/**
* virNumaGetNodeMemory:
* @node: identifier of the requested NUMA node
* @memsize: returns the total size of memory in the NUMA node
* @memfree: returns the total free memory in a NUMA node
*
* Returns the size of the memory in one NUMA node in bytes via the @size
* argument and free memory of a node in the @free argument. The caller has to
* guarantee that @node is in range (see virNumaGetMaxNode).
*
* Returns 0 on success, -1 on error. Does not report errors.
*/
int
virNumaGetNodeMemory(int node,
unsigned long long *memsize,
unsigned long long *memfree)
{
long long node_size;
long long node_free;
if (memsize)
*memsize = 0;
if (memfree)
*memfree = 0;
if ((node_size = numa_node_size64(node, &node_free)) < 0)
return -1;
if (memsize)
*memsize = node_size;
if (memfree)
*memfree = node_free;
return 0;
}
unsigned long long
nodeGetFreeMemory(virConnectPtr conn ATTRIBUTE_UNUSED)
{
unsigned long long freeMem = 0;
int n;
if (numa_available() < 0) {
nodeReportError(VIR_ERR_NO_SUPPORT,
"%s", _("NUMA not supported on this host"));
goto cleanup;
}
for (n = 0 ; n <= numa_max_node() ; n++) {
long long mem;
if (numa_node_size64(n, &mem) < 0) {
nodeReportError(VIR_ERR_INTERNAL_ERROR,
"%s", _("Failed to query NUMA free memory"));
goto cleanup;
}
freeMem += mem;
}
cleanup:
return freeMem;
}
int memkind_default_get_size(struct memkind *kind, size_t *total, size_t *free)
{
nodemask_t nodemask;
struct bitmask nodemask_bm = {NUMA_NUM_NODES, nodemask.n};
long long f;
int err = 0;
int i;
*total = 0;
*free = 0;
if (kind->ops->get_mbind_nodemask) {
err = kind->ops->get_mbind_nodemask(kind, nodemask.n, NUMA_NUM_NODES);
}
else {
copy_bitmask_to_bitmask(numa_all_nodes_ptr, &nodemask_bm);
}
if (!err) {
for (i = 0; i < NUMA_NUM_NODES; ++i) {
if (numa_bitmask_isbitset(&nodemask_bm, i)) {
*total += numa_node_size64(i, &f);
*free += f;
}
}
}
return err;
}
static int filter_nodemask_mem(nodemask_t * nodemask, unsigned long max_node)
{
#if MPOL_F_MEMS_ALLOWED
unsigned long nodemask_size = max_node / 8 + 1;
memset(nodemask, 0, nodemask_size);
/*
* avoid numa_get_mems_allowed(), because of bug in getpol()
* utility function in older versions:
* http://www.spinics.net/lists/linux-numa/msg00849.html
*/
if (ltp_syscall(__NR_get_mempolicy, NULL, nodemask->n,
max_node, 0, MPOL_F_MEMS_ALLOWED) < 0)
return -2;
#else
int i;
/*
* old libnuma/kernel don't have MPOL_F_MEMS_ALLOWED, so let's assume
* that we can use any node with memory > 0
*/
for (i = 0; i < max_node; i++) {
if (!nodemask_isset(nodemask, i))
continue;
if (numa_node_size64(i, NULL) <= 0)
nodemask_clr(nodemask, i);
}
#endif /* MPOL_F_MEMS_ALLOWED */
return 0;
}
void hardware(void)
{
int i;
int numnodes=0;
int prevnode=-1;
int skip=0;
int maxnode = numa_max_node();
for (i=0; i<=maxnode; i++)
if (numa_bitmask_isbitset(numa_nodes_ptr, i))
numnodes++;
printf("available: %d nodes (", numnodes);
for (i=0; i<=maxnode; i++) {
if (numa_bitmask_isbitset(numa_nodes_ptr, i)) {
if (prevnode == -1) {
printf("%d", i);
prevnode=i;
continue;
}
if (i > prevnode + 1) {
if (skip) {
printf("%d", prevnode);
skip=0;
}
printf(",%d", i);
prevnode=i;
continue;
}
if (i == prevnode + 1) {
if (!skip) {
printf("-");
skip=1;
}
prevnode=i;
}
if ((i == maxnode) && skip)
printf("%d", prevnode);
}
}
printf(")\n");
for (i = 0; i <= maxnode; i++) {
char buf[64];
long long fr;
unsigned long long sz = numa_node_size64(i, &fr);
if (!numa_bitmask_isbitset(numa_nodes_ptr, i))
continue;
printf("node %d cpus:", i);
print_node_cpus(i);
printf("node %d size: %s\n", i, fmt_mem(sz, buf));
printf("node %d free: %s\n", i, fmt_mem(fr, buf));
}
print_distances(maxnode);
}
// display information about total/free memory on each numa node
void DumpMemoryStats()
{
int nodes = numa_max_node(),n;
for(n = 0; n <= nodes; n++ )
{
long long freememsize;
long long memsize = numa_node_size64( n, &freememsize );
printf( "node %d has %lld bytes of memory (%lld bytes free)\n", n, memsize, freememsize );
}
}
static void print_mem_stats(pid_t pid, int node)
{
char s[64];
long long node_size, freep;
if (pid == 0)
pid = getpid();
tst_resm(TINFO, "mem_stats pid: %d, node: %d", pid, node);
/* dump pid's VM info */
sprintf(s, "cat /proc/%d/status", pid);
system(s);
sprintf(s, "cat /proc/%d/numa_maps", pid);
system(s);
/* dump node free mem */
node_size = numa_node_size64(node, &freep);
tst_resm(TINFO, "Node id: %d, size: %lld, free: %lld",
node, node_size, freep);
}
int
nodeGetCellsFreeMemory(virConnectPtr conn ATTRIBUTE_UNUSED,
unsigned long long *freeMems,
int startCell,
int maxCells)
{
int n, lastCell, numCells;
int ret = -1;
int maxCell;
if (numa_available() < 0) {
nodeReportError(VIR_ERR_NO_SUPPORT,
"%s", _("NUMA not supported on this host"));
goto cleanup;
}
maxCell = numa_max_node();
if (startCell > maxCell) {
nodeReportError(VIR_ERR_INTERNAL_ERROR,
_("start cell %d out of range (0-%d)"),
startCell, maxCell);
goto cleanup;
}
lastCell = startCell + maxCells - 1;
if (lastCell > maxCell)
lastCell = maxCell;
for (numCells = 0, n = startCell ; n <= lastCell ; n++) {
long long mem;
if (numa_node_size64(n, &mem) < 0) {
nodeReportError(VIR_ERR_INTERNAL_ERROR,
"%s", _("Failed to query NUMA free memory"));
goto cleanup;
}
freeMems[numCells++] = mem;
}
ret = numCells;
cleanup:
return ret;
}
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);
}
}
}
