Loading...
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
7 * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
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>
15#include <linux/circ_buf.h>
16#include <linux/poll.h>
17
18#include "internal.h"
19
20static void perf_output_wakeup(struct perf_output_handle *handle)
21{
22 atomic_set(&handle->rb->poll, POLLIN);
23
24 handle->event->pending_wakeup = 1;
25 irq_work_queue(&handle->event->pending);
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 /*
61 * Since the mmap() consumer (userspace) can run on a different CPU:
62 *
63 * kernel user
64 *
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 * }
71 *
72 * Where A pairs with D, and B pairs with C.
73 *
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.
77 *
78 * D needs to be a full barrier since it separates the data READ
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().
85 */
86 smp_wmb(); /* B, matches C */
87 rb->user_page->data_head = head;
88
89 /*
90 * Now check if we missed an update -- rely on previous implied
91 * compiler barriers to force a re-read.
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
105static bool __always_inline
106ring_buffer_has_space(unsigned long head, unsigned long tail,
107 unsigned long data_size, unsigned int size,
108 bool backward)
109{
110 if (!backward)
111 return CIRC_SPACE(head, tail, data_size) >= size;
112 else
113 return CIRC_SPACE(tail, head, data_size) >= size;
114}
115
116static int __always_inline
117__perf_output_begin(struct perf_output_handle *handle,
118 struct perf_event *event, unsigned int size,
119 bool backward)
120{
121 struct ring_buffer *rb;
122 unsigned long tail, offset, head;
123 int have_lost, page_shift;
124 struct {
125 struct perf_event_header header;
126 u64 id;
127 u64 lost;
128 } lost_event;
129
130 rcu_read_lock();
131 /*
132 * For inherited events we send all the output towards the parent.
133 */
134 if (event->parent)
135 event = event->parent;
136
137 rb = rcu_dereference(event->rb);
138 if (unlikely(!rb))
139 goto out;
140
141 if (unlikely(rb->paused)) {
142 if (rb->nr_pages)
143 local_inc(&rb->lost);
144 goto out;
145 }
146
147 handle->rb = rb;
148 handle->event = event;
149
150 have_lost = local_read(&rb->lost);
151 if (unlikely(have_lost)) {
152 size += sizeof(lost_event);
153 if (event->attr.sample_id_all)
154 size += event->id_header_size;
155 }
156
157 perf_output_get_handle(handle);
158
159 do {
160 tail = READ_ONCE(rb->user_page->data_tail);
161 offset = head = local_read(&rb->head);
162 if (!rb->overwrite) {
163 if (unlikely(!ring_buffer_has_space(head, tail,
164 perf_data_size(rb),
165 size, backward)))
166 goto fail;
167 }
168
169 /*
170 * The above forms a control dependency barrier separating the
171 * @tail load above from the data stores below. Since the @tail
172 * load is required to compute the branch to fail below.
173 *
174 * A, matches D; the full memory barrier userspace SHOULD issue
175 * after reading the data and before storing the new tail
176 * position.
177 *
178 * See perf_output_put_handle().
179 */
180
181 if (!backward)
182 head += size;
183 else
184 head -= size;
185 } while (local_cmpxchg(&rb->head, offset, head) != offset);
186
187 if (backward) {
188 offset = head;
189 head = (u64)(-head);
190 }
191
192 /*
193 * We rely on the implied barrier() by local_cmpxchg() to ensure
194 * none of the data stores below can be lifted up by the compiler.
195 */
196
197 if (unlikely(head - local_read(&rb->wakeup) > rb->watermark))
198 local_add(rb->watermark, &rb->wakeup);
199
200 page_shift = PAGE_SHIFT + page_order(rb);
201
202 handle->page = (offset >> page_shift) & (rb->nr_pages - 1);
203 offset &= (1UL << page_shift) - 1;
204 handle->addr = rb->data_pages[handle->page] + offset;
205 handle->size = (1UL << page_shift) - offset;
206
207 if (unlikely(have_lost)) {
208 struct perf_sample_data sample_data;
209
210 lost_event.header.size = sizeof(lost_event);
211 lost_event.header.type = PERF_RECORD_LOST;
212 lost_event.header.misc = 0;
213 lost_event.id = event->id;
214 lost_event.lost = local_xchg(&rb->lost, 0);
215
216 perf_event_header__init_id(&lost_event.header,
217 &sample_data, event);
218 perf_output_put(handle, lost_event);
219 perf_event__output_id_sample(event, handle, &sample_data);
220 }
221
222 return 0;
223
224fail:
225 local_inc(&rb->lost);
226 perf_output_put_handle(handle);
227out:
228 rcu_read_unlock();
229
230 return -ENOSPC;
231}
232
233int perf_output_begin_forward(struct perf_output_handle *handle,
234 struct perf_event *event, unsigned int size)
235{
236 return __perf_output_begin(handle, event, size, false);
237}
238
239int perf_output_begin_backward(struct perf_output_handle *handle,
240 struct perf_event *event, unsigned int size)
241{
242 return __perf_output_begin(handle, event, size, true);
243}
244
245int perf_output_begin(struct perf_output_handle *handle,
246 struct perf_event *event, unsigned int size)
247{
248
249 return __perf_output_begin(handle, event, size,
250 unlikely(is_write_backward(event)));
251}
252
253unsigned int perf_output_copy(struct perf_output_handle *handle,
254 const void *buf, unsigned int len)
255{
256 return __output_copy(handle, buf, len);
257}
258
259unsigned int perf_output_skip(struct perf_output_handle *handle,
260 unsigned int len)
261{
262 return __output_skip(handle, NULL, len);
263}
264
265void perf_output_end(struct perf_output_handle *handle)
266{
267 perf_output_put_handle(handle);
268 rcu_read_unlock();
269}
270
271static void
272ring_buffer_init(struct ring_buffer *rb, long watermark, int flags)
273{
274 long max_size = perf_data_size(rb);
275
276 if (watermark)
277 rb->watermark = min(max_size, watermark);
278
279 if (!rb->watermark)
280 rb->watermark = max_size / 2;
281
282 if (flags & RING_BUFFER_WRITABLE)
283 rb->overwrite = 0;
284 else
285 rb->overwrite = 1;
286
287 atomic_set(&rb->refcount, 1);
288
289 INIT_LIST_HEAD(&rb->event_list);
290 spin_lock_init(&rb->event_lock);
291
292 /*
293 * perf_output_begin() only checks rb->paused, therefore
294 * rb->paused must be true if we have no pages for output.
295 */
296 if (!rb->nr_pages)
297 rb->paused = 1;
298}
299
300/*
301 * This is called before hardware starts writing to the AUX area to
302 * obtain an output handle and make sure there's room in the buffer.
303 * When the capture completes, call perf_aux_output_end() to commit
304 * the recorded data to the buffer.
305 *
306 * The ordering is similar to that of perf_output_{begin,end}, with
307 * the exception of (B), which should be taken care of by the pmu
308 * driver, since ordering rules will differ depending on hardware.
309 *
310 * Call this from pmu::start(); see the comment in perf_aux_output_end()
311 * about its use in pmu callbacks. Both can also be called from the PMI
312 * handler if needed.
313 */
314void *perf_aux_output_begin(struct perf_output_handle *handle,
315 struct perf_event *event)
316{
317 struct perf_event *output_event = event;
318 unsigned long aux_head, aux_tail;
319 struct ring_buffer *rb;
320
321 if (output_event->parent)
322 output_event = output_event->parent;
323
324 /*
325 * Since this will typically be open across pmu::add/pmu::del, we
326 * grab ring_buffer's refcount instead of holding rcu read lock
327 * to make sure it doesn't disappear under us.
328 */
329 rb = ring_buffer_get(output_event);
330 if (!rb)
331 return NULL;
332
333 if (!rb_has_aux(rb))
334 goto err;
335
336 /*
337 * If aux_mmap_count is zero, the aux buffer is in perf_mmap_close(),
338 * about to get freed, so we leave immediately.
339 *
340 * Checking rb::aux_mmap_count and rb::refcount has to be done in
341 * the same order, see perf_mmap_close. Otherwise we end up freeing
342 * aux pages in this path, which is a bug, because in_atomic().
343 */
344 if (!atomic_read(&rb->aux_mmap_count))
345 goto err;
346
347 if (!atomic_inc_not_zero(&rb->aux_refcount))
348 goto err;
349
350 /*
351 * Nesting is not supported for AUX area, make sure nested
352 * writers are caught early
353 */
354 if (WARN_ON_ONCE(local_xchg(&rb->aux_nest, 1)))
355 goto err_put;
356
357 aux_head = local_read(&rb->aux_head);
358
359 handle->rb = rb;
360 handle->event = event;
361 handle->head = aux_head;
362 handle->size = 0;
363
364 /*
365 * In overwrite mode, AUX data stores do not depend on aux_tail,
366 * therefore (A) control dependency barrier does not exist. The
367 * (B) <-> (C) ordering is still observed by the pmu driver.
368 */
369 if (!rb->aux_overwrite) {
370 aux_tail = ACCESS_ONCE(rb->user_page->aux_tail);
371 handle->wakeup = local_read(&rb->aux_wakeup) + rb->aux_watermark;
372 if (aux_head - aux_tail < perf_aux_size(rb))
373 handle->size = CIRC_SPACE(aux_head, aux_tail, perf_aux_size(rb));
374
375 /*
376 * handle->size computation depends on aux_tail load; this forms a
377 * control dependency barrier separating aux_tail load from aux data
378 * store that will be enabled on successful return
379 */
380 if (!handle->size) { /* A, matches D */
381 event->pending_disable = 1;
382 perf_output_wakeup(handle);
383 local_set(&rb->aux_nest, 0);
384 goto err_put;
385 }
386 }
387
388 return handle->rb->aux_priv;
389
390err_put:
391 /* can't be last */
392 rb_free_aux(rb);
393
394err:
395 ring_buffer_put(rb);
396 handle->event = NULL;
397
398 return NULL;
399}
400
401/*
402 * Commit the data written by hardware into the ring buffer by adjusting
403 * aux_head and posting a PERF_RECORD_AUX into the perf buffer. It is the
404 * pmu driver's responsibility to observe ordering rules of the hardware,
405 * so that all the data is externally visible before this is called.
406 *
407 * Note: this has to be called from pmu::stop() callback, as the assumption
408 * of the AUX buffer management code is that after pmu::stop(), the AUX
409 * transaction must be stopped and therefore drop the AUX reference count.
410 */
411void perf_aux_output_end(struct perf_output_handle *handle, unsigned long size,
412 bool truncated)
413{
414 struct ring_buffer *rb = handle->rb;
415 bool wakeup = truncated;
416 unsigned long aux_head;
417 u64 flags = 0;
418
419 if (truncated)
420 flags |= PERF_AUX_FLAG_TRUNCATED;
421
422 /* in overwrite mode, driver provides aux_head via handle */
423 if (rb->aux_overwrite) {
424 flags |= PERF_AUX_FLAG_OVERWRITE;
425
426 aux_head = handle->head;
427 local_set(&rb->aux_head, aux_head);
428 } else {
429 aux_head = local_read(&rb->aux_head);
430 local_add(size, &rb->aux_head);
431 }
432
433 if (size || flags) {
434 /*
435 * Only send RECORD_AUX if we have something useful to communicate
436 */
437
438 perf_event_aux_event(handle->event, aux_head, size, flags);
439 }
440
441 aux_head = rb->user_page->aux_head = local_read(&rb->aux_head);
442
443 if (aux_head - local_read(&rb->aux_wakeup) >= rb->aux_watermark) {
444 wakeup = true;
445 local_add(rb->aux_watermark, &rb->aux_wakeup);
446 }
447
448 if (wakeup) {
449 if (truncated)
450 handle->event->pending_disable = 1;
451 perf_output_wakeup(handle);
452 }
453
454 handle->event = NULL;
455
456 local_set(&rb->aux_nest, 0);
457 /* can't be last */
458 rb_free_aux(rb);
459 ring_buffer_put(rb);
460}
461
462/*
463 * Skip over a given number of bytes in the AUX buffer, due to, for example,
464 * hardware's alignment constraints.
465 */
466int perf_aux_output_skip(struct perf_output_handle *handle, unsigned long size)
467{
468 struct ring_buffer *rb = handle->rb;
469 unsigned long aux_head;
470
471 if (size > handle->size)
472 return -ENOSPC;
473
474 local_add(size, &rb->aux_head);
475
476 aux_head = rb->user_page->aux_head = local_read(&rb->aux_head);
477 if (aux_head - local_read(&rb->aux_wakeup) >= rb->aux_watermark) {
478 perf_output_wakeup(handle);
479 local_add(rb->aux_watermark, &rb->aux_wakeup);
480 handle->wakeup = local_read(&rb->aux_wakeup) +
481 rb->aux_watermark;
482 }
483
484 handle->head = aux_head;
485 handle->size -= size;
486
487 return 0;
488}
489
490void *perf_get_aux(struct perf_output_handle *handle)
491{
492 /* this is only valid between perf_aux_output_begin and *_end */
493 if (!handle->event)
494 return NULL;
495
496 return handle->rb->aux_priv;
497}
498
499#define PERF_AUX_GFP (GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_NORETRY)
500
501static struct page *rb_alloc_aux_page(int node, int order)
502{
503 struct page *page;
504
505 if (order > MAX_ORDER)
506 order = MAX_ORDER;
507
508 do {
509 page = alloc_pages_node(node, PERF_AUX_GFP, order);
510 } while (!page && order--);
511
512 if (page && order) {
513 /*
514 * Communicate the allocation size to the driver:
515 * if we managed to secure a high-order allocation,
516 * set its first page's private to this order;
517 * !PagePrivate(page) means it's just a normal page.
518 */
519 split_page(page, order);
520 SetPagePrivate(page);
521 set_page_private(page, order);
522 }
523
524 return page;
525}
526
527static void rb_free_aux_page(struct ring_buffer *rb, int idx)
528{
529 struct page *page = virt_to_page(rb->aux_pages[idx]);
530
531 ClearPagePrivate(page);
532 page->mapping = NULL;
533 __free_page(page);
534}
535
536static void __rb_free_aux(struct ring_buffer *rb)
537{
538 int pg;
539
540 /*
541 * Should never happen, the last reference should be dropped from
542 * perf_mmap_close() path, which first stops aux transactions (which
543 * in turn are the atomic holders of aux_refcount) and then does the
544 * last rb_free_aux().
545 */
546 WARN_ON_ONCE(in_atomic());
547
548 if (rb->aux_priv) {
549 rb->free_aux(rb->aux_priv);
550 rb->free_aux = NULL;
551 rb->aux_priv = NULL;
552 }
553
554 if (rb->aux_nr_pages) {
555 for (pg = 0; pg < rb->aux_nr_pages; pg++)
556 rb_free_aux_page(rb, pg);
557
558 kfree(rb->aux_pages);
559 rb->aux_nr_pages = 0;
560 }
561}
562
563int rb_alloc_aux(struct ring_buffer *rb, struct perf_event *event,
564 pgoff_t pgoff, int nr_pages, long watermark, int flags)
565{
566 bool overwrite = !(flags & RING_BUFFER_WRITABLE);
567 int node = (event->cpu == -1) ? -1 : cpu_to_node(event->cpu);
568 int ret = -ENOMEM, max_order = 0;
569
570 if (!has_aux(event))
571 return -ENOTSUPP;
572
573 if (event->pmu->capabilities & PERF_PMU_CAP_AUX_NO_SG) {
574 /*
575 * We need to start with the max_order that fits in nr_pages,
576 * not the other way around, hence ilog2() and not get_order.
577 */
578 max_order = ilog2(nr_pages);
579
580 /*
581 * PMU requests more than one contiguous chunks of memory
582 * for SW double buffering
583 */
584 if ((event->pmu->capabilities & PERF_PMU_CAP_AUX_SW_DOUBLEBUF) &&
585 !overwrite) {
586 if (!max_order)
587 return -EINVAL;
588
589 max_order--;
590 }
591 }
592
593 rb->aux_pages = kzalloc_node(nr_pages * sizeof(void *), GFP_KERNEL, node);
594 if (!rb->aux_pages)
595 return -ENOMEM;
596
597 rb->free_aux = event->pmu->free_aux;
598 for (rb->aux_nr_pages = 0; rb->aux_nr_pages < nr_pages;) {
599 struct page *page;
600 int last, order;
601
602 order = min(max_order, ilog2(nr_pages - rb->aux_nr_pages));
603 page = rb_alloc_aux_page(node, order);
604 if (!page)
605 goto out;
606
607 for (last = rb->aux_nr_pages + (1 << page_private(page));
608 last > rb->aux_nr_pages; rb->aux_nr_pages++)
609 rb->aux_pages[rb->aux_nr_pages] = page_address(page++);
610 }
611
612 /*
613 * In overwrite mode, PMUs that don't support SG may not handle more
614 * than one contiguous allocation, since they rely on PMI to do double
615 * buffering. In this case, the entire buffer has to be one contiguous
616 * chunk.
617 */
618 if ((event->pmu->capabilities & PERF_PMU_CAP_AUX_NO_SG) &&
619 overwrite) {
620 struct page *page = virt_to_page(rb->aux_pages[0]);
621
622 if (page_private(page) != max_order)
623 goto out;
624 }
625
626 rb->aux_priv = event->pmu->setup_aux(event->cpu, rb->aux_pages, nr_pages,
627 overwrite);
628 if (!rb->aux_priv)
629 goto out;
630
631 ret = 0;
632
633 /*
634 * aux_pages (and pmu driver's private data, aux_priv) will be
635 * referenced in both producer's and consumer's contexts, thus
636 * we keep a refcount here to make sure either of the two can
637 * reference them safely.
638 */
639 atomic_set(&rb->aux_refcount, 1);
640
641 rb->aux_overwrite = overwrite;
642 rb->aux_watermark = watermark;
643
644 if (!rb->aux_watermark && !rb->aux_overwrite)
645 rb->aux_watermark = nr_pages << (PAGE_SHIFT - 1);
646
647out:
648 if (!ret)
649 rb->aux_pgoff = pgoff;
650 else
651 __rb_free_aux(rb);
652
653 return ret;
654}
655
656void rb_free_aux(struct ring_buffer *rb)
657{
658 if (atomic_dec_and_test(&rb->aux_refcount))
659 __rb_free_aux(rb);
660}
661
662#ifndef CONFIG_PERF_USE_VMALLOC
663
664/*
665 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
666 */
667
668static struct page *
669__perf_mmap_to_page(struct ring_buffer *rb, unsigned long pgoff)
670{
671 if (pgoff > rb->nr_pages)
672 return NULL;
673
674 if (pgoff == 0)
675 return virt_to_page(rb->user_page);
676
677 return virt_to_page(rb->data_pages[pgoff - 1]);
678}
679
680static void *perf_mmap_alloc_page(int cpu)
681{
682 struct page *page;
683 int node;
684
685 node = (cpu == -1) ? cpu : cpu_to_node(cpu);
686 page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
687 if (!page)
688 return NULL;
689
690 return page_address(page);
691}
692
693struct ring_buffer *rb_alloc(int nr_pages, long watermark, int cpu, int flags)
694{
695 struct ring_buffer *rb;
696 unsigned long size;
697 int i;
698
699 size = sizeof(struct ring_buffer);
700 size += nr_pages * sizeof(void *);
701
702 rb = kzalloc(size, GFP_KERNEL);
703 if (!rb)
704 goto fail;
705
706 rb->user_page = perf_mmap_alloc_page(cpu);
707 if (!rb->user_page)
708 goto fail_user_page;
709
710 for (i = 0; i < nr_pages; i++) {
711 rb->data_pages[i] = perf_mmap_alloc_page(cpu);
712 if (!rb->data_pages[i])
713 goto fail_data_pages;
714 }
715
716 rb->nr_pages = nr_pages;
717
718 ring_buffer_init(rb, watermark, flags);
719
720 return rb;
721
722fail_data_pages:
723 for (i--; i >= 0; i--)
724 free_page((unsigned long)rb->data_pages[i]);
725
726 free_page((unsigned long)rb->user_page);
727
728fail_user_page:
729 kfree(rb);
730
731fail:
732 return NULL;
733}
734
735static void perf_mmap_free_page(unsigned long addr)
736{
737 struct page *page = virt_to_page((void *)addr);
738
739 page->mapping = NULL;
740 __free_page(page);
741}
742
743void rb_free(struct ring_buffer *rb)
744{
745 int i;
746
747 perf_mmap_free_page((unsigned long)rb->user_page);
748 for (i = 0; i < rb->nr_pages; i++)
749 perf_mmap_free_page((unsigned long)rb->data_pages[i]);
750 kfree(rb);
751}
752
753#else
754static int data_page_nr(struct ring_buffer *rb)
755{
756 return rb->nr_pages << page_order(rb);
757}
758
759static struct page *
760__perf_mmap_to_page(struct ring_buffer *rb, unsigned long pgoff)
761{
762 /* The '>' counts in the user page. */
763 if (pgoff > data_page_nr(rb))
764 return NULL;
765
766 return vmalloc_to_page((void *)rb->user_page + pgoff * PAGE_SIZE);
767}
768
769static void perf_mmap_unmark_page(void *addr)
770{
771 struct page *page = vmalloc_to_page(addr);
772
773 page->mapping = NULL;
774}
775
776static void rb_free_work(struct work_struct *work)
777{
778 struct ring_buffer *rb;
779 void *base;
780 int i, nr;
781
782 rb = container_of(work, struct ring_buffer, work);
783 nr = data_page_nr(rb);
784
785 base = rb->user_page;
786 /* The '<=' counts in the user page. */
787 for (i = 0; i <= nr; i++)
788 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
789
790 vfree(base);
791 kfree(rb);
792}
793
794void rb_free(struct ring_buffer *rb)
795{
796 schedule_work(&rb->work);
797}
798
799struct ring_buffer *rb_alloc(int nr_pages, long watermark, int cpu, int flags)
800{
801 struct ring_buffer *rb;
802 unsigned long size;
803 void *all_buf;
804
805 size = sizeof(struct ring_buffer);
806 size += sizeof(void *);
807
808 rb = kzalloc(size, GFP_KERNEL);
809 if (!rb)
810 goto fail;
811
812 INIT_WORK(&rb->work, rb_free_work);
813
814 all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
815 if (!all_buf)
816 goto fail_all_buf;
817
818 rb->user_page = all_buf;
819 rb->data_pages[0] = all_buf + PAGE_SIZE;
820 if (nr_pages) {
821 rb->nr_pages = 1;
822 rb->page_order = ilog2(nr_pages);
823 }
824
825 ring_buffer_init(rb, watermark, flags);
826
827 return rb;
828
829fail_all_buf:
830 kfree(rb);
831
832fail:
833 return NULL;
834}
835
836#endif
837
838struct page *
839perf_mmap_to_page(struct ring_buffer *rb, unsigned long pgoff)
840{
841 if (rb->aux_nr_pages) {
842 /* above AUX space */
843 if (pgoff > rb->aux_pgoff + rb->aux_nr_pages)
844 return NULL;
845
846 /* AUX space */
847 if (pgoff >= rb->aux_pgoff)
848 return virt_to_page(rb->aux_pages[pgoff - rb->aux_pgoff]);
849 }
850
851 return __perf_mmap_to_page(rb, pgoff);
852}
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Performance events ring-buffer code:
4 *
5 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
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. <paulus@au1.ibm.com>
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);
23
24 handle->event->pending_wakeup = 1;
25 irq_work_queue(&handle->event->pending_irq);
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 perf_buffer *rb = handle->rb;
39
40 preempt_disable();
41
42 /*
43 * Avoid an explicit LOAD/STORE such that architectures with memops
44 * can use them.
45 */
46 (*(volatile unsigned int *)&rb->nest)++;
47 handle->wakeup = local_read(&rb->wakeup);
48}
49
50static void perf_output_put_handle(struct perf_output_handle *handle)
51{
52 struct perf_buffer *rb = handle->rb;
53 unsigned long head;
54 unsigned int nest;
55
56 /*
57 * If this isn't the outermost nesting, we don't have to update
58 * @rb->user_page->data_head.
59 */
60 nest = READ_ONCE(rb->nest);
61 if (nest > 1) {
62 WRITE_ONCE(rb->nest, nest - 1);
63 goto out;
64 }
65
66again:
67 /*
68 * In order to avoid publishing a head value that goes backwards,
69 * we must ensure the load of @rb->head happens after we've
70 * incremented @rb->nest.
71 *
72 * Otherwise we can observe a @rb->head value before one published
73 * by an IRQ/NMI happening between the load and the increment.
74 */
75 barrier();
76 head = local_read(&rb->head);
77
78 /*
79 * IRQ/NMI can happen here and advance @rb->head, causing our
80 * load above to be stale.
81 */
82
83 /*
84 * Since the mmap() consumer (userspace) can run on a different CPU:
85 *
86 * kernel user
87 *
88 * if (LOAD ->data_tail) { LOAD ->data_head
89 * (A) smp_rmb() (C)
90 * STORE $data LOAD $data
91 * smp_wmb() (B) smp_mb() (D)
92 * STORE ->data_head STORE ->data_tail
93 * }
94 *
95 * Where A pairs with D, and B pairs with C.
96 *
97 * In our case (A) is a control dependency that separates the load of
98 * the ->data_tail and the stores of $data. In case ->data_tail
99 * indicates there is no room in the buffer to store $data we do not.
100 *
101 * D needs to be a full barrier since it separates the data READ
102 * from the tail WRITE.
103 *
104 * For B a WMB is sufficient since it separates two WRITEs, and for C
105 * an RMB is sufficient since it separates two READs.
106 *
107 * See perf_output_begin().
108 */
109 smp_wmb(); /* B, matches C */
110 WRITE_ONCE(rb->user_page->data_head, head);
111
112 /*
113 * We must publish the head before decrementing the nest count,
114 * otherwise an IRQ/NMI can publish a more recent head value and our
115 * write will (temporarily) publish a stale value.
116 */
117 barrier();
118 WRITE_ONCE(rb->nest, 0);
119
120 /*
121 * Ensure we decrement @rb->nest before we validate the @rb->head.
122 * Otherwise we cannot be sure we caught the 'last' nested update.
123 */
124 barrier();
125 if (unlikely(head != local_read(&rb->head))) {
126 WRITE_ONCE(rb->nest, 1);
127 goto again;
128 }
129
130 if (handle->wakeup != local_read(&rb->wakeup))
131 perf_output_wakeup(handle);
132
133out:
134 preempt_enable();
135}
136
137static __always_inline bool
138ring_buffer_has_space(unsigned long head, unsigned long tail,
139 unsigned long data_size, unsigned int size,
140 bool backward)
141{
142 if (!backward)
143 return CIRC_SPACE(head, tail, data_size) >= size;
144 else
145 return CIRC_SPACE(tail, head, data_size) >= size;
146}
147
148static __always_inline int
149__perf_output_begin(struct perf_output_handle *handle,
150 struct perf_sample_data *data,
151 struct perf_event *event, unsigned int size,
152 bool backward)
153{
154 struct perf_buffer *rb;
155 unsigned long tail, offset, head;
156 int have_lost, page_shift;
157 struct {
158 struct perf_event_header header;
159 u64 id;
160 u64 lost;
161 } lost_event;
162
163 rcu_read_lock();
164 /*
165 * For inherited events we send all the output towards the parent.
166 */
167 if (event->parent)
168 event = event->parent;
169
170 rb = rcu_dereference(event->rb);
171 if (unlikely(!rb))
172 goto out;
173
174 if (unlikely(rb->paused)) {
175 if (rb->nr_pages) {
176 local_inc(&rb->lost);
177 atomic64_inc(&event->lost_samples);
178 }
179 goto out;
180 }
181
182 handle->rb = rb;
183 handle->event = event;
184
185 have_lost = local_read(&rb->lost);
186 if (unlikely(have_lost)) {
187 size += sizeof(lost_event);
188 if (event->attr.sample_id_all)
189 size += event->id_header_size;
190 }
191
192 perf_output_get_handle(handle);
193
194 offset = local_read(&rb->head);
195 do {
196 head = offset;
197 tail = READ_ONCE(rb->user_page->data_tail);
198 if (!rb->overwrite) {
199 if (unlikely(!ring_buffer_has_space(head, tail,
200 perf_data_size(rb),
201 size, backward)))
202 goto fail;
203 }
204
205 /*
206 * The above forms a control dependency barrier separating the
207 * @tail load above from the data stores below. Since the @tail
208 * load is required to compute the branch to fail below.
209 *
210 * A, matches D; the full memory barrier userspace SHOULD issue
211 * after reading the data and before storing the new tail
212 * position.
213 *
214 * See perf_output_put_handle().
215 */
216
217 if (!backward)
218 head += size;
219 else
220 head -= size;
221 } while (!local_try_cmpxchg(&rb->head, &offset, head));
222
223 if (backward) {
224 offset = head;
225 head = (u64)(-head);
226 }
227
228 /*
229 * We rely on the implied barrier() by local_cmpxchg() to ensure
230 * none of the data stores below can be lifted up by the compiler.
231 */
232
233 if (unlikely(head - local_read(&rb->wakeup) > rb->watermark))
234 local_add(rb->watermark, &rb->wakeup);
235
236 page_shift = PAGE_SHIFT + page_order(rb);
237
238 handle->page = (offset >> page_shift) & (rb->nr_pages - 1);
239 offset &= (1UL << page_shift) - 1;
240 handle->addr = rb->data_pages[handle->page] + offset;
241 handle->size = (1UL << page_shift) - offset;
242
243 if (unlikely(have_lost)) {
244 lost_event.header.size = sizeof(lost_event);
245 lost_event.header.type = PERF_RECORD_LOST;
246 lost_event.header.misc = 0;
247 lost_event.id = event->id;
248 lost_event.lost = local_xchg(&rb->lost, 0);
249
250 /* XXX mostly redundant; @data is already fully initializes */
251 perf_event_header__init_id(&lost_event.header, data, event);
252 perf_output_put(handle, lost_event);
253 perf_event__output_id_sample(event, handle, data);
254 }
255
256 return 0;
257
258fail:
259 local_inc(&rb->lost);
260 atomic64_inc(&event->lost_samples);
261 perf_output_put_handle(handle);
262out:
263 rcu_read_unlock();
264
265 return -ENOSPC;
266}
267
268int perf_output_begin_forward(struct perf_output_handle *handle,
269 struct perf_sample_data *data,
270 struct perf_event *event, unsigned int size)
271{
272 return __perf_output_begin(handle, data, event, size, false);
273}
274
275int perf_output_begin_backward(struct perf_output_handle *handle,
276 struct perf_sample_data *data,
277 struct perf_event *event, unsigned int size)
278{
279 return __perf_output_begin(handle, data, event, size, true);
280}
281
282int perf_output_begin(struct perf_output_handle *handle,
283 struct perf_sample_data *data,
284 struct perf_event *event, unsigned int size)
285{
286
287 return __perf_output_begin(handle, data, event, size,
288 unlikely(is_write_backward(event)));
289}
290
291unsigned int perf_output_copy(struct perf_output_handle *handle,
292 const void *buf, unsigned int len)
293{
294 return __output_copy(handle, buf, len);
295}
296
297unsigned int perf_output_skip(struct perf_output_handle *handle,
298 unsigned int len)
299{
300 return __output_skip(handle, NULL, len);
301}
302
303void perf_output_end(struct perf_output_handle *handle)
304{
305 perf_output_put_handle(handle);
306 rcu_read_unlock();
307}
308
309static void
310ring_buffer_init(struct perf_buffer *rb, long watermark, int flags)
311{
312 long max_size = perf_data_size(rb);
313
314 if (watermark)
315 rb->watermark = min(max_size, watermark);
316
317 if (!rb->watermark)
318 rb->watermark = max_size / 2;
319
320 if (flags & RING_BUFFER_WRITABLE)
321 rb->overwrite = 0;
322 else
323 rb->overwrite = 1;
324
325 refcount_set(&rb->refcount, 1);
326
327 INIT_LIST_HEAD(&rb->event_list);
328 spin_lock_init(&rb->event_lock);
329
330 /*
331 * perf_output_begin() only checks rb->paused, therefore
332 * rb->paused must be true if we have no pages for output.
333 */
334 if (!rb->nr_pages)
335 rb->paused = 1;
336}
337
338void perf_aux_output_flag(struct perf_output_handle *handle, u64 flags)
339{
340 /*
341 * OVERWRITE is determined by perf_aux_output_end() and can't
342 * be passed in directly.
343 */
344 if (WARN_ON_ONCE(flags & PERF_AUX_FLAG_OVERWRITE))
345 return;
346
347 handle->aux_flags |= flags;
348}
349EXPORT_SYMBOL_GPL(perf_aux_output_flag);
350
351/*
352 * This is called before hardware starts writing to the AUX area to
353 * obtain an output handle and make sure there's room in the buffer.
354 * When the capture completes, call perf_aux_output_end() to commit
355 * the recorded data to the buffer.
356 *
357 * The ordering is similar to that of perf_output_{begin,end}, with
358 * the exception of (B), which should be taken care of by the pmu
359 * driver, since ordering rules will differ depending on hardware.
360 *
361 * Call this from pmu::start(); see the comment in perf_aux_output_end()
362 * about its use in pmu callbacks. Both can also be called from the PMI
363 * handler if needed.
364 */
365void *perf_aux_output_begin(struct perf_output_handle *handle,
366 struct perf_event *event)
367{
368 struct perf_event *output_event = event;
369 unsigned long aux_head, aux_tail;
370 struct perf_buffer *rb;
371 unsigned int nest;
372
373 if (output_event->parent)
374 output_event = output_event->parent;
375
376 /*
377 * Since this will typically be open across pmu::add/pmu::del, we
378 * grab ring_buffer's refcount instead of holding rcu read lock
379 * to make sure it doesn't disappear under us.
380 */
381 rb = ring_buffer_get(output_event);
382 if (!rb)
383 return NULL;
384
385 if (!rb_has_aux(rb))
386 goto err;
387
388 /*
389 * If aux_mmap_count is zero, the aux buffer is in perf_mmap_close(),
390 * about to get freed, so we leave immediately.
391 *
392 * Checking rb::aux_mmap_count and rb::refcount has to be done in
393 * the same order, see perf_mmap_close. Otherwise we end up freeing
394 * aux pages in this path, which is a bug, because in_atomic().
395 */
396 if (!atomic_read(&rb->aux_mmap_count))
397 goto err;
398
399 if (!refcount_inc_not_zero(&rb->aux_refcount))
400 goto err;
401
402 nest = READ_ONCE(rb->aux_nest);
403 /*
404 * Nesting is not supported for AUX area, make sure nested
405 * writers are caught early
406 */
407 if (WARN_ON_ONCE(nest))
408 goto err_put;
409
410 WRITE_ONCE(rb->aux_nest, nest + 1);
411
412 aux_head = rb->aux_head;
413
414 handle->rb = rb;
415 handle->event = event;
416 handle->head = aux_head;
417 handle->size = 0;
418 handle->aux_flags = 0;
419
420 /*
421 * In overwrite mode, AUX data stores do not depend on aux_tail,
422 * therefore (A) control dependency barrier does not exist. The
423 * (B) <-> (C) ordering is still observed by the pmu driver.
424 */
425 if (!rb->aux_overwrite) {
426 aux_tail = READ_ONCE(rb->user_page->aux_tail);
427 handle->wakeup = rb->aux_wakeup + rb->aux_watermark;
428 if (aux_head - aux_tail < perf_aux_size(rb))
429 handle->size = CIRC_SPACE(aux_head, aux_tail, perf_aux_size(rb));
430
431 /*
432 * handle->size computation depends on aux_tail load; this forms a
433 * control dependency barrier separating aux_tail load from aux data
434 * store that will be enabled on successful return
435 */
436 if (!handle->size) { /* A, matches D */
437 event->pending_disable = smp_processor_id();
438 perf_output_wakeup(handle);
439 WRITE_ONCE(rb->aux_nest, 0);
440 goto err_put;
441 }
442 }
443
444 return handle->rb->aux_priv;
445
446err_put:
447 /* can't be last */
448 rb_free_aux(rb);
449
450err:
451 ring_buffer_put(rb);
452 handle->event = NULL;
453
454 return NULL;
455}
456EXPORT_SYMBOL_GPL(perf_aux_output_begin);
457
458static __always_inline bool rb_need_aux_wakeup(struct perf_buffer *rb)
459{
460 if (rb->aux_overwrite)
461 return false;
462
463 if (rb->aux_head - rb->aux_wakeup >= rb->aux_watermark) {
464 rb->aux_wakeup = rounddown(rb->aux_head, rb->aux_watermark);
465 return true;
466 }
467
468 return false;
469}
470
471/*
472 * Commit the data written by hardware into the ring buffer by adjusting
473 * aux_head and posting a PERF_RECORD_AUX into the perf buffer. It is the
474 * pmu driver's responsibility to observe ordering rules of the hardware,
475 * so that all the data is externally visible before this is called.
476 *
477 * Note: this has to be called from pmu::stop() callback, as the assumption
478 * of the AUX buffer management code is that after pmu::stop(), the AUX
479 * transaction must be stopped and therefore drop the AUX reference count.
480 */
481void perf_aux_output_end(struct perf_output_handle *handle, unsigned long size)
482{
483 bool wakeup = !!(handle->aux_flags & PERF_AUX_FLAG_TRUNCATED);
484 struct perf_buffer *rb = handle->rb;
485 unsigned long aux_head;
486
487 /* in overwrite mode, driver provides aux_head via handle */
488 if (rb->aux_overwrite) {
489 handle->aux_flags |= PERF_AUX_FLAG_OVERWRITE;
490
491 aux_head = handle->head;
492 rb->aux_head = aux_head;
493 } else {
494 handle->aux_flags &= ~PERF_AUX_FLAG_OVERWRITE;
495
496 aux_head = rb->aux_head;
497 rb->aux_head += size;
498 }
499
500 /*
501 * Only send RECORD_AUX if we have something useful to communicate
502 *
503 * Note: the OVERWRITE records by themselves are not considered
504 * useful, as they don't communicate any *new* information,
505 * aside from the short-lived offset, that becomes history at
506 * the next event sched-in and therefore isn't useful.
507 * The userspace that needs to copy out AUX data in overwrite
508 * mode should know to use user_page::aux_head for the actual
509 * offset. So, from now on we don't output AUX records that
510 * have *only* OVERWRITE flag set.
511 */
512 if (size || (handle->aux_flags & ~(u64)PERF_AUX_FLAG_OVERWRITE))
513 perf_event_aux_event(handle->event, aux_head, size,
514 handle->aux_flags);
515
516 WRITE_ONCE(rb->user_page->aux_head, rb->aux_head);
517 if (rb_need_aux_wakeup(rb))
518 wakeup = true;
519
520 if (wakeup) {
521 if (handle->aux_flags & PERF_AUX_FLAG_TRUNCATED)
522 handle->event->pending_disable = smp_processor_id();
523 perf_output_wakeup(handle);
524 }
525
526 handle->event = NULL;
527
528 WRITE_ONCE(rb->aux_nest, 0);
529 /* can't be last */
530 rb_free_aux(rb);
531 ring_buffer_put(rb);
532}
533EXPORT_SYMBOL_GPL(perf_aux_output_end);
534
535/*
536 * Skip over a given number of bytes in the AUX buffer, due to, for example,
537 * hardware's alignment constraints.
538 */
539int perf_aux_output_skip(struct perf_output_handle *handle, unsigned long size)
540{
541 struct perf_buffer *rb = handle->rb;
542
543 if (size > handle->size)
544 return -ENOSPC;
545
546 rb->aux_head += size;
547
548 WRITE_ONCE(rb->user_page->aux_head, rb->aux_head);
549 if (rb_need_aux_wakeup(rb)) {
550 perf_output_wakeup(handle);
551 handle->wakeup = rb->aux_wakeup + rb->aux_watermark;
552 }
553
554 handle->head = rb->aux_head;
555 handle->size -= size;
556
557 return 0;
558}
559EXPORT_SYMBOL_GPL(perf_aux_output_skip);
560
561void *perf_get_aux(struct perf_output_handle *handle)
562{
563 /* this is only valid between perf_aux_output_begin and *_end */
564 if (!handle->event)
565 return NULL;
566
567 return handle->rb->aux_priv;
568}
569EXPORT_SYMBOL_GPL(perf_get_aux);
570
571/*
572 * Copy out AUX data from an AUX handle.
573 */
574long perf_output_copy_aux(struct perf_output_handle *aux_handle,
575 struct perf_output_handle *handle,
576 unsigned long from, unsigned long to)
577{
578 struct perf_buffer *rb = aux_handle->rb;
579 unsigned long tocopy, remainder, len = 0;
580 void *addr;
581
582 from &= (rb->aux_nr_pages << PAGE_SHIFT) - 1;
583 to &= (rb->aux_nr_pages << PAGE_SHIFT) - 1;
584
585 do {
586 tocopy = PAGE_SIZE - offset_in_page(from);
587 if (to > from)
588 tocopy = min(tocopy, to - from);
589 if (!tocopy)
590 break;
591
592 addr = rb->aux_pages[from >> PAGE_SHIFT];
593 addr += offset_in_page(from);
594
595 remainder = perf_output_copy(handle, addr, tocopy);
596 if (remainder)
597 return -EFAULT;
598
599 len += tocopy;
600 from += tocopy;
601 from &= (rb->aux_nr_pages << PAGE_SHIFT) - 1;
602 } while (to != from);
603
604 return len;
605}
606
607#define PERF_AUX_GFP (GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_NORETRY)
608
609static struct page *rb_alloc_aux_page(int node, int order)
610{
611 struct page *page;
612
613 if (order > MAX_PAGE_ORDER)
614 order = MAX_PAGE_ORDER;
615
616 do {
617 page = alloc_pages_node(node, PERF_AUX_GFP, order);
618 } while (!page && order--);
619
620 if (page && order) {
621 /*
622 * Communicate the allocation size to the driver:
623 * if we managed to secure a high-order allocation,
624 * set its first page's private to this order;
625 * !PagePrivate(page) means it's just a normal page.
626 */
627 split_page(page, order);
628 SetPagePrivate(page);
629 set_page_private(page, order);
630 }
631
632 return page;
633}
634
635static void rb_free_aux_page(struct perf_buffer *rb, int idx)
636{
637 struct page *page = virt_to_page(rb->aux_pages[idx]);
638
639 ClearPagePrivate(page);
640 page->mapping = NULL;
641 __free_page(page);
642}
643
644static void __rb_free_aux(struct perf_buffer *rb)
645{
646 int pg;
647
648 /*
649 * Should never happen, the last reference should be dropped from
650 * perf_mmap_close() path, which first stops aux transactions (which
651 * in turn are the atomic holders of aux_refcount) and then does the
652 * last rb_free_aux().
653 */
654 WARN_ON_ONCE(in_atomic());
655
656 if (rb->aux_priv) {
657 rb->free_aux(rb->aux_priv);
658 rb->free_aux = NULL;
659 rb->aux_priv = NULL;
660 }
661
662 if (rb->aux_nr_pages) {
663 for (pg = 0; pg < rb->aux_nr_pages; pg++)
664 rb_free_aux_page(rb, pg);
665
666 kfree(rb->aux_pages);
667 rb->aux_nr_pages = 0;
668 }
669}
670
671int rb_alloc_aux(struct perf_buffer *rb, struct perf_event *event,
672 pgoff_t pgoff, int nr_pages, long watermark, int flags)
673{
674 bool overwrite = !(flags & RING_BUFFER_WRITABLE);
675 int node = (event->cpu == -1) ? -1 : cpu_to_node(event->cpu);
676 int ret = -ENOMEM, max_order;
677
678 if (!has_aux(event))
679 return -EOPNOTSUPP;
680
681 if (!overwrite) {
682 /*
683 * Watermark defaults to half the buffer, and so does the
684 * max_order, to aid PMU drivers in double buffering.
685 */
686 if (!watermark)
687 watermark = nr_pages << (PAGE_SHIFT - 1);
688
689 /*
690 * Use aux_watermark as the basis for chunking to
691 * help PMU drivers honor the watermark.
692 */
693 max_order = get_order(watermark);
694 } else {
695 /*
696 * We need to start with the max_order that fits in nr_pages,
697 * not the other way around, hence ilog2() and not get_order.
698 */
699 max_order = ilog2(nr_pages);
700 watermark = 0;
701 }
702
703 /*
704 * kcalloc_node() is unable to allocate buffer if the size is larger
705 * than: PAGE_SIZE << MAX_PAGE_ORDER; directly bail out in this case.
706 */
707 if (get_order((unsigned long)nr_pages * sizeof(void *)) > MAX_PAGE_ORDER)
708 return -ENOMEM;
709 rb->aux_pages = kcalloc_node(nr_pages, sizeof(void *), GFP_KERNEL,
710 node);
711 if (!rb->aux_pages)
712 return -ENOMEM;
713
714 rb->free_aux = event->pmu->free_aux;
715 for (rb->aux_nr_pages = 0; rb->aux_nr_pages < nr_pages;) {
716 struct page *page;
717 int last, order;
718
719 order = min(max_order, ilog2(nr_pages - rb->aux_nr_pages));
720 page = rb_alloc_aux_page(node, order);
721 if (!page)
722 goto out;
723
724 for (last = rb->aux_nr_pages + (1 << page_private(page));
725 last > rb->aux_nr_pages; rb->aux_nr_pages++)
726 rb->aux_pages[rb->aux_nr_pages] = page_address(page++);
727 }
728
729 /*
730 * In overwrite mode, PMUs that don't support SG may not handle more
731 * than one contiguous allocation, since they rely on PMI to do double
732 * buffering. In this case, the entire buffer has to be one contiguous
733 * chunk.
734 */
735 if ((event->pmu->capabilities & PERF_PMU_CAP_AUX_NO_SG) &&
736 overwrite) {
737 struct page *page = virt_to_page(rb->aux_pages[0]);
738
739 if (page_private(page) != max_order)
740 goto out;
741 }
742
743 rb->aux_priv = event->pmu->setup_aux(event, rb->aux_pages, nr_pages,
744 overwrite);
745 if (!rb->aux_priv)
746 goto out;
747
748 ret = 0;
749
750 /*
751 * aux_pages (and pmu driver's private data, aux_priv) will be
752 * referenced in both producer's and consumer's contexts, thus
753 * we keep a refcount here to make sure either of the two can
754 * reference them safely.
755 */
756 refcount_set(&rb->aux_refcount, 1);
757
758 rb->aux_overwrite = overwrite;
759 rb->aux_watermark = watermark;
760
761out:
762 if (!ret)
763 rb->aux_pgoff = pgoff;
764 else
765 __rb_free_aux(rb);
766
767 return ret;
768}
769
770void rb_free_aux(struct perf_buffer *rb)
771{
772 if (refcount_dec_and_test(&rb->aux_refcount))
773 __rb_free_aux(rb);
774}
775
776#ifndef CONFIG_PERF_USE_VMALLOC
777
778/*
779 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
780 */
781
782static struct page *
783__perf_mmap_to_page(struct perf_buffer *rb, unsigned long pgoff)
784{
785 if (pgoff > rb->nr_pages)
786 return NULL;
787
788 if (pgoff == 0)
789 return virt_to_page(rb->user_page);
790
791 return virt_to_page(rb->data_pages[pgoff - 1]);
792}
793
794static void *perf_mmap_alloc_page(int cpu)
795{
796 struct page *page;
797 int node;
798
799 node = (cpu == -1) ? cpu : cpu_to_node(cpu);
800 page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
801 if (!page)
802 return NULL;
803
804 return page_address(page);
805}
806
807static void perf_mmap_free_page(void *addr)
808{
809 struct page *page = virt_to_page(addr);
810
811 page->mapping = NULL;
812 __free_page(page);
813}
814
815struct perf_buffer *rb_alloc(int nr_pages, long watermark, int cpu, int flags)
816{
817 struct perf_buffer *rb;
818 unsigned long size;
819 int i, node;
820
821 size = sizeof(struct perf_buffer);
822 size += nr_pages * sizeof(void *);
823
824 if (order_base_2(size) > PAGE_SHIFT+MAX_PAGE_ORDER)
825 goto fail;
826
827 node = (cpu == -1) ? cpu : cpu_to_node(cpu);
828 rb = kzalloc_node(size, GFP_KERNEL, node);
829 if (!rb)
830 goto fail;
831
832 rb->user_page = perf_mmap_alloc_page(cpu);
833 if (!rb->user_page)
834 goto fail_user_page;
835
836 for (i = 0; i < nr_pages; i++) {
837 rb->data_pages[i] = perf_mmap_alloc_page(cpu);
838 if (!rb->data_pages[i])
839 goto fail_data_pages;
840 }
841
842 rb->nr_pages = nr_pages;
843
844 ring_buffer_init(rb, watermark, flags);
845
846 return rb;
847
848fail_data_pages:
849 for (i--; i >= 0; i--)
850 perf_mmap_free_page(rb->data_pages[i]);
851
852 perf_mmap_free_page(rb->user_page);
853
854fail_user_page:
855 kfree(rb);
856
857fail:
858 return NULL;
859}
860
861void rb_free(struct perf_buffer *rb)
862{
863 int i;
864
865 perf_mmap_free_page(rb->user_page);
866 for (i = 0; i < rb->nr_pages; i++)
867 perf_mmap_free_page(rb->data_pages[i]);
868 kfree(rb);
869}
870
871#else
872static struct page *
873__perf_mmap_to_page(struct perf_buffer *rb, unsigned long pgoff)
874{
875 /* The '>' counts in the user page. */
876 if (pgoff > data_page_nr(rb))
877 return NULL;
878
879 return vmalloc_to_page((void *)rb->user_page + pgoff * PAGE_SIZE);
880}
881
882static void perf_mmap_unmark_page(void *addr)
883{
884 struct page *page = vmalloc_to_page(addr);
885
886 page->mapping = NULL;
887}
888
889static void rb_free_work(struct work_struct *work)
890{
891 struct perf_buffer *rb;
892 void *base;
893 int i, nr;
894
895 rb = container_of(work, struct perf_buffer, work);
896 nr = data_page_nr(rb);
897
898 base = rb->user_page;
899 /* The '<=' counts in the user page. */
900 for (i = 0; i <= nr; i++)
901 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
902
903 vfree(base);
904 kfree(rb);
905}
906
907void rb_free(struct perf_buffer *rb)
908{
909 schedule_work(&rb->work);
910}
911
912struct perf_buffer *rb_alloc(int nr_pages, long watermark, int cpu, int flags)
913{
914 struct perf_buffer *rb;
915 unsigned long size;
916 void *all_buf;
917 int node;
918
919 size = sizeof(struct perf_buffer);
920 size += sizeof(void *);
921
922 node = (cpu == -1) ? cpu : cpu_to_node(cpu);
923 rb = kzalloc_node(size, GFP_KERNEL, node);
924 if (!rb)
925 goto fail;
926
927 INIT_WORK(&rb->work, rb_free_work);
928
929 all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
930 if (!all_buf)
931 goto fail_all_buf;
932
933 rb->user_page = all_buf;
934 rb->data_pages[0] = all_buf + PAGE_SIZE;
935 if (nr_pages) {
936 rb->nr_pages = 1;
937 rb->page_order = ilog2(nr_pages);
938 }
939
940 ring_buffer_init(rb, watermark, flags);
941
942 return rb;
943
944fail_all_buf:
945 kfree(rb);
946
947fail:
948 return NULL;
949}
950
951#endif
952
953struct page *
954perf_mmap_to_page(struct perf_buffer *rb, unsigned long pgoff)
955{
956 if (rb->aux_nr_pages) {
957 /* above AUX space */
958 if (pgoff > rb->aux_pgoff + rb->aux_nr_pages)
959 return NULL;
960
961 /* AUX space */
962 if (pgoff >= rb->aux_pgoff) {
963 int aux_pgoff = array_index_nospec(pgoff - rb->aux_pgoff, rb->aux_nr_pages);
964 return virt_to_page(rb->aux_pages[aux_pgoff]);
965 }
966 }
967
968 return __perf_mmap_to_page(rb, pgoff);
969}