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, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * libcfs/include/libcfs/libcfs_cpu.h
35 * . CPU partition is virtual processing unit
37 * . CPU partition can present 1-N cores, or 1-N NUMA nodes,
38 * in other words, CPU partition is a processors pool.
40 * CPU Partition Table (CPT)
41 * . a set of CPU partitions
43 * . There are two modes for CPT: CFS_CPU_MODE_NUMA and CFS_CPU_MODE_SMP
45 * . User can specify total number of CPU partitions while creating a
46 * CPT, ID of CPU partition is always start from 0.
48 * Example: if there are 8 cores on the system, while creating a CPT
49 * with cpu_npartitions=4:
50 * core[0, 1] = partition[0], core[2, 3] = partition[1]
51 * core[4, 5] = partition[2], core[6, 7] = partition[3]
54 * core[0, 1, ... 7] = partition[0]
56 * . User can also specify CPU partitions by string pattern
58 * Examples: cpu_partitions="0[0,1], 1[2,3]"
59 * cpu_partitions="N 0[0-3], 1[4-8]"
61 * The first character "N" means following numbers are numa ID
63 * . NUMA allocators, CPU affinity threads are built over CPU partitions,
64 * instead of HW CPUs or HW nodes.
66 * . By default, Lustre modules should refer to the global cfs_cpt_table,
67 * instead of accessing HW CPUs directly, so concurrency of Lustre can be
68 * configured by cpu_npartitions of the global cfs_cpt_table
70 * . If cpu_npartitions=1(all CPUs in one pool), lustre should work the
71 * same way as 2.2 or earlier versions
73 * Author: liang@whamcloud.com
76 #ifndef __LIBCFS_CPU_H__
77 #define __LIBCFS_CPU_H__
79 #include <linux/slab.h>
80 #include <libcfs/linux/linux-cpu.h>
82 #ifndef HAVE_LIBCFS_CPT
84 struct cfs_cpt_table {
85 /* # of CPU partitions */
90 nodemask_t ctb_nodemask;
95 #endif /* !HAVE_LIBCFS_CPT */
97 /* any CPU partition */
98 #define CFS_CPT_ANY (-1)
100 extern struct cfs_cpt_table *cfs_cpt_table;
103 * destroy a CPU partition table
105 void cfs_cpt_table_free(struct cfs_cpt_table *cptab);
107 * create a cfs_cpt_table with \a ncpt number of partitions
109 struct cfs_cpt_table *cfs_cpt_table_alloc(int ncpt);
111 * print string information of cpt-table
113 int cfs_cpt_table_print(struct cfs_cpt_table *cptab, char *buf, int len);
115 * print distance information of cpt-table
117 int cfs_cpt_distance_print(struct cfs_cpt_table *cptab, char *buf, int len);
119 * return total number of CPU partitions in \a cptab
121 int cfs_cpt_number(struct cfs_cpt_table *cptab);
123 * return number of HW cores or hypter-threadings in a CPU partition \a cpt
125 int cfs_cpt_weight(struct cfs_cpt_table *cptab, int cpt);
127 * is there any online CPU in CPU partition \a cpt
129 int cfs_cpt_online(struct cfs_cpt_table *cptab, int cpt);
131 * return cpumask of CPU partition \a cpt
133 cpumask_t *cfs_cpt_cpumask(struct cfs_cpt_table *cptab, int cpt);
135 * return nodemask of CPU partition \a cpt
137 nodemask_t *cfs_cpt_nodemask(struct cfs_cpt_table *cptab, int cpt);
139 * shadow current HW processor ID to CPU-partition ID of \a cptab
141 int cfs_cpt_current(struct cfs_cpt_table *cptab, int remap);
143 * shadow HW processor ID \a CPU to CPU-partition ID by \a cptab
145 int cfs_cpt_of_cpu(struct cfs_cpt_table *cptab, int cpu);
147 * shadow HW node ID \a NODE to CPU-partition ID by \a cptab
149 int cfs_cpt_of_node(struct cfs_cpt_table *cptab, int node);
151 * NUMA distance between \a cpt1 and \a cpt2 in \a cptab
153 unsigned cfs_cpt_distance(struct cfs_cpt_table *cptab, int cpt1, int cpt2);
155 * bind current thread on a CPU-partition \a cpt of \a cptab
157 int cfs_cpt_bind(struct cfs_cpt_table *cptab, int cpt);
159 * add \a cpu to CPU partion @cpt of \a cptab, return 1 for success,
160 * otherwise 0 is returned
162 int cfs_cpt_set_cpu(struct cfs_cpt_table *cptab, int cpt, int cpu);
164 * remove \a cpu from CPU partition \a cpt of \a cptab
166 void cfs_cpt_unset_cpu(struct cfs_cpt_table *cptab, int cpt, int cpu);
168 * add all cpus in \a mask to CPU partition \a cpt
169 * return 1 if successfully set all CPUs, otherwise return 0
172 int cfs_cpt_set_cpumask(struct cfs_cpt_table *cptab, int cpt,
173 const cpumask_t *mask);
175 * remove all cpus in \a mask from CPU partition \a cpt
177 void cfs_cpt_unset_cpumask(struct cfs_cpt_table *cptab, int cpt,
178 const cpumask_t *mask);
180 * add all cpus in NUMA node \a node to CPU partition \a cpt
181 * return 1 if successfully set all CPUs, otherwise return 0
183 int cfs_cpt_set_node(struct cfs_cpt_table *cptab, int cpt, int node);
185 * remove all cpus in NUMA node \a node from CPU partition \a cpt
187 void cfs_cpt_unset_node(struct cfs_cpt_table *cptab, int cpt, int node);
190 * add all cpus in node mask \a mask to CPU partition \a cpt
191 * return 1 if successfully set all CPUs, otherwise return 0
193 int cfs_cpt_set_nodemask(struct cfs_cpt_table *cptab, int cpt,
194 const nodemask_t *mask);
196 * remove all cpus in node mask \a mask from CPU partition \a cpt
198 void cfs_cpt_unset_nodemask(struct cfs_cpt_table *cptab, int cpt,
199 const nodemask_t *mask);
201 * convert partition id \a cpt to numa node id, if there are more than one
202 * nodes in this partition, it might return a different node id each time.
204 int cfs_cpt_spread_node(struct cfs_cpt_table *cptab, int cpt);
207 * allocate per-cpu-partition data, returned value is an array of pointers,
208 * variable can be indexed by CPU ID.
209 * cptab != NULL: size of array is number of CPU partitions
210 * cptab == NULL: size of array is number of HW cores
212 void *cfs_percpt_alloc(struct cfs_cpt_table *cptab, unsigned int size);
214 * destory per-cpu-partition variable
216 void cfs_percpt_free(void *vars);
217 int cfs_percpt_number(void *vars);
219 #define cfs_percpt_for_each(var, i, vars) \
220 for (i = 0; i < cfs_percpt_number(vars) && \
221 ((var) = (vars)[i]) != NULL; i++)
224 * percpu partition lock
226 * There are some use-cases like this in Lustre:
227 * . each CPU partition has it's own private data which is frequently changed,
228 * and mostly by the local CPU partition.
229 * . all CPU partitions share some global data, these data are rarely changed.
231 * LNet is typical example.
232 * CPU partition lock is designed for this kind of use-cases:
233 * . each CPU partition has it's own private lock
234 * . change on private data just needs to take the private lock
235 * . read on shared data just needs to take _any_ of private locks
236 * . change on shared data needs to take _all_ private locks,
237 * which is slow and should be really rare.
240 CFS_PERCPT_LOCK_EX = -1, /* negative */
243 struct cfs_percpt_lock {
244 /* cpu-partition-table for this lock */
245 struct cfs_cpt_table *pcl_cptab;
246 /* exclusively locked */
247 unsigned int pcl_locked;
248 /* private lock table */
249 spinlock_t **pcl_locks;
252 /* return number of private locks */
253 #define cfs_percpt_lock_num(pcl) cfs_cpt_number(pcl->pcl_cptab)
256 * create a cpu-partition lock based on CPU partition table \a cptab,
257 * each private lock has extra \a psize bytes padding data
259 struct cfs_percpt_lock *cfs_percpt_lock_create(struct cfs_cpt_table *cptab,
260 struct lock_class_key *keys);
261 /* destroy a cpu-partition lock */
262 void cfs_percpt_lock_free(struct cfs_percpt_lock *pcl);
264 /* lock private lock \a index of \a pcl */
265 void cfs_percpt_lock(struct cfs_percpt_lock *pcl, int index);
266 /* unlock private lock \a index of \a pcl */
267 void cfs_percpt_unlock(struct cfs_percpt_lock *pcl, int index);
269 #define CFS_PERCPT_LOCK_KEYS 256
271 /* NB: don't allocate keys dynamically, lockdep needs them to be in ".data" */
272 #define cfs_percpt_lock_alloc(cptab) \
274 static struct lock_class_key ___keys[CFS_PERCPT_LOCK_KEYS]; \
275 struct cfs_percpt_lock *___lk; \
277 if (cfs_cpt_number(cptab) > CFS_PERCPT_LOCK_KEYS) \
278 ___lk = cfs_percpt_lock_create(cptab, NULL); \
280 ___lk = cfs_percpt_lock_create(cptab, ___keys); \
285 * allocate \a nr_bytes of physical memory from a contiguous region with the
286 * properties of \a flags which are bound to the partition id \a cpt. This
287 * function should only be used for the case when only a few pages of memory
291 cfs_cpt_malloc(struct cfs_cpt_table *cptab, int cpt, size_t nr_bytes,
294 return kmalloc_node(nr_bytes, flags,
295 cfs_cpt_spread_node(cptab, cpt));
299 * allocate \a nr_bytes of virtually contiguous memory that is bound to the
300 * partition id \a cpt.
303 cfs_cpt_vzalloc(struct cfs_cpt_table *cptab, int cpt, size_t nr_bytes)
305 /* vzalloc_node() sets __GFP_FS by default but no current Kernel
306 * exported entry-point allows for both a NUMA node specification
307 * and a custom allocation flags mask. This may be an issue since
308 * __GFP_FS usage can cause some deadlock situations in our code,
309 * like when memory reclaim started, within the same context of a
310 * thread doing FS operations, that can also attempt conflicting FS
313 return vzalloc_node(nr_bytes, cfs_cpt_spread_node(cptab, cpt));
317 * allocate a single page of memory with the properties of \a flags were
318 * that page is bound to the partition id \a cpt.
320 static inline struct page *
321 cfs_page_cpt_alloc(struct cfs_cpt_table *cptab, int cpt, gfp_t flags)
323 return alloc_pages_node(cfs_cpt_spread_node(cptab, cpt), flags, 0);
327 * allocate a chunck of memory from a memory pool that is bound to the
328 * partition id \a cpt with the properites of \a flags.
331 cfs_mem_cache_cpt_alloc(struct kmem_cache *cachep, struct cfs_cpt_table *cptab,
332 int cpt, gfp_t flags)
334 return kmem_cache_alloc_node(cachep, flags,
335 cfs_cpt_spread_node(cptab, cpt));
339 * iterate over all CPU partitions in \a cptab
341 #define cfs_cpt_for_each(i, cptab) \
342 for (i = 0; i < cfs_cpt_number(cptab); i++)
344 #ifndef __read_mostly
345 # define __read_mostly
348 #ifndef ____cacheline_aligned
349 #define ____cacheline_aligned
352 int cfs_cpu_init(void);
353 void cfs_cpu_fini(void);
355 #endif /* __LIBCFS_CPU_H__ */