void* mmap_1g(void* addr /* = nullptr */, int node /* = -1 */) {
#ifdef __linux__
if (s_num1GPages >= kMaxNum1GPages) return nullptr;
if (get_huge1g_info(node).free_hugepages <= 0) return nullptr;
if (node >= 0 && !numa_node_allowed(node)) return nullptr;
#ifdef HAVE_NUMA
bitmask* memMask = nullptr;
bitmask* interleaveMask = nullptr;
if (node >= 0 && numa_num_nodes > 1) {
memMask = numa_get_membind();
interleaveMask = numa_get_interleave_mask();
bitmask* mask = numa_allocate_nodemask();
numa_bitmask_setbit(mask, node);
numa_set_membind(mask);
numa_bitmask_free(mask);
}
#endif
void* ret = mmap_1g_impl(addr);
if (ret != nullptr) {
s_1GPages[s_num1GPages++] = ret;
}
#ifdef HAVE_NUMA
if (memMask) {
assert(interleaveMask);
numa_set_membind(memMask);
numa_set_interleave_mask(interleaveMask);
numa_bitmask_free(memMask);
numa_bitmask_free(interleaveMask);
}
#endif
return ret;
#else
return nullptr;
#endif
}
static int memkind_regular_finalize(memkind_t kind)
{
if(regular_nodes_mask)
numa_bitmask_free(regular_nodes_mask);
return memkind_arena_finalize(kind);
}
/**
* mem_alloc_pages_onnode - allocates pages on a given numa node
* @nr: the number of pages
* @size: the page size (4KB, 2MB, or 1GB)
* @numa_node: the numa node to allocate the pages from
* @numa_policy: how strictly to take @numa_node
*
* Returns a pointer (virtual address) to a page or NULL if fail.
*/
void *mem_alloc_pages_onnode(int nr, int size, int node, int numa_policy)
{
void *vaddr;
struct bitmask *mask = numa_allocate_nodemask();
numa_bitmask_setbit(mask, node);
vaddr = mem_alloc_pages(nr, size, mask, numa_policy);
numa_bitmask_free(mask);
return vaddr;
}
void *__mem_alloc_pages_onnode(void *base, int nr, int size, int node)
{
void *vaddr;
struct bitmask *mask = numa_allocate_nodemask();
numa_bitmask_setbit(mask, node);
vaddr = __mem_alloc_pages(base, nr, size, mask, MPOL_BIND);
numa_bitmask_free(mask);
return vaddr;
}
///This function tries to fill bandwidth array based on knowledge about known CPU models
static int fill_bandwidth_values_heuristically(int* bandwidth, int bandwidth_len)
{
int ret = MEMKIND_ERROR_UNAVAILABLE; // Default error returned if heuristic aproach fails
int i, nodes_num, memory_only_nodes_num = 0;
struct bitmask *memory_only_nodes, *node_cpus;
if (is_cpu_xeon_phi_x200() == 0) {
log_info("Known CPU model detected: Intel(R) Xeon Phi(TM) x200.");
nodes_num = numa_num_configured_nodes();
// Check if number of numa-nodes meets expectations for
// supported configurations of Intel Xeon Phi x200
if( nodes_num != 2 && nodes_num != 4 && nodes_num!= 8 ) {
return ret;
}
memory_only_nodes = numa_allocate_nodemask();
node_cpus = numa_allocate_cpumask();
for(i=0; i<nodes_num; i++) {
numa_node_to_cpus(i, node_cpus);
if(numa_bitmask_weight(node_cpus) == 0) {
memory_only_nodes_num++;
numa_bitmask_setbit(memory_only_nodes, i);
}
}
// Check if number of memory-only nodes is equal number of memory+cpu nodes
// If it passes change ret to 0 (success) and fill bw table
if ( memory_only_nodes_num == (nodes_num - memory_only_nodes_num) ) {
ret = 0;
assign_arbitrary_bandwidth_values(bandwidth, bandwidth_len, memory_only_nodes);
}
numa_bitmask_free(memory_only_nodes);
numa_bitmask_free(node_cpus);
}
return ret;
}
static void regular_nodes_init(void)
{
int i, node = 0, nodes_num = numa_num_configured_nodes();
struct bitmask *node_cpus = numa_allocate_cpumask();
regular_nodes_mask = numa_allocate_nodemask();
for (i = 0; i < nodes_num; i++) {
numa_node_to_cpus(node, node_cpus);
if (numa_bitmask_weight(node_cpus))
numa_bitmask_setbit(regular_nodes_mask, i);
}
numa_bitmask_free(node_cpus);
}
void* mmap_2m(void* addr, int prot, int node /* = -1 */,
bool map_shared /* = false */, bool map_fixed /* = false */) {
#ifdef __linux__
if (get_huge2m_info(node).free_hugepages <= 0) return nullptr;
#ifdef HAVE_NUMA
bitmask* memMask = nullptr;
bitmask* interleaveMask = nullptr;
if (node >= 0 && numa_num_nodes > 1) {
assert(numa_node_set != 0);
if ((numa_node_set & (1u << node)) == 0) {
// Numa policy forbids allocation on the node.
return nullptr;
}
memMask = numa_get_membind();
interleaveMask = numa_get_interleave_mask();
bitmask* mask = numa_allocate_nodemask();
numa_bitmask_setbit(mask, node);
numa_set_membind(mask);
numa_bitmask_free(mask);
}
#endif
void* ret = mmap_2m_impl(addr, prot, map_shared, map_fixed);
s_num2MPages += !!ret;
#ifdef HAVE_NUMA
if (memMask) {
numa_set_membind(memMask);
numa_set_interleave_mask(interleaveMask);
numa_bitmask_free(memMask);
numa_bitmask_free(interleaveMask);
}
#endif
return ret;
#else // not linux
return nullptr;
#endif
}
static int fill_bandwidth_values_from_enviroment(int* bandwidth, int bandwidth_len, char *hbw_nodes_env)
{
struct bitmask *hbw_nodes_bm = numa_parse_nodestring(hbw_nodes_env);
if (!hbw_nodes_bm) {
log_err("Wrong MEMKIND_HBW_NODES environment value.");
return MEMKIND_ERROR_ENVIRON;
}
else {
assign_arbitrary_bandwidth_values(bandwidth, bandwidth_len, hbw_nodes_bm);
numa_bitmask_free(hbw_nodes_bm);
}
return 0;
}
int bind_cpu(int cpu)
{
cpu_set_t *cmask;
struct bitmask *bmask;
size_t ncpu, setsize;
int ret;
ncpu = get_num_cpus();
if (cpu < 0 || cpu >= (int)ncpu) {
errno = -EINVAL;
return -1;
}
cmask = CPU_ALLOC(ncpu);
if (cmask == NULL)
return -1;
setsize = CPU_ALLOC_SIZE(ncpu);
CPU_ZERO_S(setsize, cmask);
CPU_SET_S(cpu, setsize, cmask);
ret = sched_setaffinity(0, ncpu, cmask);
CPU_FREE(cmask);
/* skip NUMA stuff for UMA systems */
if (numa_max_node() == 0)
return ret;
bmask = numa_bitmask_alloc(16);
assert(bmask);
numa_bitmask_setbit(bmask, cpu % 2);
numa_set_membind(bmask);
numa_bitmask_free(bmask);
return ret;
}
int fire_worker_init(fire_worker_context_t *context)
{
buffer[context->queue_id] = (char *)malloc(buf_size);
/* init worker struct */
fire_worker_t *cc = &(workers[context->queue_id]);
cc->total_packets = 0;
cc->total_bytes = 0;
/* nids init */
nids_init(context->core_id);
#if !defined(AFFINITY_NO)
/* set schedule affinity */
unsigned long mask = 1 << context->core_id;
if (sched_setaffinity(0, sizeof(unsigned long), (cpu_set_t *)&mask) < 0) {
assert(0);
}
/* set schedule policy */
struct sched_param param;
param.sched_priority = 99;
pthread_setschedparam(pthread_self(), SCHED_FIFO, ¶m);
#endif
if (numa_max_node() == 0)
return 0;
struct bitmask *bmask;
bmask = numa_bitmask_alloc(16);
assert(bmask);
numa_bitmask_setbit(bmask, context->core_id % 2);
numa_set_membind(bmask);
numa_bitmask_free(bmask);
return 0;
}
void myhbwmalloc_init(void)
{
/* set to NULL before trying to initialize. if we return before
* successful creation of the mspace, then it will still be NULL,
* and we can use that in subsequent library calls to determine
* that the library failed to initialize. */
myhbwmalloc_mspace = NULL;
/* verbose printout? */
myhbwmalloc_verbose = 0;
{
char * env_char = getenv("HBWMALLOC_VERBOSE");
if (env_char != NULL) {
myhbwmalloc_verbose = 1;
printf("hbwmalloc: HBWMALLOC_VERBOSE set\n");
}
}
/* fail hard or soft? */
myhbwmalloc_hardfail = 1;
{
char * env_char = getenv("HBWMALLOC_SOFTFAIL");
if (env_char != NULL) {
myhbwmalloc_hardfail = 0;
printf("hbwmalloc: HBWMALLOC_SOFTFAIL set\n");
}
}
/* set the atexit handler that will destroy the mspace and free the numa allocation */
atexit(myhbwmalloc_final);
/* detect and configure use of NUMA memory nodes */
{
int max_possible_node = numa_max_possible_node();
int num_possible_nodes = numa_num_possible_nodes();
int max_numa_nodes = numa_max_node();
int num_configured_nodes = numa_num_configured_nodes();
int num_configured_cpus = numa_num_configured_cpus();
if (myhbwmalloc_verbose) {
printf("hbwmalloc: numa_max_possible_node() = %d\n", max_possible_node);
printf("hbwmalloc: numa_num_possible_nodes() = %d\n", num_possible_nodes);
printf("hbwmalloc: numa_max_node() = %d\n", max_numa_nodes);
printf("hbwmalloc: numa_num_configured_nodes() = %d\n", num_configured_nodes);
printf("hbwmalloc: numa_num_configured_cpus() = %d\n", num_configured_cpus);
}
/* FIXME this is a hack. assumes HBW is only numa node 1. */
if (num_configured_nodes <= 2) {
myhbwmalloc_numa_node = num_configured_nodes-1;
} else {
fprintf(stderr,"hbwmalloc: we support only 2 numa nodes, not %d\n", num_configured_nodes);
}
if (myhbwmalloc_verbose) {
for (int i=0; i<num_configured_nodes; i++) {
unsigned max_numa_cpus = numa_num_configured_cpus();
struct bitmask * mask = numa_bitmask_alloc( max_numa_cpus );
int rc = numa_node_to_cpus(i, mask);
if (rc != 0) {
fprintf(stderr, "hbwmalloc: numa_node_to_cpus failed\n");
} else {
printf("hbwmalloc: numa node %d cpu mask:", i);
for (unsigned j=0; j<max_numa_cpus; j++) {
int bit = numa_bitmask_isbitset(mask,j);
printf(" %d", bit);
}
printf("\n");
}
numa_bitmask_free(mask);
}
fflush(stdout);
}
}
#if 0 /* unused */
/* see if the user specifies a slab size */
size_t slab_size_requested = 0;
{
char * env_char = getenv("HBWMALLOC_BYTES");
if (env_char!=NULL) {
long units = 1L;
if ( NULL != strstr(env_char,"G") ) units = 1000000000L;
else if ( NULL != strstr(env_char,"M") ) units = 1000000L;
else if ( NULL != strstr(env_char,"K") ) units = 1000L;
else units = 1L;
int num_count = strspn(env_char, "0123456789");
memset( &env_char[num_count], ' ', strlen(env_char)-num_count);
slab_size_requested = units * atol(env_char);
}
if (myhbwmalloc_verbose) {
printf("hbwmalloc: requested slab_size_requested = %zu\n", slab_size_requested);
}
}
#endif
/* see what libnuma says is available */
size_t myhbwmalloc_slab_size;
{
int node = myhbwmalloc_numa_node;
long long freemem;
long long maxmem = numa_node_size64(node, &freemem);
if (myhbwmalloc_verbose) {
printf("hbwmalloc: numa_node_size64 says maxmem=%lld freemem=%lld for numa node %d\n",
maxmem, freemem, node);
}
myhbwmalloc_slab_size = freemem;
}
/* assume threads, disable if MPI knows otherwise, then allow user to override. */
int multithreaded = 1;
#ifdef HAVE_MPI
int nprocs;
{
int is_init, is_final;
MPI_Initialized(&is_init);
MPI_Finalized(&is_final);
if (is_init && !is_final) {
MPI_Comm_size(MPI_COMM_WORLD, &nprocs);
}
/* give equal portion to every MPI process */
myhbwmalloc_slab_size /= nprocs;
/* if the user initializes MPI with MPI_Init or
* MPI_Init_thread(MPI_THREAD_SINGLE), they assert there
* are no threads at all, which means we can skip the
* malloc mspace lock.
*
* if the user lies to MPI, they deserve any bad thing
* that comes of it. */
int provided;
MPI_Query_thread(&provided);
if (provided==MPI_THREAD_SINGLE) {
multithreaded = 0;
} else {
multithreaded = 1;
}
if (myhbwmalloc_verbose) {
printf("hbwmalloc: MPI processes = %d (threaded = %d)\n", nprocs, multithreaded);
printf("hbwmalloc: myhbwmalloc_slab_size = %d\n", myhbwmalloc_slab_size);
}
}
#endif
/* user can assert that hbwmalloc and friends need not be thread-safe */
{
char * env_char = getenv("HBWMALLOC_LOCKLESS");
if (env_char != NULL) {
multithreaded = 0;
if (myhbwmalloc_verbose) {
printf("hbwmalloc: user has disabled locking in mspaces by setting HBWMALLOC_LOCKLESS\n");
}
}
}
myhbwmalloc_slab = numa_alloc_onnode( myhbwmalloc_slab_size, myhbwmalloc_numa_node);
if (myhbwmalloc_slab==NULL) {
fprintf(stderr, "hbwmalloc: numa_alloc_onnode returned NULL for size = %zu\n", myhbwmalloc_slab_size);
return;
} else {
if (myhbwmalloc_verbose) {
printf("hbwmalloc: numa_alloc_onnode succeeded for size %zu\n", myhbwmalloc_slab_size);
}
/* part (less than 128*sizeof(size_t) bytes) of this space is used for bookkeeping,
* so the capacity must be at least this large */
if (myhbwmalloc_slab_size < 128*sizeof(size_t)) {
fprintf(stderr, "hbwmalloc: not enough space for mspace bookkeeping\n");
return;
}
/* see above regarding if the user lies to MPI. */
int locked = multithreaded;
myhbwmalloc_mspace = create_mspace_with_base( myhbwmalloc_slab, myhbwmalloc_slab_size, locked);
if (myhbwmalloc_mspace == NULL) {
fprintf(stderr, "hbwmalloc: create_mspace_with_base returned NULL\n");
return;
} else if (myhbwmalloc_verbose) {
printf("hbwmalloc: create_mspace_with_base succeeded for size %zu\n", myhbwmalloc_slab_size);
}
}
}
int main(int argc, const char **argv)
{
int num_cpus = numa_num_task_cpus();
printf("num cpus: %d\n", num_cpus);
printf("numa available: %d\n", numa_available());
numa_set_localalloc();
struct bitmask *bm = numa_bitmask_alloc(num_cpus);
for (int i=0; i<=numa_max_node(); ++i)
{
numa_node_to_cpus(i, bm);
printf("numa node %d ", i);
print_bitmask(bm);
printf(" - %g GiB\n", numa_node_size(i, 0) / (1024.*1024*1024.));
}
numa_bitmask_free(bm);
puts("");
char *x;
const size_t cache_line_size = 64;
const size_t array_size = 100*1000*1000;
size_t ntrips = 2;
#pragma omp parallel
{
assert(omp_get_num_threads() == num_cpus);
int tid = omp_get_thread_num();
pin_to_core(tid);
if(tid == 0)
x = (char *) numa_alloc_local(array_size);
// {{{ single access
#pragma omp barrier
for (size_t i = 0; i<num_cpus; ++i)
{
if (tid == i)
{
double t = measure_access(x, array_size, ntrips);
printf("sequential core %d -> core 0 : BW %g MB/s\n",
i, array_size*ntrips*cache_line_size / t / 1e6);
}
#pragma omp barrier
}
// }}}
// {{{ everybody contends for one
{
if (tid == 0) puts("");
#pragma omp barrier
double t = measure_access(x, array_size, ntrips);
#pragma omp barrier
for (size_t i = 0; i<num_cpus; ++i)
{
if (tid == i)
printf("all-contention core %d -> core 0 : BW %g MB/s\n",
tid, array_size*ntrips*cache_line_size / t / 1e6);
#pragma omp barrier
}
}
// }}}
// {{{ zero and someone else contending
if (tid == 0) puts("");
#pragma omp barrier
for (size_t i = 1; i<num_cpus; ++i)
{
double t;
if (tid == i || tid == 0)
t = measure_access(x, array_size, ntrips);
#pragma omp barrier
if (tid == 0)
{
printf("two-contention core %d -> core 0 : BW %g MB/s\n",
tid, array_size*ntrips*cache_line_size / t / 1e6);
}
#pragma omp barrier
if (tid == i)
{
printf("two-contention core %d -> core 0 : BW %g MB/s\n\n",
tid, array_size*ntrips*cache_line_size / t / 1e6);
}
#pragma omp barrier
}
}
numa_free(x, array_size);
return 0;
}
int main(int ac, char **av)
{
int c, i, nnodes=0;
long node=-1;
char *end;
char shortopts[array_len(opts)*2 + 1];
struct bitmask *mask = NULL;
get_short_opts(opts,shortopts);
while ((c = getopt_long(ac, av, shortopts, opts, NULL)) != -1) {
switch (c) {
case 's': /* --show */
show();
exit(0);
case 'H': /* --hardware */
nopolicy();
hardware();
exit(0);
case 'i': /* --interleave */
checknuma();
mask = numactl_parse_nodestring(optarg);
if (!mask) {
printf ("<%s> is invalid\n", optarg);
usage();
}
errno = 0;
setpolicy(MPOL_INTERLEAVE);
if (shmfd >= 0)
numa_interleave_memory(shmptr, shmlen, mask);
else
numa_set_interleave_mask(mask);
checkerror("setting interleave mask");
break;
case 'N': /* --cpunodebind */
case 'c': /* --cpubind */
dontshm("-c/--cpubind/--cpunodebind");
checknuma();
mask = numactl_parse_nodestring(optarg);
if (!mask) {
printf ("<%s> is invalid\n", optarg);
usage();
}
errno = 0;
check_cpubind(do_shm);
did_cpubind = 1;
numa_run_on_node_mask(mask);
checkerror("sched_setaffinity");
break;
case 'C': /* --physcpubind */
{
struct bitmask *cpubuf;
dontshm("-C/--physcpubind");
cpubuf = numa_parse_cpustring(optarg);
if (!cpubuf) {
printf ("<%s> is invalid\n", optarg);
usage();
}
errno = 0;
check_cpubind(do_shm);
did_cpubind = 1;
numa_sched_setaffinity(0, cpubuf);
checkerror("sched_setaffinity");
free(cpubuf);
break;
}
case 'm': /* --membind */
checknuma();
setpolicy(MPOL_BIND);
mask = numactl_parse_nodestring(optarg);
if (!mask) {
printf ("<%s> is invalid\n", optarg);
usage();
}
errno = 0;
numa_set_bind_policy(1);
if (shmfd >= 0) {
numa_tonodemask_memory(shmptr, shmlen, mask);
} else {
numa_set_membind(mask);
}
numa_set_bind_policy(0);
checkerror("setting membind");
break;
case 'p': /* --preferred */
checknuma();
setpolicy(MPOL_PREFERRED);
mask = numactl_parse_nodestring(optarg);
if (!mask) {
printf ("<%s> is invalid\n", optarg);
usage();
}
for (i=0; i<mask->size; i++) {
if (numa_bitmask_isbitset(mask, i)) {
node = i;
nnodes++;
}
}
if (nnodes != 1)
usage();
numa_bitmask_free(mask);
errno = 0;
numa_set_bind_policy(0);
if (shmfd >= 0)
numa_tonode_memory(shmptr, shmlen, node);
else
numa_set_preferred(node);
checkerror("setting preferred node");
break;
case 'l': /* --local */
checknuma();
setpolicy(MPOL_DEFAULT);
errno = 0;
if (shmfd >= 0)
numa_setlocal_memory(shmptr, shmlen);
else
numa_set_localalloc();
checkerror("local allocation");
break;
case 'S': /* --shm */
check_cpubind(did_cpubind);
nopolicy();
attach_sysvshm(optarg, "--shm");
shmattached = 1;
break;
case 'f': /* --file */
check_cpubind(did_cpubind);
nopolicy();
attach_shared(optarg, "--file");
shmattached = 1;
break;
case 'L': /* --length */
noshm("--length");
shmlen = memsize(optarg);
break;
case 'M': /* --shmmode */
noshm("--shmmode");
shmmode = strtoul(optarg, &end, 8);
if (end == optarg || *end)
usage();
break;
case 'd': /* --dump */
if (shmfd < 0)
complain(
"Cannot do --dump without shared memory.\n");
dump_shm();
do_dump = 1;
break;
case 'D': /* --dump-nodes */
if (shmfd < 0)
complain(
"Cannot do --dump-nodes without shared memory.\n");
dump_shm_nodes();
do_dump = 1;
break;
case 't': /* --strict */
did_strict = 1;
numa_set_strict(1);
break;
case 'I': /* --shmid */
shmid = strtoul(optarg, &end, 0);
if (end == optarg || *end)
usage();
break;
case 'u': /* --huge */
noshm("--huge");
shmflags |= SHM_HUGETLB;
break;
case 'o': /* --offset */
noshm("--offset");
shmoffset = memsize(optarg);
break;
case 'T': /* --touch */
needshm("--touch");
check_shmbeyond("--touch");
numa_police_memory(shmptr, shmlen);
break;
case 'V': /* --verify */
needshm("--verify");
if (set_policy < 0)
complain("Need a policy first to verify");
check_shmbeyond("--verify");
numa_police_memory(shmptr, shmlen);
if (!mask)
complain("Need a mask to verify");
else
verify_shm(set_policy, mask);
break;
default:
usage();
}
}
av += optind;
ac -= optind;
if (shmfd >= 0) {
if (*av)
usage();
exit(exitcode);
}
if (did_strict)
fprintf(stderr,
"numactl: warning. Strict flag for process ignored.\n");
if (do_dump)
usage_msg("cannot do --dump|--dump-shm for process");
if (shmoption)
usage_msg("shm related option %s for process", shmoption);
if (*av == NULL)
usage();
execvp(*av, av);
complain("execution of `%s': %s\n", av[0], strerror(errno));
return 0; /* not reached */
}
void
CPU_FreeSet(CPU_Set_t *cs)
{
numa_bitmask_free(cs);
}
size_t remap_interleaved_2m_pages(void* addr, size_t pages, int prot,
bool shared /* = false */) {
#ifdef __linux__
assert(reinterpret_cast<uintptr_t>(addr) % size2m == 0);
assert(addr != nullptr);
if (pages == 0) return 0;
#ifdef HAVE_NUMA
const int maxNode = numa_max_node();
bitmask* memMask = nullptr;
bitmask* interleaveMask = nullptr;
bitmask* mask = nullptr;
if (maxNode > 0) {
memMask = numa_get_membind();
interleaveMask = numa_get_interleave_mask();
mask = numa_allocate_nodemask();
}
#else
constexpr int maxNode = 0;
#endif
int node = -1;
int failed = 0; // consecutive failure count
int mapped_count = 0;
do {
#ifdef HAVE_NUMA
if (maxNode > 0) {
if (++node > maxNode) node = 0;
if (!numa_node_allowed(node)) {
// Numa policy forbids allocation on node
if (++failed > maxNode) break;
continue;
}
numa_bitmask_setbit(mask, node);
numa_set_membind(mask);
numa_bitmask_clearbit(mask, node);
}
#endif
// Fail early if we don't have huge pages reserved.
if (get_huge2m_info(node).free_hugepages > 0 &&
mmap_2m_impl(addr, prot, shared, true /* MAP_FIXED */)) {
addr = (char*)addr + size2m;
++mapped_count;
failed = 0;
continue;
}
// We failed on node, give up if we have failed on all nodes
if (++failed > maxNode) break;
} while (mapped_count < pages);
#ifdef HAVE_NUMA
if (mask) {
numa_set_membind(memMask);
numa_set_interleave_mask(interleaveMask);
numa_bitmask_free(mask);
numa_bitmask_free(interleaveMask);
numa_bitmask_free(memMask);
}
#endif
return mapped_count;
#else // not linux
return 0;
#endif
}
static void memkind_hbw_closest_numanode_init(void)
{
struct memkind_hbw_closest_numanode_t *g =
&memkind_hbw_closest_numanode_g;
int *bandwidth = NULL;
int num_unique = 0;
int high_bandwidth = 0;
int node;
struct bandwidth_nodes_t *bandwidth_nodes = NULL;
char *hbw_nodes_env;
struct bitmask *hbw_nodes_bm;
g->num_cpu = numa_num_configured_cpus();
g->closest_numanode = (int *)je_malloc(sizeof(int) * g->num_cpu);
bandwidth = (int *)je_malloc(sizeof(int) * NUMA_NUM_NODES);
if (!(g->closest_numanode && bandwidth)) {
g->init_err = MEMKIND_ERROR_MALLOC;
}
if (!g->init_err) {
hbw_nodes_env = getenv("MEMKIND_HBW_NODES");
if (hbw_nodes_env) {
hbw_nodes_bm = numa_parse_nodestring(hbw_nodes_env);
if (!hbw_nodes_bm) {
g->init_err = MEMKIND_ERROR_ENVIRON;
}
else {
for (node = 0; node < NUMA_NUM_NODES; ++node) {
if (numa_bitmask_isbitset(hbw_nodes_bm, node)) {
bandwidth[node] = 2;
}
else {
bandwidth[node] = 1;
}
}
numa_bitmask_free(hbw_nodes_bm);
}
}
else {
g->init_err = parse_node_bandwidth(NUMA_NUM_NODES, bandwidth,
MEMKIND_BANDWIDTH_PATH);
}
}
if (!g->init_err) {
g->init_err = create_bandwidth_nodes(NUMA_NUM_NODES, bandwidth,
&num_unique, &bandwidth_nodes);
}
if (!g->init_err) {
if (num_unique == 1) {
g->init_err = MEMKIND_ERROR_UNAVAILABLE;
}
}
if (!g->init_err) {
high_bandwidth = bandwidth_nodes[num_unique-1].bandwidth;
g->init_err = set_closest_numanode(num_unique, bandwidth_nodes,
high_bandwidth, g->num_cpu,
g->closest_numanode);
}
if (bandwidth_nodes) {
je_free(bandwidth_nodes);
}
if (bandwidth) {
je_free(bandwidth);
}
if (g->init_err) {
if (g->closest_numanode) {
je_free(g->closest_numanode);
g->closest_numanode = NULL;
}
}
}
/*----------------------------------------------------------------------------*/
int
mtcp_core_affinitize(int cpu)
{
#ifndef DISABLE_NUMA
struct bitmask *bmask;
#endif /* DISABLE_NUMA */
cpu_set_t cpus;
FILE *fp;
char sysfname[MAX_FILE_NAME];
int phy_id;
size_t n;
int ret;
n = GetNumCPUs();
if (cpu < 0 || cpu >= (int) n) {
errno = -EINVAL;
return -1;
}
CPU_ZERO(&cpus);
CPU_SET((unsigned)cpu, &cpus);
ret = sched_setaffinity(Gettid(), sizeof(cpus), &cpus);
#ifndef DISABLE_NUMA
if (numa_max_node() == 0)
return ret;
bmask = numa_bitmask_alloc(n);
assert(bmask);
#endif /* DISABLE_NUMA */
/* read physical id of the core from sys information */
snprintf(sysfname, MAX_FILE_NAME - 1,
"/sys/devices/system/cpu/cpu%d/topology/physical_package_id", cpu);
fp = fopen(sysfname, "r");
if (!fp) {
perror(sysfname);
errno = EFAULT;
return -1;
}
ret = fscanf(fp, "%d", &phy_id);
if (ret != 1) {
perror("Fail to read core id");
errno = EFAULT;
return -1;
}
#ifndef DISABLE_NUMA
numa_bitmask_setbit(bmask, phy_id);
numa_set_membind(bmask);
numa_bitmask_free(bmask);
#endif /* DISABLE_NUMA */
fclose(fp);
return ret;
}
