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