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