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).
19 * Copyright (c) 2010, Oracle and/or its affiliates. All rights reserved.
21 * Copyright (c) 2012, 2015 Intel Corporation.
24 * This file is part of Lustre, http://www.lustre.org/
25 * Lustre is a trademark of Sun Microsystems, Inc.
27 * Author: liang@whamcloud.com
30 #define DEBUG_SUBSYSTEM S_LNET
32 #include <linux/cpu.h>
33 #include <linux/sched.h>
34 #include "../../../include/linux/libcfs/libcfs.h"
39 * modparam for setting number of partitions
41 * 0 : estimate best value based on cores or NUMA nodes
42 * 1 : disable multiple partitions
43 * >1 : specify number of partitions
45 static int cpu_npartitions;
46 module_param(cpu_npartitions, int, 0444);
47 MODULE_PARM_DESC(cpu_npartitions, "# of CPU partitions");
50 * modparam for setting CPU partitions patterns:
52 * i.e: "0[0,1,2,3] 1[4,5,6,7]", number before bracket is CPU partition ID,
53 * number in bracket is processor ID (core or HT)
55 * i.e: "N 0[0,1] 1[2,3]" the first character 'N' means numbers in bracket
56 * are NUMA node ID, number before bracket is CPU partition ID.
58 * NB: If user specified cpu_pattern, cpu_npartitions will be ignored
60 static char *cpu_pattern = "";
61 module_param(cpu_pattern, charp, 0444);
62 MODULE_PARM_DESC(cpu_pattern, "CPU partitions pattern");
65 /* serialize hotplug etc */
67 /* reserved for hotplug */
68 unsigned long cpt_version;
69 /* mutex to protect cpt_cpumask */
70 struct mutex cpt_mutex;
71 /* scratch buffer for set/unset_node */
72 cpumask_t *cpt_cpumask;
75 static struct cfs_cpt_data cpt_data;
78 cfs_node_to_cpumask(int node, cpumask_t *mask)
80 const cpumask_t *tmp = cpumask_of_node(node);
83 cpumask_copy(mask, tmp);
89 cfs_cpt_table_free(struct cfs_cpt_table *cptab)
93 if (cptab->ctb_cpu2cpt) {
94 LIBCFS_FREE(cptab->ctb_cpu2cpt,
96 sizeof(cptab->ctb_cpu2cpt[0]));
99 for (i = 0; cptab->ctb_parts && i < cptab->ctb_nparts; i++) {
100 struct cfs_cpu_partition *part = &cptab->ctb_parts[i];
102 if (part->cpt_nodemask) {
103 LIBCFS_FREE(part->cpt_nodemask,
104 sizeof(*part->cpt_nodemask));
107 if (part->cpt_cpumask)
108 LIBCFS_FREE(part->cpt_cpumask, cpumask_size());
111 if (cptab->ctb_parts) {
112 LIBCFS_FREE(cptab->ctb_parts,
113 cptab->ctb_nparts * sizeof(cptab->ctb_parts[0]));
116 if (cptab->ctb_nodemask)
117 LIBCFS_FREE(cptab->ctb_nodemask, sizeof(*cptab->ctb_nodemask));
118 if (cptab->ctb_cpumask)
119 LIBCFS_FREE(cptab->ctb_cpumask, cpumask_size());
121 LIBCFS_FREE(cptab, sizeof(*cptab));
123 EXPORT_SYMBOL(cfs_cpt_table_free);
125 struct cfs_cpt_table *
126 cfs_cpt_table_alloc(unsigned int ncpt)
128 struct cfs_cpt_table *cptab;
131 LIBCFS_ALLOC(cptab, sizeof(*cptab));
135 cptab->ctb_nparts = ncpt;
137 LIBCFS_ALLOC(cptab->ctb_cpumask, cpumask_size());
138 LIBCFS_ALLOC(cptab->ctb_nodemask, sizeof(*cptab->ctb_nodemask));
140 if (!cptab->ctb_cpumask || !cptab->ctb_nodemask)
143 LIBCFS_ALLOC(cptab->ctb_cpu2cpt,
144 num_possible_cpus() * sizeof(cptab->ctb_cpu2cpt[0]));
145 if (!cptab->ctb_cpu2cpt)
148 memset(cptab->ctb_cpu2cpt, -1,
149 num_possible_cpus() * sizeof(cptab->ctb_cpu2cpt[0]));
151 LIBCFS_ALLOC(cptab->ctb_parts, ncpt * sizeof(cptab->ctb_parts[0]));
152 if (!cptab->ctb_parts)
155 for (i = 0; i < ncpt; i++) {
156 struct cfs_cpu_partition *part = &cptab->ctb_parts[i];
158 LIBCFS_ALLOC(part->cpt_cpumask, cpumask_size());
159 LIBCFS_ALLOC(part->cpt_nodemask, sizeof(*part->cpt_nodemask));
160 if (!part->cpt_cpumask || !part->cpt_nodemask)
164 spin_lock(&cpt_data.cpt_lock);
165 /* Reserved for hotplug */
166 cptab->ctb_version = cpt_data.cpt_version;
167 spin_unlock(&cpt_data.cpt_lock);
172 cfs_cpt_table_free(cptab);
175 EXPORT_SYMBOL(cfs_cpt_table_alloc);
178 cfs_cpt_table_print(struct cfs_cpt_table *cptab, char *buf, int len)
185 for (i = 0; i < cptab->ctb_nparts; i++) {
187 rc = snprintf(tmp, len, "%d\t: ", i);
197 for_each_cpu(j, cptab->ctb_parts[i].cpt_cpumask) {
198 rc = snprintf(tmp, len, "%d ", j);
218 EXPORT_SYMBOL(cfs_cpt_table_print);
221 cfs_cpt_number(struct cfs_cpt_table *cptab)
223 return cptab->ctb_nparts;
225 EXPORT_SYMBOL(cfs_cpt_number);
228 cfs_cpt_weight(struct cfs_cpt_table *cptab, int cpt)
230 LASSERT(cpt == CFS_CPT_ANY || (cpt >= 0 && cpt < cptab->ctb_nparts));
232 return cpt == CFS_CPT_ANY ?
233 cpumask_weight(cptab->ctb_cpumask) :
234 cpumask_weight(cptab->ctb_parts[cpt].cpt_cpumask);
236 EXPORT_SYMBOL(cfs_cpt_weight);
239 cfs_cpt_online(struct cfs_cpt_table *cptab, int cpt)
241 LASSERT(cpt == CFS_CPT_ANY || (cpt >= 0 && cpt < cptab->ctb_nparts));
243 return cpt == CFS_CPT_ANY ?
244 cpumask_any_and(cptab->ctb_cpumask,
245 cpu_online_mask) < nr_cpu_ids :
246 cpumask_any_and(cptab->ctb_parts[cpt].cpt_cpumask,
247 cpu_online_mask) < nr_cpu_ids;
249 EXPORT_SYMBOL(cfs_cpt_online);
252 cfs_cpt_cpumask(struct cfs_cpt_table *cptab, int cpt)
254 LASSERT(cpt == CFS_CPT_ANY || (cpt >= 0 && cpt < cptab->ctb_nparts));
256 return cpt == CFS_CPT_ANY ?
257 cptab->ctb_cpumask : cptab->ctb_parts[cpt].cpt_cpumask;
259 EXPORT_SYMBOL(cfs_cpt_cpumask);
262 cfs_cpt_nodemask(struct cfs_cpt_table *cptab, int cpt)
264 LASSERT(cpt == CFS_CPT_ANY || (cpt >= 0 && cpt < cptab->ctb_nparts));
266 return cpt == CFS_CPT_ANY ?
267 cptab->ctb_nodemask : cptab->ctb_parts[cpt].cpt_nodemask;
269 EXPORT_SYMBOL(cfs_cpt_nodemask);
272 cfs_cpt_set_cpu(struct cfs_cpt_table *cptab, int cpt, int cpu)
276 LASSERT(cpt >= 0 && cpt < cptab->ctb_nparts);
278 if (cpu < 0 || cpu >= nr_cpu_ids || !cpu_online(cpu)) {
279 CDEBUG(D_INFO, "CPU %d is invalid or it's offline\n", cpu);
283 if (cptab->ctb_cpu2cpt[cpu] != -1) {
284 CDEBUG(D_INFO, "CPU %d is already in partition %d\n",
285 cpu, cptab->ctb_cpu2cpt[cpu]);
289 cptab->ctb_cpu2cpt[cpu] = cpt;
291 LASSERT(!cpumask_test_cpu(cpu, cptab->ctb_cpumask));
292 LASSERT(!cpumask_test_cpu(cpu, cptab->ctb_parts[cpt].cpt_cpumask));
294 cpumask_set_cpu(cpu, cptab->ctb_cpumask);
295 cpumask_set_cpu(cpu, cptab->ctb_parts[cpt].cpt_cpumask);
297 node = cpu_to_node(cpu);
299 /* first CPU of @node in this CPT table */
300 if (!node_isset(node, *cptab->ctb_nodemask))
301 node_set(node, *cptab->ctb_nodemask);
303 /* first CPU of @node in this partition */
304 if (!node_isset(node, *cptab->ctb_parts[cpt].cpt_nodemask))
305 node_set(node, *cptab->ctb_parts[cpt].cpt_nodemask);
309 EXPORT_SYMBOL(cfs_cpt_set_cpu);
312 cfs_cpt_unset_cpu(struct cfs_cpt_table *cptab, int cpt, int cpu)
317 LASSERT(cpt == CFS_CPT_ANY || (cpt >= 0 && cpt < cptab->ctb_nparts));
319 if (cpu < 0 || cpu >= nr_cpu_ids) {
320 CDEBUG(D_INFO, "Invalid CPU id %d\n", cpu);
324 if (cpt == CFS_CPT_ANY) {
325 /* caller doesn't know the partition ID */
326 cpt = cptab->ctb_cpu2cpt[cpu];
327 if (cpt < 0) { /* not set in this CPT-table */
328 CDEBUG(D_INFO, "Try to unset cpu %d which is not in CPT-table %p\n",
333 } else if (cpt != cptab->ctb_cpu2cpt[cpu]) {
335 "CPU %d is not in cpu-partition %d\n", cpu, cpt);
339 LASSERT(cpumask_test_cpu(cpu, cptab->ctb_parts[cpt].cpt_cpumask));
340 LASSERT(cpumask_test_cpu(cpu, cptab->ctb_cpumask));
342 cpumask_clear_cpu(cpu, cptab->ctb_parts[cpt].cpt_cpumask);
343 cpumask_clear_cpu(cpu, cptab->ctb_cpumask);
344 cptab->ctb_cpu2cpt[cpu] = -1;
346 node = cpu_to_node(cpu);
348 LASSERT(node_isset(node, *cptab->ctb_parts[cpt].cpt_nodemask));
349 LASSERT(node_isset(node, *cptab->ctb_nodemask));
351 for_each_cpu(i, cptab->ctb_parts[cpt].cpt_cpumask) {
352 /* this CPT has other CPU belonging to this node? */
353 if (cpu_to_node(i) == node)
358 node_clear(node, *cptab->ctb_parts[cpt].cpt_nodemask);
360 for_each_cpu(i, cptab->ctb_cpumask) {
361 /* this CPT-table has other CPU belonging to this node? */
362 if (cpu_to_node(i) == node)
367 node_clear(node, *cptab->ctb_nodemask);
369 EXPORT_SYMBOL(cfs_cpt_unset_cpu);
372 cfs_cpt_set_cpumask(struct cfs_cpt_table *cptab, int cpt, cpumask_t *mask)
376 if (cpumask_weight(mask) == 0 ||
377 cpumask_any_and(mask, cpu_online_mask) >= nr_cpu_ids) {
378 CDEBUG(D_INFO, "No online CPU is found in the CPU mask for CPU partition %d\n",
383 for_each_cpu(i, mask) {
384 if (!cfs_cpt_set_cpu(cptab, cpt, i))
390 EXPORT_SYMBOL(cfs_cpt_set_cpumask);
393 cfs_cpt_unset_cpumask(struct cfs_cpt_table *cptab, int cpt, cpumask_t *mask)
397 for_each_cpu(i, mask)
398 cfs_cpt_unset_cpu(cptab, cpt, i);
400 EXPORT_SYMBOL(cfs_cpt_unset_cpumask);
403 cfs_cpt_set_node(struct cfs_cpt_table *cptab, int cpt, int node)
408 if (node < 0 || node >= MAX_NUMNODES) {
410 "Invalid NUMA id %d for CPU partition %d\n", node, cpt);
414 mutex_lock(&cpt_data.cpt_mutex);
416 mask = cpt_data.cpt_cpumask;
417 cfs_node_to_cpumask(node, mask);
419 rc = cfs_cpt_set_cpumask(cptab, cpt, mask);
421 mutex_unlock(&cpt_data.cpt_mutex);
425 EXPORT_SYMBOL(cfs_cpt_set_node);
428 cfs_cpt_unset_node(struct cfs_cpt_table *cptab, int cpt, int node)
432 if (node < 0 || node >= MAX_NUMNODES) {
434 "Invalid NUMA id %d for CPU partition %d\n", node, cpt);
438 mutex_lock(&cpt_data.cpt_mutex);
440 mask = cpt_data.cpt_cpumask;
441 cfs_node_to_cpumask(node, mask);
443 cfs_cpt_unset_cpumask(cptab, cpt, mask);
445 mutex_unlock(&cpt_data.cpt_mutex);
447 EXPORT_SYMBOL(cfs_cpt_unset_node);
450 cfs_cpt_set_nodemask(struct cfs_cpt_table *cptab, int cpt, nodemask_t *mask)
454 for_each_node_mask(i, *mask) {
455 if (!cfs_cpt_set_node(cptab, cpt, i))
461 EXPORT_SYMBOL(cfs_cpt_set_nodemask);
464 cfs_cpt_unset_nodemask(struct cfs_cpt_table *cptab, int cpt, nodemask_t *mask)
468 for_each_node_mask(i, *mask)
469 cfs_cpt_unset_node(cptab, cpt, i);
471 EXPORT_SYMBOL(cfs_cpt_unset_nodemask);
474 cfs_cpt_clear(struct cfs_cpt_table *cptab, int cpt)
479 if (cpt == CFS_CPT_ANY) {
480 last = cptab->ctb_nparts - 1;
486 for (; cpt <= last; cpt++) {
487 for_each_cpu(i, cptab->ctb_parts[cpt].cpt_cpumask)
488 cfs_cpt_unset_cpu(cptab, cpt, i);
491 EXPORT_SYMBOL(cfs_cpt_clear);
494 cfs_cpt_spread_node(struct cfs_cpt_table *cptab, int cpt)
501 /* convert CPU partition ID to HW node id */
503 if (cpt < 0 || cpt >= cptab->ctb_nparts) {
504 mask = cptab->ctb_nodemask;
505 rotor = cptab->ctb_spread_rotor++;
507 mask = cptab->ctb_parts[cpt].cpt_nodemask;
508 rotor = cptab->ctb_parts[cpt].cpt_spread_rotor++;
511 weight = nodes_weight(*mask);
516 for_each_node_mask(node, *mask) {
524 EXPORT_SYMBOL(cfs_cpt_spread_node);
527 cfs_cpt_current(struct cfs_cpt_table *cptab, int remap)
529 int cpu = smp_processor_id();
530 int cpt = cptab->ctb_cpu2cpt[cpu];
536 /* don't return negative value for safety of upper layer,
537 * instead we shadow the unknown cpu to a valid partition ID
539 cpt = cpu % cptab->ctb_nparts;
544 EXPORT_SYMBOL(cfs_cpt_current);
547 cfs_cpt_of_cpu(struct cfs_cpt_table *cptab, int cpu)
549 LASSERT(cpu >= 0 && cpu < nr_cpu_ids);
551 return cptab->ctb_cpu2cpt[cpu];
553 EXPORT_SYMBOL(cfs_cpt_of_cpu);
556 cfs_cpt_bind(struct cfs_cpt_table *cptab, int cpt)
559 nodemask_t *nodemask;
563 LASSERT(cpt == CFS_CPT_ANY || (cpt >= 0 && cpt < cptab->ctb_nparts));
565 if (cpt == CFS_CPT_ANY) {
566 cpumask = cptab->ctb_cpumask;
567 nodemask = cptab->ctb_nodemask;
569 cpumask = cptab->ctb_parts[cpt].cpt_cpumask;
570 nodemask = cptab->ctb_parts[cpt].cpt_nodemask;
573 if (cpumask_any_and(cpumask, cpu_online_mask) >= nr_cpu_ids) {
574 CERROR("No online CPU found in CPU partition %d, did someone do CPU hotplug on system? You might need to reload Lustre modules to keep system working well.\n",
579 for_each_online_cpu(i) {
580 if (cpumask_test_cpu(i, cpumask))
583 rc = set_cpus_allowed_ptr(current, cpumask);
584 set_mems_allowed(*nodemask);
586 schedule(); /* switch to allowed CPU */
591 /* don't need to set affinity because all online CPUs are covered */
594 EXPORT_SYMBOL(cfs_cpt_bind);
597 * Choose max to \a number CPUs from \a node and set them in \a cpt.
598 * We always prefer to choose CPU in the same core/socket.
601 cfs_cpt_choose_ncpus(struct cfs_cpt_table *cptab, int cpt,
602 cpumask_t *node, int number)
604 cpumask_t *socket = NULL;
605 cpumask_t *core = NULL;
611 if (number >= cpumask_weight(node)) {
612 while (!cpumask_empty(node)) {
613 cpu = cpumask_first(node);
615 rc = cfs_cpt_set_cpu(cptab, cpt, cpu);
618 cpumask_clear_cpu(cpu, node);
623 /* allocate scratch buffer */
624 LIBCFS_ALLOC(socket, cpumask_size());
625 LIBCFS_ALLOC(core, cpumask_size());
626 if (!socket || !core) {
631 while (!cpumask_empty(node)) {
632 cpu = cpumask_first(node);
634 /* get cpumask for cores in the same socket */
635 cpumask_copy(socket, topology_core_cpumask(cpu));
636 cpumask_and(socket, socket, node);
638 LASSERT(!cpumask_empty(socket));
640 while (!cpumask_empty(socket)) {
643 /* get cpumask for hts in the same core */
644 cpumask_copy(core, topology_sibling_cpumask(cpu));
645 cpumask_and(core, core, node);
647 LASSERT(!cpumask_empty(core));
649 for_each_cpu(i, core) {
650 cpumask_clear_cpu(i, socket);
651 cpumask_clear_cpu(i, node);
653 rc = cfs_cpt_set_cpu(cptab, cpt, i);
662 cpu = cpumask_first(socket);
668 LIBCFS_FREE(socket, cpumask_size());
670 LIBCFS_FREE(core, cpumask_size());
674 #define CPT_WEIGHT_MIN 4u
677 cfs_cpt_num_estimate(void)
679 unsigned nnode = num_online_nodes();
680 unsigned ncpu = num_online_cpus();
683 if (ncpu <= CPT_WEIGHT_MIN) {
688 /* generate reasonable number of CPU partitions based on total number
689 * of CPUs, Preferred N should be power2 and match this condition:
690 * 2 * (N - 1)^2 < NCPUS <= 2 * N^2
692 for (ncpt = 2; ncpu > 2 * ncpt * ncpt; ncpt <<= 1)
695 if (ncpt <= nnode) { /* fat numa system */
699 } else { /* ncpt > nnode */
700 while ((nnode << 1) <= ncpt)
707 #if (BITS_PER_LONG == 32)
708 /* config many CPU partitions on 32-bit system could consume
711 ncpt = min(2U, ncpt);
713 while (ncpu % ncpt != 0)
714 ncpt--; /* worst case is 1 */
719 static struct cfs_cpt_table *
720 cfs_cpt_table_create(int ncpt)
722 struct cfs_cpt_table *cptab = NULL;
723 cpumask_t *mask = NULL;
729 rc = cfs_cpt_num_estimate();
733 if (ncpt > num_online_cpus() || ncpt > 4 * rc) {
734 CWARN("CPU partition number %d is larger than suggested value (%d), your system may have performance issue or run out of memory while under pressure\n",
738 if (num_online_cpus() % ncpt != 0) {
739 CERROR("CPU number %d is not multiple of cpu_npartition %d, please try different cpu_npartitions value or set pattern string by cpu_pattern=STRING\n",
740 (int)num_online_cpus(), ncpt);
744 cptab = cfs_cpt_table_alloc(ncpt);
746 CERROR("Failed to allocate CPU map(%d)\n", ncpt);
750 num = num_online_cpus() / ncpt;
752 CERROR("CPU changed while setting CPU partition\n");
756 LIBCFS_ALLOC(mask, cpumask_size());
758 CERROR("Failed to allocate scratch cpumask\n");
762 for_each_online_node(i) {
763 cfs_node_to_cpumask(i, mask);
765 while (!cpumask_empty(mask)) {
766 struct cfs_cpu_partition *part;
770 * Each emulated NUMA node has all allowed CPUs in
772 * End loop when all partitions have assigned CPUs.
777 part = &cptab->ctb_parts[cpt];
779 n = num - cpumask_weight(part->cpt_cpumask);
782 rc = cfs_cpt_choose_ncpus(cptab, cpt, mask, n);
786 LASSERT(num >= cpumask_weight(part->cpt_cpumask));
787 if (num == cpumask_weight(part->cpt_cpumask))
793 num != cpumask_weight(cptab->ctb_parts[ncpt - 1].cpt_cpumask)) {
794 CERROR("Expect %d(%d) CPU partitions but got %d(%d), CPU hotplug/unplug while setting?\n",
795 cptab->ctb_nparts, num, cpt,
796 cpumask_weight(cptab->ctb_parts[ncpt - 1].cpt_cpumask));
800 LIBCFS_FREE(mask, cpumask_size());
805 CERROR("Failed to setup CPU-partition-table with %d CPU-partitions, online HW nodes: %d, HW cpus: %d.\n",
806 ncpt, num_online_nodes(), num_online_cpus());
809 LIBCFS_FREE(mask, cpumask_size());
812 cfs_cpt_table_free(cptab);
817 static struct cfs_cpt_table *
818 cfs_cpt_table_create_pattern(char *pattern)
820 struct cfs_cpt_table *cptab;
827 for (ncpt = 0;; ncpt++) { /* quick scan bracket */
828 str = strchr(str, '[');
834 str = cfs_trimwhite(pattern);
835 if (*str == 'n' || *str == 'N') {
841 (node && ncpt > num_online_nodes()) ||
842 (!node && ncpt > num_online_cpus())) {
843 CERROR("Invalid pattern %s, or too many partitions %d\n",
848 high = node ? MAX_NUMNODES - 1 : nr_cpu_ids - 1;
850 cptab = cfs_cpt_table_alloc(ncpt);
852 CERROR("Failed to allocate cpu partition table\n");
856 for (str = cfs_trimwhite(pattern), c = 0;; c++) {
857 struct cfs_range_expr *range;
858 struct cfs_expr_list *el;
859 char *bracket = strchr(str, '[');
867 CERROR("Invalid pattern %s\n", str);
871 CERROR("expect %d partitions but found %d\n",
878 if (sscanf(str, "%d%n", &cpt, &n) < 1) {
879 CERROR("Invalid cpu pattern %s\n", str);
883 if (cpt < 0 || cpt >= ncpt) {
884 CERROR("Invalid partition id %d, total partitions %d\n",
889 if (cfs_cpt_weight(cptab, cpt) != 0) {
890 CERROR("Partition %d has already been set.\n", cpt);
894 str = cfs_trimwhite(str + n);
895 if (str != bracket) {
896 CERROR("Invalid pattern %s\n", str);
900 bracket = strchr(str, ']');
902 CERROR("missing right bracket for cpt %d, %s\n",
907 if (cfs_expr_list_parse(str, (bracket - str) + 1,
908 0, high, &el) != 0) {
909 CERROR("Can't parse number range: %s\n", str);
913 list_for_each_entry(range, &el->el_exprs, re_link) {
914 for (i = range->re_lo; i <= range->re_hi; i++) {
915 if ((i - range->re_lo) % range->re_stride != 0)
918 rc = node ? cfs_cpt_set_node(cptab, cpt, i) :
919 cfs_cpt_set_cpu(cptab, cpt, i);
921 cfs_expr_list_free(el);
927 cfs_expr_list_free(el);
929 if (!cfs_cpt_online(cptab, cpt)) {
930 CERROR("No online CPU is found on partition %d\n", cpt);
934 str = cfs_trimwhite(bracket + 1);
940 cfs_cpt_table_free(cptab);
944 #ifdef CONFIG_HOTPLUG_CPU
946 cfs_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
948 unsigned int cpu = (unsigned long)hcpu;
953 case CPU_DEAD_FROZEN:
955 case CPU_ONLINE_FROZEN:
956 spin_lock(&cpt_data.cpt_lock);
957 cpt_data.cpt_version++;
958 spin_unlock(&cpt_data.cpt_lock);
961 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN) {
962 CDEBUG(D_INFO, "CPU changed [cpu %u action %lx]\n",
967 mutex_lock(&cpt_data.cpt_mutex);
968 /* if all HTs in a core are offline, it may break affinity */
969 cpumask_copy(cpt_data.cpt_cpumask,
970 topology_sibling_cpumask(cpu));
971 warn = cpumask_any_and(cpt_data.cpt_cpumask,
972 cpu_online_mask) >= nr_cpu_ids;
973 mutex_unlock(&cpt_data.cpt_mutex);
974 CDEBUG(warn ? D_WARNING : D_INFO,
975 "Lustre: can't support CPU plug-out well now, performance and stability could be impacted [CPU %u action: %lx]\n",
982 static struct notifier_block cfs_cpu_notifier = {
983 .notifier_call = cfs_cpu_notify,
993 cfs_cpt_table_free(cfs_cpt_table);
995 #ifdef CONFIG_HOTPLUG_CPU
996 unregister_hotcpu_notifier(&cfs_cpu_notifier);
998 if (cpt_data.cpt_cpumask)
999 LIBCFS_FREE(cpt_data.cpt_cpumask, cpumask_size());
1005 LASSERT(!cfs_cpt_table);
1007 memset(&cpt_data, 0, sizeof(cpt_data));
1009 LIBCFS_ALLOC(cpt_data.cpt_cpumask, cpumask_size());
1010 if (!cpt_data.cpt_cpumask) {
1011 CERROR("Failed to allocate scratch buffer\n");
1015 spin_lock_init(&cpt_data.cpt_lock);
1016 mutex_init(&cpt_data.cpt_mutex);
1018 #ifdef CONFIG_HOTPLUG_CPU
1019 register_hotcpu_notifier(&cfs_cpu_notifier);
1022 if (*cpu_pattern != 0) {
1023 cfs_cpt_table = cfs_cpt_table_create_pattern(cpu_pattern);
1024 if (!cfs_cpt_table) {
1025 CERROR("Failed to create cptab from pattern %s\n",
1031 cfs_cpt_table = cfs_cpt_table_create(cpu_npartitions);
1032 if (!cfs_cpt_table) {
1033 CERROR("Failed to create ptable with npartitions %d\n",
1039 spin_lock(&cpt_data.cpt_lock);
1040 if (cfs_cpt_table->ctb_version != cpt_data.cpt_version) {
1041 spin_unlock(&cpt_data.cpt_lock);
1042 CERROR("CPU hotplug/unplug during setup\n");
1045 spin_unlock(&cpt_data.cpt_lock);
1047 LCONSOLE(0, "HW CPU cores: %d, npartitions: %d\n",
1048 num_online_cpus(), cfs_cpt_number(cfs_cpt_table));