projects/linux·blob master · refs

981 lines · 23.8 KiB

1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Performance events ring-buffer code:
4 *
5 * Copyright (C) 2008 Linutronix GmbH, Thomas Gleixner <[email protected]>
6 * Copyright (C) 2008-2011 Red Hat, Inc., Ingo Molnar
7 * Copyright (C) 2008-2011 Red Hat, Inc., Peter Zijlstra
8 * Copyright © 2009 Paul Mackerras, IBM Corp. <[email protected]>
9 */
10
11#include <linux/perf_event.h>
12#include <linux/vmalloc.h>
13#include <linux/slab.h>
14#include <linux/circ_buf.h>
15#include <linux/poll.h>
16#include <linux/nospec.h>
17
18#include "internal.h"
19
20static void perf_output_wakeup(struct perf_output_handle *handle)
21{
22 atomic_set(&handle->rb->poll, EPOLLIN | EPOLLRDNORM);
23
24 handle->event->pending_wakeup = 1;
25
26 if (*perf_event_fasync(handle->event) && !handle->event->pending_kill)
27 handle->event->pending_kill = POLL_IN;
28
29 irq_work_queue(&handle->event->pending_irq);
30}
31
32/*
33 * We need to ensure a later event_id doesn't publish a head when a former
34 * event isn't done writing. However since we need to deal with NMIs we
35 * cannot fully serialize things.
36 *
37 * We only publish the head (and generate a wakeup) when the outer-most
38 * event completes.
39 */
40static void perf_output_get_handle(struct perf_output_handle *handle)
41{
42 struct perf_buffer *rb = handle->rb;
43
44 preempt_disable();
45
46 /*
47 * Avoid an explicit LOAD/STORE such that architectures with memops
48 * can use them.
49 */
50 (*(volatile unsigned int *)&rb->nest)++;
51 handle->wakeup = local_read(&rb->wakeup);
52}
53
54static void perf_output_put_handle(struct perf_output_handle *handle)
55{
56 struct perf_buffer *rb = handle->rb;
57 unsigned long head;
58 unsigned int nest;
59
60 /*
61 * If this isn't the outermost nesting, we don't have to update
62 * @rb->user_page->data_head.
63 */
64 nest = READ_ONCE(rb->nest);
65 if (nest > 1) {
66 WRITE_ONCE(rb->nest, nest - 1);
67 goto out;
68 }
69
70again:
71 /*
72 * In order to avoid publishing a head value that goes backwards,
73 * we must ensure the load of @rb->head happens after we've
74 * incremented @rb->nest.
75 *
76 * Otherwise we can observe a @rb->head value before one published
77 * by an IRQ/NMI happening between the load and the increment.
78 */
79 barrier();
80 head = local_read(&rb->head);
81
82 /*
83 * IRQ/NMI can happen here and advance @rb->head, causing our
84 * load above to be stale.
85 */
86
87 /*
88 * Since the mmap() consumer (userspace) can run on a different CPU:
89 *
90 * kernel user
91 *
92 * if (LOAD ->data_tail) { LOAD ->data_head
93 * (A) smp_rmb() (C)
94 * STORE $data LOAD $data
95 * smp_wmb() (B) smp_mb() (D)
96 * STORE ->data_head STORE ->data_tail
97 * }
98 *
99 * Where A pairs with D, and B pairs with C.
100 *
101 * In our case (A) is a control dependency that separates the load of
102 * the ->data_tail and the stores of $data. In case ->data_tail
103 * indicates there is no room in the buffer to store $data we do not.
104 *
105 * D needs to be a full barrier since it separates the data READ
106 * from the tail WRITE.
107 *
108 * For B a WMB is sufficient since it separates two WRITEs, and for C
109 * an RMB is sufficient since it separates two READs.
110 *
111 * See perf_output_begin().
112 */
113 smp_wmb(); /* B, matches C */
114 WRITE_ONCE(rb->user_page->data_head, head);
115
116 /*
117 * We must publish the head before decrementing the nest count,
118 * otherwise an IRQ/NMI can publish a more recent head value and our
119 * write will (temporarily) publish a stale value.
120 */
121 barrier();
122 WRITE_ONCE(rb->nest, 0);
123
124 /*
125 * Ensure we decrement @rb->nest before we validate the @rb->head.
126 * Otherwise we cannot be sure we caught the 'last' nested update.
127 */
128 barrier();
129 if (unlikely(head != local_read(&rb->head))) {
130 WRITE_ONCE(rb->nest, 1);
131 goto again;
132 }
133
134 if (handle->wakeup != local_read(&rb->wakeup))
135 perf_output_wakeup(handle);
136
137out:
138 preempt_enable();
139}
140
141static __always_inline bool
142ring_buffer_has_space(unsigned long head, unsigned long tail,
143 unsigned long data_size, unsigned int size,
144 bool backward)
145{
146 if (!backward)
147 return CIRC_SPACE(head, tail, data_size) >= size;
148 else
149 return CIRC_SPACE(tail, head, data_size) >= size;
150}
151
152static __always_inline int
153__perf_output_begin(struct perf_output_handle *handle,
154 struct perf_sample_data *data,
155 struct perf_event *event, unsigned int size,
156 bool backward)
157{
158 struct perf_buffer *rb;
159 unsigned long tail, offset, head;
160 int have_lost, page_shift;
161 struct {
162 struct perf_event_header header;
163 u64 id;
164 u64 lost;
165 } lost_event;
166
167 rcu_read_lock();
168 /*
169 * For inherited events we send all the output towards the parent.
170 */
171 if (event->parent)
172 event = event->parent;
173
174 rb = rcu_dereference(event->rb);
175 if (unlikely(!rb))
176 goto out;
177
178 if (unlikely(rb->paused)) {
179 if (rb->nr_pages) {
180 local_inc(&rb->lost);
181 atomic64_inc(&event->lost_samples);
182 }
183 goto out;
184 }
185
186 handle->rb = rb;
187 handle->event = event;
188 handle->flags = 0;
189
190 have_lost = local_read(&rb->lost);
191 if (unlikely(have_lost)) {
192 size += sizeof(lost_event);
193 if (event->attr.sample_id_all)
194 size += event->id_header_size;
195 }
196
197 perf_output_get_handle(handle);
198
199 offset = local_read(&rb->head);
200 do {
201 head = offset;
202 tail = READ_ONCE(rb->user_page->data_tail);
203 if (!rb->overwrite) {
204 if (unlikely(!ring_buffer_has_space(head, tail,
205 perf_data_size(rb),
206 size, backward)))
207 goto fail;
208 }
209
210 /*
211 * The above forms a control dependency barrier separating the
212 * @tail load above from the data stores below. Since the @tail
213 * load is required to compute the branch to fail below.
214 *
215 * A, matches D; the full memory barrier userspace SHOULD issue
216 * after reading the data and before storing the new tail
217 * position.
218 *
219 * See perf_output_put_handle().
220 */
221
222 if (!backward)
223 head += size;
224 else
225 head -= size;
226 } while (!local_try_cmpxchg(&rb->head, &offset, head));
227
228 if (backward) {
229 offset = head;
230 head = (u64)(-head);
231 }
232
233 /*
234 * We rely on the implied barrier() by local_cmpxchg() to ensure
235 * none of the data stores below can be lifted up by the compiler.
236 */
237
238 if (unlikely(head - local_read(&rb->wakeup) > rb->watermark))
239 local_add(rb->watermark, &rb->wakeup);
240
241 page_shift = PAGE_SHIFT + page_order(rb);
242
243 handle->page = (offset >> page_shift) & (rb->nr_pages - 1);
244 offset &= (1UL << page_shift) - 1;
245 handle->addr = rb->data_pages[handle->page] + offset;
246 handle->size = (1UL << page_shift) - offset;
247
248 if (unlikely(have_lost)) {
249 lost_event.id = event->id;
250 lost_event.lost = local_xchg(&rb->lost, 0);
251
252 /* XXX mostly redundant; @data is already fully initializes */
253 perf_event_header__init(&lost_event.header, data,
254 PERF_RECORD_LOST, /* misc= */ 0,
255 sizeof(lost_event), event);
256 perf_output_put(handle, lost_event);
257 perf_event__output_id_sample(event, handle, data);
258 }
259
260 return 0;
261
262fail:
263 local_inc(&rb->lost);
264 atomic64_inc(&event->lost_samples);
265 perf_output_put_handle(handle);
266out:
267 rcu_read_unlock();
268
269 return -ENOSPC;
270}
271
272int perf_output_begin_forward(struct perf_output_handle *handle,
273 struct perf_sample_data *data,
274 struct perf_event *event, unsigned int size)
275{
276 return __perf_output_begin(handle, data, event, size, false);
277}
278
279int perf_output_begin_backward(struct perf_output_handle *handle,
280 struct perf_sample_data *data,
281 struct perf_event *event, unsigned int size)
282{
283 return __perf_output_begin(handle, data, event, size, true);
284}
285
286int perf_output_begin(struct perf_output_handle *handle,
287 struct perf_sample_data *data,
288 struct perf_event *event, unsigned int size)
289{
290
291 return __perf_output_begin(handle, data, event, size,
292 unlikely(is_write_backward(event)));
293}
294
295unsigned int perf_output_copy(struct perf_output_handle *handle,
296 const void *buf, unsigned int len)
297{
298 return __output_copy(handle, buf, len);
299}
300
301unsigned int perf_output_skip(struct perf_output_handle *handle,
302 unsigned int len)
303{
304 return __output_skip(handle, NULL, len);
305}
306
307void perf_output_end(struct perf_output_handle *handle)
308{
309 perf_output_put_handle(handle);
310 rcu_read_unlock();
311}
312
313static void
314ring_buffer_init(struct perf_buffer *rb, long watermark, int flags)
315{
316 long max_size = perf_data_size(rb);
317
318 if (watermark)
319 rb->watermark = min(max_size, watermark);
320
321 if (!rb->watermark)
322 rb->watermark = max_size / 2;
323
324 if (flags & RING_BUFFER_WRITABLE)
325 rb->overwrite = 0;
326 else
327 rb->overwrite = 1;
328
329 refcount_set(&rb->refcount, 1);
330
331 INIT_LIST_HEAD(&rb->event_list);
332 spin_lock_init(&rb->event_lock);
333
334 /*
335 * perf_output_begin() only checks rb->paused, therefore
336 * rb->paused must be true if we have no pages for output.
337 */
338 if (!rb->nr_pages)
339 rb->paused = 1;
340
341 mutex_init(&rb->aux_mutex);
342 rb->mmap_user = get_current_user();
343 refcount_set(&rb->mmap_count, 1);
344}
345
346void perf_aux_output_flag(struct perf_output_handle *handle, u64 flags)
347{
348 /*
349 * OVERWRITE is determined by perf_aux_output_end() and can't
350 * be passed in directly.
351 */
352 if (WARN_ON_ONCE(flags & PERF_AUX_FLAG_OVERWRITE))
353 return;
354
355 handle->aux_flags |= flags;
356}
357EXPORT_SYMBOL_GPL(perf_aux_output_flag);
358
359/*
360 * This is called before hardware starts writing to the AUX area to
361 * obtain an output handle and make sure there's room in the buffer.
362 * When the capture completes, call perf_aux_output_end() to commit
363 * the recorded data to the buffer.
364 *
365 * The ordering is similar to that of perf_output_{begin,end}, with
366 * the exception of (B), which should be taken care of by the pmu
367 * driver, since ordering rules will differ depending on hardware.
368 *
369 * Call this from pmu::start(); see the comment in perf_aux_output_end()
370 * about its use in pmu callbacks. Both can also be called from the PMI
371 * handler if needed.
372 */
373void *perf_aux_output_begin(struct perf_output_handle *handle,
374 struct perf_event *event)
375{
376 struct perf_event *output_event = event;
377 unsigned long aux_head, aux_tail;
378 struct perf_buffer *rb;
379 unsigned int nest;
380
381 if (output_event->parent)
382 output_event = output_event->parent;
383
384 /*
385 * Since this will typically be open across pmu::add/pmu::del, we
386 * grab ring_buffer's refcount instead of holding rcu read lock
387 * to make sure it doesn't disappear under us.
388 */
389 rb = ring_buffer_get(output_event);
390 if (!rb)
391 return NULL;
392
393 if (!rb_has_aux(rb))
394 goto err;
395
396 /*
397 * If aux_mmap_count is zero, the aux buffer is in perf_mmap_close(),
398 * about to get freed, so we leave immediately.
399 *
400 * Checking rb::aux_mmap_count and rb::refcount has to be done in
401 * the same order, see perf_mmap_close. Otherwise we end up freeing
402 * aux pages in this path, which is a bug, because in_atomic().
403 */
404 if (!refcount_read(&rb->aux_mmap_count))
405 goto err;
406
407 if (!refcount_inc_not_zero(&rb->aux_refcount))
408 goto err;
409
410 nest = READ_ONCE(rb->aux_nest);
411 /*
412 * Nesting is not supported for AUX area, make sure nested
413 * writers are caught early
414 */
415 if (WARN_ON_ONCE(nest))
416 goto err_put;
417
418 WRITE_ONCE(rb->aux_nest, nest + 1);
419
420 aux_head = rb->aux_head;
421
422 handle->rb = rb;
423 handle->event = event;
424 handle->head = aux_head;
425 handle->size = 0;
426 handle->aux_flags = 0;
427
428 /*
429 * In overwrite mode, AUX data stores do not depend on aux_tail,
430 * therefore (A) control dependency barrier does not exist. The
431 * (B) <-> (C) ordering is still observed by the pmu driver.
432 */
433 if (!rb->aux_overwrite) {
434 aux_tail = READ_ONCE(rb->user_page->aux_tail);
435 handle->wakeup = rb->aux_wakeup + rb->aux_watermark;
436 if (aux_head - aux_tail < perf_aux_size(rb))
437 handle->size = CIRC_SPACE(aux_head, aux_tail, perf_aux_size(rb));
438
439 /*
440 * handle->size computation depends on aux_tail load; this forms a
441 * control dependency barrier separating aux_tail load from aux data
442 * store that will be enabled on successful return
443 */
444 if (!handle->size) { /* A, matches D */
445 perf_event_disable_inatomic(handle->event);
446 perf_output_wakeup(handle);
447 WRITE_ONCE(rb->aux_nest, 0);
448 goto err_put;
449 }
450 }
451
452 return handle->rb->aux_priv;
453
454err_put:
455 /* can't be last */
456 rb_free_aux(rb);
457
458err:
459 ring_buffer_put(rb);
460 handle->event = NULL;
461
462 return NULL;
463}
464EXPORT_SYMBOL_GPL(perf_aux_output_begin);
465
466static __always_inline bool rb_need_aux_wakeup(struct perf_buffer *rb)
467{
468 if (rb->aux_overwrite)
469 return false;
470
471 if (rb->aux_head - rb->aux_wakeup >= rb->aux_watermark) {
472 rb->aux_wakeup = rounddown(rb->aux_head, rb->aux_watermark);
473 return true;
474 }
475
476 return false;
477}
478
479/*
480 * Commit the data written by hardware into the ring buffer by adjusting
481 * aux_head and posting a PERF_RECORD_AUX into the perf buffer. It is the
482 * pmu driver's responsibility to observe ordering rules of the hardware,
483 * so that all the data is externally visible before this is called.
484 *
485 * Note: this has to be called from pmu::stop() callback, as the assumption
486 * of the AUX buffer management code is that after pmu::stop(), the AUX
487 * transaction must be stopped and therefore drop the AUX reference count.
488 */
489void perf_aux_output_end(struct perf_output_handle *handle, unsigned long size)
490{
491 bool wakeup = !!(handle->aux_flags & PERF_AUX_FLAG_TRUNCATED);
492 struct perf_buffer *rb = handle->rb;
493 unsigned long aux_head;
494
495 /* in overwrite mode, driver provides aux_head via handle */
496 if (rb->aux_overwrite) {
497 handle->aux_flags |= PERF_AUX_FLAG_OVERWRITE;
498
499 aux_head = handle->head;
500 rb->aux_head = aux_head;
501 } else {
502 handle->aux_flags &= ~PERF_AUX_FLAG_OVERWRITE;
503
504 aux_head = rb->aux_head;
505 rb->aux_head += size;
506 }
507
508 /*
509 * Only send RECORD_AUX if we have something useful to communicate
510 *
511 * PMU_FORMAT bits identify the PMU type rather than an AUX event
512 * has occurred, so ignore them for zero-sized records.
513 *
514 * The OVERWRITE records by themselves are not considered
515 * useful, as they don't communicate any *new* information,
516 * aside from the short-lived offset, that becomes history at
517 * the next event sched-in and therefore isn't useful.
518 * The userspace that needs to copy out AUX data in overwrite
519 * mode should know to use user_page::aux_head for the actual
520 * offset. So, from now on we don't output AUX records that
521 * have *only* OVERWRITE flag set.
522 */
523 if (size ||
524 (handle->aux_flags & ~(u64)(PERF_AUX_FLAG_PMU_FORMAT_TYPE_MASK |
525 PERF_AUX_FLAG_OVERWRITE)))
526 perf_event_aux_event(handle->event, aux_head, size,
527 handle->aux_flags);
528
529 WRITE_ONCE(rb->user_page->aux_head, rb->aux_head);
530 if (rb_need_aux_wakeup(rb))
531 wakeup = true;
532
533 if (wakeup) {
534 if (handle->aux_flags & PERF_AUX_FLAG_TRUNCATED)
535 perf_event_disable_inatomic(handle->event);
536 perf_output_wakeup(handle);
537 }
538
539 handle->event = NULL;
540
541 WRITE_ONCE(rb->aux_nest, 0);
542 /* can't be last */
543 rb_free_aux(rb);
544 ring_buffer_put(rb);
545}
546EXPORT_SYMBOL_GPL(perf_aux_output_end);
547
548/*
549 * Skip over a given number of bytes in the AUX buffer, due to, for example,
550 * hardware's alignment constraints.
551 */
552int perf_aux_output_skip(struct perf_output_handle *handle, unsigned long size)
553{
554 struct perf_buffer *rb = handle->rb;
555
556 if (size > handle->size)
557 return -ENOSPC;
558
559 rb->aux_head += size;
560
561 WRITE_ONCE(rb->user_page->aux_head, rb->aux_head);
562 if (rb_need_aux_wakeup(rb)) {
563 perf_output_wakeup(handle);
564 handle->wakeup = rb->aux_wakeup + rb->aux_watermark;
565 }
566
567 handle->head = rb->aux_head;
568 handle->size -= size;
569
570 return 0;
571}
572EXPORT_SYMBOL_GPL(perf_aux_output_skip);
573
574void *perf_get_aux(struct perf_output_handle *handle)
575{
576 /* this is only valid between perf_aux_output_begin and *_end */
577 if (!handle->event)
578 return NULL;
579
580 return handle->rb->aux_priv;
581}
582EXPORT_SYMBOL_GPL(perf_get_aux);
583
584/*
585 * Copy out AUX data from an AUX handle.
586 */
587long perf_output_copy_aux(struct perf_output_handle *aux_handle,
588 struct perf_output_handle *handle,
589 unsigned long from, unsigned long to)
590{
591 struct perf_buffer *rb = aux_handle->rb;
592 unsigned long tocopy, remainder, len = 0;
593 void *addr;
594
595 from &= (rb->aux_nr_pages << PAGE_SHIFT) - 1;
596 to &= (rb->aux_nr_pages << PAGE_SHIFT) - 1;
597
598 do {
599 tocopy = PAGE_SIZE - offset_in_page(from);
600 if (to > from)
601 tocopy = min(tocopy, to - from);
602 if (!tocopy)
603 break;
604
605 addr = rb->aux_pages[from >> PAGE_SHIFT];
606 addr += offset_in_page(from);
607
608 remainder = perf_output_copy(handle, addr, tocopy);
609 if (remainder)
610 return -EFAULT;
611
612 len += tocopy;
613 from += tocopy;
614 from &= (rb->aux_nr_pages << PAGE_SHIFT) - 1;
615 } while (to != from);
616
617 return len;
618}
619
620#define PERF_AUX_GFP (GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_NORETRY)
621
622static struct page *rb_alloc_aux_page(int node, int order)
623{
624 struct page *page;
625
626 if (order > MAX_PAGE_ORDER)
627 order = MAX_PAGE_ORDER;
628
629 do {
630 page = alloc_pages_node(node, PERF_AUX_GFP, order);
631 } while (!page && order--);
632
633 if (page && order) {
634 /*
635 * Communicate the allocation size to the driver:
636 * if we managed to secure a high-order allocation,
637 * set its first page's private to this order;
638 * !PagePrivate(page) means it's just a normal page.
639 */
640 split_page(page, order);
641 SetPagePrivate(page);
642 set_page_private(page, order);
643 }
644
645 return page;
646}
647
648static void rb_free_aux_page(struct perf_buffer *rb, int idx)
649{
650 struct page *page = virt_to_page(rb->aux_pages[idx]);
651
652 ClearPagePrivate(page);
653 __free_page(page);
654}
655
656static void __rb_free_aux(struct perf_buffer *rb)
657{
658 int pg;
659
660 /*
661 * Should never happen, the last reference should be dropped from
662 * perf_mmap_close() path, which first stops aux transactions (which
663 * in turn are the atomic holders of aux_refcount) and then does the
664 * last rb_free_aux().
665 */
666 WARN_ON_ONCE(in_atomic());
667
668 if (rb->aux_priv) {
669 rb->free_aux(rb->aux_priv);
670 rb->free_aux = NULL;
671 rb->aux_priv = NULL;
672 }
673
674 if (rb->aux_nr_pages) {
675 for (pg = 0; pg < rb->aux_nr_pages; pg++)
676 rb_free_aux_page(rb, pg);
677
678 kfree(rb->aux_pages);
679 rb->aux_nr_pages = 0;
680 }
681}
682
683int rb_alloc_aux(struct perf_buffer *rb, struct perf_event *event,
684 pgoff_t pgoff, int nr_pages, long watermark, int flags)
685{
686 bool overwrite = !(flags & RING_BUFFER_WRITABLE);
687 int node = (event->cpu == -1) ? -1 : cpu_to_node(event->cpu);
688 bool use_contiguous_pages = event->pmu->capabilities & (
689 PERF_PMU_CAP_AUX_NO_SG | PERF_PMU_CAP_AUX_PREFER_LARGE);
690 /*
691 * Initialize max_order to 0 for page allocation. This allocates single
692 * pages to minimize memory fragmentation. This is overridden if the
693 * PMU needs or prefers contiguous pages (use_contiguous_pages = true).
694 */
695 int max_order = 0;
696 int ret = -ENOMEM;
697
698 if (!has_aux(event))
699 return -EOPNOTSUPP;
700
701 if (nr_pages <= 0)
702 return -EINVAL;
703
704 if (!overwrite) {
705 /*
706 * Watermark defaults to half the buffer, to aid PMU drivers
707 * in double buffering.
708 */
709 if (!watermark)
710 watermark = min_t(unsigned long,
711 U32_MAX,
712 (unsigned long)nr_pages << (PAGE_SHIFT - 1));
713
714 /*
715 * If using contiguous pages, use aux_watermark as the basis
716 * for chunking to help PMU drivers honor the watermark.
717 */
718 if (use_contiguous_pages)
719 max_order = get_order(watermark);
720 } else {
721 /*
722 * If using contiguous pages, we need to start with the
723 * max_order that fits in nr_pages, not the other way around,
724 * hence ilog2() and not get_order.
725 */
726 if (use_contiguous_pages)
727 max_order = ilog2(nr_pages);
728 watermark = 0;
729 }
730
731 /*
732 * kcalloc_node() is unable to allocate buffer if the size is larger
733 * than: PAGE_SIZE << MAX_PAGE_ORDER; directly bail out in this case.
734 */
735 if (get_order((unsigned long)nr_pages * sizeof(void *)) > MAX_PAGE_ORDER)
736 return -ENOMEM;
737 rb->aux_pages = kcalloc_node(nr_pages, sizeof(void *), GFP_KERNEL,
738 node);
739 if (!rb->aux_pages)
740 return -ENOMEM;
741
742 rb->free_aux = event->pmu->free_aux;
743 for (rb->aux_nr_pages = 0; rb->aux_nr_pages < nr_pages;) {
744 struct page *page;
745 int last, order;
746
747 order = min(max_order, ilog2(nr_pages - rb->aux_nr_pages));
748 page = rb_alloc_aux_page(node, order);
749 if (!page)
750 goto out;
751
752 for (last = rb->aux_nr_pages + (1 << page_private(page));
753 last > rb->aux_nr_pages; rb->aux_nr_pages++)
754 rb->aux_pages[rb->aux_nr_pages] = page_address(page++);
755 }
756
757 /*
758 * In overwrite mode, PMUs that don't support SG may not handle more
759 * than one contiguous allocation, since they rely on PMI to do double
760 * buffering. In this case, the entire buffer has to be one contiguous
761 * chunk.
762 */
763 if ((event->pmu->capabilities & PERF_PMU_CAP_AUX_NO_SG) &&
764 overwrite) {
765 struct page *page = virt_to_page(rb->aux_pages[0]);
766
767 if (page_private(page) != max_order)
768 goto out;
769 }
770
771 rb->aux_priv = event->pmu->setup_aux(event, rb->aux_pages, nr_pages,
772 overwrite);
773 if (!rb->aux_priv)
774 goto out;
775
776 ret = 0;
777
778 /*
779 * aux_pages (and pmu driver's private data, aux_priv) will be
780 * referenced in both producer's and consumer's contexts, thus
781 * we keep a refcount here to make sure either of the two can
782 * reference them safely.
783 */
784 refcount_set(&rb->aux_refcount, 1);
785
786 rb->aux_overwrite = overwrite;
787 rb->aux_watermark = watermark;
788
789out:
790 if (!ret)
791 rb->aux_pgoff = pgoff;
792 else
793 __rb_free_aux(rb);
794
795 return ret;
796}
797
798void rb_free_aux(struct perf_buffer *rb)
799{
800 if (refcount_dec_and_test(&rb->aux_refcount))
801 __rb_free_aux(rb);
802}
803
804#ifndef CONFIG_PERF_USE_VMALLOC
805
806/*
807 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
808 */
809
810static struct page *
811__perf_mmap_to_page(struct perf_buffer *rb, unsigned long pgoff)
812{
813 if (pgoff > rb->nr_pages)
814 return NULL;
815
816 if (pgoff == 0)
817 return virt_to_page(rb->user_page);
818
819 return virt_to_page(rb->data_pages[pgoff - 1]);
820}
821
822static void *perf_mmap_alloc_page(int cpu)
823{
824 struct page *page;
825 int node;
826
827 node = (cpu == -1) ? cpu : cpu_to_node(cpu);
828 page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
829 if (!page)
830 return NULL;
831
832 return page_address(page);
833}
834
835static void perf_mmap_free_page(void *addr)
836{
837 struct page *page = virt_to_page(addr);
838
839 __free_page(page);
840}
841
842struct perf_buffer *rb_alloc(int nr_pages, long watermark, int cpu, int flags)
843{
844 struct perf_buffer *rb;
845 unsigned long size;
846 int i, node;
847
848 size = sizeof(struct perf_buffer);
849 size += nr_pages * sizeof(void *);
850
851 if (order_base_2(size) > PAGE_SHIFT+MAX_PAGE_ORDER)
852 goto fail;
853
854 node = (cpu == -1) ? cpu : cpu_to_node(cpu);
855 rb = kzalloc_node(size, GFP_KERNEL, node);
856 if (!rb)
857 goto fail;
858
859 rb->user_page = perf_mmap_alloc_page(cpu);
860 if (!rb->user_page)
861 goto fail_user_page;
862
863 for (i = 0; i < nr_pages; i++) {
864 rb->data_pages[i] = perf_mmap_alloc_page(cpu);
865 if (!rb->data_pages[i])
866 goto fail_data_pages;
867 }
868
869 rb->nr_pages = nr_pages;
870
871 ring_buffer_init(rb, watermark, flags);
872
873 return rb;
874
875fail_data_pages:
876 for (i--; i >= 0; i--)
877 perf_mmap_free_page(rb->data_pages[i]);
878
879 perf_mmap_free_page(rb->user_page);
880
881fail_user_page:
882 kfree(rb);
883
884fail:
885 return NULL;
886}
887
888void rb_free(struct perf_buffer *rb)
889{
890 int i;
891
892 perf_mmap_free_page(rb->user_page);
893 for (i = 0; i < rb->nr_pages; i++)
894 perf_mmap_free_page(rb->data_pages[i]);
895 kfree(rb);
896}
897
898#else
899static struct page *
900__perf_mmap_to_page(struct perf_buffer *rb, unsigned long pgoff)
901{
902 /* The '>' counts in the user page. */
903 if (pgoff > data_page_nr(rb))
904 return NULL;
905
906 return vmalloc_to_page((void *)rb->user_page + pgoff * PAGE_SIZE);
907}
908
909static void rb_free_work(struct work_struct *work)
910{
911 struct perf_buffer *rb;
912
913 rb = container_of(work, struct perf_buffer, work);
914
915 vfree(rb->user_page);
916 kfree(rb);
917}
918
919void rb_free(struct perf_buffer *rb)
920{
921 schedule_work(&rb->work);
922}
923
924struct perf_buffer *rb_alloc(int nr_pages, long watermark, int cpu, int flags)
925{
926 struct perf_buffer *rb;
927 unsigned long size;
928 void *all_buf;
929 int node;
930
931 size = sizeof(struct perf_buffer);
932 size += sizeof(void *);
933
934 node = (cpu == -1) ? cpu : cpu_to_node(cpu);
935 rb = kzalloc_node(size, GFP_KERNEL, node);
936 if (!rb)
937 goto fail;
938
939 INIT_WORK(&rb->work, rb_free_work);
940
941 all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
942 if (!all_buf)
943 goto fail_all_buf;
944
945 rb->user_page = all_buf;
946 rb->data_pages[0] = all_buf + PAGE_SIZE;
947 if (nr_pages) {
948 rb->nr_pages = 1;
949 rb->page_order = ilog2(nr_pages);
950 }
951
952 ring_buffer_init(rb, watermark, flags);
953
954 return rb;
955
956fail_all_buf:
957 kfree(rb);
958
959fail:
960 return NULL;
961}
962
963#endif
964
965struct page *
966perf_mmap_to_page(struct perf_buffer *rb, unsigned long pgoff)
967{
968 if (rb->aux_nr_pages) {
969 /* above AUX space */
970 if (pgoff > rb->aux_pgoff + rb->aux_nr_pages)
971 return NULL;
972
973 /* AUX space */
974 if (pgoff >= rb->aux_pgoff) {
975 int aux_pgoff = array_index_nospec(pgoff - rb->aux_pgoff, rb->aux_nr_pages);
976 return virt_to_page(rb->aux_pages[aux_pgoff]);
977 }
978 }
979
980 return __perf_mmap_to_page(rb, pgoff);
981}