void *
hwloc_alloc_heap(hwloc_topology_t topology __hwloc_attribute_unused, size_t len)
{
void *p = NULL;
#if defined(hwloc_getpagesize) && defined(HAVE_POSIX_MEMALIGN)
errno = posix_memalign(&p, hwloc_getpagesize(), len);
if (errno)
p = NULL;
#elif defined(hwloc_getpagesize) && defined(HAVE_MEMALIGN)
p = memalign(hwloc_getpagesize(), len);
#else
p = malloc(len);
#endif
return p;
}
static void
look_rset(int sdl, hwloc_obj_type_t type, struct hwloc_topology *topology, int level)
{
rsethandle_t rset, rad;
int i,maxcpus,j;
int nbnodes;
struct hwloc_obj *obj;
if ((topology->flags & HWLOC_TOPOLOGY_FLAG_WHOLE_SYSTEM))
rset = rs_alloc(RS_ALL);
else
rset = rs_alloc(RS_PARTITION);
rad = rs_alloc(RS_EMPTY);
nbnodes = rs_numrads(rset, sdl, 0);
if (nbnodes == -1) {
perror("rs_numrads");
return;
}
for (i = 0; i < nbnodes; i++) {
if (rs_getrad(rset, rad, sdl, i, 0)) {
fprintf(stderr,"rs_getrad(%d) failed: %s\n", i, strerror(errno));
continue;
}
if (!rs_getinfo(rad, R_NUMPROCS, 0))
continue;
/* It seems logical processors are numbered from 1 here, while the
* bindprocessor functions numbers them from 0... */
obj = hwloc_alloc_setup_object(type, i - (type == HWLOC_OBJ_PU));
obj->cpuset = hwloc_bitmap_alloc();
obj->os_level = sdl;
maxcpus = rs_getinfo(rad, R_MAXPROCS, 0);
for (j = 0; j < maxcpus; j++) {
if (rs_op(RS_TESTRESOURCE, rad, NULL, R_PROCS, j))
hwloc_bitmap_set(obj->cpuset, j);
}
switch(type) {
case HWLOC_OBJ_NODE:
obj->nodeset = hwloc_bitmap_alloc();
hwloc_bitmap_set(obj->nodeset, i);
obj->memory.local_memory = 0; /* TODO: odd, rs_getinfo(rad, R_MEMSIZE, 0) << 10 returns the total memory ... */
obj->memory.page_types_len = 2;
obj->memory.page_types = malloc(2*sizeof(*obj->memory.page_types));
memset(obj->memory.page_types, 0, 2*sizeof(*obj->memory.page_types));
obj->memory.page_types[0].size = hwloc_getpagesize();
#ifdef HAVE__SC_LARGE_PAGESIZE
obj->memory.page_types[1].size = sysconf(_SC_LARGE_PAGESIZE);
#endif
/* TODO: obj->memory.page_types[1].count = rs_getinfo(rset, R_LGPGFREE, 0) / hugepagesize */
break;
case HWLOC_OBJ_CACHE:
obj->attr->cache.size = _system_configuration.L2_cache_size;
obj->attr->cache.associativity = _system_configuration.L2_cache_asc;
obj->attr->cache.linesize = 0; /* TODO: ? */
obj->attr->cache.depth = 2;
obj->attr->cache.type = HWLOC_OBJ_CACHE_UNIFIED; /* FIXME? */
break;
case HWLOC_OBJ_GROUP:
obj->attr->group.depth = level;
break;
case HWLOC_OBJ_CORE:
{
hwloc_obj_t obj2, obj3;
obj2 = hwloc_alloc_setup_object(HWLOC_OBJ_CACHE, i);
obj2->cpuset = hwloc_bitmap_dup(obj->cpuset);
obj2->attr->cache.size = _system_configuration.dcache_size;
obj2->attr->cache.associativity = _system_configuration.dcache_asc;
obj2->attr->cache.linesize = _system_configuration.dcache_line;
obj2->attr->cache.depth = 1;
if (_system_configuration.cache_attrib & (1<<30)) {
/* Unified cache */
obj2->attr->cache.type = HWLOC_OBJ_CACHE_UNIFIED;
hwloc_debug("Adding an L1u cache for core %d\n", i);
hwloc_insert_object_by_cpuset(topology, obj2);
} else {
/* Separate Instruction and Data caches */
obj2->attr->cache.type = HWLOC_OBJ_CACHE_DATA;
hwloc_debug("Adding an L1d cache for core %d\n", i);
hwloc_insert_object_by_cpuset(topology, obj2);
obj3 = hwloc_alloc_setup_object(HWLOC_OBJ_CACHE, i);
obj3->cpuset = hwloc_bitmap_dup(obj->cpuset);
obj3->attr->cache.size = _system_configuration.icache_size;
obj3->attr->cache.associativity = _system_configuration.icache_asc;
obj3->attr->cache.linesize = _system_configuration.icache_line;
obj3->attr->cache.depth = 1;
obj3->attr->cache.type = HWLOC_OBJ_CACHE_INSTRUCTION;
hwloc_debug("Adding an L1i cache for core %d\n", i);
hwloc_insert_object_by_cpuset(topology, obj3);
}
break;
}
default:
break;
}
hwloc_debug_2args_bitmap("%s %d has cpuset %s\n",
hwloc_obj_type_string(type),
i, obj->cpuset);
hwloc_insert_object_by_cpuset(topology, obj);
}
rs_free(rset);
rs_free(rad);
}
static int test(hwloc_topology_t orig, const char *callname) {
unsigned long forced_addr;
unsigned long fileoffset;
size_t shmem_length;
int synthetic_with_distances = (hwloc_obj_get_info_by_name(hwloc_get_root_obj(orig), "SyntheticDescription") != NULL);
char tmpname[] = "/tmp/hwloc_test_shmem.XXXXXX";
char cmd[512];
struct stat st;
int fd, err;
int ret = EXIT_SKIP;
printf("opening temporary file\n");
fd = mkstemp(tmpname);
if (fd < 0) {
perror("mkstemp");
goto out;
}
printf("opened %s\n", tmpname);
printf("exporting XML\n");
err = hwloc_topology_export_xml(orig, tmpname, 0);
assert(!err);
err = stat(tmpname, &st);
assert(!err);
printf("exported %lu bytes\n", (unsigned long) st.st_size);
fileoffset = st.st_size+1; /* skip a couple bytes to make sure the XML is don" */
fileoffset = (fileoffset + hwloc_getpagesize() - 1) &~(hwloc_getpagesize() - 1);
printf("will mmap at file offset %lu\n", fileoffset);
err = hwloc_shmem_topology_get_length(orig, &shmem_length, 0);
assert(!err);
printf("need mmap length %lu\n", (unsigned long) shmem_length);
#if SIZEOF_VOID_P == 8
forced_addr = 0x300000000000UL;
#else
forced_addr = 0xb0000000UL;
#endif
printf("write to shmem at address %lx in file %s offset %lu\n", forced_addr, tmpname, fileoffset);
err = hwloc_shmem_topology_write(orig, fd, fileoffset, (void*)(uintptr_t)forced_addr, shmem_length, 0);
if (err == -1 && errno == EBUSY) {
fprintf(stderr, "Failed to shmem write, requested mapping is busy\n");
goto out_with_fd;
}
assert(!err);
printf("wrote length %lu\n", (unsigned long) shmem_length);
printf("adopting locally\n");
ret = adopt(fd, fileoffset, forced_addr, shmem_length, synthetic_with_distances);
assert(ret == EXIT_SUCCESS || ret == EXIT_SKIP);
printf("adopting in other child process\n");
snprintf(cmd, sizeof(cmd), "%s %s %lu %lu %lu %d", callname, tmpname, fileoffset, forced_addr, (unsigned long) shmem_length, synthetic_with_distances);
printf("running command %s\n", cmd);
err = system(cmd);
assert(WIFEXITED(err));
printf("child process returned %d\n", WEXITSTATUS(err));
assert(WEXITSTATUS(err) == EXIT_SUCCESS || WEXITSTATUS(err) == EXIT_SKIP);
/* we caught errors above.
* return SKIP if both returned SKIP. otherwise SUCCESS
*/
if (WEXITSTATUS(err) == EXIT_SKIP && ret == EXIT_SKIP)
ret = EXIT_SKIP;
else
ret = EXIT_SUCCESS;
out_with_fd:
close(fd);
unlink(tmpname);
out:
return ret;
}
static int
hwloc_look_osf(struct hwloc_backend *backend)
{
struct hwloc_topology *topology = backend->topology;
cpu_cursor_t cursor;
unsigned nbnodes;
radid_t radid, radid2;
radset_t radset, radset2;
cpuid_t cpuid;
cpuset_t cpuset;
struct hwloc_obj *obj;
unsigned distance;
if (topology->levels[0][0]->cpuset)
/* somebody discovered things */
return 0;
hwloc_alloc_obj_cpusets(topology->levels[0][0]);
nbnodes = rad_get_num();
cpusetcreate(&cpuset);
radsetcreate(&radset);
radsetcreate(&radset2);
{
hwloc_obj_t *nodes = calloc(nbnodes, sizeof(hwloc_obj_t));
unsigned *indexes = calloc(nbnodes, sizeof(unsigned));
float *distances = calloc(nbnodes*nbnodes, sizeof(float));
unsigned nfound;
numa_attr_t attr;
attr.nattr_type = R_RAD;
attr.nattr_descr.rd_radset = radset;
attr.nattr_flags = 0;
for (radid = 0; radid < (radid_t) nbnodes; radid++) {
rademptyset(radset);
radaddset(radset, radid);
cpuemptyset(cpuset);
if (rad_get_cpus(radid, cpuset)==-1) {
fprintf(stderr,"rad_get_cpus(%d) failed: %s\n",radid,strerror(errno));
continue;
}
indexes[radid] = radid;
nodes[radid] = obj = hwloc_alloc_setup_object(HWLOC_OBJ_NODE, radid);
obj->cpuset = hwloc_bitmap_alloc();
obj->memory.local_memory = rad_get_physmem(radid) * hwloc_getpagesize();
obj->memory.page_types_len = 2;
obj->memory.page_types = malloc(2*sizeof(*obj->memory.page_types));
memset(obj->memory.page_types, 0, 2*sizeof(*obj->memory.page_types));
obj->memory.page_types[0].size = hwloc_getpagesize();
#ifdef HAVE__SC_LARGE_PAGESIZE
obj->memory.page_types[1].size = sysconf(_SC_LARGE_PAGESIZE);
#endif
cursor = SET_CURSOR_INIT;
while((cpuid = cpu_foreach(cpuset, 0, &cursor)) != CPU_NONE)
hwloc_bitmap_set(obj->cpuset, cpuid);
hwloc_debug_1arg_bitmap("node %d has cpuset %s\n",
radid, obj->cpuset);
hwloc_insert_object_by_cpuset(topology, obj);
nfound = 0;
for (radid2 = 0; radid2 < (radid_t) nbnodes; radid2++)
distances[radid*nbnodes+radid2] = RAD_DIST_REMOTE;
for (distance = RAD_DIST_LOCAL; distance < RAD_DIST_REMOTE; distance++) {
attr.nattr_distance = distance;
/* get set of NUMA nodes at distance <= DISTANCE */
if (nloc(&attr, radset2)) {
fprintf(stderr,"nloc failed: %s\n", strerror(errno));
continue;
}
cursor = SET_CURSOR_INIT;
while ((radid2 = rad_foreach(radset2, 0, &cursor)) != RAD_NONE) {
if (distances[radid*nbnodes+radid2] == RAD_DIST_REMOTE) {
distances[radid*nbnodes+radid2] = (float) distance;
nfound++;
}
}
if (nfound == nbnodes)
/* Finished finding distances, no need to go up to RAD_DIST_REMOTE */
break;
}
}
hwloc_distances_set(topology, HWLOC_OBJ_NODE, nbnodes, indexes, nodes, distances, 0 /* OS cannot force */);
}
radsetdestroy(&radset2);
radsetdestroy(&radset);
cpusetdestroy(&cpuset);
/* add PU objects */
hwloc_setup_pu_level(topology, hwloc_fallback_nbprocessors(topology));
hwloc_obj_add_info(topology->levels[0][0], "Backend", "OSF");
if (topology->is_thissystem)
hwloc_add_uname_info(topology);
return 1;
}
static int
hwloc_look_darwin(struct hwloc_backend *backend)
{
struct hwloc_topology *topology = backend->topology;
int64_t _nprocs;
unsigned nprocs;
int64_t _npackages;
unsigned i, j, cpu;
struct hwloc_obj *obj;
size_t size;
int64_t l1dcachesize, l1icachesize;
int64_t cacheways[2];
int64_t l2cachesize;
int64_t cachelinesize;
int64_t memsize;
char cpumodel[64];
if (topology->levels[0][0]->cpuset)
/* somebody discovered things */
return -1;
hwloc_alloc_obj_cpusets(topology->levels[0][0]);
if (hwloc_get_sysctlbyname("hw.ncpu", &_nprocs) || _nprocs <= 0)
return -1;
nprocs = _nprocs;
topology->support.discovery->pu = 1;
hwloc_debug("%u procs\n", nprocs);
size = sizeof(cpumodel);
if (sysctlbyname("machdep.cpu.brand_string", cpumodel, &size, NULL, 0))
cpumodel[0] = '\0';
if (!hwloc_get_sysctlbyname("hw.packages", &_npackages) && _npackages > 0) {
unsigned npackages = _npackages;
int64_t _cores_per_package;
int64_t _logical_per_package;
unsigned logical_per_package;
hwloc_debug("%u packages\n", npackages);
if (!hwloc_get_sysctlbyname("machdep.cpu.logical_per_package", &_logical_per_package) && _logical_per_package > 0)
logical_per_package = _logical_per_package;
else
/* Assume the trivia. */
logical_per_package = nprocs / npackages;
hwloc_debug("%u threads per package\n", logical_per_package);
if (nprocs == npackages * logical_per_package
&& hwloc_filter_check_keep_object_type(topology, HWLOC_OBJ_PACKAGE))
for (i = 0; i < npackages; i++) {
obj = hwloc_alloc_setup_object(HWLOC_OBJ_PACKAGE, i);
obj->cpuset = hwloc_bitmap_alloc();
for (cpu = i*logical_per_package; cpu < (i+1)*logical_per_package; cpu++)
hwloc_bitmap_set(obj->cpuset, cpu);
hwloc_debug_1arg_bitmap("package %u has cpuset %s\n",
i, obj->cpuset);
if (cpumodel[0] != '\0')
hwloc_obj_add_info(obj, "CPUModel", cpumodel);
hwloc_insert_object_by_cpuset(topology, obj);
}
else
if (cpumodel[0] != '\0')
hwloc_obj_add_info(topology->levels[0][0], "CPUModel", cpumodel);
if (!hwloc_get_sysctlbyname("machdep.cpu.cores_per_package", &_cores_per_package) && _cores_per_package > 0
&& hwloc_filter_check_keep_object_type(topology, HWLOC_OBJ_CORE)) {
unsigned cores_per_package = _cores_per_package;
hwloc_debug("%u cores per package\n", cores_per_package);
if (!(logical_per_package % cores_per_package))
for (i = 0; i < npackages * cores_per_package; i++) {
obj = hwloc_alloc_setup_object(HWLOC_OBJ_CORE, i);
obj->cpuset = hwloc_bitmap_alloc();
for (cpu = i*(logical_per_package/cores_per_package);
cpu < (i+1)*(logical_per_package/cores_per_package);
cpu++)
hwloc_bitmap_set(obj->cpuset, cpu);
hwloc_debug_1arg_bitmap("core %u has cpuset %s\n",
i, obj->cpuset);
hwloc_insert_object_by_cpuset(topology, obj);
}
}
} else
if (cpumodel[0] != '\0')
hwloc_obj_add_info(topology->levels[0][0], "CPUModel", cpumodel);
if (hwloc_get_sysctlbyname("hw.l1dcachesize", &l1dcachesize))
l1dcachesize = 0;
if (hwloc_get_sysctlbyname("hw.l1icachesize", &l1icachesize))
l1icachesize = 0;
if (hwloc_get_sysctlbyname("hw.l2cachesize", &l2cachesize))
l2cachesize = 0;
if (hwloc_get_sysctlbyname("machdep.cpu.cache.L1_associativity", &cacheways[0]))
cacheways[0] = 0;
else if (cacheways[0] == 0xff)
cacheways[0] = -1;
if (hwloc_get_sysctlbyname("machdep.cpu.cache.L2_associativity", &cacheways[1]))
cacheways[1] = 0;
else if (cacheways[1] == 0xff)
cacheways[1] = -1;
if (hwloc_get_sysctlbyname("hw.cachelinesize", &cachelinesize))
cachelinesize = 0;
if (hwloc_get_sysctlbyname("hw.memsize", &memsize))
memsize = 0;
if (!sysctlbyname("hw.cacheconfig", NULL, &size, NULL, 0)) {
unsigned n = size / sizeof(uint32_t);
uint64_t *cacheconfig = NULL;
uint64_t *cachesize = NULL;
uint32_t *cacheconfig32 = NULL;
cacheconfig = malloc(sizeof(uint64_t) * n);
if (NULL == cacheconfig) {
goto out;
}
cachesize = malloc(sizeof(uint64_t) * n);
if (NULL == cachesize) {
goto out;
}
cacheconfig32 = malloc(sizeof(uint32_t) * n);
if (NULL == cacheconfig32) {
goto out;
}
if ((!sysctlbyname("hw.cacheconfig", cacheconfig, &size, NULL, 0))) {
/* Yeech. Darwin seemingly has changed from 32bit to 64bit integers for
* cacheconfig, with apparently no way for detection. Assume the machine
* won't have more than 4 billion cpus */
if (cacheconfig[0] > 0xFFFFFFFFUL) {
memcpy(cacheconfig32, cacheconfig, size);
for (i = 0 ; i < size / sizeof(uint32_t); i++)
cacheconfig[i] = cacheconfig32[i];
}
memset(cachesize, 0, sizeof(uint64_t) * n);
size = sizeof(uint64_t) * n;
if (sysctlbyname("hw.cachesize", cachesize, &size, NULL, 0)) {
if (n > 0)
cachesize[0] = memsize;
if (n > 1)
cachesize[1] = l1dcachesize;
if (n > 2)
cachesize[2] = l2cachesize;
}
hwloc_debug("%s", "caches");
for (i = 0; i < n && cacheconfig[i]; i++)
hwloc_debug(" %"PRIu64"(%"PRIu64"kB)", cacheconfig[i], cachesize[i] / 1024);
/* Now we know how many caches there are */
n = i;
hwloc_debug("\n%u cache levels\n", n - 1);
/* For each cache level (0 is memory) */
for (i = 0; i < n; i++) {
/* cacheconfig tells us how many cpus share it, let's iterate on each cache */
for (j = 0; j < (nprocs / cacheconfig[i]); j++) {
if (!i) {
obj = hwloc_alloc_setup_object(HWLOC_OBJ_NUMANODE, j);
obj->nodeset = hwloc_bitmap_alloc();
hwloc_bitmap_set(obj->nodeset, j);
} else {
obj = hwloc_alloc_setup_object(HWLOC_OBJ_L1CACHE+i-1, -1);
}
obj->cpuset = hwloc_bitmap_alloc();
for (cpu = j*cacheconfig[i];
cpu < ((j+1)*cacheconfig[i]);
cpu++)
hwloc_bitmap_set(obj->cpuset, cpu);
if (i == 1 && l1icachesize
&& hwloc_filter_check_keep_object_type(topology, HWLOC_OBJ_L1ICACHE)) {
/* FIXME assuming that L1i and L1d are shared the same way. Darwin
* does not yet provide a way to know. */
hwloc_obj_t l1i = hwloc_alloc_setup_object(HWLOC_OBJ_L1ICACHE, -1);
l1i->cpuset = hwloc_bitmap_dup(obj->cpuset);
hwloc_debug_1arg_bitmap("L1icache %u has cpuset %s\n",
j, l1i->cpuset);
l1i->attr->cache.depth = i;
l1i->attr->cache.size = l1icachesize;
l1i->attr->cache.linesize = cachelinesize;
l1i->attr->cache.associativity = 0;
l1i->attr->cache.type = HWLOC_OBJ_CACHE_INSTRUCTION;
hwloc_insert_object_by_cpuset(topology, l1i);
}
if (i) {
hwloc_debug_2args_bitmap("L%ucache %u has cpuset %s\n",
i, j, obj->cpuset);
obj->attr->cache.depth = i;
obj->attr->cache.size = cachesize[i];
obj->attr->cache.linesize = cachelinesize;
if (i <= sizeof(cacheways) / sizeof(cacheways[0]))
obj->attr->cache.associativity = cacheways[i-1];
else
obj->attr->cache.associativity = 0;
if (i == 1 && l1icachesize)
obj->attr->cache.type = HWLOC_OBJ_CACHE_DATA;
else
obj->attr->cache.type = HWLOC_OBJ_CACHE_UNIFIED;
} else {
hwloc_debug_1arg_bitmap("node %u has cpuset %s\n",
j, obj->cpuset);
obj->memory.local_memory = cachesize[i];
obj->memory.page_types_len = 2;
obj->memory.page_types = malloc(2*sizeof(*obj->memory.page_types));
memset(obj->memory.page_types, 0, 2*sizeof(*obj->memory.page_types));
obj->memory.page_types[0].size = hwloc_getpagesize();
#ifdef HAVE__SC_LARGE_PAGESIZE
obj->memory.page_types[1].size = sysconf(_SC_LARGE_PAGESIZE);
#endif
}
if (hwloc_filter_check_keep_object_type(topology, obj->type))
hwloc_insert_object_by_cpuset(topology, obj);
else
hwloc_free_unlinked_object(obj); /* FIXME: don't built at all, just build the cpuset in case l1i needs it */
}
}
}
out:
free(cacheconfig);
free(cachesize);
free(cacheconfig32);
}
/* add PU objects */
hwloc_setup_pu_level(topology, nprocs);
hwloc_obj_add_info(topology->levels[0][0], "Backend", "Darwin");
hwloc_add_uname_info(topology, NULL);
return 0;
}
{
size_t *tma_length = tma->data;
*tma_length += (length + HWLOC_SHMEM_MALLOC_ALIGN - 1) & ~(HWLOC_SHMEM_MALLOC_ALIGN - 1);
return malloc(length);
}
int
hwloc_shmem_topology_get_length(hwloc_topology_t topology,
size_t *lengthp,
unsigned long flags)
{
hwloc_topology_t new;
struct hwloc_tma tma;
size_t length = 0;
unsigned long pagesize = hwloc_getpagesize(); /* round-up to full page for mmap() */
int err;
if (flags) {
errno = EINVAL;
return -1;
}
tma.malloc = tma_get_length_malloc;
tma.dontfree = 0;
tma.data = &length;
err = hwloc__topology_dup(&new, topology, &tma);
if (err < 0)
return err;
hwloc_topology_destroy(new);
static void
browse(struct hwloc_topology *topology, lgrp_cookie_t cookie, lgrp_id_t lgrp, hwloc_obj_t *glob_lgrps, unsigned *curlgrp)
{
int n;
hwloc_obj_t obj;
lgrp_mem_size_t mem_size;
n = lgrp_cpus(cookie, lgrp, NULL, 0, LGRP_CONTENT_HIERARCHY);
if (n == -1)
return;
/* Is this lgrp a NUMA node? */
if ((mem_size = lgrp_mem_size(cookie, lgrp, LGRP_MEM_SZ_INSTALLED, LGRP_CONTENT_DIRECT)) > 0)
{
int i;
processorid_t *cpuids;
cpuids = malloc(sizeof(processorid_t) * n);
assert(cpuids != NULL);
obj = hwloc_alloc_setup_object(HWLOC_OBJ_NUMANODE, lgrp);
obj->nodeset = hwloc_bitmap_alloc();
hwloc_bitmap_set(obj->nodeset, lgrp);
obj->cpuset = hwloc_bitmap_alloc();
glob_lgrps[(*curlgrp)++] = obj;
lgrp_cpus(cookie, lgrp, cpuids, n, LGRP_CONTENT_HIERARCHY);
for (i = 0; i < n ; i++) {
hwloc_debug("node %ld's cpu %d is %d\n", lgrp, i, cpuids[i]);
hwloc_bitmap_set(obj->cpuset, cpuids[i]);
}
hwloc_debug_1arg_bitmap("node %ld has cpuset %s\n",
lgrp, obj->cpuset);
/* or LGRP_MEM_SZ_FREE */
hwloc_debug("node %ld has %lldkB\n", lgrp, mem_size/1024);
obj->memory.local_memory = mem_size;
obj->memory.page_types_len = 2;
obj->memory.page_types = malloc(2*sizeof(*obj->memory.page_types));
memset(obj->memory.page_types, 0, 2*sizeof(*obj->memory.page_types));
obj->memory.page_types[0].size = hwloc_getpagesize();
#if HAVE_DECL__SC_LARGE_PAGESIZE
obj->memory.page_types[1].size = sysconf(_SC_LARGE_PAGESIZE);
#endif
hwloc_insert_object_by_cpuset(topology, obj);
free(cpuids);
}
n = lgrp_children(cookie, lgrp, NULL, 0);
{
lgrp_id_t *lgrps;
int i;
lgrps = malloc(sizeof(lgrp_id_t) * n);
assert(lgrps != NULL);
lgrp_children(cookie, lgrp, lgrps, n);
hwloc_debug("lgrp %ld has %d children\n", lgrp, n);
for (i = 0; i < n ; i++)
{
browse(topology, cookie, lgrps[i], glob_lgrps, curlgrp);
}
hwloc_debug("lgrp %ld's children done\n", lgrp);
free(lgrps);
}
}
static void
look_rset(int sdl, hwloc_obj_type_t type, struct hwloc_topology *topology, int level)
{
rsethandle_t rset, rad;
int i,maxcpus,j;
int nbnodes;
struct hwloc_obj *obj;
if ((topology->flags & HWLOC_TOPOLOGY_FLAG_WHOLE_SYSTEM))
rset = rs_alloc(RS_ALL);
else
rset = rs_alloc(RS_PARTITION);
rad = rs_alloc(RS_EMPTY);
nbnodes = rs_numrads(rset, sdl, 0);
if (nbnodes == -1) {
perror("rs_numrads");
return;
}
for (i = 0; i < nbnodes; i++) {
hwloc_bitmap_t cpuset;
unsigned os_index = (unsigned) -1; /* no os_index except for PU and NUMANODE below */
if (rs_getrad(rset, rad, sdl, i, 0)) {
fprintf(stderr,"rs_getrad(%d) failed: %s\n", i, strerror(errno));
continue;
}
if (!rs_getinfo(rad, R_NUMPROCS, 0))
continue;
maxcpus = rs_getinfo(rad, R_MAXPROCS, 0);
cpuset = hwloc_bitmap_alloc();
for (j = 0; j < maxcpus; j++) {
if (rs_op(RS_TESTRESOURCE, rad, NULL, R_PROCS, j))
hwloc_bitmap_set(cpuset, j);
}
if (type == HWLOC_OBJ_PU) {
os_index = hwloc_bitmap_first(cpuset);
hwloc_debug("Found PU #%u inside node %d for sdl %d\n", os_index, i, sdl);
assert(hwloc_bitmap_weight(cpuset) == 1);
} else if (type == HWLOC_OBJ_NUMANODE) {
/* NUMA node os_index isn't used for binding, just use the rad number to get unique values.
* Note that we'll use that fact in hwloc_aix_prepare_membind(). */
os_index = i;
hwloc_debug("Using os_index #%u for NUMA node inside node %d for sdl %d\n", os_index, i, sdl);
}
obj = hwloc_alloc_setup_object(type, os_index);
obj->cpuset = cpuset;
obj->os_level = sdl;
switch(type) {
case HWLOC_OBJ_NUMANODE:
obj->nodeset = hwloc_bitmap_alloc();
hwloc_bitmap_set(obj->nodeset, i);
obj->memory.local_memory = 0; /* TODO: odd, rs_getinfo(rad, R_MEMSIZE, 0) << 10 returns the total memory ... */
obj->memory.page_types_len = 2;
obj->memory.page_types = malloc(2*sizeof(*obj->memory.page_types));
memset(obj->memory.page_types, 0, 2*sizeof(*obj->memory.page_types));
obj->memory.page_types[0].size = hwloc_getpagesize();
#ifdef HAVE__SC_LARGE_PAGESIZE
obj->memory.page_types[1].size = sysconf(_SC_LARGE_PAGESIZE);
#endif
/* TODO: obj->memory.page_types[1].count = rs_getinfo(rset, R_LGPGFREE, 0) / hugepagesize */
break;
case HWLOC_OBJ_CACHE:
obj->attr->cache.size = _system_configuration.L2_cache_size;
obj->attr->cache.associativity = _system_configuration.L2_cache_asc;
obj->attr->cache.linesize = 0; /* unknown by default */
if (__power_pc())
if (__power_4() || __power_5() || __power_6() || __power_7())
obj->attr->cache.linesize = 128;
obj->attr->cache.depth = 2;
obj->attr->cache.type = HWLOC_OBJ_CACHE_UNIFIED; /* OK for power[4567], unknown for others */
break;
case HWLOC_OBJ_GROUP:
obj->attr->group.depth = level;
break;
case HWLOC_OBJ_CORE:
{
hwloc_obj_t obj2, obj3;
obj2 = hwloc_alloc_setup_object(HWLOC_OBJ_CACHE, i);
obj2->cpuset = hwloc_bitmap_dup(obj->cpuset);
obj2->attr->cache.size = _system_configuration.dcache_size;
obj2->attr->cache.associativity = _system_configuration.dcache_asc;
obj2->attr->cache.linesize = _system_configuration.dcache_line;
obj2->attr->cache.depth = 1;
if (_system_configuration.cache_attrib & (1<<30)) {
/* Unified cache */
obj2->attr->cache.type = HWLOC_OBJ_CACHE_UNIFIED;
hwloc_debug("Adding an L1u cache for core %d\n", i);
hwloc_insert_object_by_cpuset(topology, obj2);
} else {
/* Separate Instruction and Data caches */
obj2->attr->cache.type = HWLOC_OBJ_CACHE_DATA;
hwloc_debug("Adding an L1d cache for core %d\n", i);
hwloc_insert_object_by_cpuset(topology, obj2);
obj3 = hwloc_alloc_setup_object(HWLOC_OBJ_CACHE, i);
obj3->cpuset = hwloc_bitmap_dup(obj->cpuset);
obj3->attr->cache.size = _system_configuration.icache_size;
obj3->attr->cache.associativity = _system_configuration.icache_asc;
obj3->attr->cache.linesize = _system_configuration.icache_line;
obj3->attr->cache.depth = 1;
obj3->attr->cache.type = HWLOC_OBJ_CACHE_INSTRUCTION;
hwloc_debug("Adding an L1i cache for core %d\n", i);
hwloc_insert_object_by_cpuset(topology, obj3);
}
break;
}
default:
break;
}
hwloc_debug_2args_bitmap("%s %d has cpuset %s\n",
hwloc_obj_type_string(type),
i, obj->cpuset);
hwloc_insert_object_by_cpuset(topology, obj);
}
rs_free(rset);
rs_free(rad);
}
static int
hwloc_look_darwin(struct hwloc_backend *backend)
{
struct hwloc_topology *topology = backend->topology;
int64_t _nprocs;
unsigned nprocs;
int64_t _npackages;
unsigned i, j, cpu;
struct hwloc_obj *obj;
size_t size;
int64_t l1dcachesize, l1icachesize;
int64_t cacheways[2];
int64_t l2cachesize;
int64_t l3cachesize;
int64_t cachelinesize;
int64_t memsize;
int64_t _tmp;
char cpumodel[64];
char cpuvendor[64];
char cpufamilynumber[20], cpumodelnumber[20], cpustepping[20];
int gotnuma = 0;
int gotnumamemory = 0;
if (topology->levels[0][0]->cpuset)
/* somebody discovered things */
return -1;
hwloc_alloc_root_sets(topology->levels[0][0]);
/* Don't use hwloc_fallback_nbprocessors() because it would return online cpus only,
* while we need all cpus when computing logical_per_package, etc below.
* We don't know which CPUs are offline, but Darwin doesn't support binding anyway.
*
* TODO: try hw.logicalcpu_max
*/
if (hwloc_get_sysctlbyname("hw.logicalcpu", &_nprocs) || _nprocs <= 0)
/* fallback to deprecated way */
if (hwloc_get_sysctlbyname("hw.ncpu", &_nprocs) || _nprocs <= 0)
return -1;
nprocs = _nprocs;
topology->support.discovery->pu = 1;
hwloc_debug("%u procs\n", nprocs);
size = sizeof(cpuvendor);
if (sysctlbyname("machdep.cpu.vendor", cpuvendor, &size, NULL, 0))
cpuvendor[0] = '\0';
size = sizeof(cpumodel);
if (sysctlbyname("machdep.cpu.brand_string", cpumodel, &size, NULL, 0))
cpumodel[0] = '\0';
if (hwloc_get_sysctlbyname("machdep.cpu.family", &_tmp))
cpufamilynumber[0] = '\0';
else
snprintf(cpufamilynumber, sizeof(cpufamilynumber), "%lld", (long long) _tmp);
if (hwloc_get_sysctlbyname("machdep.cpu.model", &_tmp))
cpumodelnumber[0] = '\0';
else
snprintf(cpumodelnumber, sizeof(cpumodelnumber), "%lld", (long long) _tmp);
/* .extfamily and .extmodel are already added to .family and .model */
if (hwloc_get_sysctlbyname("machdep.cpu.stepping", &_tmp))
cpustepping[0] = '\0';
else
snprintf(cpustepping, sizeof(cpustepping), "%lld", (long long) _tmp);
if (!hwloc_get_sysctlbyname("hw.packages", &_npackages) && _npackages > 0) {
unsigned npackages = _npackages;
int64_t _cores_per_package;
unsigned cores_per_package;
int64_t _logical_per_package;
unsigned logical_per_package;
hwloc_debug("%u packages\n", npackages);
if (!hwloc_get_sysctlbyname("machdep.cpu.thread_count", &_logical_per_package) && _logical_per_package > 0)
/* official/modern way */
logical_per_package = _logical_per_package;
else if (!hwloc_get_sysctlbyname("machdep.cpu.logical_per_package", &_logical_per_package) && _logical_per_package > 0)
/* old way, gives the max supported by this "kind" of processor,
* can be larger than the actual number for this model.
*/
logical_per_package = _logical_per_package;
else
/* Assume the trivia. */
logical_per_package = nprocs / npackages;
hwloc_debug("%u threads per package\n", logical_per_package);
if (nprocs == npackages * logical_per_package
&& hwloc_filter_check_keep_object_type(topology, HWLOC_OBJ_PACKAGE))
for (i = 0; i < npackages; i++) {
obj = hwloc_alloc_setup_object(topology, HWLOC_OBJ_PACKAGE, i);
obj->cpuset = hwloc_bitmap_alloc();
for (cpu = i*logical_per_package; cpu < (i+1)*logical_per_package; cpu++)
hwloc_bitmap_set(obj->cpuset, cpu);
hwloc_debug_1arg_bitmap("package %u has cpuset %s\n",
i, obj->cpuset);
if (cpuvendor[0] != '\0')
hwloc_obj_add_info(obj, "CPUVendor", cpuvendor);
if (cpumodel[0] != '\0')
hwloc_obj_add_info(obj, "CPUModel", cpumodel);
if (cpufamilynumber[0] != '\0')
hwloc_obj_add_info(obj, "CPUFamilyNumber", cpufamilynumber);
if (cpumodelnumber[0] != '\0')
hwloc_obj_add_info(obj, "CPUModelNumber", cpumodelnumber);
if (cpustepping[0] != '\0')
hwloc_obj_add_info(obj, "CPUStepping", cpustepping);
hwloc_insert_object_by_cpuset(topology, obj);
}
else {
if (cpuvendor[0] != '\0')
hwloc_obj_add_info(topology->levels[0][0], "CPUVendor", cpuvendor);
if (cpumodel[0] != '\0')
hwloc_obj_add_info(topology->levels[0][0], "CPUModel", cpumodel);
if (cpufamilynumber[0] != '\0')
hwloc_obj_add_info(topology->levels[0][0], "CPUFamilyNumber", cpufamilynumber);
if (cpumodelnumber[0] != '\0')
hwloc_obj_add_info(topology->levels[0][0], "CPUModelNumber", cpumodelnumber);
if (cpustepping[0] != '\0')
hwloc_obj_add_info(topology->levels[0][0], "CPUStepping", cpustepping);
}
if (!hwloc_get_sysctlbyname("machdep.cpu.core_count", &_cores_per_package) && _cores_per_package > 0)
/* official/modern way */
cores_per_package = _cores_per_package;
else if (!hwloc_get_sysctlbyname("machdep.cpu.cores_per_package", &_cores_per_package) && _cores_per_package > 0)
/* old way, gives the max supported by this "kind" of processor,
* can be larger than the actual number for this model.
*/
cores_per_package = _cores_per_package;
else
/* no idea */
cores_per_package = 0;
if (cores_per_package > 0
&& hwloc_filter_check_keep_object_type(topology, HWLOC_OBJ_CORE)) {
hwloc_debug("%u cores per package\n", cores_per_package);
if (!(logical_per_package % cores_per_package))
for (i = 0; i < npackages * cores_per_package; i++) {
obj = hwloc_alloc_setup_object(topology, HWLOC_OBJ_CORE, i);
obj->cpuset = hwloc_bitmap_alloc();
for (cpu = i*(logical_per_package/cores_per_package);
cpu < (i+1)*(logical_per_package/cores_per_package);
cpu++)
hwloc_bitmap_set(obj->cpuset, cpu);
hwloc_debug_1arg_bitmap("core %u has cpuset %s\n",
i, obj->cpuset);
hwloc_insert_object_by_cpuset(topology, obj);
}
}
} else {
if (cpuvendor[0] != '\0')
hwloc_obj_add_info(topology->levels[0][0], "CPUVendor", cpuvendor);
if (cpumodel[0] != '\0')
hwloc_obj_add_info(topology->levels[0][0], "CPUModel", cpumodel);
if (cpufamilynumber[0] != '\0')
hwloc_obj_add_info(topology->levels[0][0], "CPUFamilyNumber", cpufamilynumber);
if (cpumodelnumber[0] != '\0')
hwloc_obj_add_info(topology->levels[0][0], "CPUModelNumber", cpumodelnumber);
if (cpustepping[0] != '\0')
hwloc_obj_add_info(topology->levels[0][0], "CPUStepping", cpustepping);
}
if (hwloc_get_sysctlbyname("hw.l1dcachesize", &l1dcachesize))
l1dcachesize = 0;
if (hwloc_get_sysctlbyname("hw.l1icachesize", &l1icachesize))
l1icachesize = 0;
if (hwloc_get_sysctlbyname("hw.l2cachesize", &l2cachesize))
l2cachesize = 0;
if (hwloc_get_sysctlbyname("hw.l3cachesize", &l3cachesize))
l3cachesize = 0;
if (hwloc_get_sysctlbyname("machdep.cpu.cache.L1_associativity", &cacheways[0]))
cacheways[0] = 0;
else if (cacheways[0] == 0xff)
cacheways[0] = -1;
if (hwloc_get_sysctlbyname("machdep.cpu.cache.L2_associativity", &cacheways[1]))
cacheways[1] = 0;
else if (cacheways[1] == 0xff)
cacheways[1] = -1;
if (hwloc_get_sysctlbyname("hw.cachelinesize", &cachelinesize))
cachelinesize = 0;
if (hwloc_get_sysctlbyname("hw.memsize", &memsize))
memsize = 0;
if (!sysctlbyname("hw.cacheconfig", NULL, &size, NULL, 0)) {
unsigned n = size / sizeof(uint32_t);
uint64_t cacheconfig[n];
uint64_t cachesize[n];
uint32_t cacheconfig32[n];
if ((!sysctlbyname("hw.cacheconfig", cacheconfig, &size, NULL, 0))) {
/* Yeech. Darwin seemingly has changed from 32bit to 64bit integers for
* cacheconfig, with apparently no way for detection. Assume the machine
* won't have more than 4 billion cpus */
if (cacheconfig[0] > 0xFFFFFFFFUL) {
memcpy(cacheconfig32, cacheconfig, size);
for (i = 0 ; i < size / sizeof(uint32_t); i++)
cacheconfig[i] = cacheconfig32[i];
}
memset(cachesize, 0, sizeof(uint64_t) * n);
size = sizeof(uint64_t) * n;
if (sysctlbyname("hw.cachesize", cachesize, &size, NULL, 0)) {
if (n > 0)
cachesize[0] = memsize;
if (n > 1)
cachesize[1] = l1dcachesize;
if (n > 2)
cachesize[2] = l2cachesize;
if (n > 3)
cachesize[3] = l3cachesize;
}
hwloc_debug("%s", "caches");
for (i = 0; i < n && cacheconfig[i]; i++)
hwloc_debug(" %"PRIu64"(%"PRIu64"kB)", cacheconfig[i], cachesize[i] / 1024);
/* Now we know how many caches there are */
n = i;
hwloc_debug("\n%u cache levels\n", n - 1);
/* For each cache level (0 is memory) */
for (i = 0; i < n; i++) {
/* cacheconfig tells us how many cpus share it, let's iterate on each cache */
for (j = 0; j < (nprocs / cacheconfig[i]); j++) {
if (!i) {
obj = hwloc_alloc_setup_object(topology, HWLOC_OBJ_NUMANODE, j);
obj->nodeset = hwloc_bitmap_alloc();
hwloc_bitmap_set(obj->nodeset, j);
gotnuma++;
} else {
obj = hwloc_alloc_setup_object(topology, HWLOC_OBJ_L1CACHE+i-1, HWLOC_UNKNOWN_INDEX);
}
obj->cpuset = hwloc_bitmap_alloc();
for (cpu = j*cacheconfig[i];
cpu < ((j+1)*cacheconfig[i]);
cpu++)
hwloc_bitmap_set(obj->cpuset, cpu);
if (i == 1 && l1icachesize
&& hwloc_filter_check_keep_object_type(topology, HWLOC_OBJ_L1ICACHE)) {
/* FIXME assuming that L1i and L1d are shared the same way. Darwin
* does not yet provide a way to know. */
hwloc_obj_t l1i = hwloc_alloc_setup_object(topology, HWLOC_OBJ_L1ICACHE, HWLOC_UNKNOWN_INDEX);
l1i->cpuset = hwloc_bitmap_dup(obj->cpuset);
hwloc_debug_1arg_bitmap("L1icache %u has cpuset %s\n",
j, l1i->cpuset);
l1i->attr->cache.depth = i;
l1i->attr->cache.size = l1icachesize;
l1i->attr->cache.linesize = cachelinesize;
l1i->attr->cache.associativity = 0;
l1i->attr->cache.type = HWLOC_OBJ_CACHE_INSTRUCTION;
hwloc_insert_object_by_cpuset(topology, l1i);
}
if (i) {
hwloc_debug_2args_bitmap("L%ucache %u has cpuset %s\n",
i, j, obj->cpuset);
obj->attr->cache.depth = i;
obj->attr->cache.size = cachesize[i];
obj->attr->cache.linesize = cachelinesize;
if (i <= sizeof(cacheways) / sizeof(cacheways[0]))
obj->attr->cache.associativity = cacheways[i-1];
else
obj->attr->cache.associativity = 0;
if (i == 1 && l1icachesize)
obj->attr->cache.type = HWLOC_OBJ_CACHE_DATA;
else
obj->attr->cache.type = HWLOC_OBJ_CACHE_UNIFIED;
} else {
hwloc_debug_1arg_bitmap("node %u has cpuset %s\n",
j, obj->cpuset);
if (cachesize[i]) {
obj->attr->numanode.local_memory = cachesize[i];
gotnumamemory++;
}
obj->attr->numanode.page_types_len = 2;
obj->attr->numanode.page_types = malloc(2*sizeof(*obj->attr->numanode.page_types));
memset(obj->attr->numanode.page_types, 0, 2*sizeof(*obj->attr->numanode.page_types));
obj->attr->numanode.page_types[0].size = hwloc_getpagesize();
#if HAVE_DECL__SC_LARGE_PAGESIZE
obj->attr->numanode.page_types[1].size = sysconf(_SC_LARGE_PAGESIZE);
#endif
}
if (hwloc_filter_check_keep_object_type(topology, obj->type))
hwloc_insert_object_by_cpuset(topology, obj);
else
hwloc_free_unlinked_object(obj); /* FIXME: don't built at all, just build the cpuset in case l1i needs it */
}
}
}
}
if (gotnuma)
topology->support.discovery->numa = 1;
if (gotnumamemory)
topology->support.discovery->numa = 1;
/* add PU objects */
hwloc_setup_pu_level(topology, nprocs);
hwloc_obj_add_info(topology->levels[0][0], "Backend", "Darwin");
hwloc_add_uname_info(topology, NULL);
return 0;
}
