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.
25 * Copyright (c) 2012, Intel Corporation.
28 * This file is part of Lustre, http://www.lustre.org/
29 * Lustre is a trademark of Sun Microsystems, Inc.
31 * Author: liang@whamcloud.com
34 #define DEBUG_SUBSYSTEM S_LNET
36 #include <linux/cpu.h>
37 #include <linux/sched.h>
38 #include "../../../include/linux/libcfs/libcfs.h"
43 * modparam for setting number of partitions
45 * 0 : estimate best value based on cores or NUMA nodes
46 * 1 : disable multiple partitions
47 * >1 : specify number of partitions
49 static int cpu_npartitions;
50 module_param(cpu_npartitions, int, 0444);
51 MODULE_PARM_DESC(cpu_npartitions, "# 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 * NB: If user specified cpu_pattern, cpu_npartitions will be ignored
64 static char *cpu_pattern = "";
65 module_param(cpu_pattern, charp, 0444);
66 MODULE_PARM_DESC(cpu_pattern, "CPU partitions pattern");
69 /* serialize hotplug etc */
71 /* reserved for hotplug */
72 unsigned long cpt_version;
73 /* mutex to protect cpt_cpumask */
74 struct mutex cpt_mutex;
75 /* scratch buffer for set/unset_node */
76 cpumask_t *cpt_cpumask;
79 static struct cfs_cpt_data cpt_data;
82 cfs_cpt_table_free(struct cfs_cpt_table *cptab)
86 if (cptab->ctb_cpu2cpt != NULL) {
87 LIBCFS_FREE(cptab->ctb_cpu2cpt,
89 sizeof(cptab->ctb_cpu2cpt[0]));
92 for (i = 0; cptab->ctb_parts != NULL && i < cptab->ctb_nparts; i++) {
93 struct cfs_cpu_partition *part = &cptab->ctb_parts[i];
95 if (part->cpt_nodemask != NULL) {
96 LIBCFS_FREE(part->cpt_nodemask,
97 sizeof(*part->cpt_nodemask));
100 if (part->cpt_cpumask != NULL)
101 LIBCFS_FREE(part->cpt_cpumask, cpumask_size());
104 if (cptab->ctb_parts != NULL) {
105 LIBCFS_FREE(cptab->ctb_parts,
106 cptab->ctb_nparts * sizeof(cptab->ctb_parts[0]));
109 if (cptab->ctb_nodemask != NULL)
110 LIBCFS_FREE(cptab->ctb_nodemask, sizeof(*cptab->ctb_nodemask));
111 if (cptab->ctb_cpumask != NULL)
112 LIBCFS_FREE(cptab->ctb_cpumask, cpumask_size());
114 LIBCFS_FREE(cptab, sizeof(*cptab));
116 EXPORT_SYMBOL(cfs_cpt_table_free);
118 struct cfs_cpt_table *
119 cfs_cpt_table_alloc(unsigned int ncpt)
121 struct cfs_cpt_table *cptab;
124 LIBCFS_ALLOC(cptab, sizeof(*cptab));
128 cptab->ctb_nparts = ncpt;
130 LIBCFS_ALLOC(cptab->ctb_cpumask, cpumask_size());
131 LIBCFS_ALLOC(cptab->ctb_nodemask, sizeof(*cptab->ctb_nodemask));
133 if (cptab->ctb_cpumask == NULL || cptab->ctb_nodemask == NULL)
136 LIBCFS_ALLOC(cptab->ctb_cpu2cpt,
137 num_possible_cpus() * sizeof(cptab->ctb_cpu2cpt[0]));
138 if (cptab->ctb_cpu2cpt == NULL)
141 memset(cptab->ctb_cpu2cpt, -1,
142 num_possible_cpus() * sizeof(cptab->ctb_cpu2cpt[0]));
144 LIBCFS_ALLOC(cptab->ctb_parts, ncpt * sizeof(cptab->ctb_parts[0]));
145 if (cptab->ctb_parts == NULL)
148 for (i = 0; i < ncpt; i++) {
149 struct cfs_cpu_partition *part = &cptab->ctb_parts[i];
151 LIBCFS_ALLOC(part->cpt_cpumask, cpumask_size());
152 LIBCFS_ALLOC(part->cpt_nodemask, sizeof(*part->cpt_nodemask));
153 if (part->cpt_cpumask == NULL || part->cpt_nodemask == NULL)
157 spin_lock(&cpt_data.cpt_lock);
158 /* Reserved for hotplug */
159 cptab->ctb_version = cpt_data.cpt_version;
160 spin_unlock(&cpt_data.cpt_lock);
165 cfs_cpt_table_free(cptab);
168 EXPORT_SYMBOL(cfs_cpt_table_alloc);
171 cfs_cpt_table_print(struct cfs_cpt_table *cptab, char *buf, int len)
178 for (i = 0; i < cptab->ctb_nparts; i++) {
180 rc = snprintf(tmp, len, "%d\t: ", i);
190 for_each_cpu(j, cptab->ctb_parts[i].cpt_cpumask) {
191 rc = snprintf(tmp, len, "%d ", j);
211 EXPORT_SYMBOL(cfs_cpt_table_print);
214 cfs_cpt_number(struct cfs_cpt_table *cptab)
216 return cptab->ctb_nparts;
218 EXPORT_SYMBOL(cfs_cpt_number);
221 cfs_cpt_weight(struct cfs_cpt_table *cptab, int cpt)
223 LASSERT(cpt == CFS_CPT_ANY || (cpt >= 0 && cpt < cptab->ctb_nparts));
225 return cpt == CFS_CPT_ANY ?
226 cpumask_weight(cptab->ctb_cpumask) :
227 cpumask_weight(cptab->ctb_parts[cpt].cpt_cpumask);
229 EXPORT_SYMBOL(cfs_cpt_weight);
232 cfs_cpt_online(struct cfs_cpt_table *cptab, int cpt)
234 LASSERT(cpt == CFS_CPT_ANY || (cpt >= 0 && cpt < cptab->ctb_nparts));
236 return cpt == CFS_CPT_ANY ?
237 cpumask_any_and(cptab->ctb_cpumask,
238 cpu_online_mask) < nr_cpu_ids :
239 cpumask_any_and(cptab->ctb_parts[cpt].cpt_cpumask,
240 cpu_online_mask) < nr_cpu_ids;
242 EXPORT_SYMBOL(cfs_cpt_online);
245 cfs_cpt_cpumask(struct cfs_cpt_table *cptab, int cpt)
247 LASSERT(cpt == CFS_CPT_ANY || (cpt >= 0 && cpt < cptab->ctb_nparts));
249 return cpt == CFS_CPT_ANY ?
250 cptab->ctb_cpumask : cptab->ctb_parts[cpt].cpt_cpumask;
252 EXPORT_SYMBOL(cfs_cpt_cpumask);
255 cfs_cpt_nodemask(struct cfs_cpt_table *cptab, int cpt)
257 LASSERT(cpt == CFS_CPT_ANY || (cpt >= 0 && cpt < cptab->ctb_nparts));
259 return cpt == CFS_CPT_ANY ?
260 cptab->ctb_nodemask : cptab->ctb_parts[cpt].cpt_nodemask;
262 EXPORT_SYMBOL(cfs_cpt_nodemask);
265 cfs_cpt_set_cpu(struct cfs_cpt_table *cptab, int cpt, int cpu)
269 LASSERT(cpt >= 0 && cpt < cptab->ctb_nparts);
271 if (cpu < 0 || cpu >= nr_cpu_ids || !cpu_online(cpu)) {
272 CDEBUG(D_INFO, "CPU %d is invalid or it's offline\n", cpu);
276 if (cptab->ctb_cpu2cpt[cpu] != -1) {
277 CDEBUG(D_INFO, "CPU %d is already in partition %d\n",
278 cpu, cptab->ctb_cpu2cpt[cpu]);
282 cptab->ctb_cpu2cpt[cpu] = cpt;
284 LASSERT(!cpumask_test_cpu(cpu, cptab->ctb_cpumask));
285 LASSERT(!cpumask_test_cpu(cpu, cptab->ctb_parts[cpt].cpt_cpumask));
287 cpumask_set_cpu(cpu, cptab->ctb_cpumask);
288 cpumask_set_cpu(cpu, cptab->ctb_parts[cpt].cpt_cpumask);
290 node = cpu_to_node(cpu);
292 /* first CPU of @node in this CPT table */
293 if (!node_isset(node, *cptab->ctb_nodemask))
294 node_set(node, *cptab->ctb_nodemask);
296 /* first CPU of @node in this partition */
297 if (!node_isset(node, *cptab->ctb_parts[cpt].cpt_nodemask))
298 node_set(node, *cptab->ctb_parts[cpt].cpt_nodemask);
302 EXPORT_SYMBOL(cfs_cpt_set_cpu);
305 cfs_cpt_unset_cpu(struct cfs_cpt_table *cptab, int cpt, int cpu)
310 LASSERT(cpt == CFS_CPT_ANY || (cpt >= 0 && cpt < cptab->ctb_nparts));
312 if (cpu < 0 || cpu >= nr_cpu_ids) {
313 CDEBUG(D_INFO, "Invalid CPU id %d\n", cpu);
317 if (cpt == CFS_CPT_ANY) {
318 /* caller doesn't know the partition ID */
319 cpt = cptab->ctb_cpu2cpt[cpu];
320 if (cpt < 0) { /* not set in this CPT-table */
321 CDEBUG(D_INFO, "Try to unset cpu %d which is not in CPT-table %p\n",
326 } else if (cpt != cptab->ctb_cpu2cpt[cpu]) {
328 "CPU %d is not in cpu-partition %d\n", cpu, cpt);
332 LASSERT(cpumask_test_cpu(cpu, cptab->ctb_parts[cpt].cpt_cpumask));
333 LASSERT(cpumask_test_cpu(cpu, cptab->ctb_cpumask));
335 cpumask_clear_cpu(cpu, cptab->ctb_parts[cpt].cpt_cpumask);
336 cpumask_clear_cpu(cpu, cptab->ctb_cpumask);
337 cptab->ctb_cpu2cpt[cpu] = -1;
339 node = cpu_to_node(cpu);
341 LASSERT(node_isset(node, *cptab->ctb_parts[cpt].cpt_nodemask));
342 LASSERT(node_isset(node, *cptab->ctb_nodemask));
344 for_each_cpu(i, cptab->ctb_parts[cpt].cpt_cpumask) {
345 /* this CPT has other CPU belonging to this node? */
346 if (cpu_to_node(i) == node)
351 node_clear(node, *cptab->ctb_parts[cpt].cpt_nodemask);
353 for_each_cpu(i, cptab->ctb_cpumask) {
354 /* this CPT-table has other CPU belonging to this node? */
355 if (cpu_to_node(i) == node)
360 node_clear(node, *cptab->ctb_nodemask);
364 EXPORT_SYMBOL(cfs_cpt_unset_cpu);
367 cfs_cpt_set_cpumask(struct cfs_cpt_table *cptab, int cpt, cpumask_t *mask)
371 if (cpumask_weight(mask) == 0 ||
372 cpumask_any_and(mask, cpu_online_mask) >= nr_cpu_ids) {
373 CDEBUG(D_INFO, "No online CPU is found in the CPU mask for CPU partition %d\n",
378 for_each_cpu(i, mask) {
379 if (!cfs_cpt_set_cpu(cptab, cpt, i))
385 EXPORT_SYMBOL(cfs_cpt_set_cpumask);
388 cfs_cpt_unset_cpumask(struct cfs_cpt_table *cptab, int cpt, cpumask_t *mask)
392 for_each_cpu(i, mask)
393 cfs_cpt_unset_cpu(cptab, cpt, i);
395 EXPORT_SYMBOL(cfs_cpt_unset_cpumask);
398 cfs_cpt_set_node(struct cfs_cpt_table *cptab, int cpt, int node)
403 if (node < 0 || node >= MAX_NUMNODES) {
405 "Invalid NUMA id %d for CPU partition %d\n", node, cpt);
409 mutex_lock(&cpt_data.cpt_mutex);
411 mask = cpt_data.cpt_cpumask;
412 cpumask_copy(mask, cpumask_of_node(node));
414 rc = cfs_cpt_set_cpumask(cptab, cpt, mask);
416 mutex_unlock(&cpt_data.cpt_mutex);
420 EXPORT_SYMBOL(cfs_cpt_set_node);
423 cfs_cpt_unset_node(struct cfs_cpt_table *cptab, int cpt, int node)
427 if (node < 0 || node >= MAX_NUMNODES) {
429 "Invalid NUMA id %d for CPU partition %d\n", node, cpt);
433 mutex_lock(&cpt_data.cpt_mutex);
435 mask = cpt_data.cpt_cpumask;
436 cpumask_copy(mask, cpumask_of_node(node));
438 cfs_cpt_unset_cpumask(cptab, cpt, mask);
440 mutex_unlock(&cpt_data.cpt_mutex);
442 EXPORT_SYMBOL(cfs_cpt_unset_node);
445 cfs_cpt_set_nodemask(struct cfs_cpt_table *cptab, int cpt, nodemask_t *mask)
449 for_each_node_mask(i, *mask) {
450 if (!cfs_cpt_set_node(cptab, cpt, i))
456 EXPORT_SYMBOL(cfs_cpt_set_nodemask);
459 cfs_cpt_unset_nodemask(struct cfs_cpt_table *cptab, int cpt, nodemask_t *mask)
463 for_each_node_mask(i, *mask)
464 cfs_cpt_unset_node(cptab, cpt, i);
466 EXPORT_SYMBOL(cfs_cpt_unset_nodemask);
469 cfs_cpt_clear(struct cfs_cpt_table *cptab, int cpt)
474 if (cpt == CFS_CPT_ANY) {
475 last = cptab->ctb_nparts - 1;
481 for (; cpt <= last; cpt++) {
482 for_each_cpu(i, cptab->ctb_parts[cpt].cpt_cpumask)
483 cfs_cpt_unset_cpu(cptab, cpt, i);
486 EXPORT_SYMBOL(cfs_cpt_clear);
489 cfs_cpt_spread_node(struct cfs_cpt_table *cptab, int cpt)
496 /* convert CPU partition ID to HW node id */
498 if (cpt < 0 || cpt >= cptab->ctb_nparts) {
499 mask = cptab->ctb_nodemask;
500 rotor = cptab->ctb_spread_rotor++;
502 mask = cptab->ctb_parts[cpt].cpt_nodemask;
503 rotor = cptab->ctb_parts[cpt].cpt_spread_rotor++;
506 weight = nodes_weight(*mask);
511 for_each_node_mask(node, *mask) {
519 EXPORT_SYMBOL(cfs_cpt_spread_node);
522 cfs_cpt_current(struct cfs_cpt_table *cptab, int remap)
524 int cpu = smp_processor_id();
525 int cpt = cptab->ctb_cpu2cpt[cpu];
531 /* don't return negative value for safety of upper layer,
532 * instead we shadow the unknown cpu to a valid partition ID */
533 cpt = cpu % cptab->ctb_nparts;
538 EXPORT_SYMBOL(cfs_cpt_current);
541 cfs_cpt_of_cpu(struct cfs_cpt_table *cptab, int cpu)
543 LASSERT(cpu >= 0 && cpu < nr_cpu_ids);
545 return cptab->ctb_cpu2cpt[cpu];
547 EXPORT_SYMBOL(cfs_cpt_of_cpu);
550 cfs_cpt_bind(struct cfs_cpt_table *cptab, int cpt)
553 nodemask_t *nodemask;
557 LASSERT(cpt == CFS_CPT_ANY || (cpt >= 0 && cpt < cptab->ctb_nparts));
559 if (cpt == CFS_CPT_ANY) {
560 cpumask = cptab->ctb_cpumask;
561 nodemask = cptab->ctb_nodemask;
563 cpumask = cptab->ctb_parts[cpt].cpt_cpumask;
564 nodemask = cptab->ctb_parts[cpt].cpt_nodemask;
567 if (cpumask_any_and(cpumask, cpu_online_mask) >= nr_cpu_ids) {
568 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",
573 for_each_online_cpu(i) {
574 if (cpumask_test_cpu(i, cpumask))
577 rc = set_cpus_allowed_ptr(current, cpumask);
578 set_mems_allowed(*nodemask);
580 schedule(); /* switch to allowed CPU */
585 /* don't need to set affinity because all online CPUs are covered */
588 EXPORT_SYMBOL(cfs_cpt_bind);
591 * Choose max to \a number CPUs from \a node and set them in \a cpt.
592 * We always prefer to choose CPU in the same core/socket.
595 cfs_cpt_choose_ncpus(struct cfs_cpt_table *cptab, int cpt,
596 cpumask_t *node, int number)
598 cpumask_t *socket = NULL;
599 cpumask_t *core = NULL;
605 if (number >= cpumask_weight(node)) {
606 while (!cpumask_empty(node)) {
607 cpu = cpumask_first(node);
609 rc = cfs_cpt_set_cpu(cptab, cpt, cpu);
612 cpumask_clear_cpu(cpu, node);
617 /* allocate scratch buffer */
618 LIBCFS_ALLOC(socket, cpumask_size());
619 LIBCFS_ALLOC(core, cpumask_size());
620 if (socket == NULL || core == NULL) {
625 while (!cpumask_empty(node)) {
626 cpu = cpumask_first(node);
628 /* get cpumask for cores in the same socket */
629 cpumask_copy(socket, topology_core_cpumask(cpu));
630 cpumask_and(socket, socket, node);
632 LASSERT(!cpumask_empty(socket));
634 while (!cpumask_empty(socket)) {
637 /* get cpumask for hts in the same core */
638 cpumask_copy(core, topology_sibling_cpumask(cpu));
639 cpumask_and(core, core, node);
641 LASSERT(!cpumask_empty(core));
643 for_each_cpu(i, core) {
644 cpumask_clear_cpu(i, socket);
645 cpumask_clear_cpu(i, node);
647 rc = cfs_cpt_set_cpu(cptab, cpt, i);
656 cpu = cpumask_first(socket);
662 LIBCFS_FREE(socket, cpumask_size());
664 LIBCFS_FREE(core, cpumask_size());
668 #define CPT_WEIGHT_MIN 4u
671 cfs_cpt_num_estimate(void)
673 unsigned nnode = num_online_nodes();
674 unsigned ncpu = num_online_cpus();
677 if (ncpu <= CPT_WEIGHT_MIN) {
682 /* generate reasonable number of CPU partitions based on total number
683 * of CPUs, Preferred N should be power2 and match this condition:
684 * 2 * (N - 1)^2 < NCPUS <= 2 * N^2 */
685 for (ncpt = 2; ncpu > 2 * ncpt * ncpt; ncpt <<= 1)
688 if (ncpt <= nnode) { /* fat numa system */
692 } else { /* ncpt > nnode */
693 while ((nnode << 1) <= ncpt)
700 #if (BITS_PER_LONG == 32)
701 /* config many CPU partitions on 32-bit system could consume
703 ncpt = min(2U, ncpt);
705 while (ncpu % ncpt != 0)
706 ncpt--; /* worst case is 1 */
711 static struct cfs_cpt_table *
712 cfs_cpt_table_create(int ncpt)
714 struct cfs_cpt_table *cptab = NULL;
715 cpumask_t *mask = NULL;
721 rc = cfs_cpt_num_estimate();
725 if (ncpt > num_online_cpus() || ncpt > 4 * rc) {
726 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",
730 if (num_online_cpus() % ncpt != 0) {
731 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",
732 (int)num_online_cpus(), ncpt);
736 cptab = cfs_cpt_table_alloc(ncpt);
738 CERROR("Failed to allocate CPU map(%d)\n", ncpt);
742 num = num_online_cpus() / ncpt;
744 CERROR("CPU changed while setting CPU partition\n");
748 LIBCFS_ALLOC(mask, cpumask_size());
750 CERROR("Failed to allocate scratch cpumask\n");
754 for_each_online_node(i) {
755 cpumask_copy(mask, cpumask_of_node(i));
757 while (!cpumask_empty(mask)) {
758 struct cfs_cpu_partition *part;
764 part = &cptab->ctb_parts[cpt];
766 n = num - cpumask_weight(part->cpt_cpumask);
769 rc = cfs_cpt_choose_ncpus(cptab, cpt, mask, n);
773 LASSERT(num >= cpumask_weight(part->cpt_cpumask));
774 if (num == cpumask_weight(part->cpt_cpumask))
780 num != cpumask_weight(cptab->ctb_parts[ncpt - 1].cpt_cpumask)) {
781 CERROR("Expect %d(%d) CPU partitions but got %d(%d), CPU hotplug/unplug while setting?\n",
782 cptab->ctb_nparts, num, cpt,
783 cpumask_weight(cptab->ctb_parts[ncpt - 1].cpt_cpumask));
787 LIBCFS_FREE(mask, cpumask_size());
792 CERROR("Failed to setup CPU-partition-table with %d CPU-partitions, online HW nodes: %d, HW cpus: %d.\n",
793 ncpt, num_online_nodes(), num_online_cpus());
796 LIBCFS_FREE(mask, cpumask_size());
799 cfs_cpt_table_free(cptab);
804 static struct cfs_cpt_table *
805 cfs_cpt_table_create_pattern(char *pattern)
807 struct cfs_cpt_table *cptab;
814 for (ncpt = 0;; ncpt++) { /* quick scan bracket */
815 str = strchr(str, '[');
821 str = cfs_trimwhite(pattern);
822 if (*str == 'n' || *str == 'N') {
828 (node && ncpt > num_online_nodes()) ||
829 (!node && ncpt > num_online_cpus())) {
830 CERROR("Invalid pattern %s, or too many partitions %d\n",
835 high = node ? MAX_NUMNODES - 1 : nr_cpu_ids - 1;
837 cptab = cfs_cpt_table_alloc(ncpt);
839 CERROR("Failed to allocate cpu partition table\n");
843 for (str = cfs_trimwhite(pattern), c = 0;; c++) {
844 struct cfs_range_expr *range;
845 struct cfs_expr_list *el;
846 char *bracket = strchr(str, '[');
852 if (bracket == NULL) {
854 CERROR("Invalid pattern %s\n", str);
856 } else if (c != ncpt) {
857 CERROR("expect %d partitions but found %d\n",
864 if (sscanf(str, "%d%n", &cpt, &n) < 1) {
865 CERROR("Invalid cpu pattern %s\n", str);
869 if (cpt < 0 || cpt >= ncpt) {
870 CERROR("Invalid partition id %d, total partitions %d\n",
875 if (cfs_cpt_weight(cptab, cpt) != 0) {
876 CERROR("Partition %d has already been set.\n", cpt);
880 str = cfs_trimwhite(str + n);
881 if (str != bracket) {
882 CERROR("Invalid pattern %s\n", str);
886 bracket = strchr(str, ']');
887 if (bracket == NULL) {
888 CERROR("missing right bracket for cpt %d, %s\n",
893 if (cfs_expr_list_parse(str, (bracket - str) + 1,
894 0, high, &el) != 0) {
895 CERROR("Can't parse number range: %s\n", str);
899 list_for_each_entry(range, &el->el_exprs, re_link) {
900 for (i = range->re_lo; i <= range->re_hi; i++) {
901 if ((i - range->re_lo) % range->re_stride != 0)
904 rc = node ? cfs_cpt_set_node(cptab, cpt, i) :
905 cfs_cpt_set_cpu(cptab, cpt, i);
907 cfs_expr_list_free(el);
913 cfs_expr_list_free(el);
915 if (!cfs_cpt_online(cptab, cpt)) {
916 CERROR("No online CPU is found on partition %d\n", cpt);
920 str = cfs_trimwhite(bracket + 1);
926 cfs_cpt_table_free(cptab);
930 #ifdef CONFIG_HOTPLUG_CPU
932 cfs_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
934 unsigned int cpu = (unsigned long)hcpu;
939 case CPU_DEAD_FROZEN:
941 case CPU_ONLINE_FROZEN:
942 spin_lock(&cpt_data.cpt_lock);
943 cpt_data.cpt_version++;
944 spin_unlock(&cpt_data.cpt_lock);
946 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN) {
947 CDEBUG(D_INFO, "CPU changed [cpu %u action %lx]\n",
952 mutex_lock(&cpt_data.cpt_mutex);
953 /* if all HTs in a core are offline, it may break affinity */
954 cpumask_copy(cpt_data.cpt_cpumask,
955 topology_sibling_cpumask(cpu));
956 warn = cpumask_any_and(cpt_data.cpt_cpumask,
957 cpu_online_mask) >= nr_cpu_ids;
958 mutex_unlock(&cpt_data.cpt_mutex);
959 CDEBUG(warn ? D_WARNING : D_INFO,
960 "Lustre: can't support CPU plug-out well now, performance and stability could be impacted [CPU %u action: %lx]\n",
967 static struct notifier_block cfs_cpu_notifier = {
968 .notifier_call = cfs_cpu_notify,
977 if (cfs_cpt_table != NULL)
978 cfs_cpt_table_free(cfs_cpt_table);
980 #ifdef CONFIG_HOTPLUG_CPU
981 unregister_hotcpu_notifier(&cfs_cpu_notifier);
983 if (cpt_data.cpt_cpumask != NULL)
984 LIBCFS_FREE(cpt_data.cpt_cpumask, cpumask_size());
990 LASSERT(cfs_cpt_table == NULL);
992 memset(&cpt_data, 0, sizeof(cpt_data));
994 LIBCFS_ALLOC(cpt_data.cpt_cpumask, cpumask_size());
995 if (cpt_data.cpt_cpumask == NULL) {
996 CERROR("Failed to allocate scratch buffer\n");
1000 spin_lock_init(&cpt_data.cpt_lock);
1001 mutex_init(&cpt_data.cpt_mutex);
1003 #ifdef CONFIG_HOTPLUG_CPU
1004 register_hotcpu_notifier(&cfs_cpu_notifier);
1007 if (*cpu_pattern != 0) {
1008 cfs_cpt_table = cfs_cpt_table_create_pattern(cpu_pattern);
1009 if (cfs_cpt_table == NULL) {
1010 CERROR("Failed to create cptab from pattern %s\n",
1016 cfs_cpt_table = cfs_cpt_table_create(cpu_npartitions);
1017 if (cfs_cpt_table == NULL) {
1018 CERROR("Failed to create ptable with npartitions %d\n",
1024 spin_lock(&cpt_data.cpt_lock);
1025 if (cfs_cpt_table->ctb_version != cpt_data.cpt_version) {
1026 spin_unlock(&cpt_data.cpt_lock);
1027 CERROR("CPU hotplug/unplug during setup\n");
1030 spin_unlock(&cpt_data.cpt_lock);
1032 LCONSOLE(0, "HW CPU cores: %d, npartitions: %d\n",
1033 num_online_cpus(), cfs_cpt_number(cfs_cpt_table));