*)dest+block.first, (uint8_t*)src+block.first, block.second-block.first); + } +} - simgrid::kernel::activity::Comm *comm = dynamic_cast(synchro); +static void check_blocks(std::vector> &private_blocks, size_t buff_size) { + for(auto block : private_blocks) { + xbt_assert(block.first <= block.second && block.second <= buff_size, "Oops, bug in shared malloc."); + } +} +void smpi_comm_copy_buffer_callback(smx_activity_t synchro, void *buff, size_t buff_size) +{ + simgrid::kernel::activity::Comm *comm = dynamic_cast(synchro); + int src_shared = 0; + int dst_shared = 0; + size_t src_offset = 0; + size_t dst_offset = 0; + std::vector> src_private_blocks; + std::vector> dst_private_blocks; XBT_DEBUG("Copy the data over"); - if(smpi_is_shared(buff)){ + if((src_shared=smpi_is_shared(buff, src_private_blocks, &src_offset))) { XBT_DEBUG("Sender %p is shared. Let's ignore it.", buff); - }else if(smpi_is_shared((char*)comm->dst_buff)){ + src_private_blocks = shift_and_frame_private_blocks(src_private_blocks, src_offset, buff_size); + } + else { + src_private_blocks.clear(); + src_private_blocks.push_back(std::make_pair(0, buff_size)); + } + if((dst_shared=smpi_is_shared((char*)comm->dst_buff, dst_private_blocks, &dst_offset))) { XBT_DEBUG("Receiver %p is shared. Let's ignore it.", (char*)comm->dst_buff); - }else{ - void* tmpbuff=buff; - if((smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP) && (static_cast(buff) >= smpi_start_data_exe) - && (static_cast(buff) < smpi_start_data_exe + smpi_size_data_exe ) - ){ - XBT_DEBUG("Privatization : We are copying from a zone inside global memory... Saving data to temp buffer !"); - - smpi_switch_data_segment( - (static_cast((static_cast(comm->src_proc->data)->data))->index())); - tmpbuff = static_cast(xbt_malloc(buff_size)); - memcpy(tmpbuff, buff, buff_size); - } - - if((smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP) && ((char*)comm->dst_buff >= smpi_start_data_exe) - && ((char*)comm->dst_buff < smpi_start_data_exe + smpi_size_data_exe )){ - XBT_DEBUG("Privatization : We are copying to a zone inside global memory - Switch data segment"); - smpi_switch_data_segment( - (static_cast((static_cast(comm->dst_proc->data)->data))->index())); - } - - XBT_DEBUG("Copying %zu bytes from %p to %p", buff_size, tmpbuff,comm->dst_buff); - memcpy(comm->dst_buff, tmpbuff, buff_size); + dst_private_blocks = shift_and_frame_private_blocks(dst_private_blocks, dst_offset, buff_size); + } + else { + dst_private_blocks.clear(); + dst_private_blocks.push_back(std::make_pair(0, buff_size)); + } + check_blocks(src_private_blocks, buff_size); + check_blocks(dst_private_blocks, buff_size); + auto private_blocks = merge_private_blocks(src_private_blocks, dst_private_blocks); + check_blocks(private_blocks, buff_size); + void* tmpbuff=buff; + if((smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP) && (static_cast(buff) >= smpi_start_data_exe) + && (static_cast(buff) < smpi_start_data_exe + smpi_size_data_exe ) + ){ + XBT_DEBUG("Privatization : We are copying from a zone inside global memory... Saving data to temp buffer !"); + + smpi_switch_data_segment( + static_cast((static_cast(comm->src_proc->data)->data)) + ->index()); + tmpbuff = static_cast(xbt_malloc(buff_size)); + memcpy_private(tmpbuff, buff, private_blocks); + } - if (comm->detached) { - // if this is a detached send, the source buffer was duplicated by SMPI - // sender to make the original buffer available to the application ASAP - xbt_free(buff); - //It seems that the request is used after the call there this should be free somewhere else but where??? - //xbt_free(comm->comm.src_data);// inside SMPI the request is kept inside the user data and should be free - comm->src_buff = nullptr; - } - if(tmpbuff!=buff)xbt_free(tmpbuff); + if((smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP) && ((char*)comm->dst_buff >= smpi_start_data_exe) + && ((char*)comm->dst_buff < smpi_start_data_exe + smpi_size_data_exe )){ + XBT_DEBUG("Privatization : We are copying to a zone inside global memory - Switch data segment"); + smpi_switch_data_segment( + static_cast((static_cast(comm->dst_proc->data)->data)) + ->index()); } + XBT_DEBUG("Copying %zu bytes from %p to %p", buff_size, tmpbuff,comm->dst_buff); + memcpy_private(comm->dst_buff, tmpbuff, private_blocks); + + if (comm->detached) { + // if this is a detached send, the source buffer was duplicated by SMPI + // sender to make the original buffer available to the application ASAP + xbt_free(buff); + //It seems that the request is used after the call there this should be free somewhere else but where??? + //xbt_free(comm->comm.src_data);// inside SMPI the request is kept inside the user data and should be free + comm->src_buff = nullptr; + } + if(tmpbuff!=buff)xbt_free(tmpbuff); } @@ -394,12 +434,14 @@ static void smpi_init_logs(){ } static void smpi_init_options(){ - + //return if already called + if (smpi_cpu_threshold > -1) + return; simgrid::smpi::Colls::set_collectives(); simgrid::smpi::Colls::smpi_coll_cleanup_callback=nullptr; smpi_cpu_threshold = xbt_cfg_get_double("smpi/cpu-threshold"); smpi_host_speed = xbt_cfg_get_double("smpi/host-speed"); - const char* smpi_privatize_option = xbt_cfg_get_string("smpi/privatize-global-variables"); + const char* smpi_privatize_option = xbt_cfg_get_string("smpi/privatization"); if (std::strcmp(smpi_privatize_option, "no") == 0) smpi_privatize_global_variables = SMPI_PRIVATIZE_NONE; else if (std::strcmp(smpi_privatize_option, "yes") == 0) @@ -416,8 +458,14 @@ static void smpi_init_options(){ smpi_privatize_global_variables = SMPI_PRIVATIZE_NONE; else - xbt_die("Invalid value for smpi/privatize-global-variables: %s", - smpi_privatize_option); + xbt_die("Invalid value for smpi/privatization: '%s'", smpi_privatize_option); + +#if defined(__FreeBSD__) + if (smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP) { + XBT_INFO("Mixing mmap privatization is broken on FreeBSD, switching to dlopen privatization instead."); + smpi_privatize_global_variables = SMPI_PRIVATIZE_DLOPEN; + } +#endif if (smpi_cpu_threshold < 0) smpi_cpu_threshold = DBL_MAX; @@ -435,27 +483,16 @@ static void smpi_init_options(){ } } -static int execute_command(const char * const argv[]) -{ - pid_t pid; - int status; - if (posix_spawnp(&pid, argv[0], nullptr, nullptr, (char* const*) argv, environ) != 0) - return 127; - if (waitpid(pid, &status, 0) != pid) - return 127; - return status; -} - typedef std::function smpi_entry_point_type; typedef int (* smpi_c_entry_point_type)(int argc, char **argv); -typedef void (* smpi_fortran_entry_point_type)(void); +typedef void (*smpi_fortran_entry_point_type)(); static int smpi_run_entry_point(smpi_entry_point_type entry_point, std::vector args) { const int argc = args.size(); std::unique_ptr argv(new char*[argc + 1]); for (int i = 0; i != argc; ++i) - argv[i] = args[i].empty() ? const_cast(""): &args[i].front(); + argv[i] = args[i].empty() ? '\0' : &args[i].front(); argv[argc] = nullptr; int res = entry_point(argc, argv.get()); @@ -469,20 +506,17 @@ static int smpi_run_entry_point(smpi_entry_point_type entry_point, std::vectordefault_function = [executable_copy](std::vector args) { - return std::function([executable_copy, args] { + + // Prepare the copy of the binary (get its size) + struct stat fdin_stat; + stat(executable_copy.c_str(), &fdin_stat); + off_t fdin_size = fdin_stat.st_size; + + simix_global->default_function = [executable_copy, fdin_size](std::vector args) { + return std::function([executable_copy, fdin_size, args] { // Copy the dynamic library: std::string target_executable = executable_copy + "_" + std::to_string(getpid()) + "_" + std::to_string(rank++) + ".so"; - // TODO, execute directly instead of relying on cp - const char* command1 [] = { - "cp", "--reflink=auto", "--", executable_copy.c_str(), target_executable.c_str(), - nullptr - }; - const char* command2 [] = { - "cp", "--", executable_copy.c_str(), target_executable.c_str(), - nullptr - }; - if (execute_command(command1) != 0 && execute_command(command2) != 0) - xbt_die("copy failed"); + + int fdin = open(executable_copy.c_str(), O_RDONLY); + xbt_assert(fdin >= 0, "Cannot read from %s", executable_copy.c_str()); + int fdout = open(target_executable.c_str(), O_CREAT | O_RDWR, S_IRWXU); + xbt_assert(fdout >= 0, "Cannot write into %s", target_executable.c_str()); + +#if HAVE_SENDFILE + ssize_t sent_size = sendfile(fdout, fdin, NULL, fdin_size); + xbt_assert(sent_size == fdin_size, + "Error while copying %s: only %zd bytes copied instead of %ld (errno: %d -- %s)", + target_executable.c_str(), sent_size, fdin_size, errno, strerror(errno)); +#else + XBT_VERB("Copy %d bytes into %s", static_cast(fdin_size), target_executable.c_str()); + const int bufsize = 1024 * 1024 * 4; + char buf[bufsize]; + while (int got = read(fdin, buf, bufsize)) { + if (got == -1) { + xbt_assert(errno == EINTR, "Cannot read from %s", executable_copy.c_str()); + } else { + char* p = buf; + int todo = got; + while (int done = write(fdout, p, todo)) { + if (done == -1) { + xbt_assert(errno == EINTR, "Cannot write into %s", target_executable.c_str()); + } else { + p += done; + todo -= done; + } + } + } + } +#endif + close(fdin); + close(fdout); // Load the copy and resolve the entry point: void* handle = dlopen(target_executable.c_str(), RTLD_LAZY | RTLD_LOCAL | RTLD_DEEPBIND); - unlink(target_executable.c_str()); + if (xbt_cfg_get_boolean("smpi/keep-temps") == false) + unlink(target_executable.c_str()); if (handle == nullptr) - xbt_die("dlopen failed"); + xbt_die("dlopen failed: %s (errno: %d -- %s)", dlerror(), errno, strerror(errno)); smpi_entry_point_type entry_point = smpi_resolve_function(handle); if (!entry_point) xbt_die("Could not resolve entry point"); - smpi_run_entry_point(entry_point, args); + smpi_run_entry_point(entry_point, args); }); }; @@ -553,7 +623,7 @@ int smpi_main(const char* executable, int argc, char *argv[]) // Load the dynamic library and resolve the entry point: void* handle = dlopen(executable, RTLD_LAZY | RTLD_LOCAL | RTLD_DEEPBIND); if (handle == nullptr) - xbt_die("dlopen failed for %s", executable); + xbt_die("dlopen failed for %s: %s (errno: %d -- %s)", executable, dlerror(), errno, strerror(errno)); smpi_entry_point_type entry_point = smpi_resolve_function(handle); if (!entry_point) xbt_die("main not found in %s", executable); @@ -596,8 +666,8 @@ int smpi_main(const char* executable, int argc, char *argv[]) } } int count = smpi_process_count(); - int i, ret=0; - for (i = 0; i < count; i++) { + int ret = 0; + for (int i = 0; i < count; i++) { if(process_data[i]->return_value()!=0){ ret=process_data[i]->return_value();//return first non 0 value break;