Merge remote-tracking branch 'asoc/topic/core' into asoc-next
[linux-2.6-block.git] / kernel / events / ring_buffer.c
CommitLineData
76369139
FW
1/*
2 * Performance events ring-buffer code:
3 *
4 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2011 Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2008-2011 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
d36b6910 7 * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
76369139
FW
8 *
9 * For licensing details see kernel-base/COPYING
10 */
11
12#include <linux/perf_event.h>
13#include <linux/vmalloc.h>
14#include <linux/slab.h>
26c86da8 15#include <linux/circ_buf.h>
7c60fc0e 16#include <linux/poll.h>
76369139
FW
17
18#include "internal.h"
19
76369139
FW
20static void perf_output_wakeup(struct perf_output_handle *handle)
21{
7c60fc0e 22 atomic_set(&handle->rb->poll, POLLIN);
76369139 23
a8b0ca17
PZ
24 handle->event->pending_wakeup = 1;
25 irq_work_queue(&handle->event->pending);
76369139
FW
26}
27
28/*
29 * We need to ensure a later event_id doesn't publish a head when a former
30 * event isn't done writing. However since we need to deal with NMIs we
31 * cannot fully serialize things.
32 *
33 * We only publish the head (and generate a wakeup) when the outer-most
34 * event completes.
35 */
36static void perf_output_get_handle(struct perf_output_handle *handle)
37{
38 struct ring_buffer *rb = handle->rb;
39
40 preempt_disable();
41 local_inc(&rb->nest);
42 handle->wakeup = local_read(&rb->wakeup);
43}
44
45static void perf_output_put_handle(struct perf_output_handle *handle)
46{
47 struct ring_buffer *rb = handle->rb;
48 unsigned long head;
49
50again:
51 head = local_read(&rb->head);
52
53 /*
54 * IRQ/NMI can happen here, which means we can miss a head update.
55 */
56
57 if (!local_dec_and_test(&rb->nest))
58 goto out;
59
60 /*
bf378d34
PZ
61 * Since the mmap() consumer (userspace) can run on a different CPU:
62 *
63 * kernel user
64 *
c7f2e3cd
PZ
65 * if (LOAD ->data_tail) { LOAD ->data_head
66 * (A) smp_rmb() (C)
67 * STORE $data LOAD $data
68 * smp_wmb() (B) smp_mb() (D)
69 * STORE ->data_head STORE ->data_tail
70 * }
bf378d34
PZ
71 *
72 * Where A pairs with D, and B pairs with C.
73 *
c7f2e3cd
PZ
74 * In our case (A) is a control dependency that separates the load of
75 * the ->data_tail and the stores of $data. In case ->data_tail
76 * indicates there is no room in the buffer to store $data we do not.
bf378d34 77 *
c7f2e3cd 78 * D needs to be a full barrier since it separates the data READ
bf378d34
PZ
79 * from the tail WRITE.
80 *
81 * For B a WMB is sufficient since it separates two WRITEs, and for C
82 * an RMB is sufficient since it separates two READs.
83 *
84 * See perf_output_begin().
76369139 85 */
c7f2e3cd 86 smp_wmb(); /* B, matches C */
76369139
FW
87 rb->user_page->data_head = head;
88
89 /*
394570b7
PZ
90 * Now check if we missed an update -- rely on previous implied
91 * compiler barriers to force a re-read.
76369139
FW
92 */
93 if (unlikely(head != local_read(&rb->head))) {
94 local_inc(&rb->nest);
95 goto again;
96 }
97
98 if (handle->wakeup != local_read(&rb->wakeup))
99 perf_output_wakeup(handle);
100
101out:
102 preempt_enable();
103}
104
105int perf_output_begin(struct perf_output_handle *handle,
a7ac67ea 106 struct perf_event *event, unsigned int size)
76369139
FW
107{
108 struct ring_buffer *rb;
109 unsigned long tail, offset, head;
524feca5 110 int have_lost, page_shift;
76369139
FW
111 struct {
112 struct perf_event_header header;
113 u64 id;
114 u64 lost;
115 } lost_event;
116
117 rcu_read_lock();
118 /*
119 * For inherited events we send all the output towards the parent.
120 */
121 if (event->parent)
122 event = event->parent;
123
124 rb = rcu_dereference(event->rb);
c72b42a3 125 if (unlikely(!rb))
76369139
FW
126 goto out;
127
c72b42a3 128 if (unlikely(!rb->nr_pages))
76369139
FW
129 goto out;
130
c72b42a3
PZ
131 handle->rb = rb;
132 handle->event = event;
133
76369139 134 have_lost = local_read(&rb->lost);
c72b42a3 135 if (unlikely(have_lost)) {
d20a973f
PZ
136 size += sizeof(lost_event);
137 if (event->attr.sample_id_all)
138 size += event->id_header_size;
76369139
FW
139 }
140
141 perf_output_get_handle(handle);
142
143 do {
5af4692a 144 tail = READ_ONCE_CTRL(rb->user_page->data_tail);
76369139 145 offset = head = local_read(&rb->head);
26c86da8
PZ
146 if (!rb->overwrite &&
147 unlikely(CIRC_SPACE(head, tail, perf_data_size(rb)) < size))
76369139 148 goto fail;
c7f2e3cd
PZ
149
150 /*
151 * The above forms a control dependency barrier separating the
152 * @tail load above from the data stores below. Since the @tail
153 * load is required to compute the branch to fail below.
154 *
155 * A, matches D; the full memory barrier userspace SHOULD issue
156 * after reading the data and before storing the new tail
157 * position.
158 *
159 * See perf_output_put_handle().
160 */
161
26c86da8 162 head += size;
76369139
FW
163 } while (local_cmpxchg(&rb->head, offset, head) != offset);
164
85f59edf 165 /*
c7f2e3cd
PZ
166 * We rely on the implied barrier() by local_cmpxchg() to ensure
167 * none of the data stores below can be lifted up by the compiler.
85f59edf 168 */
85f59edf 169
c72b42a3 170 if (unlikely(head - local_read(&rb->wakeup) > rb->watermark))
76369139
FW
171 local_add(rb->watermark, &rb->wakeup);
172
524feca5
PZ
173 page_shift = PAGE_SHIFT + page_order(rb);
174
175 handle->page = (offset >> page_shift) & (rb->nr_pages - 1);
176 offset &= (1UL << page_shift) - 1;
177 handle->addr = rb->data_pages[handle->page] + offset;
178 handle->size = (1UL << page_shift) - offset;
76369139 179
c72b42a3 180 if (unlikely(have_lost)) {
d20a973f
PZ
181 struct perf_sample_data sample_data;
182
183 lost_event.header.size = sizeof(lost_event);
76369139
FW
184 lost_event.header.type = PERF_RECORD_LOST;
185 lost_event.header.misc = 0;
186 lost_event.id = event->id;
187 lost_event.lost = local_xchg(&rb->lost, 0);
188
d20a973f
PZ
189 perf_event_header__init_id(&lost_event.header,
190 &sample_data, event);
76369139
FW
191 perf_output_put(handle, lost_event);
192 perf_event__output_id_sample(event, handle, &sample_data);
193 }
194
195 return 0;
196
197fail:
198 local_inc(&rb->lost);
199 perf_output_put_handle(handle);
200out:
201 rcu_read_unlock();
202
203 return -ENOSPC;
204}
205
91d7753a 206unsigned int perf_output_copy(struct perf_output_handle *handle,
76369139
FW
207 const void *buf, unsigned int len)
208{
91d7753a 209 return __output_copy(handle, buf, len);
76369139
FW
210}
211
5685e0ff
JO
212unsigned int perf_output_skip(struct perf_output_handle *handle,
213 unsigned int len)
214{
215 return __output_skip(handle, NULL, len);
216}
217
76369139
FW
218void perf_output_end(struct perf_output_handle *handle)
219{
76369139
FW
220 perf_output_put_handle(handle);
221 rcu_read_unlock();
222}
223
57ffc5ca
PZ
224static void rb_irq_work(struct irq_work *work);
225
76369139
FW
226static void
227ring_buffer_init(struct ring_buffer *rb, long watermark, int flags)
228{
229 long max_size = perf_data_size(rb);
230
231 if (watermark)
232 rb->watermark = min(max_size, watermark);
233
234 if (!rb->watermark)
235 rb->watermark = max_size / 2;
236
237 if (flags & RING_BUFFER_WRITABLE)
dd9c086d
SE
238 rb->overwrite = 0;
239 else
240 rb->overwrite = 1;
76369139
FW
241
242 atomic_set(&rb->refcount, 1);
10c6db11
PZ
243
244 INIT_LIST_HEAD(&rb->event_list);
245 spin_lock_init(&rb->event_lock);
57ffc5ca
PZ
246 init_irq_work(&rb->irq_work, rb_irq_work);
247}
248
249static void ring_buffer_put_async(struct ring_buffer *rb)
250{
251 if (!atomic_dec_and_test(&rb->refcount))
252 return;
253
254 rb->rcu_head.next = (void *)rb;
255 irq_work_queue(&rb->irq_work);
76369139
FW
256}
257
fdc26706
AS
258/*
259 * This is called before hardware starts writing to the AUX area to
260 * obtain an output handle and make sure there's room in the buffer.
261 * When the capture completes, call perf_aux_output_end() to commit
262 * the recorded data to the buffer.
263 *
264 * The ordering is similar to that of perf_output_{begin,end}, with
265 * the exception of (B), which should be taken care of by the pmu
266 * driver, since ordering rules will differ depending on hardware.
267 */
268void *perf_aux_output_begin(struct perf_output_handle *handle,
269 struct perf_event *event)
270{
271 struct perf_event *output_event = event;
272 unsigned long aux_head, aux_tail;
273 struct ring_buffer *rb;
274
275 if (output_event->parent)
276 output_event = output_event->parent;
277
278 /*
279 * Since this will typically be open across pmu::add/pmu::del, we
280 * grab ring_buffer's refcount instead of holding rcu read lock
281 * to make sure it doesn't disappear under us.
282 */
283 rb = ring_buffer_get(output_event);
284 if (!rb)
285 return NULL;
286
287 if (!rb_has_aux(rb) || !atomic_inc_not_zero(&rb->aux_refcount))
288 goto err;
289
290 /*
291 * Nesting is not supported for AUX area, make sure nested
292 * writers are caught early
293 */
294 if (WARN_ON_ONCE(local_xchg(&rb->aux_nest, 1)))
295 goto err_put;
296
297 aux_head = local_read(&rb->aux_head);
fdc26706
AS
298
299 handle->rb = rb;
300 handle->event = event;
301 handle->head = aux_head;
2023a0d2 302 handle->size = 0;
fdc26706
AS
303
304 /*
2023a0d2
AS
305 * In overwrite mode, AUX data stores do not depend on aux_tail,
306 * therefore (A) control dependency barrier does not exist. The
307 * (B) <-> (C) ordering is still observed by the pmu driver.
fdc26706 308 */
2023a0d2
AS
309 if (!rb->aux_overwrite) {
310 aux_tail = ACCESS_ONCE(rb->user_page->aux_tail);
1a594131 311 handle->wakeup = local_read(&rb->aux_wakeup) + rb->aux_watermark;
2023a0d2
AS
312 if (aux_head - aux_tail < perf_aux_size(rb))
313 handle->size = CIRC_SPACE(aux_head, aux_tail, perf_aux_size(rb));
314
315 /*
316 * handle->size computation depends on aux_tail load; this forms a
317 * control dependency barrier separating aux_tail load from aux data
318 * store that will be enabled on successful return
319 */
320 if (!handle->size) { /* A, matches D */
321 event->pending_disable = 1;
322 perf_output_wakeup(handle);
323 local_set(&rb->aux_nest, 0);
324 goto err_put;
325 }
fdc26706
AS
326 }
327
328 return handle->rb->aux_priv;
329
330err_put:
331 rb_free_aux(rb);
332
333err:
57ffc5ca 334 ring_buffer_put_async(rb);
fdc26706
AS
335 handle->event = NULL;
336
337 return NULL;
338}
339
340/*
341 * Commit the data written by hardware into the ring buffer by adjusting
342 * aux_head and posting a PERF_RECORD_AUX into the perf buffer. It is the
343 * pmu driver's responsibility to observe ordering rules of the hardware,
344 * so that all the data is externally visible before this is called.
345 */
346void perf_aux_output_end(struct perf_output_handle *handle, unsigned long size,
347 bool truncated)
348{
349 struct ring_buffer *rb = handle->rb;
2023a0d2 350 unsigned long aux_head;
fdc26706
AS
351 u64 flags = 0;
352
353 if (truncated)
354 flags |= PERF_AUX_FLAG_TRUNCATED;
355
2023a0d2
AS
356 /* in overwrite mode, driver provides aux_head via handle */
357 if (rb->aux_overwrite) {
358 flags |= PERF_AUX_FLAG_OVERWRITE;
359
360 aux_head = handle->head;
361 local_set(&rb->aux_head, aux_head);
362 } else {
363 aux_head = local_read(&rb->aux_head);
364 local_add(size, &rb->aux_head);
365 }
fdc26706
AS
366
367 if (size || flags) {
368 /*
369 * Only send RECORD_AUX if we have something useful to communicate
370 */
371
372 perf_event_aux_event(handle->event, aux_head, size, flags);
373 }
374
1a594131 375 aux_head = rb->user_page->aux_head = local_read(&rb->aux_head);
fdc26706 376
1a594131
AS
377 if (aux_head - local_read(&rb->aux_wakeup) >= rb->aux_watermark) {
378 perf_output_wakeup(handle);
379 local_add(rb->aux_watermark, &rb->aux_wakeup);
380 }
fdc26706
AS
381 handle->event = NULL;
382
383 local_set(&rb->aux_nest, 0);
384 rb_free_aux(rb);
57ffc5ca 385 ring_buffer_put_async(rb);
fdc26706
AS
386}
387
388/*
389 * Skip over a given number of bytes in the AUX buffer, due to, for example,
390 * hardware's alignment constraints.
391 */
392int perf_aux_output_skip(struct perf_output_handle *handle, unsigned long size)
393{
394 struct ring_buffer *rb = handle->rb;
395 unsigned long aux_head;
396
397 if (size > handle->size)
398 return -ENOSPC;
399
400 local_add(size, &rb->aux_head);
401
1a594131
AS
402 aux_head = rb->user_page->aux_head = local_read(&rb->aux_head);
403 if (aux_head - local_read(&rb->aux_wakeup) >= rb->aux_watermark) {
404 perf_output_wakeup(handle);
405 local_add(rb->aux_watermark, &rb->aux_wakeup);
406 handle->wakeup = local_read(&rb->aux_wakeup) +
407 rb->aux_watermark;
408 }
409
fdc26706
AS
410 handle->head = aux_head;
411 handle->size -= size;
412
413 return 0;
414}
415
416void *perf_get_aux(struct perf_output_handle *handle)
417{
418 /* this is only valid between perf_aux_output_begin and *_end */
419 if (!handle->event)
420 return NULL;
421
422 return handle->rb->aux_priv;
423}
424
0a4e38e6
AS
425#define PERF_AUX_GFP (GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_NORETRY)
426
427static struct page *rb_alloc_aux_page(int node, int order)
428{
429 struct page *page;
430
431 if (order > MAX_ORDER)
432 order = MAX_ORDER;
433
434 do {
435 page = alloc_pages_node(node, PERF_AUX_GFP, order);
436 } while (!page && order--);
437
438 if (page && order) {
439 /*
440 * Communicate the allocation size to the driver
441 */
442 split_page(page, order);
443 SetPagePrivate(page);
444 set_page_private(page, order);
445 }
446
447 return page;
448}
449
450static void rb_free_aux_page(struct ring_buffer *rb, int idx)
451{
452 struct page *page = virt_to_page(rb->aux_pages[idx]);
453
454 ClearPagePrivate(page);
455 page->mapping = NULL;
456 __free_page(page);
457}
458
45bfb2e5 459int rb_alloc_aux(struct ring_buffer *rb, struct perf_event *event,
1a594131 460 pgoff_t pgoff, int nr_pages, long watermark, int flags)
45bfb2e5
PZ
461{
462 bool overwrite = !(flags & RING_BUFFER_WRITABLE);
463 int node = (event->cpu == -1) ? -1 : cpu_to_node(event->cpu);
0a4e38e6 464 int ret = -ENOMEM, max_order = 0;
45bfb2e5
PZ
465
466 if (!has_aux(event))
467 return -ENOTSUPP;
468
6a279230 469 if (event->pmu->capabilities & PERF_PMU_CAP_AUX_NO_SG) {
0a4e38e6
AS
470 /*
471 * We need to start with the max_order that fits in nr_pages,
472 * not the other way around, hence ilog2() and not get_order.
473 */
474 max_order = ilog2(nr_pages);
475
6a279230
AS
476 /*
477 * PMU requests more than one contiguous chunks of memory
478 * for SW double buffering
479 */
480 if ((event->pmu->capabilities & PERF_PMU_CAP_AUX_SW_DOUBLEBUF) &&
481 !overwrite) {
482 if (!max_order)
483 return -EINVAL;
484
485 max_order--;
486 }
487 }
488
45bfb2e5
PZ
489 rb->aux_pages = kzalloc_node(nr_pages * sizeof(void *), GFP_KERNEL, node);
490 if (!rb->aux_pages)
491 return -ENOMEM;
492
493 rb->free_aux = event->pmu->free_aux;
0a4e38e6 494 for (rb->aux_nr_pages = 0; rb->aux_nr_pages < nr_pages;) {
45bfb2e5 495 struct page *page;
0a4e38e6 496 int last, order;
45bfb2e5 497
0a4e38e6
AS
498 order = min(max_order, ilog2(nr_pages - rb->aux_nr_pages));
499 page = rb_alloc_aux_page(node, order);
45bfb2e5
PZ
500 if (!page)
501 goto out;
502
0a4e38e6
AS
503 for (last = rb->aux_nr_pages + (1 << page_private(page));
504 last > rb->aux_nr_pages; rb->aux_nr_pages++)
505 rb->aux_pages[rb->aux_nr_pages] = page_address(page++);
45bfb2e5
PZ
506 }
507
aa319bcd
AS
508 /*
509 * In overwrite mode, PMUs that don't support SG may not handle more
510 * than one contiguous allocation, since they rely on PMI to do double
511 * buffering. In this case, the entire buffer has to be one contiguous
512 * chunk.
513 */
514 if ((event->pmu->capabilities & PERF_PMU_CAP_AUX_NO_SG) &&
515 overwrite) {
516 struct page *page = virt_to_page(rb->aux_pages[0]);
517
518 if (page_private(page) != max_order)
519 goto out;
520 }
521
45bfb2e5
PZ
522 rb->aux_priv = event->pmu->setup_aux(event->cpu, rb->aux_pages, nr_pages,
523 overwrite);
524 if (!rb->aux_priv)
525 goto out;
526
527 ret = 0;
528
529 /*
530 * aux_pages (and pmu driver's private data, aux_priv) will be
531 * referenced in both producer's and consumer's contexts, thus
532 * we keep a refcount here to make sure either of the two can
533 * reference them safely.
534 */
535 atomic_set(&rb->aux_refcount, 1);
536
2023a0d2 537 rb->aux_overwrite = overwrite;
1a594131
AS
538 rb->aux_watermark = watermark;
539
540 if (!rb->aux_watermark && !rb->aux_overwrite)
541 rb->aux_watermark = nr_pages << (PAGE_SHIFT - 1);
2023a0d2 542
45bfb2e5
PZ
543out:
544 if (!ret)
545 rb->aux_pgoff = pgoff;
546 else
547 rb_free_aux(rb);
548
549 return ret;
550}
551
552static void __rb_free_aux(struct ring_buffer *rb)
553{
554 int pg;
555
556 if (rb->aux_priv) {
557 rb->free_aux(rb->aux_priv);
558 rb->free_aux = NULL;
559 rb->aux_priv = NULL;
560 }
561
ee9397a6
BH
562 if (rb->aux_nr_pages) {
563 for (pg = 0; pg < rb->aux_nr_pages; pg++)
564 rb_free_aux_page(rb, pg);
45bfb2e5 565
ee9397a6
BH
566 kfree(rb->aux_pages);
567 rb->aux_nr_pages = 0;
568 }
45bfb2e5
PZ
569}
570
571void rb_free_aux(struct ring_buffer *rb)
572{
573 if (atomic_dec_and_test(&rb->aux_refcount))
57ffc5ca
PZ
574 irq_work_queue(&rb->irq_work);
575}
576
577static void rb_irq_work(struct irq_work *work)
578{
579 struct ring_buffer *rb = container_of(work, struct ring_buffer, irq_work);
580
581 if (!atomic_read(&rb->aux_refcount))
45bfb2e5 582 __rb_free_aux(rb);
57ffc5ca
PZ
583
584 if (rb->rcu_head.next == (void *)rb)
585 call_rcu(&rb->rcu_head, rb_free_rcu);
45bfb2e5
PZ
586}
587
76369139
FW
588#ifndef CONFIG_PERF_USE_VMALLOC
589
590/*
591 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
592 */
593
45bfb2e5
PZ
594static struct page *
595__perf_mmap_to_page(struct ring_buffer *rb, unsigned long pgoff)
76369139
FW
596{
597 if (pgoff > rb->nr_pages)
598 return NULL;
599
600 if (pgoff == 0)
601 return virt_to_page(rb->user_page);
602
603 return virt_to_page(rb->data_pages[pgoff - 1]);
604}
605
606static void *perf_mmap_alloc_page(int cpu)
607{
608 struct page *page;
609 int node;
610
611 node = (cpu == -1) ? cpu : cpu_to_node(cpu);
612 page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
613 if (!page)
614 return NULL;
615
616 return page_address(page);
617}
618
619struct ring_buffer *rb_alloc(int nr_pages, long watermark, int cpu, int flags)
620{
621 struct ring_buffer *rb;
622 unsigned long size;
623 int i;
624
625 size = sizeof(struct ring_buffer);
626 size += nr_pages * sizeof(void *);
627
628 rb = kzalloc(size, GFP_KERNEL);
629 if (!rb)
630 goto fail;
631
632 rb->user_page = perf_mmap_alloc_page(cpu);
633 if (!rb->user_page)
634 goto fail_user_page;
635
636 for (i = 0; i < nr_pages; i++) {
637 rb->data_pages[i] = perf_mmap_alloc_page(cpu);
638 if (!rb->data_pages[i])
639 goto fail_data_pages;
640 }
641
642 rb->nr_pages = nr_pages;
643
644 ring_buffer_init(rb, watermark, flags);
645
646 return rb;
647
648fail_data_pages:
649 for (i--; i >= 0; i--)
650 free_page((unsigned long)rb->data_pages[i]);
651
652 free_page((unsigned long)rb->user_page);
653
654fail_user_page:
655 kfree(rb);
656
657fail:
658 return NULL;
659}
660
661static void perf_mmap_free_page(unsigned long addr)
662{
663 struct page *page = virt_to_page((void *)addr);
664
665 page->mapping = NULL;
666 __free_page(page);
667}
668
669void rb_free(struct ring_buffer *rb)
670{
671 int i;
672
673 perf_mmap_free_page((unsigned long)rb->user_page);
674 for (i = 0; i < rb->nr_pages; i++)
675 perf_mmap_free_page((unsigned long)rb->data_pages[i]);
676 kfree(rb);
677}
678
679#else
5919b309
JO
680static int data_page_nr(struct ring_buffer *rb)
681{
682 return rb->nr_pages << page_order(rb);
683}
76369139 684
45bfb2e5
PZ
685static struct page *
686__perf_mmap_to_page(struct ring_buffer *rb, unsigned long pgoff)
76369139 687{
5919b309
JO
688 /* The '>' counts in the user page. */
689 if (pgoff > data_page_nr(rb))
76369139
FW
690 return NULL;
691
692 return vmalloc_to_page((void *)rb->user_page + pgoff * PAGE_SIZE);
693}
694
695static void perf_mmap_unmark_page(void *addr)
696{
697 struct page *page = vmalloc_to_page(addr);
698
699 page->mapping = NULL;
700}
701
702static void rb_free_work(struct work_struct *work)
703{
704 struct ring_buffer *rb;
705 void *base;
706 int i, nr;
707
708 rb = container_of(work, struct ring_buffer, work);
5919b309 709 nr = data_page_nr(rb);
76369139
FW
710
711 base = rb->user_page;
5919b309
JO
712 /* The '<=' counts in the user page. */
713 for (i = 0; i <= nr; i++)
76369139
FW
714 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
715
716 vfree(base);
717 kfree(rb);
718}
719
720void rb_free(struct ring_buffer *rb)
721{
722 schedule_work(&rb->work);
723}
724
725struct ring_buffer *rb_alloc(int nr_pages, long watermark, int cpu, int flags)
726{
727 struct ring_buffer *rb;
728 unsigned long size;
729 void *all_buf;
730
731 size = sizeof(struct ring_buffer);
732 size += sizeof(void *);
733
734 rb = kzalloc(size, GFP_KERNEL);
735 if (!rb)
736 goto fail;
737
738 INIT_WORK(&rb->work, rb_free_work);
739
740 all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
741 if (!all_buf)
742 goto fail_all_buf;
743
744 rb->user_page = all_buf;
745 rb->data_pages[0] = all_buf + PAGE_SIZE;
746 rb->page_order = ilog2(nr_pages);
5919b309 747 rb->nr_pages = !!nr_pages;
76369139
FW
748
749 ring_buffer_init(rb, watermark, flags);
750
751 return rb;
752
753fail_all_buf:
754 kfree(rb);
755
756fail:
757 return NULL;
758}
759
760#endif
45bfb2e5
PZ
761
762struct page *
763perf_mmap_to_page(struct ring_buffer *rb, unsigned long pgoff)
764{
765 if (rb->aux_nr_pages) {
766 /* above AUX space */
767 if (pgoff > rb->aux_pgoff + rb->aux_nr_pages)
768 return NULL;
769
770 /* AUX space */
771 if (pgoff >= rb->aux_pgoff)
772 return virt_to_page(rb->aux_pages[pgoff - rb->aux_pgoff]);
773 }
774
775 return __perf_mmap_to_page(rb, pgoff);
776}