static int _str_to_memset(nodemask_t *mask, const char* str)
{
int len = strlen(str);
const char *ptr = str + len - 1;
int base = 0;
/* skip 0x, it's all hex anyway */
if (len > 1 && !memcmp(str, "0x", 2L))
str += 2;
nodemask_zero(mask);
while (ptr >= str) {
char val = char_to_val(*ptr);
if (val == (char) -1)
return -1;
if (val & 1)
nodemask_set(mask, base);
if (val & 2)
nodemask_set(mask, base+1);
if (val & 4)
nodemask_set(mask, base+2);
if (val & 8)
nodemask_set(mask, base+3);
len--;
ptr--;
base += 4;
}
return 0;
}
int main(void)
{
#if HAVE_NUMA_H
nodemask_t nodemask;
void hardware();
if (numa_available() < 0) {
printf("This system does not support NUMA policy\n");
numa_error("numa_available");
numa_exit_on_error = 1;
exit(numa_exit_on_error);
}
nodemask_zero(&nodemask);
nodemask_set(&nodemask, 1);
numa_bind(&nodemask);
hardware();
return numa_exit_on_error;
#else
printf("NUMA is not available\n");
return 1;
#endif
}
int get_memset(nodemask_t *mask, slurmd_job_t *job)
{
int nummasks, maskid, i, threads;
char *curstr, *selstr;
char mstr[1 + NUMA_NUM_NODES / 4];
int local_id = job->envtp->localid;
debug3("get_memset (%d) %s", job->mem_bind_type, job->mem_bind);
if (job->mem_bind_type & MEM_BIND_LOCAL) {
*mask = numa_get_run_node_mask();
return true;
}
nodemask_zero(mask);
if (job->mem_bind_type & MEM_BIND_NONE) {
return true;
}
if (job->mem_bind_type & MEM_BIND_RANK) {
threads = MAX(conf->threads, 1);
nodemask_set(mask, job->envtp->localid % (job->cpus*threads));
return true;
}
if (!job->mem_bind)
return false;
nummasks = 1;
maskid = 0;
selstr = NULL;
/* get number of strings present in mem_bind */
curstr = job->mem_bind;
while (*curstr) {
if (nummasks == local_id+1) {
selstr = curstr;
maskid = local_id;
break;
}
if (*curstr == ',')
nummasks++;
curstr++;
}
/* if we didn't already find the mask... */
if (!selstr) {
/* ...select mask string by wrapping task ID into list */
maskid = local_id % nummasks;
i = maskid;
curstr = job->mem_bind;
while (*curstr && i) {
if (*curstr == ',')
i--;
curstr++;
}
if (!*curstr) {
return false;
}
selstr = curstr;
}
/* extract the selected mask from the list */
i = 0;
curstr = mstr;
while (*selstr && *selstr != ',' && i++ < (NUMA_NUM_NODES/4))
*curstr++ = *selstr++;
*curstr = '\0';
if (job->mem_bind_type & MEM_BIND_MASK) {
/* convert mask string into nodemask_t mask */
if (_str_to_memset(mask, mstr) < 0) {
error("_str_to_memset %s", mstr);
return false;
}
return true;
}
if (job->mem_bind_type & MEM_BIND_MAP) {
unsigned int my_node = 0;
if (strncmp(mstr, "0x", 2) == 0) {
my_node = strtoul (&(mstr[2]), NULL, 16);
} else {
my_node = strtoul (mstr, NULL, 10);
}
nodemask_set(mask, my_node);
return true;
}
return false;
}
int
virNumaSetupMemoryPolicy(virDomainNumatuneMemMode mode,
virBitmapPtr nodeset)
{
nodemask_t mask;
int node = -1;
int ret = -1;
int bit = 0;
size_t i;
int maxnode = 0;
if (!nodeset)
return 0;
if (!virNumaNodesetIsAvailable(nodeset))
return -1;
maxnode = numa_max_node();
maxnode = maxnode < NUMA_NUM_NODES ? maxnode : NUMA_NUM_NODES;
/* Convert nodemask to NUMA bitmask. */
nodemask_zero(&mask);
bit = -1;
while ((bit = virBitmapNextSetBit(nodeset, bit)) >= 0) {
if (bit > maxnode) {
virReportError(VIR_ERR_INTERNAL_ERROR,
_("NUMA node %d is out of range"), bit);
return -1;
}
nodemask_set(&mask, bit);
}
switch (mode) {
case VIR_DOMAIN_NUMATUNE_MEM_STRICT:
numa_set_bind_policy(1);
numa_set_membind(&mask);
numa_set_bind_policy(0);
break;
case VIR_DOMAIN_NUMATUNE_MEM_PREFERRED:
{
int nnodes = 0;
for (i = 0; i < NUMA_NUM_NODES; i++) {
if (nodemask_isset(&mask, i)) {
node = i;
nnodes++;
}
}
if (nnodes != 1) {
virReportError(VIR_ERR_INTERNAL_ERROR,
"%s", _("NUMA memory tuning in 'preferred' mode "
"only supports single node"));
goto cleanup;
}
numa_set_bind_policy(0);
numa_set_preferred(node);
}
break;
case VIR_DOMAIN_NUMATUNE_MEM_INTERLEAVE:
numa_set_interleave_mask(&mask);
break;
case VIR_DOMAIN_NUMATUNE_MEM_LAST:
break;
}
ret = 0;
cleanup:
return ret;
}
int
virNumaSetupMemoryPolicy(virNumaTuneDef numatune,
virBitmapPtr nodemask)
{
nodemask_t mask;
int mode = -1;
int node = -1;
int ret = -1;
int i = 0;
int maxnode = 0;
virBitmapPtr tmp_nodemask = NULL;
if (numatune.memory.placement_mode ==
VIR_NUMA_TUNE_MEM_PLACEMENT_MODE_STATIC) {
if (!numatune.memory.nodemask)
return 0;
VIR_DEBUG("Set NUMA memory policy with specified nodeset");
tmp_nodemask = numatune.memory.nodemask;
} else if (numatune.memory.placement_mode ==
VIR_NUMA_TUNE_MEM_PLACEMENT_MODE_AUTO) {
VIR_DEBUG("Set NUMA memory policy with advisory nodeset from numad");
tmp_nodemask = nodemask;
} else {
return 0;
}
if (numa_available() < 0) {
virReportError(VIR_ERR_INTERNAL_ERROR,
"%s", _("Host kernel is not aware of NUMA."));
return -1;
}
maxnode = numa_max_node() + 1;
/* Convert nodemask to NUMA bitmask. */
nodemask_zero(&mask);
i = -1;
while ((i = virBitmapNextSetBit(tmp_nodemask, i)) >= 0) {
if (i > maxnode || i > NUMA_NUM_NODES) {
virReportError(VIR_ERR_INTERNAL_ERROR,
_("Nodeset is out of range, host cannot support "
"NUMA node bigger than %d"), i);
return -1;
}
nodemask_set(&mask, i);
}
mode = numatune.memory.mode;
if (mode == VIR_DOMAIN_NUMATUNE_MEM_STRICT) {
numa_set_bind_policy(1);
numa_set_membind(&mask);
numa_set_bind_policy(0);
} else if (mode == VIR_DOMAIN_NUMATUNE_MEM_PREFERRED) {
int nnodes = 0;
for (i = 0; i < NUMA_NUM_NODES; i++) {
if (nodemask_isset(&mask, i)) {
node = i;
nnodes++;
}
}
if (nnodes != 1) {
virReportError(VIR_ERR_INTERNAL_ERROR,
"%s", _("NUMA memory tuning in 'preferred' mode "
"only supports single node"));
goto cleanup;
}
numa_set_bind_policy(0);
numa_set_preferred(node);
} else if (mode == VIR_DOMAIN_NUMATUNE_MEM_INTERLEAVE) {
numa_set_interleave_mask(&mask);
} else {
/* XXX: Shouldn't go here, as we already do checking when
* parsing domain XML.
*/
virReportError(VIR_ERR_XML_ERROR,
"%s", _("Invalid mode for memory NUMA tuning."));
goto cleanup;
}
ret = 0;
cleanup:
return ret;
}
int INTERNAL qt_affinity_gendists(qthread_shepherd_t *sheps,
qthread_shepherd_id_t nshepherds)
{ /*{{{ */
const size_t num_extant_nodes = numa_max_node() + 1;
nodemask_t bmask;
qthread_debug(AFFINITY_FUNCTIONS, "sheps(%p), nshepherds(%u), num_extant_nodes:%u\n", sheps, nshepherds, (unsigned)num_extant_nodes);
if (numa_available() == -1) {
return QTHREAD_THIRD_PARTY_ERROR;
}
nodemask_zero(&bmask);
/* assign nodes */
qthread_debug(AFFINITY_DETAILS, "assign nodes...\n");
for (size_t i = 0; i < nshepherds; ++i) {
sheps[i].node = i % num_extant_nodes;
qthread_debug(AFFINITY_DETAILS, "set bit %u in bmask\n", i % num_extant_nodes);
nodemask_set(&bmask, i % num_extant_nodes);
}
qthread_debug(AFFINITY_DETAILS, "numa_set_interleave_mask\n");
numa_set_interleave_mask(&bmask);
qthread_debug(AFFINITY_DETAILS, "querying distances...\n");
/* truly ancient versions of libnuma (in the changelog, this is
* considered "pre-history") do not have numa_distance() */
for (qthread_shepherd_id_t i = 0; i < nshepherds; i++) {
qthread_debug(AFFINITY_DETAILS, "i = %u < %u...\n", i, nshepherds);
const unsigned int node_i = sheps[i].node;
size_t j, k;
sheps[i].shep_dists = calloc(nshepherds, sizeof(unsigned int));
sheps[i].sorted_sheplist = calloc(nshepherds - 1, sizeof(qthread_shepherd_id_t));
qthread_debug(AFFINITY_DETAILS, "allocs %p %p\n", sheps[i].shep_dists, sheps[i].sorted_sheplist);
assert(sheps[i].shep_dists);
assert(sheps[i].sorted_sheplist);
for (j = 0; j < nshepherds; j++) {
const unsigned int node_j = sheps[j].node;
#if QTHREAD_NUMA_DISTANCE_WORKING
if ((node_i != QTHREAD_NO_NODE) && (node_j != QTHREAD_NO_NODE) && (node_i != node_j)) {
sheps[i].shep_dists[j] = numa_distance(node_i, node_j);
} else {
#endif
/* XXX too arbitrary */
if (i == j) {
sheps[i].shep_dists[j] = 0;
} else {
sheps[i].shep_dists[j] = 20;
}
#if QTHREAD_NUMA_DISTANCE_WORKING
}
#endif
qthread_debug(AFFINITY_DETAILS, "shep %u to shep %u distance: %u\n", i, j, sheps[i].shep_dists[j]);
}
k = 0;
for (j = 0; j < nshepherds; j++) {
if (j != i) {
sheps[i].sorted_sheplist[k++] = j;
}
}
if (nshepherds > 1) {
sort_sheps(sheps[i].shep_dists, sheps[i].sorted_sheplist, nshepherds);
}
}
return QTHREAD_SUCCESS;
} /*}}} */
static int virLXCControllerSetupNUMAPolicy(virLXCControllerPtr ctrl)
{
nodemask_t mask;
int mode = -1;
int node = -1;
int ret = -1;
int i = 0;
int maxnode = 0;
bool warned = false;
if (!ctrl->def->numatune.memory.nodemask)
return 0;
VIR_DEBUG("Setting NUMA memory policy");
if (numa_available() < 0) {
virReportError(VIR_ERR_CONFIG_UNSUPPORTED,
"%s", _("Host kernel is not aware of NUMA."));
return -1;
}
maxnode = numa_max_node() + 1;
/* Convert nodemask to NUMA bitmask. */
nodemask_zero(&mask);
i = -1;
while ((i = virBitmapNextSetBit(ctrl->def->numatune.memory.nodemask, i)) >= 0) {
if (i > NUMA_NUM_NODES) {
virReportError(VIR_ERR_CONFIG_UNSUPPORTED,
_("Host cannot support NUMA node %d"), i);
return -1;
}
if (i > maxnode && !warned) {
VIR_WARN("nodeset is out of range, there is only %d NUMA "
"nodes on host", maxnode);
warned = true;
}
nodemask_set(&mask, i);
}
mode = ctrl->def->numatune.memory.mode;
if (mode == VIR_DOMAIN_NUMATUNE_MEM_STRICT) {
numa_set_bind_policy(1);
numa_set_membind(&mask);
numa_set_bind_policy(0);
} else if (mode == VIR_DOMAIN_NUMATUNE_MEM_PREFERRED) {
int nnodes = 0;
for (i = 0; i < NUMA_NUM_NODES; i++) {
if (nodemask_isset(&mask, i)) {
node = i;
nnodes++;
}
}
if (nnodes != 1) {
virReportError(VIR_ERR_CONFIG_UNSUPPORTED,
"%s", _("NUMA memory tuning in 'preferred' mode "
"only supports single node"));
goto cleanup;
}
numa_set_bind_policy(0);
numa_set_preferred(node);
} else if (mode == VIR_DOMAIN_NUMATUNE_MEM_INTERLEAVE) {
numa_set_interleave_mask(&mask);
} else {
virReportError(VIR_ERR_CONFIG_UNSUPPORTED,
_("Unable to set NUMA policy %s"),
virDomainNumatuneMemModeTypeToString(mode));
goto cleanup;
}
ret = 0;
cleanup:
return ret;
}