static void test(hwloc_const_bitmap_t cpuset, int flags)
{
hwloc_bitmap_t new_cpuset = hwloc_bitmap_alloc();
result_set("Bind this singlethreaded process", hwloc_set_cpubind(topology, cpuset, flags), support->cpubind->set_thisproc_cpubind || support->cpubind->set_thisthread_cpubind);
result_get("Get this singlethreaded process", cpuset, new_cpuset, hwloc_get_cpubind(topology, new_cpuset, flags), support->cpubind->get_thisproc_cpubind || support->cpubind->get_thisthread_cpubind);
result_set("Bind this thread", hwloc_set_cpubind(topology, cpuset, flags | HWLOC_CPUBIND_THREAD), support->cpubind->set_thisthread_cpubind);
result_get("Get this thread", cpuset, new_cpuset, hwloc_get_cpubind(topology, new_cpuset, flags | HWLOC_CPUBIND_THREAD), support->cpubind->get_thisthread_cpubind);
result_set("Bind this whole process", hwloc_set_cpubind(topology, cpuset, flags | HWLOC_CPUBIND_PROCESS), support->cpubind->set_thisproc_cpubind);
result_get("Get this whole process", cpuset, new_cpuset, hwloc_get_cpubind(topology, new_cpuset, flags | HWLOC_CPUBIND_PROCESS), support->cpubind->get_thisproc_cpubind);
#ifdef HWLOC_WIN_SYS
result_set("Bind process", hwloc_set_proc_cpubind(topology, GetCurrentProcess(), cpuset, flags | HWLOC_CPUBIND_PROCESS), support->cpubind->set_proc_cpubind);
result_get("Get process", cpuset, new_cpuset, hwloc_get_proc_cpubind(topology, GetCurrentProcess(), new_cpuset, flags | HWLOC_CPUBIND_PROCESS), support->cpubind->get_proc_cpubind);
result_set("Bind thread", hwloc_set_thread_cpubind(topology, GetCurrentThread(), cpuset, flags | HWLOC_CPUBIND_THREAD), support->cpubind->set_thread_cpubind);
result_get("Get thread", cpuset, new_cpuset, hwloc_get_thread_cpubind(topology, GetCurrentThread(), new_cpuset, flags | HWLOC_CPUBIND_THREAD), support->cpubind->get_thread_cpubind);
#else /* !HWLOC_WIN_SYS */
result_set("Bind whole process", hwloc_set_proc_cpubind(topology, getpid(), cpuset, flags | HWLOC_CPUBIND_PROCESS), support->cpubind->set_proc_cpubind);
result_get("Get whole process", cpuset, new_cpuset, hwloc_get_proc_cpubind(topology, getpid(), new_cpuset, flags | HWLOC_CPUBIND_PROCESS), support->cpubind->get_proc_cpubind);
result_set("Bind process", hwloc_set_proc_cpubind(topology, getpid(), cpuset, flags), support->cpubind->set_proc_cpubind);
result_get("Get process", cpuset, new_cpuset, hwloc_get_proc_cpubind(topology, getpid(), new_cpuset, flags), support->cpubind->get_proc_cpubind);
#ifdef hwloc_thread_t
result_set("Bind thread", hwloc_set_thread_cpubind(topology, pthread_self(), cpuset, flags), support->cpubind->set_thread_cpubind);
result_get("Get thread", cpuset, new_cpuset, hwloc_get_thread_cpubind(topology, pthread_self(), new_cpuset, flags), support->cpubind->get_thread_cpubind);
#endif
#endif /* !HWLOC_WIN_SYS */
printf("\n");
hwloc_bitmap_free(new_cpuset);
}
/*
* Get the core where the current process is running
* param flag set of cores is possible
* return the core
*/
int hw_get_proc_core(int flag)
{
int core;
hwloc_cpuset_t set;
set = hwloc_bitmap_alloc();
hwloc_get_proc_cpubind (topology,0,set,HWLOC_CPUBIND_PROCESS);
core = hwloc_bitmap_first(set);
hwloc_bitmap_free(set);
//hwloc can't detect the thread core - use linux based syscall
if (flag)
core = sched_getcpu();
return core;
}
/*
* Get the first core of a node
* return the core
*/
int hw_get_firstcore()
{
int core;
hwloc_cpuset_t set;
if (local_topo->nnodes != 0 ){
set = hwloc_bitmap_alloc();
hwloc_get_proc_cpubind (topology,0,set,HWLOC_CPUBIND_PROCESS);
core = hwloc_bitmap_first(set);
hwloc_bitmap_free(set);
}
else
core = -1;
return core;
}
/*
* Get the node where the current process is running
* return the node of the core
*/
int hw_get_proc_node()
{
int node;
hwloc_cpuset_t set;
hwloc_nodeset_t nset;
if (local_topo->nnodes != 0 ){
set = hwloc_bitmap_alloc();
nset = hwloc_bitmap_alloc();
hwloc_get_proc_cpubind (topology,0,set,HWLOC_CPUBIND_PROCESS);
hwloc_cpuset_to_nodeset(topology,set,nset);
node = hwloc_bitmap_first(nset);
hwloc_bitmap_free(set);
hwloc_bitmap_free(nset);
}
else
node = -1;
return node;
}
static void add_process_objects(hwloc_topology_t topology)
{
#ifdef HAVE_DIRENT_H
hwloc_obj_t root;
hwloc_bitmap_t cpuset;
#ifdef HWLOC_LINUX_SYS
hwloc_bitmap_t task_cpuset;
#endif /* HWLOC_LINUX_SYS */
DIR *dir;
struct dirent *dirent;
const struct hwloc_topology_support *support;
root = hwloc_get_root_obj(topology);
support = hwloc_topology_get_support(topology);
if (!support->cpubind->get_proc_cpubind)
return;
dir = opendir("/proc");
if (!dir)
return;
cpuset = hwloc_bitmap_alloc();
#ifdef HWLOC_LINUX_SYS
task_cpuset = hwloc_bitmap_alloc();
#endif /* HWLOC_LINUX_SYS */
while ((dirent = readdir(dir))) {
long local_pid_number;
hwloc_pid_t local_pid;
char *end;
char name[80];
int proc_cpubind;
local_pid_number = strtol(dirent->d_name, &end, 10);
if (*end)
/* Not a number */
continue;
snprintf(name, sizeof(name), "%ld", local_pid_number);
local_pid = hwloc_pid_from_number(local_pid_number, 0);
proc_cpubind = hwloc_get_proc_cpubind(topology, local_pid, cpuset, 0) != -1;
#ifdef HWLOC_LINUX_SYS
{
char comm[16];
char *path;
size_t pathlen = 6 + strlen(dirent->d_name) + 1 + 7 + 1;
path = malloc(pathlen);
{
/* Get the process name */
char cmd[64];
int file;
ssize_t n;
snprintf(path, pathlen, "/proc/%s/cmdline", dirent->d_name);
file = open(path, O_RDONLY);
if (file < 0) {
/* Ignore errors */
free(path);
continue;
}
n = read(file, cmd, sizeof(cmd));
close(file);
if (n <= 0) {
/* Ignore kernel threads and errors */
free(path);
continue;
}
snprintf(path, pathlen, "/proc/%s/comm", dirent->d_name);
file = open(path, O_RDONLY);
if (file >= 0) {
n = read(file, comm, sizeof(comm) - 1);
close(file);
if (n > 0) {
comm[n] = 0;
if (n > 1 && comm[n-1] == '\n')
comm[n-1] = 0;
} else {
snprintf(comm, sizeof(comm), "(unknown)");
}
} else {
/* Old kernel, have to look at old file */
char stats[32];
char *parenl = NULL, *parenr;
snprintf(path, pathlen, "/proc/%s/stat", dirent->d_name);
file = open(path, O_RDONLY);
if (file < 0) {
/* Ignore errors */
free(path);
continue;
}
/* "pid (comm) ..." */
n = read(file, stats, sizeof(stats) - 1);
close(file);
if (n > 0) {
stats[n] = 0;
parenl = strchr(stats, '(');
parenr = strchr(stats, ')');
if (!parenr)
parenr = &stats[sizeof(stats)-1];
*parenr = 0;
}
if (!parenl) {
snprintf(comm, sizeof(comm), "(unknown)");
} else {
snprintf(comm, sizeof(comm), "%s", parenl+1);
}
}
snprintf(name, sizeof(name), "%ld %s", local_pid_number, comm);
}
{
/* Get threads */
DIR *task_dir;
struct dirent *task_dirent;
snprintf(path, pathlen, "/proc/%s/task", dirent->d_name);
task_dir = opendir(path);
if (task_dir) {
while ((task_dirent = readdir(task_dir))) {
long local_tid;
char *task_end;
const size_t tid_len = sizeof(local_tid)*3+1;
size_t task_pathlen = 6 + strlen(dirent->d_name) + 1 + 4 + 1
+ strlen(task_dirent->d_name) + 1 + 4 + 1;
char *task_path;
int comm_file;
char task_comm[16] = "";
char task_name[sizeof(name) + 1 + tid_len + 1 + sizeof(task_comm) + 1];
ssize_t n;
local_tid = strtol(task_dirent->d_name, &task_end, 10);
if (*task_end)
/* Not a number, or the main task */
continue;
task_path = malloc(task_pathlen);
snprintf(task_path, task_pathlen, "/proc/%s/task/%s/comm",
dirent->d_name, task_dirent->d_name);
comm_file = open(task_path, O_RDONLY);
free(task_path);
if (comm_file >= 0) {
n = read(comm_file, task_comm, sizeof(task_comm) - 1);
if (n < 0)
n = 0;
close(comm_file);
task_comm[n] = 0;
if (n > 1 && task_comm[n-1] == '\n')
task_comm[n-1] = 0;
if (!strcmp(comm, task_comm))
/* Same as process comm, do not show it again */
n = 0;
} else {
n = 0;
}
if (hwloc_linux_get_tid_cpubind(topology, local_tid, task_cpuset))
continue;
if (proc_cpubind && hwloc_bitmap_isequal(task_cpuset, cpuset))
continue;
if (n) {
snprintf(task_name, sizeof(task_name), "%s %li %s", name, local_tid, task_comm);
} else {
snprintf(task_name, sizeof(task_name), "%s %li", name, local_tid);
}
insert_task(topology, task_cpuset, task_name);
}
closedir(task_dir);
}
}
free(path);
}
#endif /* HWLOC_LINUX_SYS */
if (!proc_cpubind)
continue;
if (hwloc_bitmap_isincluded(root->cpuset, cpuset))
continue;
insert_task(topology, cpuset, name);
}
hwloc_bitmap_free(cpuset);
#ifdef HWLOC_LINUX_SYS
hwloc_bitmap_free(task_cpuset);
#endif /* HWLOC_LINUX_SYS */
closedir(dir);
#endif /* HAVE_DIRENT_H */
}
int main(int argc, char *argv[])
{
const struct hwloc_topology_support *support;
hwloc_topology_t topology;
hwloc_const_bitmap_t topocpuset;
hwloc_bitmap_t cpuset;
unsigned long flags = 0;
DIR *dir;
struct dirent *dirent;
int show_all = 0;
int show_threads = 0;
int get_last_cpu_location = 0;
char *callname;
char *pidcmd = NULL;
int err;
int opt;
callname = strrchr(argv[0], '/');
if (!callname)
callname = argv[0];
else
callname++;
/* skip argv[0], handle options */
argc--;
argv++;
hwloc_utils_check_api_version(callname);
while (argc >= 1) {
opt = 0;
if (!strcmp(argv[0], "-a"))
show_all = 1;
else if (!strcmp(argv[0], "-l") || !strcmp(argv[0], "--logical")) {
logical = 1;
} else if (!strcmp(argv[0], "-p") || !strcmp(argv[0], "--physical")) {
logical = 0;
} else if (!strcmp(argv[0], "-c") || !strcmp(argv[0], "--cpuset")) {
show_cpuset = 1;
} else if (!strcmp(argv[0], "-e") || !strncmp(argv[0], "--get-last-cpu-location", 10)) {
get_last_cpu_location = 1;
} else if (!strcmp(argv[0], "-t") || !strcmp(argv[0], "--threads")) {
#ifdef HWLOC_LINUX_SYS
show_threads = 1;
#else
fprintf (stderr, "Listing threads is currently only supported on Linux\n");
#endif
} else if (!strcmp (argv[0], "--whole-system")) {
flags |= HWLOC_TOPOLOGY_FLAG_WHOLE_SYSTEM;
} else if (!strcmp (argv[0], "--pid-cmd")) {
if (argc < 2) {
usage(callname, stdout);
exit(EXIT_FAILURE);
}
pidcmd = argv[1];
opt = 1;
} else {
fprintf (stderr, "Unrecognized option: %s\n", argv[0]);
usage (callname, stderr);
exit(EXIT_FAILURE);
}
argc -= opt+1;
argv += opt+1;
}
err = hwloc_topology_init(&topology);
if (err)
goto out;
hwloc_topology_set_flags(topology, flags);
err = hwloc_topology_load(topology);
if (err)
goto out_with_topology;
support = hwloc_topology_get_support(topology);
if (get_last_cpu_location) {
if (!support->cpubind->get_proc_last_cpu_location)
goto out_with_topology;
} else {
if (!support->cpubind->get_proc_cpubind)
goto out_with_topology;
}
topocpuset = hwloc_topology_get_topology_cpuset(topology);
dir = opendir("/proc");
if (!dir)
goto out_with_topology;
cpuset = hwloc_bitmap_alloc();
if (!cpuset)
goto out_with_dir;
while ((dirent = readdir(dir))) {
long pid_number;
hwloc_pid_t pid;
char pidoutput[1024];
char *end;
char name[64] = "";
/* management of threads */
unsigned boundthreads = 0, i;
long *tids = NULL; /* NULL if process is not threaded */
hwloc_bitmap_t *tidcpusets = NULL;
pid_number = strtol(dirent->d_name, &end, 10);
if (*end)
/* Not a number */
continue;
pid = hwloc_pid_from_number(pid_number, 0);
#ifdef HWLOC_LINUX_SYS
{
unsigned pathlen = 6 + strlen(dirent->d_name) + 1 + 7 + 1;
char *path;
int file;
ssize_t n;
path = malloc(pathlen);
snprintf(path, pathlen, "/proc/%s/cmdline", dirent->d_name);
file = open(path, O_RDONLY);
free(path);
if (file >= 0) {
n = read(file, name, sizeof(name) - 1);
close(file);
if (n <= 0)
/* Ignore kernel threads and errors */
continue;
name[n] = 0;
}
}
#endif /* HWLOC_LINUX_SYS */
if (show_threads) {
#ifdef HWLOC_LINUX_SYS
/* check if some threads must be displayed */
unsigned pathlen = 6 + strlen(dirent->d_name) + 1 + 4 + 1;
char *path;
DIR *taskdir;
path = malloc(pathlen);
snprintf(path, pathlen, "/proc/%s/task", dirent->d_name);
taskdir = opendir(path);
if (taskdir) {
struct dirent *taskdirent;
long tid;
unsigned n = 0;
/* count threads */
while ((taskdirent = readdir(taskdir))) {
tid = strtol(taskdirent->d_name, &end, 10);
if (*end)
/* Not a number */
continue;
n++;
}
if (n > 1) {
/* if there's more than one thread, see if some are bound */
tids = malloc(n * sizeof(*tids));
tidcpusets = calloc(n+1, sizeof(*tidcpusets));
if (tids && tidcpusets) {
/* reread the directory but gather info now */
rewinddir(taskdir);
i = 0;
while ((taskdirent = readdir(taskdir))) {
tid = strtol(taskdirent->d_name, &end, 10);
if (*end)
/* Not a number */
continue;
if (get_last_cpu_location) {
if (hwloc_linux_get_tid_last_cpu_location(topology, tid, cpuset))
continue;
} else {
if (hwloc_linux_get_tid_cpubind(topology, tid, cpuset))
continue;
}
hwloc_bitmap_and(cpuset, cpuset, topocpuset);
tids[i] = tid;
tidcpusets[i] = hwloc_bitmap_dup(cpuset);
i++;
if (hwloc_bitmap_iszero(cpuset))
continue;
if (hwloc_bitmap_isequal(cpuset, topocpuset) && !show_all)
continue;
boundthreads++;
}
} else {
/* failed to alloc, behave as if there were no threads */
free(tids);
tids = NULL;
free(tidcpusets);
tidcpusets = NULL;
}
}
closedir(taskdir);
}
#endif /* HWLOC_LINUX_SYS */
}
if (get_last_cpu_location) {
if (hwloc_get_proc_last_cpu_location(topology, pid, cpuset, 0))
continue;
} else {
if (hwloc_get_proc_cpubind(topology, pid, cpuset, 0))
continue;
}
hwloc_bitmap_and(cpuset, cpuset, topocpuset);
if (hwloc_bitmap_iszero(cpuset))
continue;
/* don't print anything if the process isn't bound and if no threads are bound and if not showing all */
if (hwloc_bitmap_isequal(cpuset, topocpuset) && (!tids || !boundthreads) && !show_all)
continue;
pidoutput[0] = '\0';
if (pidcmd) {
char *cmd;
FILE *file;
cmd = malloc(strlen(pidcmd)+1+5+2+1);
sprintf(cmd, "%s %u", pidcmd, pid);
file = popen(cmd, "r");
if (file) {
if (fgets(pidoutput, sizeof(pidoutput), file)) {
end = strchr(pidoutput, '\n');
if (end)
*end = '\0';
}
pclose(file);
}
free(cmd);
}
/* print the process */
print_task(topology, pid_number, name, cpuset, pidoutput[0] == '\0' ? NULL : pidoutput, 0);
if (tids)
/* print each tid we found (it's tidcpuset isn't NULL anymore) */
for(i=0; tidcpusets[i] != NULL; i++) {
print_task(topology, tids[i], "", tidcpusets[i], NULL, 1);
hwloc_bitmap_free(tidcpusets[i]);
}
/* free threads stuff */
free(tidcpusets);
free(tids);
}
err = 0;
hwloc_bitmap_free(cpuset);
out_with_dir:
closedir(dir);
out_with_topology:
hwloc_topology_destroy(topology);
out:
return err;
}
int
main (int argc, char *argv[])
{
int err;
hwloc_topology_t topology;
const char *filename = NULL;
unsigned long flags = HWLOC_TOPOLOGY_FLAG_IO_DEVICES | HWLOC_TOPOLOGY_FLAG_IO_BRIDGES | HWLOC_TOPOLOGY_FLAG_ICACHES;
unsigned long restrict_flags = 0;
int merge = 0;
int ignorecache = 0;
char * callname;
char * input = NULL;
enum hwloc_utils_input_format input_format = HWLOC_UTILS_INPUT_DEFAULT;
enum output_format output_format = LSTOPO_OUTPUT_DEFAULT;
char *restrictstring = NULL;
struct lstopo_output loutput;
int opt;
unsigned i;
loutput.overwrite = 0;
loutput.logical = -1;
loutput.legend = 1;
loutput.verbose_mode = LSTOPO_VERBOSE_MODE_DEFAULT;
for(i=0; i<HWLOC_OBJ_TYPE_MAX; i++)
force_orient[i] = LSTOPO_ORIENT_NONE;
force_orient[HWLOC_OBJ_PU] = LSTOPO_ORIENT_HORIZ;
force_orient[HWLOC_OBJ_CACHE] = LSTOPO_ORIENT_HORIZ;
force_orient[HWLOC_OBJ_NUMANODE] = LSTOPO_ORIENT_HORIZ;
/* enable verbose backends */
putenv("HWLOC_XML_VERBOSE=1");
putenv("HWLOC_SYNTHETIC_VERBOSE=1");
#ifdef HAVE_SETLOCALE
setlocale(LC_ALL, "");
#endif
callname = strrchr(argv[0], '/');
if (!callname)
callname = argv[0];
else
callname++;
/* skip argv[0], handle options */
argc--;
argv++;
err = hwloc_topology_init (&topology);
if (err)
return EXIT_FAILURE;
while (argc >= 1)
{
opt = 0;
if (!strcmp (argv[0], "-v") || !strcmp (argv[0], "--verbose")) {
loutput.verbose_mode++;
} else if (!strcmp (argv[0], "-s") || !strcmp (argv[0], "--silent")) {
loutput.verbose_mode--;
} else if (!strcmp (argv[0], "-h") || !strcmp (argv[0], "--help")) {
usage(callname, stdout);
exit(EXIT_SUCCESS);
} else if (!strcmp (argv[0], "-f") || !strcmp (argv[0], "--force"))
loutput.overwrite = 1;
else if (!strcmp (argv[0], "-l") || !strcmp (argv[0], "--logical"))
loutput.logical = 1;
else if (!strcmp (argv[0], "-p") || !strcmp (argv[0], "--physical"))
loutput.logical = 0;
else if (!strcmp (argv[0], "-c") || !strcmp (argv[0], "--cpuset"))
lstopo_show_cpuset = 1;
else if (!strcmp (argv[0], "-C") || !strcmp (argv[0], "--cpuset-only"))
lstopo_show_cpuset = 2;
else if (!strcmp (argv[0], "--taskset")) {
lstopo_show_taskset = 1;
if (!lstopo_show_cpuset)
lstopo_show_cpuset = 1;
} else if (!strcmp (argv[0], "--only")) {
if (argc < 2) {
usage (callname, stderr);
exit(EXIT_FAILURE);
}
if (hwloc_obj_type_sscanf(argv[1], &lstopo_show_only, NULL, NULL, 0) < 0)
fprintf(stderr, "Unsupported type `%s' passed to --only, ignoring.\n", argv[1]);
opt = 1;
}
else if (!strcmp (argv[0], "--ignore")) {
hwloc_obj_type_t type;
if (argc < 2) {
usage (callname, stderr);
exit(EXIT_FAILURE);
}
if (hwloc_obj_type_sscanf(argv[1], &type, NULL, NULL, 0) < 0)
fprintf(stderr, "Unsupported type `%s' passed to --ignore, ignoring.\n", argv[1]);
else if (type == HWLOC_OBJ_PU)
lstopo_ignore_pus = 1;
else
hwloc_topology_ignore_type(topology, type);
opt = 1;
}
else if (!strcmp (argv[0], "--no-caches"))
ignorecache = 2;
else if (!strcmp (argv[0], "--no-useless-caches"))
ignorecache = 1;
else if (!strcmp (argv[0], "--no-icaches"))
flags &= ~HWLOC_TOPOLOGY_FLAG_ICACHES;
else if (!strcmp (argv[0], "--whole-system"))
flags |= HWLOC_TOPOLOGY_FLAG_WHOLE_SYSTEM;
else if (!strcmp (argv[0], "--no-io"))
flags &= ~(HWLOC_TOPOLOGY_FLAG_IO_DEVICES | HWLOC_TOPOLOGY_FLAG_IO_BRIDGES);
else if (!strcmp (argv[0], "--no-bridges"))
flags &= ~(HWLOC_TOPOLOGY_FLAG_IO_BRIDGES);
else if (!strcmp (argv[0], "--whole-io"))
flags |= HWLOC_TOPOLOGY_FLAG_WHOLE_IO;
else if (!strcmp (argv[0], "--merge"))
merge = 1;
else if (!strcmp (argv[0], "--no-collapse"))
lstopo_collapse = 0;
else if (!strcmp (argv[0], "--thissystem"))
flags |= HWLOC_TOPOLOGY_FLAG_IS_THISSYSTEM;
else if (!strcmp (argv[0], "--restrict")) {
if (argc < 2) {
usage (callname, stderr);
exit(EXIT_FAILURE);
}
restrictstring = strdup(argv[1]);
opt = 1;
}
else if (!strcmp (argv[0], "--restrict-flags")) {
if (argc < 2) {
usage (callname, stderr);
exit(EXIT_FAILURE);
}
restrict_flags = (unsigned long) strtoull(argv[1], NULL, 0);
opt = 1;
}
else if (!strcmp (argv[0], "--export-synthetic-flags")) {
if (argc < 2) {
usage (callname, stderr);
exit(EXIT_FAILURE);
}
lstopo_export_synthetic_flags = (unsigned long) strtoull(argv[1], NULL, 0);
opt = 1;
}
else if (!strcmp (argv[0], "--horiz"))
for(i=0; i<HWLOC_OBJ_TYPE_MAX; i++)
force_orient[i] = LSTOPO_ORIENT_HORIZ;
else if (!strcmp (argv[0], "--vert"))
for(i=0; i<HWLOC_OBJ_TYPE_MAX; i++)
force_orient[i] = LSTOPO_ORIENT_VERT;
else if (!strcmp (argv[0], "--rect"))
for(i=0; i<HWLOC_OBJ_TYPE_MAX; i++)
force_orient[i] = LSTOPO_ORIENT_RECT;
else if (!strncmp (argv[0], "--horiz=", 8)
|| !strncmp (argv[0], "--vert=", 7)
|| !strncmp (argv[0], "--rect=", 7)) {
enum lstopo_orient_e orient = (argv[0][2] == 'h') ? LSTOPO_ORIENT_HORIZ : (argv[0][2] == 'v') ? LSTOPO_ORIENT_VERT : LSTOPO_ORIENT_RECT;
char *tmp = argv[0] + ((argv[0][2] == 'h') ? 8 : 7);
while (tmp) {
char *end = strchr(tmp, ',');
hwloc_obj_type_t type;
if (end)
*end = '\0';
if (hwloc_obj_type_sscanf(tmp, &type, NULL, NULL, 0) < 0)
fprintf(stderr, "Unsupported type `%s' passed to %s, ignoring.\n", tmp, argv[0]);
else
force_orient[type] = orient;
if (!end)
break;
tmp = end+1;
}
}
else if (!strcmp (argv[0], "--fontsize")) {
if (argc < 2) {
usage (callname, stderr);
exit(EXIT_FAILURE);
}
fontsize = atoi(argv[1]);
opt = 1;
}
else if (!strcmp (argv[0], "--gridsize")) {
if (argc < 2) {
usage (callname, stderr);
exit(EXIT_FAILURE);
}
gridsize = atoi(argv[1]);
opt = 1;
}
else if (!strcmp (argv[0], "--no-legend")) {
loutput.legend = 0;
}
else if (!strcmp (argv[0], "--append-legend")) {
char **tmp;
if (argc < 2) {
usage (callname, stderr);
exit(EXIT_FAILURE);
}
tmp = realloc(lstopo_append_legends, (lstopo_append_legends_nr+1) * sizeof(*lstopo_append_legends));
if (!tmp) {
fprintf(stderr, "Failed to realloc legend append array, legend ignored.\n");
} else {
lstopo_append_legends = tmp;
lstopo_append_legends[lstopo_append_legends_nr] = strdup(argv[1]);
lstopo_append_legends_nr++;
}
opt = 1;
}
else if (hwloc_utils_lookup_input_option(argv, argc, &opt,
&input, &input_format,
callname)) {
/* nothing to do anymore */
} else if (!strcmp (argv[0], "--pid")) {
if (argc < 2) {
usage (callname, stderr);
exit(EXIT_FAILURE);
}
lstopo_pid_number = atoi(argv[1]); opt = 1;
} else if (!strcmp (argv[0], "--ps") || !strcmp (argv[0], "--top"))
top = 1;
else if (!strcmp (argv[0], "--version")) {
printf("%s %s\n", callname, HWLOC_VERSION);
exit(EXIT_SUCCESS);
} else if (!strcmp (argv[0], "--output-format") || !strcmp (argv[0], "--of")) {
if (argc < 2) {
usage (callname, stderr);
exit(EXIT_FAILURE);
}
output_format = parse_output_format(argv[1], callname);
opt = 1;
} else {
if (filename) {
fprintf (stderr, "Unrecognized option: %s\n", argv[0]);
usage (callname, stderr);
exit(EXIT_FAILURE);
} else
filename = argv[0];
}
argc -= opt+1;
argv += opt+1;
}
if (lstopo_show_only != (hwloc_obj_type_t)-1)
merge = 0;
hwloc_topology_set_flags(topology, flags);
if (ignorecache > 1) {
hwloc_topology_ignore_type(topology, HWLOC_OBJ_CACHE);
} else if (ignorecache) {
hwloc_topology_ignore_type_keep_structure(topology, HWLOC_OBJ_CACHE);
}
if (merge)
hwloc_topology_ignore_all_keep_structure(topology);
if (input) {
err = hwloc_utils_enable_input_format(topology, input, &input_format, loutput.verbose_mode > 1, callname);
if (err)
return err;
}
if (lstopo_pid_number > 0) {
lstopo_pid = hwloc_pid_from_number(lstopo_pid_number, 0);
if (hwloc_topology_set_pid(topology, lstopo_pid)) {
perror("Setting target pid");
return EXIT_FAILURE;
}
}
/* if the output format wasn't enforced, look at the filename */
if (filename && output_format == LSTOPO_OUTPUT_DEFAULT) {
if (!strcmp(filename, "-")
|| !strcmp(filename, "/dev/stdout")) {
output_format = LSTOPO_OUTPUT_CONSOLE;
} else {
char *dot = strrchr(filename, '.');
if (dot)
output_format = parse_output_format(dot+1, callname);
else {
fprintf(stderr, "Cannot infer output type for file `%s' without any extension, using default output.\n", filename);
filename = NULL;
}
}
}
/* if the output format wasn't enforced, think a bit about what the user probably want */
if (output_format == LSTOPO_OUTPUT_DEFAULT) {
if (lstopo_show_cpuset
|| lstopo_show_only != (hwloc_obj_type_t)-1
|| loutput.verbose_mode != LSTOPO_VERBOSE_MODE_DEFAULT)
output_format = LSTOPO_OUTPUT_CONSOLE;
}
if (input_format == HWLOC_UTILS_INPUT_XML
&& output_format == LSTOPO_OUTPUT_XML) {
/* must be after parsing output format and before loading the topology */
putenv("HWLOC_XML_USERDATA_NOT_DECODED=1");
hwloc_topology_set_userdata_import_callback(topology, hwloc_utils_userdata_import_cb);
hwloc_topology_set_userdata_export_callback(topology, hwloc_utils_userdata_export_cb);
}
err = hwloc_topology_load (topology);
if (err) {
fprintf(stderr, "hwloc_topology_load() failed (%s).\n", strerror(errno));
return EXIT_FAILURE;
}
if (top)
add_process_objects(topology);
if (restrictstring) {
hwloc_bitmap_t restrictset = hwloc_bitmap_alloc();
if (!strcmp (restrictstring, "binding")) {
if (lstopo_pid_number > 0)
hwloc_get_proc_cpubind(topology, lstopo_pid, restrictset, HWLOC_CPUBIND_PROCESS);
else
hwloc_get_cpubind(topology, restrictset, HWLOC_CPUBIND_PROCESS);
} else {
hwloc_bitmap_sscanf(restrictset, restrictstring);
}
err = hwloc_topology_restrict (topology, restrictset, restrict_flags);
if (err) {
perror("Restricting the topology");
/* fallthrough */
}
hwloc_bitmap_free(restrictset);
free(restrictstring);
}
if (loutput.logical == -1) {
if (output_format == LSTOPO_OUTPUT_CONSOLE)
loutput.logical = 1;
else if (output_format != LSTOPO_OUTPUT_DEFAULT)
loutput.logical = 0;
}
loutput.topology = topology;
loutput.file = NULL;
lstopo_populate_userdata(hwloc_get_root_obj(topology));
if (output_format != LSTOPO_OUTPUT_XML && lstopo_collapse)
lstopo_add_collapse_attributes(topology);
switch (output_format) {
case LSTOPO_OUTPUT_DEFAULT:
#ifdef LSTOPO_HAVE_GRAPHICS
#if CAIRO_HAS_XLIB_SURFACE && defined HWLOC_HAVE_X11_KEYSYM
if (getenv("DISPLAY")) {
if (loutput.logical == -1)
loutput.logical = 0;
output_x11(&loutput, NULL);
} else
#endif /* CAIRO_HAS_XLIB_SURFACE */
#ifdef HWLOC_WIN_SYS
{
if (loutput.logical == -1)
loutput.logical = 0;
output_windows(&loutput, NULL);
}
#endif
#endif /* !LSTOPO_HAVE_GRAPHICS */
#if !defined HWLOC_WIN_SYS || !defined LSTOPO_HAVE_GRAPHICS
{
if (loutput.logical == -1)
loutput.logical = 1;
output_console(&loutput, NULL);
}
#endif
break;
case LSTOPO_OUTPUT_CONSOLE:
output_console(&loutput, filename);
break;
case LSTOPO_OUTPUT_SYNTHETIC:
output_synthetic(&loutput, filename);
break;
case LSTOPO_OUTPUT_ASCII:
output_ascii(&loutput, filename);
break;
case LSTOPO_OUTPUT_FIG:
output_fig(&loutput, filename);
break;
#ifdef LSTOPO_HAVE_GRAPHICS
# if CAIRO_HAS_PNG_FUNCTIONS
case LSTOPO_OUTPUT_PNG:
output_png(&loutput, filename);
break;
# endif /* CAIRO_HAS_PNG_FUNCTIONS */
# if CAIRO_HAS_PDF_SURFACE
case LSTOPO_OUTPUT_PDF:
output_pdf(&loutput, filename);
break;
# endif /* CAIRO_HAS_PDF_SURFACE */
# if CAIRO_HAS_PS_SURFACE
case LSTOPO_OUTPUT_PS:
output_ps(&loutput, filename);
break;
#endif /* CAIRO_HAS_PS_SURFACE */
#if CAIRO_HAS_SVG_SURFACE
case LSTOPO_OUTPUT_SVG:
output_svg(&loutput, filename);
break;
#endif /* CAIRO_HAS_SVG_SURFACE */
#endif /* LSTOPO_HAVE_GRAPHICS */
case LSTOPO_OUTPUT_XML:
output_xml(&loutput, filename);
break;
default:
fprintf(stderr, "file format not supported\n");
usage(callname, stderr);
exit(EXIT_FAILURE);
}
lstopo_destroy_userdata(hwloc_get_root_obj(topology));
hwloc_utils_userdata_free_recursive(hwloc_get_root_obj(topology));
hwloc_topology_destroy (topology);
for(i=0; i<lstopo_append_legends_nr; i++)
free(lstopo_append_legends[i]);
free(lstopo_append_legends);
return EXIT_SUCCESS;
}
static void add_process_objects(hwloc_topology_t topology)
{
#ifdef HAVE_DIRENT_H
hwloc_obj_t root;
hwloc_bitmap_t cpuset;
#ifdef HWLOC_LINUX_SYS
hwloc_bitmap_t task_cpuset;
#endif /* HWLOC_LINUX_SYS */
DIR *dir;
struct dirent *dirent;
const struct hwloc_topology_support *support;
root = hwloc_get_root_obj(topology);
support = hwloc_topology_get_support(topology);
if (!support->cpubind->get_proc_cpubind)
return;
dir = opendir("/proc");
if (!dir)
return;
cpuset = hwloc_bitmap_alloc();
#ifdef HWLOC_LINUX_SYS
task_cpuset = hwloc_bitmap_alloc();
#endif /* HWLOC_LINUX_SYS */
while ((dirent = readdir(dir))) {
long local_pid_number;
hwloc_pid_t local_pid;
char *end;
char name[64];
int proc_cpubind;
local_pid_number = strtol(dirent->d_name, &end, 10);
if (*end)
/* Not a number */
continue;
snprintf(name, sizeof(name), "%ld", local_pid_number);
local_pid = hwloc_pid_from_number(local_pid_number, 0);
proc_cpubind = hwloc_get_proc_cpubind(topology, local_pid, cpuset, 0) != -1;
#ifdef HWLOC_LINUX_SYS
{
/* Get the process name */
char *path;
unsigned pathlen = 6 + strlen(dirent->d_name) + 1 + 7 + 1;
char cmd[64], *c;
int file;
ssize_t n;
path = malloc(pathlen);
snprintf(path, pathlen, "/proc/%s/cmdline", dirent->d_name);
file = open(path, O_RDONLY);
free(path);
if (file >= 0) {
n = read(file, cmd, sizeof(cmd) - 1);
close(file);
if (n <= 0)
/* Ignore kernel threads and errors */
continue;
cmd[n] = 0;
if ((c = strchr(cmd, ' ')))
*c = 0;
snprintf(name, sizeof(name), "%ld %s", local_pid_number, cmd);
}
}
{
/* Get threads */
char *path;
unsigned pathlen = 6+strlen(dirent->d_name) + 1 + 4 + 1;
DIR *task_dir;
struct dirent *task_dirent;
path = malloc(pathlen);
snprintf(path, pathlen, "/proc/%s/task", dirent->d_name);
task_dir = opendir(path);
free(path);
if (task_dir) {
while ((task_dirent = readdir(task_dir))) {
long local_tid;
char *task_end;
char task_name[64];
local_tid = strtol(task_dirent->d_name, &task_end, 10);
if (*task_end)
/* Not a number, or the main task */
continue;
if (hwloc_linux_get_tid_cpubind(topology, local_tid, task_cpuset))
continue;
if (proc_cpubind && hwloc_bitmap_isequal(task_cpuset, cpuset))
continue;
snprintf(task_name, sizeof(task_name), "%s %li", name, local_tid);
insert_task(topology, task_cpuset, task_name);
}
closedir(task_dir);
}
}
#endif /* HWLOC_LINUX_SYS */
if (!proc_cpubind)
continue;
if (hwloc_bitmap_isincluded(root->cpuset, cpuset))
continue;
insert_task(topology, cpuset, name);
}
hwloc_bitmap_free(cpuset);
#ifdef HWLOC_LINUX_SYS
hwloc_bitmap_free(task_cpuset);
#endif /* HWLOC_LINUX_SYS */
closedir(dir);
#endif /* HAVE_DIRENT_H */
}
int main(int argc, char *argv[])
{
hwloc_topology_t topology;
int loaded = 0;
unsigned depth;
hwloc_bitmap_t cpubind_set, membind_set;
int got_cpubind = 0, got_membind = 0;
int working_on_cpubind = 1; /* membind if 0 */
int get_binding = 0;
int get_last_cpu_location = 0;
unsigned long flags = 0;
int force = 0;
int single = 0;
int verbose = 0;
int logical = 1;
int taskset = 0;
int cpubind_flags = 0;
hwloc_membind_policy_t membind_policy = HWLOC_MEMBIND_BIND;
int membind_flags = 0;
int opt;
int ret;
int pid_number = -1;
hwloc_pid_t pid = 0; /* only valid when pid_number > 0, but gcc-4.8 still reports uninitialized warnings */
char *callname;
cpubind_set = hwloc_bitmap_alloc();
membind_set = hwloc_bitmap_alloc();
/* don't load now, in case some options change the config before the topology is actually used */
#define LOADED() (loaded)
#define ENSURE_LOADED() do { \
if (!loaded) { \
hwloc_topology_init(&topology); \
hwloc_topology_set_all_types_filter(topology, HWLOC_TYPE_FILTER_KEEP_ALL); \
hwloc_topology_set_flags(topology, flags); \
hwloc_topology_load(topology); \
depth = hwloc_topology_get_depth(topology); \
loaded = 1; \
} \
} while (0)
callname = argv[0];
/* skip argv[0], handle options */
argv++;
argc--;
while (argc >= 1) {
if (!strcmp(argv[0], "--")) {
argc--;
argv++;
break;
}
opt = 0;
if (*argv[0] == '-') {
if (!strcmp(argv[0], "-v") || !strcmp(argv[0], "--verbose")) {
verbose++;
goto next;
}
if (!strcmp(argv[0], "-q") || !strcmp(argv[0], "--quiet")) {
verbose--;
goto next;
}
if (!strcmp(argv[0], "--help")) {
usage("hwloc-bind", stdout);
return EXIT_SUCCESS;
}
if (!strcmp(argv[0], "--single")) {
single = 1;
goto next;
}
if (!strcmp(argv[0], "-f") || !strcmp(argv[0], "--force")) {
force = 1;
goto next;
}
if (!strcmp(argv[0], "--strict")) {
cpubind_flags |= HWLOC_CPUBIND_STRICT;
membind_flags |= HWLOC_MEMBIND_STRICT;
goto next;
}
if (!strcmp(argv[0], "--pid")) {
if (argc < 2) {
usage ("hwloc-bind", stderr);
exit(EXIT_FAILURE);
}
pid_number = atoi(argv[1]);
opt = 1;
goto next;
}
if (!strcmp (argv[0], "--version")) {
printf("%s %s\n", callname, HWLOC_VERSION);
exit(EXIT_SUCCESS);
}
if (!strcmp(argv[0], "-l") || !strcmp(argv[0], "--logical")) {
logical = 1;
goto next;
}
if (!strcmp(argv[0], "-p") || !strcmp(argv[0], "--physical")) {
logical = 0;
goto next;
}
if (!strcmp(argv[0], "--taskset")) {
taskset = 1;
goto next;
}
if (!strcmp (argv[0], "-e") || !strncmp (argv[0], "--get-last-cpu-location", 10)) {
get_last_cpu_location = 1;
goto next;
}
if (!strcmp (argv[0], "--get")) {
get_binding = 1;
goto next;
}
if (!strcmp (argv[0], "--cpubind")) {
working_on_cpubind = 1;
goto next;
}
if (!strcmp (argv[0], "--membind")) {
working_on_cpubind = 0;
goto next;
}
if (!strcmp (argv[0], "--mempolicy")) {
if (!strncmp(argv[1], "default", 2))
membind_policy = HWLOC_MEMBIND_DEFAULT;
else if (!strncmp(argv[1], "firsttouch", 2))
membind_policy = HWLOC_MEMBIND_FIRSTTOUCH;
else if (!strncmp(argv[1], "bind", 2))
membind_policy = HWLOC_MEMBIND_BIND;
else if (!strncmp(argv[1], "interleave", 2))
membind_policy = HWLOC_MEMBIND_INTERLEAVE;
else if (!strncmp(argv[1], "nexttouch", 2))
membind_policy = HWLOC_MEMBIND_NEXTTOUCH;
else {
fprintf(stderr, "Unrecognized memory binding policy %s\n", argv[1]);
usage ("hwloc-bind", stderr);
exit(EXIT_FAILURE);
}
opt = 1;
goto next;
}
if (!strcmp (argv[0], "--whole-system")) {
if (loaded) {
fprintf(stderr, "Input option %s disallowed after options using the topology\n", argv[0]);
exit(EXIT_FAILURE);
}
flags |= HWLOC_TOPOLOGY_FLAG_WHOLE_SYSTEM;
goto next;
}
if (!strcmp (argv[0], "--restrict")) {
hwloc_bitmap_t restrictset;
int err;
if (argc < 2) {
usage (callname, stdout);
exit(EXIT_FAILURE);
}
restrictset = hwloc_bitmap_alloc();
hwloc_bitmap_sscanf(restrictset, argv[1]);
ENSURE_LOADED();
err = hwloc_topology_restrict (topology, restrictset, 0);
if (err) {
perror("Restricting the topology");
/* fallthrough */
}
hwloc_bitmap_free(restrictset);
argc--;
argv++;
goto next;
}
fprintf (stderr, "Unrecognized option: %s\n", argv[0]);
usage("hwloc-bind", stderr);
return EXIT_FAILURE;
}
ENSURE_LOADED();
ret = hwloc_calc_process_arg(topology, depth, argv[0], logical,
working_on_cpubind ? cpubind_set : membind_set,
verbose);
if (ret < 0) {
if (verbose > 0)
fprintf(stderr, "assuming the command starts at %s\n", argv[0]);
break;
}
if (working_on_cpubind)
got_cpubind = 1;
else
got_membind = 1;
next:
argc -= opt+1;
argv += opt+1;
}
ENSURE_LOADED();
if (pid_number > 0) {
pid = hwloc_pid_from_number(pid_number, !(get_binding || get_last_cpu_location));
/* no need to set_pid()
* the doc just says we're operating on pid, not that we're retrieving the topo/cpuset as seen from inside pid
*/
}
if (get_last_cpu_location && !working_on_cpubind) {
fprintf(stderr, "Options --membind and --get-last-cpu-location cannot be combined.\n");
return EXIT_FAILURE;
}
if ((get_binding || get_last_cpu_location) && (got_cpubind || got_membind)) {
/* doesn't work because get_binding/get_last_cpu_location overwrites cpubind_set */
fprintf(stderr, "Cannot display and set binding at the same time.\n");
return EXIT_FAILURE;
}
if (get_binding || get_last_cpu_location) {
char *s;
const char *policystr = NULL;
int err;
if (working_on_cpubind) {
if (get_last_cpu_location) {
if (pid_number > 0)
err = hwloc_get_proc_last_cpu_location(topology, pid, cpubind_set, 0);
else
err = hwloc_get_last_cpu_location(topology, cpubind_set, 0);
} else {
if (pid_number > 0)
err = hwloc_get_proc_cpubind(topology, pid, cpubind_set, 0);
else
err = hwloc_get_cpubind(topology, cpubind_set, 0);
}
if (err) {
const char *errmsg = strerror(errno);
if (pid_number > 0)
fprintf(stderr, "hwloc_get_proc_%s %d failed (errno %d %s)\n", get_last_cpu_location ? "last_cpu_location" : "cpubind", pid_number, errno, errmsg);
else
fprintf(stderr, "hwloc_get_%s failed (errno %d %s)\n", get_last_cpu_location ? "last_cpu_location" : "cpubind", errno, errmsg);
return EXIT_FAILURE;
}
if (taskset)
hwloc_bitmap_taskset_asprintf(&s, cpubind_set);
else
hwloc_bitmap_asprintf(&s, cpubind_set);
} else {
hwloc_membind_policy_t policy;
if (pid_number > 0)
err = hwloc_get_proc_membind(topology, pid, membind_set, &policy, 0);
else
err = hwloc_get_membind(topology, membind_set, &policy, 0);
if (err) {
const char *errmsg = strerror(errno);
if (pid_number > 0)
fprintf(stderr, "hwloc_get_proc_membind %d failed (errno %d %s)\n", pid_number, errno, errmsg);
else
fprintf(stderr, "hwloc_get_membind failed (errno %d %s)\n", errno, errmsg);
return EXIT_FAILURE;
}
if (taskset)
hwloc_bitmap_taskset_asprintf(&s, membind_set);
else
hwloc_bitmap_asprintf(&s, membind_set);
switch (policy) {
case HWLOC_MEMBIND_DEFAULT: policystr = "default"; break;
case HWLOC_MEMBIND_FIRSTTOUCH: policystr = "firsttouch"; break;
case HWLOC_MEMBIND_BIND: policystr = "bind"; break;
case HWLOC_MEMBIND_INTERLEAVE: policystr = "interleave"; break;
case HWLOC_MEMBIND_NEXTTOUCH: policystr = "nexttouch"; break;
default: fprintf(stderr, "unknown memory policy %d\n", policy); assert(0); break;
}
}
if (policystr)
printf("%s (%s)\n", s, policystr);
else
printf("%s\n", s);
free(s);
}
if (got_membind) {
if (hwloc_bitmap_iszero(membind_set)) {
if (verbose >= 0)
fprintf(stderr, "cannot membind to empty set\n");
if (!force)
goto failed_binding;
}
if (verbose > 0) {
char *s;
hwloc_bitmap_asprintf(&s, membind_set);
fprintf(stderr, "binding on memory set %s\n", s);
free(s);
}
if (single)
hwloc_bitmap_singlify(membind_set);
if (pid_number > 0)
ret = hwloc_set_proc_membind(topology, pid, membind_set, membind_policy, membind_flags);
else
ret = hwloc_set_membind(topology, membind_set, membind_policy, membind_flags);
if (ret && verbose >= 0) {
int bind_errno = errno;
const char *errmsg = strerror(bind_errno);
char *s;
hwloc_bitmap_asprintf(&s, membind_set);
if (pid_number > 0)
fprintf(stderr, "hwloc_set_proc_membind %s %d failed (errno %d %s)\n", s, pid_number, bind_errno, errmsg);
else
fprintf(stderr, "hwloc_set_membind %s failed (errno %d %s)\n", s, bind_errno, errmsg);
free(s);
}
if (ret && !force)
goto failed_binding;
}
if (got_cpubind) {
if (hwloc_bitmap_iszero(cpubind_set)) {
if (verbose >= 0)
fprintf(stderr, "cannot cpubind to empty set\n");
if (!force)
goto failed_binding;
}
if (verbose > 0) {
char *s;
hwloc_bitmap_asprintf(&s, cpubind_set);
fprintf(stderr, "binding on cpu set %s\n", s);
free(s);
}
if (single)
hwloc_bitmap_singlify(cpubind_set);
if (pid_number > 0)
ret = hwloc_set_proc_cpubind(topology, pid, cpubind_set, cpubind_flags);
else
ret = hwloc_set_cpubind(topology, cpubind_set, cpubind_flags);
if (ret && verbose >= 0) {
int bind_errno = errno;
const char *errmsg = strerror(bind_errno);
char *s;
hwloc_bitmap_asprintf(&s, cpubind_set);
if (pid_number > 0)
fprintf(stderr, "hwloc_set_proc_cpubind %s %d failed (errno %d %s)\n", s, pid_number, bind_errno, errmsg);
else
fprintf(stderr, "hwloc_set_cpubind %s failed (errno %d %s)\n", s, bind_errno, errmsg);
free(s);
}
if (ret && !force)
goto failed_binding;
}
hwloc_bitmap_free(cpubind_set);
hwloc_bitmap_free(membind_set);
hwloc_topology_destroy(topology);
if (pid_number > 0)
return EXIT_SUCCESS;
if (0 == argc) {
if (get_binding || get_last_cpu_location)
return EXIT_SUCCESS;
fprintf(stderr, "%s: nothing to do!\n", callname);
return EXIT_FAILURE;
}
/* FIXME: check whether Windows execvp() passes INHERIT_PARENT_AFFINITY to CreateProcess()
* because we need to propagate processor group affinity. However process-wide affinity
* isn't supported with processor groups so far.
*/
ret = execvp(argv[0], argv);
if (ret) {
fprintf(stderr, "%s: Failed to launch executable \"%s\"\n",
callname, argv[0]);
perror("execvp");
}
return EXIT_FAILURE;
failed_binding:
hwloc_bitmap_free(cpubind_set);
hwloc_bitmap_free(membind_set);
hwloc_topology_destroy(topology);
return EXIT_FAILURE;
}
int main(int argc, char *argv[])
{
hwloc_topology_t topology;
unsigned depth;
hwloc_bitmap_t cpubind_set, membind_set;
int got_cpubind = 0, got_membind = 0;
int working_on_cpubind = 1; /* membind if 0 */
int get_binding = 0;
int get_last_cpu_location = 0;
unsigned long flags = HWLOC_TOPOLOGY_FLAG_WHOLE_IO|HWLOC_TOPOLOGY_FLAG_ICACHES;
int force = 0;
int single = 0;
int verbose = 0;
int logical = 1;
int taskset = 0;
int cpubind_flags = 0;
hwloc_membind_policy_t membind_policy = HWLOC_MEMBIND_BIND;
int membind_flags = 0;
int opt;
int ret;
int pid_number = 0;
hwloc_pid_t pid;
char *callname;
cpubind_set = hwloc_bitmap_alloc();
membind_set = hwloc_bitmap_alloc();
hwloc_topology_init(&topology);
hwloc_topology_set_flags(topology, flags);
hwloc_topology_load(topology);
depth = hwloc_topology_get_depth(topology);
callname = argv[0];
/* skip argv[0], handle options */
argv++;
argc--;
while (argc >= 1) {
if (!strcmp(argv[0], "--")) {
argc--;
argv++;
break;
}
opt = 0;
if (*argv[0] == '-') {
if (!strcmp(argv[0], "-v") || !strcmp(argv[0], "--verbose")) {
verbose++;
goto next;
}
else if (!strcmp(argv[0], "-q") || !strcmp(argv[0], "--quiet")) {
verbose--;
goto next;
}
else if (!strcmp(argv[0], "--help")) {
usage("hwloc-bind", stdout);
return EXIT_SUCCESS;
}
else if (!strcmp(argv[0], "--single")) {
single = 1;
goto next;
}
else if (!strcmp(argv[0], "-f") || !strcmp(argv[0], "--force")) {
force = 1;
goto next;
}
else if (!strcmp(argv[0], "--strict")) {
cpubind_flags |= HWLOC_CPUBIND_STRICT;
membind_flags |= HWLOC_MEMBIND_STRICT;
goto next;
}
else if (!strcmp(argv[0], "--pid")) {
if (argc < 2) {
usage ("hwloc-bind", stderr);
exit(EXIT_FAILURE);
}
pid_number = atoi(argv[1]);
opt = 1;
goto next;
}
else if (!strcmp (argv[0], "--version")) {
printf("%s %s\n", callname, VERSION);
exit(EXIT_SUCCESS);
}
if (!strcmp(argv[0], "-l") || !strcmp(argv[0], "--logical")) {
logical = 1;
goto next;
}
if (!strcmp(argv[0], "-p") || !strcmp(argv[0], "--physical")) {
logical = 0;
goto next;
}
if (!strcmp(argv[0], "--taskset")) {
taskset = 1;
goto next;
}
else if (!strcmp (argv[0], "-e") || !strncmp (argv[0], "--get-last-cpu-location", 10)) {
get_last_cpu_location = 1;
goto next;
}
else if (!strcmp (argv[0], "--get")) {
get_binding = 1;
goto next;
}
else if (!strcmp (argv[0], "--cpubind")) {
working_on_cpubind = 1;
goto next;
}
else if (!strcmp (argv[0], "--membind")) {
working_on_cpubind = 0;
goto next;
}
else if (!strcmp (argv[0], "--mempolicy")) {
if (!strncmp(argv[1], "default", 2))
membind_policy = HWLOC_MEMBIND_DEFAULT;
else if (!strncmp(argv[1], "firsttouch", 2))
membind_policy = HWLOC_MEMBIND_FIRSTTOUCH;
else if (!strncmp(argv[1], "bind", 2))
membind_policy = HWLOC_MEMBIND_BIND;
else if (!strncmp(argv[1], "interleave", 2))
membind_policy = HWLOC_MEMBIND_INTERLEAVE;
else if (!strncmp(argv[1], "replicate", 2))
membind_policy = HWLOC_MEMBIND_REPLICATE;
else if (!strncmp(argv[1], "nexttouch", 2))
membind_policy = HWLOC_MEMBIND_NEXTTOUCH;
else {
fprintf(stderr, "Unrecognized memory binding policy %s\n", argv[1]);
usage ("hwloc-bind", stderr);
exit(EXIT_FAILURE);
}
opt = 1;
goto next;
}
else if (!strcmp (argv[0], "--whole-system")) {
flags |= HWLOC_TOPOLOGY_FLAG_WHOLE_SYSTEM;
hwloc_topology_destroy(topology);
hwloc_topology_init(&topology);
hwloc_topology_set_flags(topology, flags);
hwloc_topology_load(topology);
depth = hwloc_topology_get_depth(topology);
goto next;
}
else if (!strcmp (argv[0], "--restrict")) {
hwloc_bitmap_t restrictset;
int err;
if (argc < 2) {
usage (callname, stdout);
exit(EXIT_FAILURE);
}
restrictset = hwloc_bitmap_alloc();
hwloc_bitmap_sscanf(restrictset, argv[1]);
err = hwloc_topology_restrict (topology, restrictset, 0);
if (err) {
perror("Restricting the topology");
/* fallthrough */
}
hwloc_bitmap_free(restrictset);
argc--;
argv++;
goto next;
}
fprintf (stderr, "Unrecognized option: %s\n", argv[0]);
usage("hwloc-bind", stderr);
return EXIT_FAILURE;
}
ret = hwloc_calc_process_arg(topology, depth, argv[0], logical,
working_on_cpubind ? cpubind_set : membind_set,
verbose);
if (ret < 0) {
if (verbose > 0)
fprintf(stderr, "assuming the command starts at %s\n", argv[0]);
break;
}
if (working_on_cpubind)
got_cpubind = 1;
else
got_membind = 1;
next:
argc -= opt+1;
argv += opt+1;
}
pid = hwloc_pid_from_number(pid_number, !(get_binding || get_last_cpu_location));
if (get_binding || get_last_cpu_location) {
char *s;
const char *policystr = NULL;
int err;
if (working_on_cpubind) {
if (get_last_cpu_location) {
if (pid_number)
err = hwloc_get_proc_last_cpu_location(topology, pid, cpubind_set, 0);
else
err = hwloc_get_last_cpu_location(topology, cpubind_set, 0);
} else {
if (pid_number)
err = hwloc_get_proc_cpubind(topology, pid, cpubind_set, 0);
else
err = hwloc_get_cpubind(topology, cpubind_set, 0);
}
if (err) {
const char *errmsg = strerror(errno);
if (pid_number)
fprintf(stderr, "hwloc_get_proc_%s %d failed (errno %d %s)\n", get_last_cpu_location ? "last_cpu_location" : "cpubind", pid_number, errno, errmsg);
else
fprintf(stderr, "hwloc_get_%s failed (errno %d %s)\n", get_last_cpu_location ? "last_cpu_location" : "cpubind", errno, errmsg);
return EXIT_FAILURE;
}
if (taskset)
hwloc_bitmap_taskset_asprintf(&s, cpubind_set);
else
hwloc_bitmap_asprintf(&s, cpubind_set);
} else {
hwloc_membind_policy_t policy;
if (pid_number)
err = hwloc_get_proc_membind(topology, pid, membind_set, &policy, 0);
else
err = hwloc_get_membind(topology, membind_set, &policy, 0);
if (err) {
const char *errmsg = strerror(errno);
if (pid_number)
fprintf(stderr, "hwloc_get_proc_membind %d failed (errno %d %s)\n", pid_number, errno, errmsg);
else
fprintf(stderr, "hwloc_get_membind failed (errno %d %s)\n", errno, errmsg);
return EXIT_FAILURE;
}
if (taskset)
hwloc_bitmap_taskset_asprintf(&s, membind_set);
else
hwloc_bitmap_asprintf(&s, membind_set);
switch (policy) {
case HWLOC_MEMBIND_DEFAULT: policystr = "default"; break;
case HWLOC_MEMBIND_FIRSTTOUCH: policystr = "firsttouch"; break;
case HWLOC_MEMBIND_BIND: policystr = "bind"; break;
case HWLOC_MEMBIND_INTERLEAVE: policystr = "interleave"; break;
case HWLOC_MEMBIND_REPLICATE: policystr = "replicate"; break;
case HWLOC_MEMBIND_NEXTTOUCH: policystr = "nexttouch"; break;
default: fprintf(stderr, "unknown memory policy %d\n", policy); assert(0); break;
}
}
if (policystr)
printf("%s (%s)\n", s, policystr);
else
printf("%s\n", s);
free(s);
return EXIT_SUCCESS;
}
if (got_membind) {
if (hwloc_bitmap_iszero(membind_set)) {
if (verbose >= 0)
fprintf(stderr, "cannot membind to empty set\n");
if (!force)
goto failed_binding;
}
if (verbose > 0) {
char *s;
hwloc_bitmap_asprintf(&s, membind_set);
fprintf(stderr, "binding on memory set %s\n", s);
free(s);
}
if (single)
hwloc_bitmap_singlify(membind_set);
if (pid_number)
ret = hwloc_set_proc_membind(topology, pid, membind_set, membind_policy, membind_flags);
else
ret = hwloc_set_membind(topology, membind_set, membind_policy, membind_flags);
if (ret && verbose >= 0) {
int bind_errno = errno;
const char *errmsg = strerror(bind_errno);
char *s;
hwloc_bitmap_asprintf(&s, membind_set);
if (pid_number)
fprintf(stderr, "hwloc_set_proc_membind %s %d failed (errno %d %s)\n", s, pid_number, bind_errno, errmsg);
else
fprintf(stderr, "hwloc_set_membind %s failed (errno %d %s)\n", s, bind_errno, errmsg);
free(s);
}
if (ret && !force)
goto failed_binding;
}
if (got_cpubind) {
if (hwloc_bitmap_iszero(cpubind_set)) {
if (verbose >= 0)
fprintf(stderr, "cannot cpubind to empty set\n");
if (!force)
goto failed_binding;
}
if (verbose > 0) {
char *s;
hwloc_bitmap_asprintf(&s, cpubind_set);
fprintf(stderr, "binding on cpu set %s\n", s);
free(s);
}
if (single)
hwloc_bitmap_singlify(cpubind_set);
if (pid_number)
ret = hwloc_set_proc_cpubind(topology, pid, cpubind_set, cpubind_flags);
else
ret = hwloc_set_cpubind(topology, cpubind_set, cpubind_flags);
if (ret && verbose >= 0) {
int bind_errno = errno;
const char *errmsg = strerror(bind_errno);
char *s;
hwloc_bitmap_asprintf(&s, cpubind_set);
if (pid_number)
fprintf(stderr, "hwloc_set_proc_cpubind %s %d failed (errno %d %s)\n", s, pid_number, bind_errno, errmsg);
else
fprintf(stderr, "hwloc_set_cpubind %s failed (errno %d %s)\n", s, bind_errno, errmsg);
free(s);
}
if (ret && !force)
goto failed_binding;
}
hwloc_bitmap_free(cpubind_set);
hwloc_bitmap_free(membind_set);
hwloc_topology_destroy(topology);
if (pid_number)
return EXIT_SUCCESS;
if (0 == argc) {
fprintf(stderr, "%s: nothing to do!\n", callname);
return EXIT_FAILURE;
}
ret = execvp(argv[0], argv);
if (ret) {
fprintf(stderr, "%s: Failed to launch executable \"%s\"\n",
callname, argv[0]);
perror("execvp");
}
return EXIT_FAILURE;
failed_binding:
hwloc_bitmap_free(cpubind_set);
hwloc_bitmap_free(membind_set);
hwloc_topology_destroy(topology);
return EXIT_FAILURE;
}
int MPIR_Init_thread(int *argc, char ***argv, int required, int *provided)
{
int mpi_errno = MPI_SUCCESS;
int has_args;
int has_env;
int thread_provided = 0;
int exit_init_cs_on_failure = 0;
MPIR_Info *info_ptr;
#if defined(MPICH_IS_THREADED)
bool cs_initialized = false;
#endif
/* The threading library must be initialized at the very beginning because
* it manages all synchronization objects (e.g., mutexes) that will be
* initialized later */
{
int thread_err;
MPL_thread_init(&thread_err);
if (thread_err)
goto fn_fail;
}
#ifdef HAVE_HWLOC
MPIR_Process.bindset = hwloc_bitmap_alloc();
hwloc_topology_init(&MPIR_Process.hwloc_topology);
MPIR_Process.bindset_is_valid = 0;
hwloc_topology_set_io_types_filter(MPIR_Process.hwloc_topology, HWLOC_TYPE_FILTER_KEEP_ALL);
if (!hwloc_topology_load(MPIR_Process.hwloc_topology)) {
MPIR_Process.bindset_is_valid =
!hwloc_get_proc_cpubind(MPIR_Process.hwloc_topology, getpid(), MPIR_Process.bindset,
HWLOC_CPUBIND_PROCESS);
}
#endif
#ifdef HAVE_NETLOC
MPIR_Process.network_attr.u.tree.node_levels = NULL;
MPIR_Process.network_attr.network_endpoint = NULL;
MPIR_Process.netloc_topology = NULL;
MPIR_Process.network_attr.type = MPIR_NETLOC_NETWORK_TYPE__INVALID;
if (strlen(MPIR_CVAR_NETLOC_NODE_FILE)) {
mpi_errno =
netloc_parse_topology(&MPIR_Process.netloc_topology, MPIR_CVAR_NETLOC_NODE_FILE);
if (mpi_errno == NETLOC_SUCCESS) {
MPIR_Netloc_parse_topology(MPIR_Process.netloc_topology, &MPIR_Process.network_attr);
}
}
#endif
/* For any code in the device that wants to check for runtime
* decisions on the value of isThreaded, set a provisional
* value here. We could let the MPID_Init routine override this */
#if defined MPICH_IS_THREADED
MPIR_ThreadInfo.isThreaded = required == MPI_THREAD_MULTIPLE;
#endif /* MPICH_IS_THREADED */
#if defined(MPICH_IS_THREADED)
mpi_errno = thread_cs_init();
cs_initialized = true;
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
#endif
/* FIXME: Move to os-dependent interface? */
#ifdef HAVE_WINDOWS_H
/* prevent the process from bringing up an error message window if mpich
* asserts */
_CrtSetReportMode(_CRT_ASSERT, _CRTDBG_MODE_FILE);
_CrtSetReportFile(_CRT_ASSERT, _CRTDBG_FILE_STDERR);
_CrtSetReportHook2(_CRT_RPTHOOK_INSTALL, assert_hook);
#ifdef _WIN64
{
/* FIXME: (Windows) This severly degrades performance but fixes alignment
* issues with the datatype code. */
/* Prevent misaligned faults on Win64 machines */
UINT mode, old_mode;
old_mode = SetErrorMode(SEM_NOALIGNMENTFAULTEXCEPT);
mode = old_mode | SEM_NOALIGNMENTFAULTEXCEPT;
SetErrorMode(mode);
}
#endif
#endif
/* We need this inorder to implement IS_THREAD_MAIN */
#if (MPICH_THREAD_LEVEL >= MPI_THREAD_SERIALIZED) && defined(MPICH_IS_THREADED)
{
MPID_Thread_self(&MPIR_ThreadInfo.master_thread);
}
#endif
#ifdef HAVE_ERROR_CHECKING
/* Because the PARAM system has not been initialized, temporarily
* uncondtionally enable error checks. Once the PARAM system is
* initialized, this may be reset */
MPIR_Process.do_error_checks = 1;
#else
MPIR_Process.do_error_checks = 0;
#endif
/* Initialize necessary subsystems and setup the predefined attribute
* values. Subsystems may change these values. */
MPIR_Process.attrs.appnum = -1;
MPIR_Process.attrs.host = MPI_PROC_NULL;
MPIR_Process.attrs.io = MPI_PROC_NULL;
MPIR_Process.attrs.lastusedcode = MPI_ERR_LASTCODE;
MPIR_Process.attrs.universe = MPIR_UNIVERSE_SIZE_NOT_SET;
MPIR_Process.attrs.wtime_is_global = 0;
/* Set the functions used to duplicate attributes. These are
* when the first corresponding keyval is created */
MPIR_Process.attr_dup = 0;
MPIR_Process.attr_free = 0;
#ifdef HAVE_CXX_BINDING
/* Set the functions used to call functions in the C++ binding
* for reductions and attribute operations. These are null
* until a C++ operation is defined. This allows the C code
* that implements these operations to not invoke a C++ code
* directly, which may force the inclusion of symbols known only
* to the C++ compiler (e.g., under more non-GNU compilers, including
* Solaris and IRIX). */
MPIR_Process.cxx_call_op_fn = 0;
#endif
#ifdef HAVE_F08_BINDING
MPIR_C_MPI_UNWEIGHTED = MPI_UNWEIGHTED;
MPIR_C_MPI_WEIGHTS_EMPTY = MPI_WEIGHTS_EMPTY;
#endif
/* This allows the device to select an alternative function for
* dimsCreate */
MPIR_Process.dimsCreate = 0;
/* "Allocate" from the reserved space for builtin communicators and
* (partially) initialize predefined communicators. comm_parent is
* intially NULL and will be allocated by the device if the process group
* was started using one of the MPI_Comm_spawn functions. */
MPIR_Process.comm_world = MPIR_Comm_builtin + 0;
MPII_Comm_init(MPIR_Process.comm_world);
MPIR_Process.comm_world->handle = MPI_COMM_WORLD;
MPIR_Process.comm_world->context_id = 0 << MPIR_CONTEXT_PREFIX_SHIFT;
MPIR_Process.comm_world->recvcontext_id = 0 << MPIR_CONTEXT_PREFIX_SHIFT;
MPIR_Process.comm_world->comm_kind = MPIR_COMM_KIND__INTRACOMM;
/* This initialization of the comm name could be done only when
* comm_get_name is called */
MPL_strncpy(MPIR_Process.comm_world->name, "MPI_COMM_WORLD", MPI_MAX_OBJECT_NAME);
MPIR_Process.comm_self = MPIR_Comm_builtin + 1;
MPII_Comm_init(MPIR_Process.comm_self);
MPIR_Process.comm_self->handle = MPI_COMM_SELF;
MPIR_Process.comm_self->context_id = 1 << MPIR_CONTEXT_PREFIX_SHIFT;
MPIR_Process.comm_self->recvcontext_id = 1 << MPIR_CONTEXT_PREFIX_SHIFT;
MPIR_Process.comm_self->comm_kind = MPIR_COMM_KIND__INTRACOMM;
MPL_strncpy(MPIR_Process.comm_self->name, "MPI_COMM_SELF", MPI_MAX_OBJECT_NAME);
#ifdef MPID_NEEDS_ICOMM_WORLD
MPIR_Process.icomm_world = MPIR_Comm_builtin + 2;
MPII_Comm_init(MPIR_Process.icomm_world);
MPIR_Process.icomm_world->handle = MPIR_ICOMM_WORLD;
MPIR_Process.icomm_world->context_id = 2 << MPIR_CONTEXT_PREFIX_SHIFT;
MPIR_Process.icomm_world->recvcontext_id = 2 << MPIR_CONTEXT_PREFIX_SHIFT;
MPIR_Process.icomm_world->comm_kind = MPIR_COMM_KIND__INTRACOMM;
MPL_strncpy(MPIR_Process.icomm_world->name, "MPI_ICOMM_WORLD", MPI_MAX_OBJECT_NAME);
/* Note that these communicators are not ready for use - MPID_Init
* will setup self and world, and icomm_world if it desires it. */
#endif
MPIR_Process.comm_parent = NULL;
/* Setup the initial communicator list in case we have
* enabled the debugger message-queue interface */
MPII_COMML_REMEMBER(MPIR_Process.comm_world);
MPII_COMML_REMEMBER(MPIR_Process.comm_self);
/* MPIU_Timer_pre_init(); */
/* Wait for debugger to attach if requested. */
if (MPIR_CVAR_DEBUG_HOLD) {
volatile int hold = 1;
while (hold)
#ifdef HAVE_USLEEP
usleep(100);
#endif
;
}
#if defined(HAVE_ERROR_CHECKING) && (HAVE_ERROR_CHECKING == MPID_ERROR_LEVEL_RUNTIME)
MPIR_Process.do_error_checks = MPIR_CVAR_ERROR_CHECKING;
#endif
/* define MPI as initialized so that we can use MPI functions within
* MPID_Init if necessary */
OPA_store_int(&MPIR_Process.mpich_state, MPICH_MPI_STATE__IN_INIT);
/* We can't acquire any critical sections until this point. Any
* earlier the basic data structures haven't been initialized */
MPID_THREAD_CS_ENTER(GLOBAL, MPIR_THREAD_GLOBAL_ALLFUNC_MUTEX);
exit_init_cs_on_failure = 1;
/* create MPI_INFO_NULL object */
/* FIXME: Currently this info object is empty, we need to add data to this
* as defined by the standard. */
info_ptr = MPIR_Info_builtin + 1;
info_ptr->handle = MPI_INFO_ENV;
MPIR_Object_set_ref(info_ptr, 1);
info_ptr->next = NULL;
info_ptr->key = NULL;
info_ptr->value = NULL;
#ifdef USE_MEMORY_TRACING
MPL_trinit();
#endif
/* Set the number of tag bits. The device may override this value. */
MPIR_Process.tag_bits = MPIR_TAG_BITS_DEFAULT;
/* Create complete request to return in the event of immediately complete
* operations. Use a SEND request to cover all possible use-cases. */
MPIR_Process.lw_req = MPIR_Request_create(MPIR_REQUEST_KIND__SEND);
MPIR_ERR_CHKANDSTMT(MPIR_Process.lw_req == NULL, mpi_errno, MPIX_ERR_NOREQ, goto fn_fail,
"**nomemreq");
MPIR_cc_set(&MPIR_Process.lw_req->cc, 0);
mpi_errno = MPID_Init(argc, argv, required, &thread_provided, &has_args, &has_env);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
/* Initialize collectives infrastructure */
mpi_errno = MPII_Coll_init();
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
/* Set tag_ub as function of tag_bits set by the device */
MPIR_Process.attrs.tag_ub = MPIR_TAG_USABLE_BITS;
/* Assert: tag_ub should be a power of 2 minus 1 */
MPIR_Assert(((unsigned) MPIR_Process.
attrs.tag_ub & ((unsigned) MPIR_Process.attrs.tag_ub + 1)) == 0);
/* Assert: tag_ub is at least the minimum asked for in the MPI spec */
MPIR_Assert(MPIR_Process.attrs.tag_ub >= 32767);
/* Capture the level of thread support provided */
MPIR_ThreadInfo.thread_provided = thread_provided;
if (provided)
*provided = thread_provided;
#if defined MPICH_IS_THREADED
MPIR_ThreadInfo.isThreaded = (thread_provided == MPI_THREAD_MULTIPLE);
#endif /* MPICH_IS_THREADED */
/* FIXME: Define these in the interface. Does Timer init belong here? */
MPII_Timer_init(MPIR_Process.comm_world->rank, MPIR_Process.comm_world->local_size);
#ifdef USE_MEMORY_TRACING
#ifdef MPICH_IS_THREADED
MPL_trconfig(MPIR_Process.comm_world->rank, MPIR_ThreadInfo.isThreaded);
#else
MPL_trconfig(MPIR_Process.comm_world->rank, 0);
#endif
/* Indicate that we are near the end of the init step; memory
* allocated already will have an id of zero; this helps
* separate memory leaks in the initialization code from
* leaks in the "active" code */
#endif
#ifdef MPL_USE_DBG_LOGGING
/* FIXME: This is a hack to handle the common case of two worlds.
* If the parent comm is not NULL, we always give the world number
* as "1" (false). */
#ifdef MPICH_IS_THREADED
MPL_dbg_init(argc, argv, has_args, has_env,
MPIR_Process.comm_parent != NULL, MPIR_Process.comm_world->rank,
MPIR_ThreadInfo.isThreaded);
#else
MPL_dbg_init(argc, argv, has_args, has_env,
MPIR_Process.comm_parent != NULL, MPIR_Process.comm_world->rank, 0);
#endif
MPIR_DBG_INIT = MPL_dbg_class_alloc("INIT", "init");
MPIR_DBG_PT2PT = MPL_dbg_class_alloc("PT2PT", "pt2pt");
MPIR_DBG_THREAD = MPL_dbg_class_alloc("THREAD", "thread");
MPIR_DBG_DATATYPE = MPL_dbg_class_alloc("DATATYPE", "datatype");
MPIR_DBG_HANDLE = MPL_dbg_class_alloc("HANDLE", "handle");
MPIR_DBG_COMM = MPL_dbg_class_alloc("COMM", "comm");
MPIR_DBG_BSEND = MPL_dbg_class_alloc("BSEND", "bsend");
MPIR_DBG_ERRHAND = MPL_dbg_class_alloc("ERRHAND", "errhand");
MPIR_DBG_OTHER = MPL_dbg_class_alloc("OTHER", "other");
MPIR_DBG_REQUEST = MPL_dbg_class_alloc("REQUEST", "request");
MPIR_DBG_COLL = MPL_dbg_class_alloc("COLL", "coll");
MPIR_DBG_ASSERT = MPL_dbg_class_alloc("ASSERT", "assert");
MPIR_DBG_STRING = MPL_dbg_class_alloc("STRING", "string");
#endif
/* Initialize the C versions of the Fortran link-time constants.
*
* We now initialize the Fortran symbols from within the Fortran
* interface in the routine that first needs the symbols.
* This fixes a problem with symbols added by a Fortran compiler that
* are not part of the C runtime environment (the Portland group
* compilers would do this)
*/
#if defined(HAVE_FORTRAN_BINDING) && defined(HAVE_MPI_F_INIT_WORKS_WITH_C)
mpirinitf_();
#endif
/* FIXME: Does this need to come before the call to MPID_InitComplete?
* For some debugger support, MPII_Wait_for_debugger may want to use
* MPI communication routines to collect information for the debugger */
#ifdef HAVE_DEBUGGER_SUPPORT
MPII_Wait_for_debugger();
#endif
/* Let the device know that the rest of the init process is completed */
if (mpi_errno == MPI_SUCCESS)
mpi_errno = MPID_InitCompleted();
MPID_THREAD_CS_EXIT(GLOBAL, MPIR_THREAD_GLOBAL_ALLFUNC_MUTEX);
/* Make fields of MPIR_Process global visible and set mpich_state
* atomically so that MPI_Initialized() etc. are thread safe */
OPA_write_barrier();
OPA_store_int(&MPIR_Process.mpich_state, MPICH_MPI_STATE__POST_INIT);
return mpi_errno;
fn_fail:
/* --BEGIN ERROR HANDLING-- */
/* signal to error handling routines that core services are unavailable */
OPA_store_int(&MPIR_Process.mpich_state, MPICH_MPI_STATE__PRE_INIT);
if (exit_init_cs_on_failure) {
MPID_THREAD_CS_EXIT(GLOBAL, MPIR_THREAD_GLOBAL_ALLFUNC_MUTEX);
}
#if defined(MPICH_IS_THREADED)
if (cs_initialized) {
MPIR_Thread_CS_Finalize();
}
#endif
return mpi_errno;
/* --END ERROR HANDLING-- */
}
static
void getNumCPUs(void) {
//
// accessible cores
//
//
// Hwloc can't tell us the number of accessible cores directly, so
// get that by counting the parent cores of the accessible PUs.
//
//
// We could seemingly use hwloc_topology_get_allowed_cpuset() to get
// the set of accessible PUs here. But that seems not to reflect the
// schedaffinity settings, so use hwloc_get_proc_cpubind() instead.
//
hwloc_cpuset_t logAccSet;
CHK_ERR_ERRNO((logAccSet = hwloc_bitmap_alloc()) != NULL);
if (hwloc_get_proc_cpubind(topology, getpid(), logAccSet, 0) != 0) {
#ifdef __APPLE__
const int errRecoverable = (errno == ENOSYS); // no cpubind on macOS
#else
const int errRecoverable = 0;
#endif
if (errRecoverable) {
hwloc_bitmap_fill(logAccSet);
} else {
REPORT_ERR_ERRNO(hwloc_get_proc_cpubind(topology, getpid(), logAccSet, 0)
== 0);
}
}
hwloc_bitmap_and(logAccSet, logAccSet,
hwloc_topology_get_online_cpuset(topology));
hwloc_cpuset_t physAccSet;
CHK_ERR_ERRNO((physAccSet = hwloc_bitmap_alloc()) != NULL);
#define NEXT_PU(pu) \
hwloc_get_next_obj_inside_cpuset_by_type(topology, logAccSet, \
HWLOC_OBJ_PU, pu)
for (hwloc_obj_t pu = NEXT_PU(NULL); pu != NULL; pu = NEXT_PU(pu)) {
hwloc_obj_t core;
CHK_ERR_ERRNO((core = hwloc_get_ancestor_obj_by_type(topology,
HWLOC_OBJ_CORE,
pu))
!= NULL);
hwloc_bitmap_set(physAccSet, core->logical_index);
}
#undef NEXT_PU
numCPUsPhysAcc = hwloc_bitmap_weight(physAccSet);
hwloc_bitmap_free(physAccSet);
CHK_ERR(numCPUsPhysAcc > 0);
//
// all cores
//
numCPUsPhysAll = hwloc_get_nbobjs_by_type(topology, HWLOC_OBJ_CORE);
CHK_ERR(numCPUsPhysAll > 0);
//
// accessible PUs
//
numCPUsLogAcc = hwloc_bitmap_weight(logAccSet);
CHK_ERR(numCPUsLogAcc > 0);
hwloc_bitmap_free(logAccSet);
//
// all PUs
//
numCPUsLogAll = hwloc_get_nbobjs_by_type(topology, HWLOC_OBJ_PU);
CHK_ERR(numCPUsLogAll > 0);
}
