ring_buffer.c

  1// SPDX-License-Identifier: GPL-2.0
  2/*
  3 * Performance events ring-buffer code:
  4 *
  5 *  Copyright (C) 2008 Linutronix GmbH, Thomas Gleixner <tglx@kernel.org>
  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 | 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.header.size = sizeof(lost_event);
250		lost_event.header.type = PERF_RECORD_LOST;
251		lost_event.header.misc = 0;
252		lost_event.id          = event->id;
253		lost_event.lost        = local_xchg(&rb->lost, 0);
254
255		/* XXX mostly redundant; @data is already fully initializes */
256		perf_event_header__init_id(&lost_event.header, data, event);
257		perf_output_put(handle, lost_event);
258		perf_event__output_id_sample(event, handle, data);
259	}
260
261	return 0;
262
263fail:
264	local_inc(&rb->lost);
265	atomic64_inc(&event->lost_samples);
266	perf_output_put_handle(handle);
267out:
268	rcu_read_unlock();
269
270	return -ENOSPC;
271}
272
273int perf_output_begin_forward(struct perf_output_handle *handle,
274			      struct perf_sample_data *data,
275			      struct perf_event *event, unsigned int size)
276{
277	return __perf_output_begin(handle, data, event, size, false);
278}
279
280int perf_output_begin_backward(struct perf_output_handle *handle,
281			       struct perf_sample_data *data,
282			       struct perf_event *event, unsigned int size)
283{
284	return __perf_output_begin(handle, data, event, size, true);
285}
286
287int perf_output_begin(struct perf_output_handle *handle,
288		      struct perf_sample_data *data,
289		      struct perf_event *event, unsigned int size)
290{
291
292	return __perf_output_begin(handle, data, event, size,
293				   unlikely(is_write_backward(event)));
294}
295
296unsigned int perf_output_copy(struct perf_output_handle *handle,
297		      const void *buf, unsigned int len)
298{
299	return __output_copy(handle, buf, len);
300}
301
302unsigned int perf_output_skip(struct perf_output_handle *handle,
303			      unsigned int len)
304{
305	return __output_skip(handle, NULL, len);
306}
307
308void perf_output_end(struct perf_output_handle *handle)
309{
310	perf_output_put_handle(handle);
311	rcu_read_unlock();
312}
313
314static void
315ring_buffer_init(struct perf_buffer *rb, long watermark, int flags)
316{
317	long max_size = perf_data_size(rb);
318
319	if (watermark)
320		rb->watermark = min(max_size, watermark);
321
322	if (!rb->watermark)
323		rb->watermark = max_size / 2;
324
325	if (flags & RING_BUFFER_WRITABLE)
326		rb->overwrite = 0;
327	else
328		rb->overwrite = 1;
329
330	refcount_set(&rb->refcount, 1);
331
332	INIT_LIST_HEAD(&rb->event_list);
333	spin_lock_init(&rb->event_lock);
334
335	/*
336	 * perf_output_begin() only checks rb->paused, therefore
337	 * rb->paused must be true if we have no pages for output.
338	 */
339	if (!rb->nr_pages)
340		rb->paused = 1;
341
342	mutex_init(&rb->aux_mutex);
343}
344
345void perf_aux_output_flag(struct perf_output_handle *handle, u64 flags)
346{
347	/*
348	 * OVERWRITE is determined by perf_aux_output_end() and can't
349	 * be passed in directly.
350	 */
351	if (WARN_ON_ONCE(flags & PERF_AUX_FLAG_OVERWRITE))
352		return;
353
354	handle->aux_flags |= flags;
355}
356EXPORT_SYMBOL_GPL(perf_aux_output_flag);
357
358/*
359 * This is called before hardware starts writing to the AUX area to
360 * obtain an output handle and make sure there's room in the buffer.
361 * When the capture completes, call perf_aux_output_end() to commit
362 * the recorded data to the buffer.
363 *
364 * The ordering is similar to that of perf_output_{begin,end}, with
365 * the exception of (B), which should be taken care of by the pmu
366 * driver, since ordering rules will differ depending on hardware.
367 *
368 * Call this from pmu::start(); see the comment in perf_aux_output_end()
369 * about its use in pmu callbacks. Both can also be called from the PMI
370 * handler if needed.
371 */
372void *perf_aux_output_begin(struct perf_output_handle *handle,
373			    struct perf_event *event)
374{
375	struct perf_event *output_event = event;
376	unsigned long aux_head, aux_tail;
377	struct perf_buffer *rb;
378	unsigned int nest;
379
380	if (output_event->parent)
381		output_event = output_event->parent;
382
383	/*
384	 * Since this will typically be open across pmu::add/pmu::del, we
385	 * grab ring_buffer's refcount instead of holding rcu read lock
386	 * to make sure it doesn't disappear under us.
387	 */
388	rb = ring_buffer_get(output_event);
389	if (!rb)
390		return NULL;
391
392	if (!rb_has_aux(rb))
393		goto err;
394
395	/*
396	 * If aux_mmap_count is zero, the aux buffer is in perf_mmap_close(),
397	 * about to get freed, so we leave immediately.
398	 *
399	 * Checking rb::aux_mmap_count and rb::refcount has to be done in
400	 * the same order, see perf_mmap_close. Otherwise we end up freeing
401	 * aux pages in this path, which is a bug, because in_atomic().
402	 */
403	if (!refcount_read(&rb->aux_mmap_count))
404		goto err;
405
406	if (!refcount_inc_not_zero(&rb->aux_refcount))
407		goto err;
408
409	nest = READ_ONCE(rb->aux_nest);
410	/*
411	 * Nesting is not supported for AUX area, make sure nested
412	 * writers are caught early
413	 */
414	if (WARN_ON_ONCE(nest))
415		goto err_put;
416
417	WRITE_ONCE(rb->aux_nest, nest + 1);
418
419	aux_head = rb->aux_head;
420
421	handle->rb = rb;
422	handle->event = event;
423	handle->head = aux_head;
424	handle->size = 0;
425	handle->aux_flags = 0;
426
427	/*
428	 * In overwrite mode, AUX data stores do not depend on aux_tail,
429	 * therefore (A) control dependency barrier does not exist. The
430	 * (B) <-> (C) ordering is still observed by the pmu driver.
431	 */
432	if (!rb->aux_overwrite) {
433		aux_tail = READ_ONCE(rb->user_page->aux_tail);
434		handle->wakeup = rb->aux_wakeup + rb->aux_watermark;
435		if (aux_head - aux_tail < perf_aux_size(rb))
436			handle->size = CIRC_SPACE(aux_head, aux_tail, perf_aux_size(rb));
437
438		/*
439		 * handle->size computation depends on aux_tail load; this forms a
440		 * control dependency barrier separating aux_tail load from aux data
441		 * store that will be enabled on successful return
442		 */
443		if (!handle->size) { /* A, matches D */
444			perf_event_disable_inatomic(handle->event);
445			perf_output_wakeup(handle);
446			WRITE_ONCE(rb->aux_nest, 0);
447			goto err_put;
448		}
449	}
450
451	return handle->rb->aux_priv;
452
453err_put:
454	/* can't be last */
455	rb_free_aux(rb);
456
457err:
458	ring_buffer_put(rb);
459	handle->event = NULL;
460
461	return NULL;
462}
463EXPORT_SYMBOL_GPL(perf_aux_output_begin);
464
465static __always_inline bool rb_need_aux_wakeup(struct perf_buffer *rb)
466{
467	if (rb->aux_overwrite)
468		return false;
469
470	if (rb->aux_head - rb->aux_wakeup >= rb->aux_watermark) {
471		rb->aux_wakeup = rounddown(rb->aux_head, rb->aux_watermark);
472		return true;
473	}
474
475	return false;
476}
477
478/*
479 * Commit the data written by hardware into the ring buffer by adjusting
480 * aux_head and posting a PERF_RECORD_AUX into the perf buffer. It is the
481 * pmu driver's responsibility to observe ordering rules of the hardware,
482 * so that all the data is externally visible before this is called.
483 *
484 * Note: this has to be called from pmu::stop() callback, as the assumption
485 * of the AUX buffer management code is that after pmu::stop(), the AUX
486 * transaction must be stopped and therefore drop the AUX reference count.
487 */
488void perf_aux_output_end(struct perf_output_handle *handle, unsigned long size)
489{
490	bool wakeup = !!(handle->aux_flags & PERF_AUX_FLAG_TRUNCATED);
491	struct perf_buffer *rb = handle->rb;
492	unsigned long aux_head;
493
494	/* in overwrite mode, driver provides aux_head via handle */
495	if (rb->aux_overwrite) {
496		handle->aux_flags |= PERF_AUX_FLAG_OVERWRITE;
497
498		aux_head = handle->head;
499		rb->aux_head = aux_head;
500	} else {
501		handle->aux_flags &= ~PERF_AUX_FLAG_OVERWRITE;
502
503		aux_head = rb->aux_head;
504		rb->aux_head += size;
505	}
506
507	/*
508	 * Only send RECORD_AUX if we have something useful to communicate
509	 *
510	 * Note: the OVERWRITE records by themselves are not considered
511	 * useful, as they don't communicate any *new* information,
512	 * aside from the short-lived offset, that becomes history at
513	 * the next event sched-in and therefore isn't useful.
514	 * The userspace that needs to copy out AUX data in overwrite
515	 * mode should know to use user_page::aux_head for the actual
516	 * offset. So, from now on we don't output AUX records that
517	 * have *only* OVERWRITE flag set.
518	 */
519	if (size || (handle->aux_flags & ~(u64)PERF_AUX_FLAG_OVERWRITE))
520		perf_event_aux_event(handle->event, aux_head, size,
521				     handle->aux_flags);
522
523	WRITE_ONCE(rb->user_page->aux_head, rb->aux_head);
524	if (rb_need_aux_wakeup(rb))
525		wakeup = true;
526
527	if (wakeup) {
528		if (handle->aux_flags & PERF_AUX_FLAG_TRUNCATED)
529			perf_event_disable_inatomic(handle->event);
530		perf_output_wakeup(handle);
531	}
532
533	handle->event = NULL;
534
535	WRITE_ONCE(rb->aux_nest, 0);
536	/* can't be last */
537	rb_free_aux(rb);
538	ring_buffer_put(rb);
539}
540EXPORT_SYMBOL_GPL(perf_aux_output_end);
541
542/*
543 * Skip over a given number of bytes in the AUX buffer, due to, for example,
544 * hardware's alignment constraints.
545 */
546int perf_aux_output_skip(struct perf_output_handle *handle, unsigned long size)
547{
548	struct perf_buffer *rb = handle->rb;
549
550	if (size > handle->size)
551		return -ENOSPC;
552
553	rb->aux_head += size;
554
555	WRITE_ONCE(rb->user_page->aux_head, rb->aux_head);
556	if (rb_need_aux_wakeup(rb)) {
557		perf_output_wakeup(handle);
558		handle->wakeup = rb->aux_wakeup + rb->aux_watermark;
559	}
560
561	handle->head = rb->aux_head;
562	handle->size -= size;
563
564	return 0;
565}
566EXPORT_SYMBOL_GPL(perf_aux_output_skip);
567
568void *perf_get_aux(struct perf_output_handle *handle)
569{
570	/* this is only valid between perf_aux_output_begin and *_end */
571	if (!handle->event)
572		return NULL;
573
574	return handle->rb->aux_priv;
575}
576EXPORT_SYMBOL_GPL(perf_get_aux);
577
578/*
579 * Copy out AUX data from an AUX handle.
580 */
581long perf_output_copy_aux(struct perf_output_handle *aux_handle,
582			  struct perf_output_handle *handle,
583			  unsigned long from, unsigned long to)
584{
585	struct perf_buffer *rb = aux_handle->rb;
586	unsigned long tocopy, remainder, len = 0;
587	void *addr;
588
589	from &= (rb->aux_nr_pages << PAGE_SHIFT) - 1;
590	to &= (rb->aux_nr_pages << PAGE_SHIFT) - 1;
591
592	do {
593		tocopy = PAGE_SIZE - offset_in_page(from);
594		if (to > from)
595			tocopy = min(tocopy, to - from);
596		if (!tocopy)
597			break;
598
599		addr = rb->aux_pages[from >> PAGE_SHIFT];
600		addr += offset_in_page(from);
601
602		remainder = perf_output_copy(handle, addr, tocopy);
603		if (remainder)
604			return -EFAULT;
605
606		len += tocopy;
607		from += tocopy;
608		from &= (rb->aux_nr_pages << PAGE_SHIFT) - 1;
609	} while (to != from);
610
611	return len;
612}
613
614#define PERF_AUX_GFP	(GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_NORETRY)
615
616static struct page *rb_alloc_aux_page(int node, int order)
617{
618	struct page *page;
619
620	if (order > MAX_PAGE_ORDER)
621		order = MAX_PAGE_ORDER;
622
623	do {
624		page = alloc_pages_node(node, PERF_AUX_GFP, order);
625	} while (!page && order--);
626
627	if (page && order) {
628		/*
629		 * Communicate the allocation size to the driver:
630		 * if we managed to secure a high-order allocation,
631		 * set its first page's private to this order;
632		 * !PagePrivate(page) means it's just a normal page.
633		 */
634		split_page(page, order);
635		SetPagePrivate(page);
636		set_page_private(page, order);
637	}
638
639	return page;
640}
641
642static void rb_free_aux_page(struct perf_buffer *rb, int idx)
643{
644	struct page *page = virt_to_page(rb->aux_pages[idx]);
645
646	ClearPagePrivate(page);
647	__free_page(page);
648}
649
650static void __rb_free_aux(struct perf_buffer *rb)
651{
652	int pg;
653
654	/*
655	 * Should never happen, the last reference should be dropped from
656	 * perf_mmap_close() path, which first stops aux transactions (which
657	 * in turn are the atomic holders of aux_refcount) and then does the
658	 * last rb_free_aux().
659	 */
660	WARN_ON_ONCE(in_atomic());
661
662	if (rb->aux_priv) {
663		rb->free_aux(rb->aux_priv);
664		rb->free_aux = NULL;
665		rb->aux_priv = NULL;
666	}
667
668	if (rb->aux_nr_pages) {
669		for (pg = 0; pg < rb->aux_nr_pages; pg++)
670			rb_free_aux_page(rb, pg);
671
672		kfree(rb->aux_pages);
673		rb->aux_nr_pages = 0;
674	}
675}
676
677int rb_alloc_aux(struct perf_buffer *rb, struct perf_event *event,
678		 pgoff_t pgoff, int nr_pages, long watermark, int flags)
679{
680	bool overwrite = !(flags & RING_BUFFER_WRITABLE);
681	int node = (event->cpu == -1) ? -1 : cpu_to_node(event->cpu);
682	bool use_contiguous_pages = event->pmu->capabilities & (
683		PERF_PMU_CAP_AUX_NO_SG | PERF_PMU_CAP_AUX_PREFER_LARGE);
684	/*
685	 * Initialize max_order to 0 for page allocation. This allocates single
686	 * pages to minimize memory fragmentation. This is overridden if the
687	 * PMU needs or prefers contiguous pages (use_contiguous_pages = true).
688	 */
689	int max_order = 0;
690	int ret = -ENOMEM;
691
692	if (!has_aux(event))
693		return -EOPNOTSUPP;
694
695	if (nr_pages <= 0)
696		return -EINVAL;
697
698	if (!overwrite) {
699		/*
700		 * Watermark defaults to half the buffer, to aid PMU drivers
701		 * in double buffering.
702		 */
703		if (!watermark)
704			watermark = min_t(unsigned long,
705					  U32_MAX,
706					  (unsigned long)nr_pages << (PAGE_SHIFT - 1));
707
708		/*
709		 * If using contiguous pages, use aux_watermark as the basis
710		 * for chunking to help PMU drivers honor the watermark.
711		 */
712		if (use_contiguous_pages)
713			max_order = get_order(watermark);
714	} else {
715		/*
716		 * If using contiguous pages, we need to start with the
717		 * max_order that fits in nr_pages, not the other way around,
718		 * hence ilog2() and not get_order.
719		 */
720		if (use_contiguous_pages)
721			max_order = ilog2(nr_pages);
722		watermark = 0;
723	}
724
725	/*
726	 * kcalloc_node() is unable to allocate buffer if the size is larger
727	 * than: PAGE_SIZE << MAX_PAGE_ORDER; directly bail out in this case.
728	 */
729	if (get_order((unsigned long)nr_pages * sizeof(void *)) > MAX_PAGE_ORDER)
730		return -ENOMEM;
731	rb->aux_pages = kcalloc_node(nr_pages, sizeof(void *), GFP_KERNEL,
732				     node);
733	if (!rb->aux_pages)
734		return -ENOMEM;
735
736	rb->free_aux = event->pmu->free_aux;
737	for (rb->aux_nr_pages = 0; rb->aux_nr_pages < nr_pages;) {
738		struct page *page;
739		int last, order;
740
741		order = min(max_order, ilog2(nr_pages - rb->aux_nr_pages));
742		page = rb_alloc_aux_page(node, order);
743		if (!page)
744			goto out;
745
746		for (last = rb->aux_nr_pages + (1 << page_private(page));
747		     last > rb->aux_nr_pages; rb->aux_nr_pages++)
748			rb->aux_pages[rb->aux_nr_pages] = page_address(page++);
749	}
750
751	/*
752	 * In overwrite mode, PMUs that don't support SG may not handle more
753	 * than one contiguous allocation, since they rely on PMI to do double
754	 * buffering. In this case, the entire buffer has to be one contiguous
755	 * chunk.
756	 */
757	if ((event->pmu->capabilities & PERF_PMU_CAP_AUX_NO_SG) &&
758	    overwrite) {
759		struct page *page = virt_to_page(rb->aux_pages[0]);
760
761		if (page_private(page) != max_order)
762			goto out;
763	}
764
765	rb->aux_priv = event->pmu->setup_aux(event, rb->aux_pages, nr_pages,
766					     overwrite);
767	if (!rb->aux_priv)
768		goto out;
769
770	ret = 0;
771
772	/*
773	 * aux_pages (and pmu driver's private data, aux_priv) will be
774	 * referenced in both producer's and consumer's contexts, thus
775	 * we keep a refcount here to make sure either of the two can
776	 * reference them safely.
777	 */
778	refcount_set(&rb->aux_refcount, 1);
779
780	rb->aux_overwrite = overwrite;
781	rb->aux_watermark = watermark;
782
783out:
784	if (!ret)
785		rb->aux_pgoff = pgoff;
786	else
787		__rb_free_aux(rb);
788
789	return ret;
790}
791
792void rb_free_aux(struct perf_buffer *rb)
793{
794	if (refcount_dec_and_test(&rb->aux_refcount))
795		__rb_free_aux(rb);
796}
797
798#ifndef CONFIG_PERF_USE_VMALLOC
799
800/*
801 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
802 */
803
804static struct page *
805__perf_mmap_to_page(struct perf_buffer *rb, unsigned long pgoff)
806{
807	if (pgoff > rb->nr_pages)
808		return NULL;
809
810	if (pgoff == 0)
811		return virt_to_page(rb->user_page);
812
813	return virt_to_page(rb->data_pages[pgoff - 1]);
814}
815
816static void *perf_mmap_alloc_page(int cpu)
817{
818	struct page *page;
819	int node;
820
821	node = (cpu == -1) ? cpu : cpu_to_node(cpu);
822	page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
823	if (!page)
824		return NULL;
825
826	return page_address(page);
827}
828
829static void perf_mmap_free_page(void *addr)
830{
831	struct page *page = virt_to_page(addr);
832
833	__free_page(page);
834}
835
836struct perf_buffer *rb_alloc(int nr_pages, long watermark, int cpu, int flags)
837{
838	struct perf_buffer *rb;
839	unsigned long size;
840	int i, node;
841
842	size = sizeof(struct perf_buffer);
843	size += nr_pages * sizeof(void *);
844
845	if (order_base_2(size) > PAGE_SHIFT+MAX_PAGE_ORDER)
846		goto fail;
847
848	node = (cpu == -1) ? cpu : cpu_to_node(cpu);
849	rb = kzalloc_node(size, GFP_KERNEL, node);
850	if (!rb)
851		goto fail;
852
853	rb->user_page = perf_mmap_alloc_page(cpu);
854	if (!rb->user_page)
855		goto fail_user_page;
856
857	for (i = 0; i < nr_pages; i++) {
858		rb->data_pages[i] = perf_mmap_alloc_page(cpu);
859		if (!rb->data_pages[i])
860			goto fail_data_pages;
861	}
862
863	rb->nr_pages = nr_pages;
864
865	ring_buffer_init(rb, watermark, flags);
866
867	return rb;
868
869fail_data_pages:
870	for (i--; i >= 0; i--)
871		perf_mmap_free_page(rb->data_pages[i]);
872
873	perf_mmap_free_page(rb->user_page);
874
875fail_user_page:
876	kfree(rb);
877
878fail:
879	return NULL;
880}
881
882void rb_free(struct perf_buffer *rb)
883{
884	int i;
885
886	perf_mmap_free_page(rb->user_page);
887	for (i = 0; i < rb->nr_pages; i++)
888		perf_mmap_free_page(rb->data_pages[i]);
889	kfree(rb);
890}
891
892#else
893static struct page *
894__perf_mmap_to_page(struct perf_buffer *rb, unsigned long pgoff)
895{
896	/* The '>' counts in the user page. */
897	if (pgoff > data_page_nr(rb))
898		return NULL;
899
900	return vmalloc_to_page((void *)rb->user_page + pgoff * PAGE_SIZE);
901}
902
903static void rb_free_work(struct work_struct *work)
904{
905	struct perf_buffer *rb;
906
907	rb = container_of(work, struct perf_buffer, work);
908
909	vfree(rb->user_page);
910	kfree(rb);
911}
912
913void rb_free(struct perf_buffer *rb)
914{
915	schedule_work(&rb->work);
916}
917
918struct perf_buffer *rb_alloc(int nr_pages, long watermark, int cpu, int flags)
919{
920	struct perf_buffer *rb;
921	unsigned long size;
922	void *all_buf;
923	int node;
924
925	size = sizeof(struct perf_buffer);
926	size += sizeof(void *);
927
928	node = (cpu == -1) ? cpu : cpu_to_node(cpu);
929	rb = kzalloc_node(size, GFP_KERNEL, node);
930	if (!rb)
931		goto fail;
932
933	INIT_WORK(&rb->work, rb_free_work);
934
935	all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
936	if (!all_buf)
937		goto fail_all_buf;
938
939	rb->user_page = all_buf;
940	rb->data_pages[0] = all_buf + PAGE_SIZE;
941	if (nr_pages) {
942		rb->nr_pages = 1;
943		rb->page_order = ilog2(nr_pages);
944	}
945
946	ring_buffer_init(rb, watermark, flags);
947
948	return rb;
949
950fail_all_buf:
951	kfree(rb);
952
953fail:
954	return NULL;
955}
956
957#endif
958
959struct page *
960perf_mmap_to_page(struct perf_buffer *rb, unsigned long pgoff)
961{
962	if (rb->aux_nr_pages) {
963		/* above AUX space */
964		if (pgoff > rb->aux_pgoff + rb->aux_nr_pages)
965			return NULL;
966
967		/* AUX space */
968		if (pgoff >= rb->aux_pgoff) {
969			int aux_pgoff = array_index_nospec(pgoff - rb->aux_pgoff, rb->aux_nr_pages);
970			return virt_to_page(rb->aux_pages[aux_pgoff]);
971		}
972	}
973
974	return __perf_mmap_to_page(rb, pgoff);
975}