2 * Copyright (c) 2007 Apple Inc. All rights reserved.
4 * @APPLE_OSREFERENCE_LICENSE_HEADER_START@
6 * This file contains Original Code and/or Modifications of Original Code
7 * as defined in and that are subject to the Apple Public Source License
8 * Version 2.0 (the 'License'). You may not use this file except in
9 * compliance with the License. The rights granted to you under the License
10 * may not be used to create, or enable the creation or redistribution of,
11 * unlawful or unlicensed copies of an Apple operating system, or to
12 * circumvent, violate, or enable the circumvention or violation of, any
13 * terms of an Apple operating system software license agreement.
15 * Please obtain a copy of the License at
16 * http://www.opensource.apple.com/apsl/ and read it before using this file.
18 * The Original Code and all software distributed under the License are
19 * distributed on an 'AS IS' basis, WITHOUT WARRANTY OF ANY KIND, EITHER
20 * EXPRESS OR IMPLIED, AND APPLE HEREBY DISCLAIMS ALL SUCH WARRANTIES,
21 * INCLUDING WITHOUT LIMITATION, ANY WARRANTIES OF MERCHANTABILITY,
22 * FITNESS FOR A PARTICULAR PURPOSE, QUIET ENJOYMENT OR NON-INFRINGEMENT.
23 * Please see the License for the specific language governing rights and
24 * limitations under the License.
26 * @APPLE_OSREFERENCE_LICENSE_HEADER_END@
29 #include <kern/affinity.h>
30 #include <kern/task.h>
31 #include <kern/kalloc.h>
32 #include <machine/cpu_affinity.h>
35 * Affinity involves 2 objects:
36 * - affinity namespace:
37 * shared by a task family, this controls affinity tag lookup and
38 * allocation; it anchors all affinity sets in one namespace
40 * anchors all threads with membership of this affinity set
41 * and which share an affinity tag in the owning namespace.
44 * - The task lock protects the creation of an affinity namespace.
45 * - The affinity namespace mutex protects the inheritance of a namespace
46 * and its thread membership. This includes its destruction when the task
47 * reference count goes to zero.
48 * - The thread mutex protects a thread's affinity set membership, but in
49 * addition, the thread_lock is taken to write thread->affinity_set since this
50 * field (representng the active affinity set) is read by the scheduler.
52 * The lock ordering is: task lock, thread mutex, namespace mutex, thread lock.
56 #define DBG(x...) kprintf("DBG: " x)
61 struct affinity_space
{
63 uint32_t aspc_task_count
;
64 queue_head_t aspc_affinities
;
66 typedef struct affinity_space
*affinity_space_t
;
68 static affinity_space_t
affinity_space_alloc(void);
69 static void affinity_space_free(affinity_space_t aspc
);
70 static affinity_set_t
affinity_set_alloc(void);
71 static void affinity_set_free(affinity_set_t aset
);
72 static affinity_set_t
affinity_set_find(affinity_space_t aspc
, uint32_t tag
);
73 static void affinity_set_place(affinity_space_t aspc
, affinity_set_t aset
);
74 static void affinity_set_add(affinity_set_t aset
, thread_t thread
);
75 static affinity_set_t
affinity_set_remove(affinity_set_t aset
, thread_t thread
);
78 * The following globals may be modified by the sysctls
79 * kern.affinity_sets_enabled - disables hinting if cleared
80 * kern.affinity_sets_mapping - controls cache distribution policy
81 * See bsd/kern_sysctl.c
83 * Affinity sets are not used on embedded, which typically only
84 * has a single pset, and last-processor affinity is
85 * more important than pset affinity.
88 boolean_t affinity_sets_enabled
= FALSE
;
89 int affinity_sets_mapping
= 0;
90 #else /* !CONFIG_EMBEDDED */
91 boolean_t affinity_sets_enabled
= TRUE
;
92 int affinity_sets_mapping
= 1;
93 #endif /* !CONFIG_EMBEDDED */
96 thread_affinity_is_supported(void)
98 return (ml_get_max_affinity_sets() != 0);
103 * thread_affinity_get()
104 * Return the affinity tag for a thread.
105 * Called with the thread mutex held.
108 thread_affinity_get(thread_t thread
)
112 if (thread
->affinity_set
!= NULL
)
113 tag
= thread
->affinity_set
->aset_tag
;
115 tag
= THREAD_AFFINITY_TAG_NULL
;
122 * thread_affinity_set()
123 * Place a thread in an affinity set identified by a tag.
124 * Called with thread referenced but not locked.
127 thread_affinity_set(thread_t thread
, uint32_t tag
)
130 affinity_set_t empty_aset
= NULL
;
131 affinity_space_t aspc
;
132 affinity_space_t new_aspc
= NULL
;
134 DBG("thread_affinity_set(%p,%u)\n", thread
, tag
);
136 task_lock(thread
->task
);
137 aspc
= thread
->task
->affinity_space
;
139 task_unlock(thread
->task
);
140 new_aspc
= affinity_space_alloc();
141 if (new_aspc
== NULL
)
142 return KERN_RESOURCE_SHORTAGE
;
143 task_lock(thread
->task
);
144 if (thread
->task
->affinity_space
== NULL
) {
145 thread
->task
->affinity_space
= new_aspc
;
148 aspc
= thread
->task
->affinity_space
;
150 task_unlock(thread
->task
);
152 affinity_space_free(new_aspc
);
154 thread_mtx_lock(thread
);
155 if (!thread
->active
) {
156 /* Beaten to lock and the thread is dead */
157 thread_mtx_unlock(thread
);
158 return KERN_TERMINATED
;
161 lck_mtx_lock(&aspc
->aspc_lock
);
162 aset
= thread
->affinity_set
;
165 * Remove thread from current affinity set
167 DBG("thread_affinity_set(%p,%u) removing from aset %p\n",
169 empty_aset
= affinity_set_remove(aset
, thread
);
172 if (tag
!= THREAD_AFFINITY_TAG_NULL
) {
173 aset
= affinity_set_find(aspc
, tag
);
176 * Add thread to existing affinity set
178 DBG("thread_affinity_set(%p,%u) found aset %p\n",
182 * Use the new affinity set, add this thread
183 * and place it in a suitable processor set.
185 if (empty_aset
!= NULL
) {
189 aset
= affinity_set_alloc();
191 lck_mtx_unlock(&aspc
->aspc_lock
);
192 thread_mtx_unlock(thread
);
193 return KERN_RESOURCE_SHORTAGE
;
196 DBG("thread_affinity_set(%p,%u) (re-)using aset %p\n",
198 aset
->aset_tag
= tag
;
199 affinity_set_place(aspc
, aset
);
201 affinity_set_add(aset
, thread
);
204 lck_mtx_unlock(&aspc
->aspc_lock
);
205 thread_mtx_unlock(thread
);
208 * If we wound up not using an empty aset we created,
211 if (empty_aset
!= NULL
)
212 affinity_set_free(empty_aset
);
214 if (thread
== current_thread())
215 thread_block(THREAD_CONTINUE_NULL
);
221 * task_affinity_create()
222 * Called from task create.
225 task_affinity_create(task_t parent_task
, task_t child_task
)
227 affinity_space_t aspc
= parent_task
->affinity_space
;
229 DBG("task_affinity_create(%p,%p)\n", parent_task
, child_task
);
234 * Bump the task reference count on the shared namespace and
235 * give it to the child.
237 lck_mtx_lock(&aspc
->aspc_lock
);
238 aspc
->aspc_task_count
++;
239 child_task
->affinity_space
= aspc
;
240 lck_mtx_unlock(&aspc
->aspc_lock
);
244 * task_affinity_deallocate()
245 * Called from task_deallocate() when there's a namespace to dereference.
248 task_affinity_deallocate(task_t task
)
250 affinity_space_t aspc
= task
->affinity_space
;
252 DBG("task_affinity_deallocate(%p) aspc %p task_count %d\n",
253 task
, aspc
, aspc
->aspc_task_count
);
255 lck_mtx_lock(&aspc
->aspc_lock
);
256 if (--(aspc
->aspc_task_count
) == 0) {
257 assert(queue_empty(&aspc
->aspc_affinities
));
258 lck_mtx_unlock(&aspc
->aspc_lock
);
259 affinity_space_free(aspc
);
261 lck_mtx_unlock(&aspc
->aspc_lock
);
266 * task_affinity_info()
267 * Return affinity tag info (number, min, max) for the task.
269 * Conditions: task is locked.
274 task_info_t task_info_out
,
275 mach_msg_type_number_t
*task_info_count
)
278 affinity_space_t aspc
;
279 task_affinity_tag_info_t info
;
281 *task_info_count
= TASK_AFFINITY_TAG_INFO_COUNT
;
282 info
= (task_affinity_tag_info_t
) task_info_out
;
284 info
->task_count
= 0;
285 info
->min
= THREAD_AFFINITY_TAG_NULL
;
286 info
->max
= THREAD_AFFINITY_TAG_NULL
;
288 aspc
= task
->affinity_space
;
290 lck_mtx_lock(&aspc
->aspc_lock
);
291 queue_iterate(&aspc
->aspc_affinities
,
292 aset
, affinity_set_t
, aset_affinities
) {
294 if (info
->min
== THREAD_AFFINITY_TAG_NULL
||
295 aset
->aset_tag
< (uint32_t) info
->min
)
296 info
->min
= aset
->aset_tag
;
297 if (info
->max
== THREAD_AFFINITY_TAG_NULL
||
298 aset
->aset_tag
> (uint32_t) info
->max
)
299 info
->max
= aset
->aset_tag
;
301 info
->task_count
= aspc
->aspc_task_count
;
302 lck_mtx_unlock(&aspc
->aspc_lock
);
308 * Called from thread_dup() during fork() with child's mutex held.
309 * Set the child into the parent's affinity set.
310 * Note the affinity space is shared.
313 thread_affinity_dup(thread_t parent
, thread_t child
)
316 affinity_space_t aspc
;
318 thread_mtx_lock(parent
);
319 aset
= parent
->affinity_set
;
320 DBG("thread_affinity_dup(%p,%p) aset %p\n", parent
, child
, aset
);
322 thread_mtx_unlock(parent
);
326 aspc
= aset
->aset_space
;
327 assert(aspc
== parent
->task
->affinity_space
);
328 assert(aspc
== child
->task
->affinity_space
);
330 lck_mtx_lock(&aspc
->aspc_lock
);
331 affinity_set_add(aset
, child
);
332 lck_mtx_unlock(&aspc
->aspc_lock
);
334 thread_mtx_unlock(parent
);
338 * thread_affinity_terminate()
339 * Remove thread from any affinity set.
340 * Called with the thread mutex locked.
343 thread_affinity_terminate(thread_t thread
)
345 affinity_set_t aset
= thread
->affinity_set
;
346 affinity_space_t aspc
;
348 DBG("thread_affinity_terminate(%p)\n", thread
);
350 aspc
= aset
->aset_space
;
351 lck_mtx_lock(&aspc
->aspc_lock
);
352 if (affinity_set_remove(aset
, thread
)) {
353 affinity_set_free(aset
);
355 lck_mtx_unlock(&aspc
->aspc_lock
);
359 * thread_affinity_exec()
360 * Called from execve() to cancel any current affinity - a new image implies
361 * the calling thread terminates any expressed or inherited affinity.
364 thread_affinity_exec(thread_t thread
)
366 if (thread
->affinity_set
!= AFFINITY_SET_NULL
)
367 thread_affinity_terminate(thread
);
371 * Create an empty affinity namespace data structure.
373 static affinity_space_t
374 affinity_space_alloc(void)
376 affinity_space_t aspc
;
378 aspc
= (affinity_space_t
) kalloc(sizeof(struct affinity_space
));
382 lck_mtx_init(&aspc
->aspc_lock
, &task_lck_grp
, &task_lck_attr
);
383 queue_init(&aspc
->aspc_affinities
);
384 aspc
->aspc_task_count
= 1;
386 DBG("affinity_space_create() returns %p\n", aspc
);
391 * Destroy the given empty affinity namespace data structure.
394 affinity_space_free(affinity_space_t aspc
)
396 assert(queue_empty(&aspc
->aspc_affinities
));
398 lck_mtx_destroy(&aspc
->aspc_lock
, &task_lck_grp
);
399 DBG("affinity_space_free(%p)\n", aspc
);
400 kfree(aspc
, sizeof(struct affinity_space
));
405 * Create an empty affinity set data structure
406 * entering it into a list anchored by the owning task.
408 static affinity_set_t
409 affinity_set_alloc(void)
413 aset
= (affinity_set_t
) kalloc(sizeof(struct affinity_set
));
417 aset
->aset_thread_count
= 0;
418 queue_init(&aset
->aset_affinities
);
419 queue_init(&aset
->aset_threads
);
421 aset
->aset_pset
= PROCESSOR_SET_NULL
;
422 aset
->aset_space
= NULL
;
424 DBG("affinity_set_create() returns %p\n", aset
);
429 * Destroy the given empty affinity set data structure
430 * after removing it from the parent task.
433 affinity_set_free(affinity_set_t aset
)
435 assert(queue_empty(&aset
->aset_threads
));
437 DBG("affinity_set_free(%p)\n", aset
);
438 kfree(aset
, sizeof(struct affinity_set
));
442 * Add a thread to an affinity set.
443 * The caller must have the thread mutex and space locked.
446 affinity_set_add(affinity_set_t aset
, thread_t thread
)
450 DBG("affinity_set_add(%p,%p)\n", aset
, thread
);
451 queue_enter(&aset
->aset_threads
,
452 thread
, thread_t
, affinity_threads
);
453 aset
->aset_thread_count
++;
456 thread
->affinity_set
= affinity_sets_enabled
? aset
: NULL
;
457 thread_unlock(thread
);
462 * Remove a thread from an affinity set returning the set if now empty.
463 * The caller must have the thread mutex and space locked.
465 static affinity_set_t
466 affinity_set_remove(affinity_set_t aset
, thread_t thread
)
472 thread
->affinity_set
= NULL
;
473 thread_unlock(thread
);
476 aset
->aset_thread_count
--;
477 queue_remove(&aset
->aset_threads
,
478 thread
, thread_t
, affinity_threads
);
479 if (queue_empty(&aset
->aset_threads
)) {
480 queue_remove(&aset
->aset_space
->aspc_affinities
,
481 aset
, affinity_set_t
, aset_affinities
);
482 assert(aset
->aset_thread_count
== 0);
483 aset
->aset_tag
= THREAD_AFFINITY_TAG_NULL
;
485 aset
->aset_pset
= PROCESSOR_SET_NULL
;
486 aset
->aset_space
= NULL
;
487 DBG("affinity_set_remove(%p,%p) set now empty\n", aset
, thread
);
490 DBG("affinity_set_remove(%p,%p)\n", aset
, thread
);
496 * Find an affinity set in the parent task with the given affinity tag.
497 * The caller must have the space locked.
499 static affinity_set_t
500 affinity_set_find(affinity_space_t space
, uint32_t tag
)
504 queue_iterate(&space
->aspc_affinities
,
505 aset
, affinity_set_t
, aset_affinities
) {
506 if (aset
->aset_tag
== tag
) {
507 DBG("affinity_set_find(%p,%u) finds %p\n",
512 DBG("affinity_set_find(%p,%u) not found\n", space
, tag
);
517 * affinity_set_place() assigns an affinity set to a suitable processor_set.
518 * The selection criteria is:
519 * - the set currently occupied by the least number of affinities
520 * belonging to the owning the task.
521 * The caller must have the space locked.
524 affinity_set_place(affinity_space_t aspc
, affinity_set_t new_aset
)
526 unsigned int num_cpu_asets
= ml_get_max_affinity_sets();
527 unsigned int set_occupancy
[num_cpu_asets
];
529 unsigned int i_least_occupied
;
532 for (i
= 0; i
< num_cpu_asets
; i
++)
533 set_occupancy
[i
] = 0;
536 * Scan the affinity sets calculating the number of sets
537 * occupy the available physical affinities.
539 queue_iterate(&aspc
->aspc_affinities
,
540 aset
, affinity_set_t
, aset_affinities
) {
541 if(aset
->aset_num
< num_cpu_asets
)
542 set_occupancy
[aset
->aset_num
]++;
544 panic("aset_num = %d in %s\n", aset
->aset_num
, __FUNCTION__
);
548 * Find the least occupied set (or the first empty set).
549 * To distribute placements somewhat, start searching from
550 * a cpu affinity chosen randomly per namespace:
551 * [(unsigned int)aspc % 127] % num_cpu_asets
552 * unless this mapping policy is overridden.
554 if (affinity_sets_mapping
== 0)
555 i_least_occupied
= 0;
557 i_least_occupied
= (unsigned int)(((uintptr_t)aspc
% 127) % num_cpu_asets
);
558 for (i
= 0; i
< num_cpu_asets
; i
++) {
559 unsigned int j
= (i_least_occupied
+ i
) % num_cpu_asets
;
560 if (set_occupancy
[j
] == 0) {
561 i_least_occupied
= j
;
564 if (set_occupancy
[j
] < set_occupancy
[i_least_occupied
])
565 i_least_occupied
= j
;
567 new_aset
->aset_num
= i_least_occupied
;
568 new_aset
->aset_pset
= ml_affinity_to_pset(i_least_occupied
);
570 /* Add the new affinity set to the group */
571 new_aset
->aset_space
= aspc
;
572 queue_enter(&aspc
->aspc_affinities
,
573 new_aset
, affinity_set_t
, aset_affinities
);
575 DBG("affinity_set_place(%p,%p) selected affinity %u pset %p\n",
576 aspc
, new_aset
, new_aset
->aset_num
, new_aset
->aset_pset
);