4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; if not, write to the
18 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
19 * Boston, MA 021110-1307, USA
24 * Copyright (c) 2010, Oracle and/or its affiliates. All rights reserved.
26 * Copyright (c) 2012, 2014, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * Author: liang@whamcloud.com
35 #define DEBUG_SUBSYSTEM S_LNET
37 #include <linux/cpu.h>
38 #include <linux/sched.h>
39 #include <libcfs/libcfs.h>
44 * modparam for setting number of partitions
46 * 0 : estimate best value based on cores or NUMA nodes
47 * 1 : disable multiple partitions
48 * >1 : specify number of partitions
50 static int cpu_npartitions;
51 CFS_MODULE_PARM(cpu_npartitions, "i", int, 0444, "# of CPU partitions");
54 * modparam for setting CPU partitions patterns:
56 * i.e: "0[0,1,2,3] 1[4,5,6,7]", number before bracket is CPU partition ID,
57 * number in bracket is processor ID (core or HT)
59 * i.e: "N 0[0,1] 1[2,3]" the first character 'N' means numbers in bracket
60 * are NUMA node ID, number before bracket is CPU partition ID.
62 * i.e: "N", shortcut expression to create CPT from NUMA & CPU topology
64 * NB: If user specified cpu_pattern, cpu_npartitions will be ignored
66 static char *cpu_pattern = "";
67 CFS_MODULE_PARM(cpu_pattern, "s", charp, 0444, "CPU partitions pattern");
70 /* serialize hotplug etc */
72 /* reserved for hotplug */
73 unsigned long cpt_version;
74 /* mutex to protect cpt_cpumask */
75 struct mutex cpt_mutex;
76 /* scratch buffer for set/unset_node */
77 cpumask_t *cpt_cpumask;
80 static struct cfs_cpt_data cpt_data;
83 cfs_cpu_core_siblings(int cpu, cpumask_t *mask)
85 /* return cpumask of cores in the same socket */
86 cpumask_copy(mask, topology_core_cpumask(cpu));
89 /* return number of cores in the same socket of \a cpu */
91 cfs_cpu_core_nsiblings(int cpu)
95 mutex_lock(&cpt_data.cpt_mutex);
97 cfs_cpu_core_siblings(cpu, cpt_data.cpt_cpumask);
98 num = cpumask_weight(cpt_data.cpt_cpumask);
100 mutex_unlock(&cpt_data.cpt_mutex);
105 /* return cpumask of HTs in the same core */
107 cfs_cpu_ht_siblings(int cpu, cpumask_t *mask)
109 cpumask_copy(mask, topology_sibling_cpumask(cpu));
112 /* return number of HTs in the same core of \a cpu */
114 cfs_cpu_ht_nsiblings(int cpu)
118 num = cpumask_weight(topology_sibling_cpumask(cpu));
122 EXPORT_SYMBOL(cfs_cpu_ht_nsiblings);
125 cfs_node_to_cpumask(int node, cpumask_t *mask)
127 const cpumask_t *tmp = cpumask_of_node(node);
130 cpumask_copy(mask, tmp);
136 cfs_cpt_table_free(struct cfs_cpt_table *cptab)
140 if (cptab->ctb_cpu2cpt != NULL) {
141 LIBCFS_FREE(cptab->ctb_cpu2cpt,
142 num_possible_cpus() *
143 sizeof(cptab->ctb_cpu2cpt[0]));
146 for (i = 0; cptab->ctb_parts != NULL && i < cptab->ctb_nparts; i++) {
147 struct cfs_cpu_partition *part = &cptab->ctb_parts[i];
149 if (part->cpt_nodemask != NULL) {
150 LIBCFS_FREE(part->cpt_nodemask,
151 sizeof(*part->cpt_nodemask));
154 if (part->cpt_cpumask != NULL)
155 LIBCFS_FREE(part->cpt_cpumask, cpumask_size());
158 if (cptab->ctb_parts != NULL) {
159 LIBCFS_FREE(cptab->ctb_parts,
160 cptab->ctb_nparts * sizeof(cptab->ctb_parts[0]));
163 if (cptab->ctb_nodemask != NULL)
164 LIBCFS_FREE(cptab->ctb_nodemask, sizeof(*cptab->ctb_nodemask));
165 if (cptab->ctb_cpumask != NULL)
166 LIBCFS_FREE(cptab->ctb_cpumask, cpumask_size());
168 LIBCFS_FREE(cptab, sizeof(*cptab));
170 EXPORT_SYMBOL(cfs_cpt_table_free);
172 struct cfs_cpt_table *
173 cfs_cpt_table_alloc(unsigned int ncpt)
175 struct cfs_cpt_table *cptab;
178 LIBCFS_ALLOC(cptab, sizeof(*cptab));
182 cptab->ctb_nparts = ncpt;
184 LIBCFS_ALLOC(cptab->ctb_cpumask, cpumask_size());
185 LIBCFS_ALLOC(cptab->ctb_nodemask, sizeof(*cptab->ctb_nodemask));
187 if (cptab->ctb_cpumask == NULL || cptab->ctb_nodemask == NULL)
190 LIBCFS_ALLOC(cptab->ctb_cpu2cpt,
191 num_possible_cpus() * sizeof(cptab->ctb_cpu2cpt[0]));
192 if (cptab->ctb_cpu2cpt == NULL)
195 memset(cptab->ctb_cpu2cpt, -1,
196 num_possible_cpus() * sizeof(cptab->ctb_cpu2cpt[0]));
198 LIBCFS_ALLOC(cptab->ctb_parts, ncpt * sizeof(cptab->ctb_parts[0]));
199 if (cptab->ctb_parts == NULL)
202 for (i = 0; i < ncpt; i++) {
203 struct cfs_cpu_partition *part = &cptab->ctb_parts[i];
205 LIBCFS_ALLOC(part->cpt_cpumask, cpumask_size());
206 LIBCFS_ALLOC(part->cpt_nodemask, sizeof(*part->cpt_nodemask));
207 if (part->cpt_cpumask == NULL || part->cpt_nodemask == NULL)
211 spin_lock(&cpt_data.cpt_lock);
212 /* Reserved for hotplug */
213 cptab->ctb_version = cpt_data.cpt_version;
214 spin_unlock(&cpt_data.cpt_lock);
219 cfs_cpt_table_free(cptab);
222 EXPORT_SYMBOL(cfs_cpt_table_alloc);
225 cfs_cpt_table_print(struct cfs_cpt_table *cptab, char *buf, int len)
232 for (i = 0; i < cptab->ctb_nparts; i++) {
234 rc = snprintf(tmp, len, "%d\t: ", i);
244 for_each_cpu(j, cptab->ctb_parts[i].cpt_cpumask) {
245 rc = snprintf(tmp, len, "%d ", j);
265 EXPORT_SYMBOL(cfs_cpt_table_print);
268 cfs_cpt_number(struct cfs_cpt_table *cptab)
270 return cptab->ctb_nparts;
272 EXPORT_SYMBOL(cfs_cpt_number);
275 cfs_cpt_weight(struct cfs_cpt_table *cptab, int cpt)
277 LASSERT(cpt == CFS_CPT_ANY || (cpt >= 0 && cpt < cptab->ctb_nparts));
279 return cpt == CFS_CPT_ANY ?
280 cpumask_weight(cptab->ctb_cpumask) :
281 cpumask_weight(cptab->ctb_parts[cpt].cpt_cpumask);
283 EXPORT_SYMBOL(cfs_cpt_weight);
286 cfs_cpt_online(struct cfs_cpt_table *cptab, int cpt)
288 LASSERT(cpt == CFS_CPT_ANY || (cpt >= 0 && cpt < cptab->ctb_nparts));
290 return cpt == CFS_CPT_ANY ?
291 cpumask_any_and(cptab->ctb_cpumask,
292 cpu_online_mask) < nr_cpu_ids :
293 cpumask_any_and(cptab->ctb_parts[cpt].cpt_cpumask,
294 cpu_online_mask) < nr_cpu_ids;
296 EXPORT_SYMBOL(cfs_cpt_online);
299 cfs_cpt_cpumask(struct cfs_cpt_table *cptab, int cpt)
301 LASSERT(cpt == CFS_CPT_ANY || (cpt >= 0 && cpt < cptab->ctb_nparts));
303 return cpt == CFS_CPT_ANY ?
304 cptab->ctb_cpumask : cptab->ctb_parts[cpt].cpt_cpumask;
306 EXPORT_SYMBOL(cfs_cpt_cpumask);
309 cfs_cpt_nodemask(struct cfs_cpt_table *cptab, int cpt)
311 LASSERT(cpt == CFS_CPT_ANY || (cpt >= 0 && cpt < cptab->ctb_nparts));
313 return cpt == CFS_CPT_ANY ?
314 cptab->ctb_nodemask : cptab->ctb_parts[cpt].cpt_nodemask;
316 EXPORT_SYMBOL(cfs_cpt_nodemask);
319 cfs_cpt_set_cpu(struct cfs_cpt_table *cptab, int cpt, int cpu)
323 LASSERT(cpt >= 0 && cpt < cptab->ctb_nparts);
325 if (cpu < 0 || cpu >= nr_cpu_ids || !cpu_online(cpu)) {
326 CDEBUG(D_INFO, "CPU %d is invalid or it's offline\n", cpu);
330 if (cptab->ctb_cpu2cpt[cpu] != -1) {
331 CDEBUG(D_INFO, "CPU %d is already in partition %d\n",
332 cpu, cptab->ctb_cpu2cpt[cpu]);
336 cptab->ctb_cpu2cpt[cpu] = cpt;
338 LASSERT(!cpumask_test_cpu(cpu, cptab->ctb_cpumask));
339 LASSERT(!cpumask_test_cpu(cpu, cptab->ctb_parts[cpt].cpt_cpumask));
341 cpumask_set_cpu(cpu, cptab->ctb_cpumask);
342 cpumask_set_cpu(cpu, cptab->ctb_parts[cpt].cpt_cpumask);
344 node = cpu_to_node(cpu);
346 /* first CPU of @node in this CPT table */
347 if (!node_isset(node, *cptab->ctb_nodemask))
348 node_set(node, *cptab->ctb_nodemask);
350 /* first CPU of @node in this partition */
351 if (!node_isset(node, *cptab->ctb_parts[cpt].cpt_nodemask))
352 node_set(node, *cptab->ctb_parts[cpt].cpt_nodemask);
356 EXPORT_SYMBOL(cfs_cpt_set_cpu);
359 cfs_cpt_unset_cpu(struct cfs_cpt_table *cptab, int cpt, int cpu)
364 LASSERT(cpt == CFS_CPT_ANY || (cpt >= 0 && cpt < cptab->ctb_nparts));
366 if (cpu < 0 || cpu >= nr_cpu_ids) {
367 CDEBUG(D_INFO, "Invalid CPU id %d\n", cpu);
371 if (cpt == CFS_CPT_ANY) {
372 /* caller doesn't know the partition ID */
373 cpt = cptab->ctb_cpu2cpt[cpu];
374 if (cpt < 0) { /* not set in this CPT-table */
375 CDEBUG(D_INFO, "Try to unset cpu %d which is "
376 "not in CPT-table %p\n", cpt, cptab);
380 } else if (cpt != cptab->ctb_cpu2cpt[cpu]) {
382 "CPU %d is not in cpu-partition %d\n", cpu, cpt);
386 LASSERT(cpumask_test_cpu(cpu, cptab->ctb_parts[cpt].cpt_cpumask));
387 LASSERT(cpumask_test_cpu(cpu, cptab->ctb_cpumask));
389 cpumask_clear_cpu(cpu, cptab->ctb_parts[cpt].cpt_cpumask);
390 cpumask_clear_cpu(cpu, cptab->ctb_cpumask);
391 cptab->ctb_cpu2cpt[cpu] = -1;
393 node = cpu_to_node(cpu);
395 LASSERT(node_isset(node, *cptab->ctb_parts[cpt].cpt_nodemask));
396 LASSERT(node_isset(node, *cptab->ctb_nodemask));
398 for_each_cpu(i, cptab->ctb_parts[cpt].cpt_cpumask) {
399 /* this CPT has other CPU belonging to this node? */
400 if (cpu_to_node(i) == node)
405 node_clear(node, *cptab->ctb_parts[cpt].cpt_nodemask);
407 for_each_cpu(i, cptab->ctb_cpumask) {
408 /* this CPT-table has other CPU belonging to this node? */
409 if (cpu_to_node(i) == node)
414 node_clear(node, *cptab->ctb_nodemask);
418 EXPORT_SYMBOL(cfs_cpt_unset_cpu);
421 cfs_cpt_set_cpumask(struct cfs_cpt_table *cptab, int cpt, cpumask_t *mask)
425 if (cpumask_weight(mask) == 0 ||
426 cpumask_any_and(mask, cpu_online_mask) >= nr_cpu_ids) {
427 CDEBUG(D_INFO, "No online CPU is found in the CPU mask "
428 "for CPU partition %d\n", cpt);
432 for_each_cpu(i, mask) {
433 if (!cfs_cpt_set_cpu(cptab, cpt, i))
439 EXPORT_SYMBOL(cfs_cpt_set_cpumask);
442 cfs_cpt_unset_cpumask(struct cfs_cpt_table *cptab, int cpt, cpumask_t *mask)
446 for_each_cpu(i, mask)
447 cfs_cpt_unset_cpu(cptab, cpt, i);
449 EXPORT_SYMBOL(cfs_cpt_unset_cpumask);
452 cfs_cpt_set_node(struct cfs_cpt_table *cptab, int cpt, int node)
457 if (node < 0 || node >= MAX_NUMNODES) {
459 "Invalid NUMA id %d for CPU partition %d\n", node, cpt);
463 mutex_lock(&cpt_data.cpt_mutex);
465 mask = cpt_data.cpt_cpumask;
466 cfs_node_to_cpumask(node, mask);
468 rc = cfs_cpt_set_cpumask(cptab, cpt, mask);
470 mutex_unlock(&cpt_data.cpt_mutex);
474 EXPORT_SYMBOL(cfs_cpt_set_node);
477 cfs_cpt_unset_node(struct cfs_cpt_table *cptab, int cpt, int node)
481 if (node < 0 || node >= MAX_NUMNODES) {
483 "Invalid NUMA id %d for CPU partition %d\n", node, cpt);
487 mutex_lock(&cpt_data.cpt_mutex);
489 mask = cpt_data.cpt_cpumask;
490 cfs_node_to_cpumask(node, mask);
492 cfs_cpt_unset_cpumask(cptab, cpt, mask);
494 mutex_unlock(&cpt_data.cpt_mutex);
496 EXPORT_SYMBOL(cfs_cpt_unset_node);
499 cfs_cpt_set_nodemask(struct cfs_cpt_table *cptab, int cpt, nodemask_t *mask)
503 for_each_node_mask(i, *mask) {
504 if (!cfs_cpt_set_node(cptab, cpt, i))
510 EXPORT_SYMBOL(cfs_cpt_set_nodemask);
513 cfs_cpt_unset_nodemask(struct cfs_cpt_table *cptab, int cpt, nodemask_t *mask)
517 for_each_node_mask(i, *mask)
518 cfs_cpt_unset_node(cptab, cpt, i);
520 EXPORT_SYMBOL(cfs_cpt_unset_nodemask);
523 cfs_cpt_clear(struct cfs_cpt_table *cptab, int cpt)
528 if (cpt == CFS_CPT_ANY) {
529 last = cptab->ctb_nparts - 1;
535 for (; cpt <= last; cpt++) {
536 for_each_cpu(i, cptab->ctb_parts[cpt].cpt_cpumask)
537 cfs_cpt_unset_cpu(cptab, cpt, i);
540 EXPORT_SYMBOL(cfs_cpt_clear);
543 cfs_cpt_spread_node(struct cfs_cpt_table *cptab, int cpt)
550 /* convert CPU partition ID to HW node id */
552 if (cpt < 0 || cpt >= cptab->ctb_nparts) {
553 mask = cptab->ctb_nodemask;
554 rotor = cptab->ctb_spread_rotor++;
556 mask = cptab->ctb_parts[cpt].cpt_nodemask;
557 rotor = cptab->ctb_parts[cpt].cpt_spread_rotor++;
560 weight = nodes_weight(*mask);
565 for_each_node_mask(node, *mask) {
573 EXPORT_SYMBOL(cfs_cpt_spread_node);
576 cfs_cpt_current(struct cfs_cpt_table *cptab, int remap)
578 int cpu = smp_processor_id();
579 int cpt = cptab->ctb_cpu2cpt[cpu];
585 /* don't return negative value for safety of upper layer,
586 * instead we shadow the unknown cpu to a valid partition ID */
587 cpt = cpu % cptab->ctb_nparts;
592 EXPORT_SYMBOL(cfs_cpt_current);
595 cfs_cpt_of_cpu(struct cfs_cpt_table *cptab, int cpu)
597 LASSERT(cpu >= 0 && cpu < nr_cpu_ids);
599 return cptab->ctb_cpu2cpt[cpu];
601 EXPORT_SYMBOL(cfs_cpt_of_cpu);
604 cfs_cpt_bind(struct cfs_cpt_table *cptab, int cpt)
607 nodemask_t *nodemask;
611 LASSERT(cpt == CFS_CPT_ANY || (cpt >= 0 && cpt < cptab->ctb_nparts));
613 if (cpt == CFS_CPT_ANY) {
614 cpumask = cptab->ctb_cpumask;
615 nodemask = cptab->ctb_nodemask;
617 cpumask = cptab->ctb_parts[cpt].cpt_cpumask;
618 nodemask = cptab->ctb_parts[cpt].cpt_nodemask;
621 if (cpumask_any_and(cpumask, cpu_online_mask) >= nr_cpu_ids) {
622 CERROR("No online CPU found in CPU partition %d, did someone "
623 "do CPU hotplug on system? You might need to reload "
624 "Lustre modules to keep system working well.\n", cpt);
628 for_each_online_cpu(i) {
629 if (cpumask_test_cpu(i, cpumask))
632 rc = set_cpus_allowed_ptr(current, cpumask);
633 set_mems_allowed(*nodemask);
635 schedule(); /* switch to allowed CPU */
640 /* don't need to set affinity because all online CPUs are covered */
643 EXPORT_SYMBOL(cfs_cpt_bind);
646 * Choose max to \a number CPUs from \a node and set them in \a cpt.
647 * We always prefer to choose CPU in the same core/socket.
650 cfs_cpt_choose_ncpus(struct cfs_cpt_table *cptab, int cpt,
651 cpumask_t *node, int number)
653 cpumask_t *socket = NULL;
654 cpumask_t *core = NULL;
660 if (number >= cpumask_weight(node)) {
661 while (!cpumask_empty(node)) {
662 cpu = cpumask_first(node);
664 rc = cfs_cpt_set_cpu(cptab, cpt, cpu);
667 cpumask_clear_cpu(cpu, node);
672 /* allocate scratch buffer */
673 LIBCFS_ALLOC(socket, cpumask_size());
674 LIBCFS_ALLOC(core, cpumask_size());
675 if (socket == NULL || core == NULL) {
680 while (!cpumask_empty(node)) {
681 cpu = cpumask_first(node);
683 /* get cpumask for cores in the same socket */
684 cfs_cpu_core_siblings(cpu, socket);
685 cpumask_and(socket, socket, node);
687 LASSERT(!cpumask_empty(socket));
689 while (!cpumask_empty(socket)) {
692 /* get cpumask for hts in the same core */
693 cfs_cpu_ht_siblings(cpu, core);
694 cpumask_and(core, core, node);
696 LASSERT(!cpumask_empty(core));
698 for_each_cpu(i, core) {
699 cpumask_clear_cpu(i, socket);
700 cpumask_clear_cpu(i, node);
702 rc = cfs_cpt_set_cpu(cptab, cpt, i);
711 cpu = cpumask_first(socket);
717 LIBCFS_FREE(socket, cpumask_size());
719 LIBCFS_FREE(core, cpumask_size());
723 #define CPT_WEIGHT_MIN 4u
726 cfs_cpt_num_estimate(void)
728 unsigned nnode = num_online_nodes();
729 unsigned ncpu = num_online_cpus();
732 if (ncpu <= CPT_WEIGHT_MIN) {
737 /* generate reasonable number of CPU partitions based on total number
738 * of CPUs, Preferred N should be power2 and match this condition:
739 * 2 * (N - 1)^2 < NCPUS <= 2 * N^2 */
740 for (ncpt = 2; ncpu > 2 * ncpt * ncpt; ncpt <<= 1) {}
742 if (ncpt <= nnode) { /* fat numa system */
746 } else { /* ncpt > nnode */
747 while ((nnode << 1) <= ncpt)
754 #if (BITS_PER_LONG == 32)
755 /* config many CPU partitions on 32-bit system could consume
757 ncpt = min(2U, ncpt);
759 while (ncpu % ncpt != 0)
760 ncpt--; /* worst case is 1 */
765 static struct cfs_cpt_table *
766 cfs_cpt_table_create(int ncpt)
768 struct cfs_cpt_table *cptab = NULL;
769 cpumask_t *mask = NULL;
775 rc = cfs_cpt_num_estimate();
779 if (ncpt > num_online_cpus() || ncpt > 4 * rc) {
780 CWARN("CPU partition number %d is larger than suggested "
781 "value (%d), your system may have performance"
782 "issue or run out of memory while under pressure\n",
786 if (num_online_cpus() % ncpt != 0) {
787 CERROR("CPU number %d is not multiple of cpu_npartition %d, "
788 "please try different cpu_npartitions value or"
789 "set pattern string by cpu_pattern=STRING\n",
790 (int)num_online_cpus(), ncpt);
794 cptab = cfs_cpt_table_alloc(ncpt);
796 CERROR("Failed to allocate CPU map(%d)\n", ncpt);
800 num = num_online_cpus() / ncpt;
802 CERROR("CPU changed while setting CPU partition\n");
806 LIBCFS_ALLOC(mask, cpumask_size());
808 CERROR("Failed to allocate scratch cpumask\n");
812 for_each_online_node(i) {
813 cfs_node_to_cpumask(i, mask);
815 while (!cpumask_empty(mask)) {
816 struct cfs_cpu_partition *part;
819 /* Each emulated NUMA node has all allowed CPUs in
821 * End loop when all partitions have assigned CPUs.
826 part = &cptab->ctb_parts[cpt];
828 n = num - cpumask_weight(part->cpt_cpumask);
831 rc = cfs_cpt_choose_ncpus(cptab, cpt, mask, n);
835 LASSERT(num >= cpumask_weight(part->cpt_cpumask));
836 if (num == cpumask_weight(part->cpt_cpumask))
842 num != cpumask_weight(cptab->ctb_parts[ncpt - 1].cpt_cpumask)) {
843 CERROR("Expect %d(%d) CPU partitions but got %d(%d), "
844 "CPU hotplug/unplug while setting?\n",
845 cptab->ctb_nparts, num, cpt,
846 cpumask_weight(cptab->ctb_parts[ncpt - 1].cpt_cpumask));
850 LIBCFS_FREE(mask, cpumask_size());
855 CERROR("Failed to setup CPU-partition-table with %d "
856 "CPU-partitions, online HW nodes: %d, HW cpus: %d.\n",
857 ncpt, num_online_nodes(), num_online_cpus());
860 LIBCFS_FREE(mask, cpumask_size());
863 cfs_cpt_table_free(cptab);
868 static struct cfs_cpt_table *
869 cfs_cpt_table_create_pattern(char *pattern)
871 struct cfs_cpt_table *cptab;
881 str = cfs_trimwhite(pattern);
882 if (*str == 'n' || *str == 'N') {
884 if (*pattern != '\0') {
885 node = 1; /* numa pattern */
887 } else { /* shortcut to create CPT from NUMA & CPU topology */
889 ncpt = num_online_nodes();
893 if (ncpt == 0) { /* scanning bracket which is mark of partition */
894 for (str = pattern;; str++, ncpt++) {
895 str = strchr(str, '[');
902 (node && ncpt > num_online_nodes()) ||
903 (!node && ncpt > num_online_cpus())) {
904 CERROR("Invalid pattern %s, or too many partitions %d\n",
909 cptab = cfs_cpt_table_alloc(ncpt);
911 CERROR("Failed to allocate cpu partition table\n");
915 if (node < 0) { /* shortcut to create CPT from NUMA & CPU topology */
917 for_each_online_node(i) {
919 CERROR("CPU changed while setting CPU "
920 "partition table, %d/%d\n", cpt, ncpt);
924 rc = cfs_cpt_set_node(cptab, cpt++, i);
931 high = node ? MAX_NUMNODES - 1 : nr_cpu_ids - 1;
933 for (str = cfs_trimwhite(pattern), c = 0;; c++) {
934 struct cfs_range_expr *range;
935 struct cfs_expr_list *el;
936 char *bracket = strchr(str, '[');
939 if (bracket == NULL) {
941 CERROR("Invalid pattern %s\n", str);
943 } else if (c != ncpt) {
944 CERROR("expect %d partitions but found %d\n",
951 if (sscanf(str, "%d%n", &cpt, &n) < 1) {
952 CERROR("Invalid cpu pattern %s\n", str);
956 if (cpt < 0 || cpt >= ncpt) {
957 CERROR("Invalid partition id %d, total partitions %d\n",
962 if (cfs_cpt_weight(cptab, cpt) != 0) {
963 CERROR("Partition %d has already been set.\n", cpt);
967 str = cfs_trimwhite(str + n);
968 if (str != bracket) {
969 CERROR("Invalid pattern %s\n", str);
973 bracket = strchr(str, ']');
974 if (bracket == NULL) {
975 CERROR("missing right bracket for cpt %d, %s\n",
980 if (cfs_expr_list_parse(str, (bracket - str) + 1,
981 0, high, &el) != 0) {
982 CERROR("Can't parse number range: %s\n", str);
986 list_for_each_entry(range, &el->el_exprs, re_link) {
987 for (i = range->re_lo; i <= range->re_hi; i++) {
988 if ((i - range->re_lo) % range->re_stride != 0)
991 rc = node ? cfs_cpt_set_node(cptab, cpt, i) :
992 cfs_cpt_set_cpu(cptab, cpt, i);
994 cfs_expr_list_free(el);
1000 cfs_expr_list_free(el);
1002 if (!cfs_cpt_online(cptab, cpt)) {
1003 CERROR("No online CPU is found on partition %d\n", cpt);
1007 str = cfs_trimwhite(bracket + 1);
1013 cfs_cpt_table_free(cptab);
1017 #ifdef CONFIG_HOTPLUG_CPU
1019 cfs_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
1021 unsigned int cpu = (unsigned long)hcpu;
1026 case CPU_DEAD_FROZEN:
1028 case CPU_ONLINE_FROZEN:
1029 spin_lock(&cpt_data.cpt_lock);
1030 cpt_data.cpt_version++;
1031 spin_unlock(&cpt_data.cpt_lock);
1033 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN) {
1034 CDEBUG(D_INFO, "CPU changed [cpu %u action %lx]\n",
1039 mutex_lock(&cpt_data.cpt_mutex);
1040 /* if all HTs in a core are offline, it may break affinity */
1041 cfs_cpu_ht_siblings(cpu, cpt_data.cpt_cpumask);
1042 warn = cpumask_any_and(cpt_data.cpt_cpumask,
1043 cpu_online_mask) >= nr_cpu_ids;
1044 mutex_unlock(&cpt_data.cpt_mutex);
1045 CDEBUG(warn ? D_WARNING : D_INFO,
1046 "Lustre: can't support CPU plug-out well now, "
1047 "performance and stability could be impacted"
1048 "[CPU %u action: %lx]\n", cpu, action);
1054 static struct notifier_block cfs_cpu_notifier = {
1055 .notifier_call = cfs_cpu_notify,
1064 if (cfs_cpt_table != NULL)
1065 cfs_cpt_table_free(cfs_cpt_table);
1067 #ifdef CONFIG_HOTPLUG_CPU
1068 unregister_hotcpu_notifier(&cfs_cpu_notifier);
1070 if (cpt_data.cpt_cpumask != NULL)
1071 LIBCFS_FREE(cpt_data.cpt_cpumask, cpumask_size());
1077 LASSERT(cfs_cpt_table == NULL);
1079 memset(&cpt_data, 0, sizeof(cpt_data));
1081 LIBCFS_ALLOC(cpt_data.cpt_cpumask, cpumask_size());
1082 if (cpt_data.cpt_cpumask == NULL) {
1083 CERROR("Failed to allocate scratch buffer\n");
1087 spin_lock_init(&cpt_data.cpt_lock);
1088 mutex_init(&cpt_data.cpt_mutex);
1090 #ifdef CONFIG_HOTPLUG_CPU
1091 register_hotcpu_notifier(&cfs_cpu_notifier);
1094 if (*cpu_pattern != 0) {
1095 cfs_cpt_table = cfs_cpt_table_create_pattern(cpu_pattern);
1096 if (cfs_cpt_table == NULL) {
1097 CERROR("Failed to create cptab from pattern %s\n",
1103 cfs_cpt_table = cfs_cpt_table_create(cpu_npartitions);
1104 if (cfs_cpt_table == NULL) {
1105 CERROR("Failed to create ptable with npartitions %d\n",
1111 spin_lock(&cpt_data.cpt_lock);
1112 if (cfs_cpt_table->ctb_version != cpt_data.cpt_version) {
1113 spin_unlock(&cpt_data.cpt_lock);
1114 CERROR("CPU hotplug/unplug during setup\n");
1117 spin_unlock(&cpt_data.cpt_lock);
1119 LCONSOLE(0, "HW CPU cores: %d, npartitions: %d\n",
1120 num_online_cpus(), cfs_cpt_number(cfs_cpt_table));