2 * Performance events ring-buffer code:
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>
9 * For licensing details see kernel-base/COPYING
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>
20 static void perf_output_wakeup(struct perf_output_handle *handle)
22 atomic_set(&handle->rb->poll, POLLIN);
24 handle->event->pending_wakeup = 1;
25 irq_work_queue(&handle->event->pending);
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.
33 * We only publish the head (and generate a wakeup) when the outer-most
36 static void perf_output_get_handle(struct perf_output_handle *handle)
38 struct ring_buffer *rb = handle->rb;
42 handle->wakeup = local_read(&rb->wakeup);
45 static void perf_output_put_handle(struct perf_output_handle *handle)
47 struct ring_buffer *rb = handle->rb;
51 head = local_read(&rb->head);
54 * IRQ/NMI can happen here, which means we can miss a head update.
57 if (!local_dec_and_test(&rb->nest))
61 * Since the mmap() consumer (userspace) can run on a different CPU:
65 * if (LOAD ->data_tail) { LOAD ->data_head
67 * STORE $data LOAD $data
68 * smp_wmb() (B) smp_mb() (D)
69 * STORE ->data_head STORE ->data_tail
72 * Where A pairs with D, and B pairs with C.
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.
78 * D needs to be a full barrier since it separates the data READ
79 * from the tail WRITE.
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.
84 * See perf_output_begin().
86 smp_wmb(); /* B, matches C */
87 rb->user_page->data_head = head;
90 * Now check if we missed an update -- rely on previous implied
91 * compiler barriers to force a re-read.
93 if (unlikely(head != local_read(&rb->head))) {
98 if (handle->wakeup != local_read(&rb->wakeup))
99 perf_output_wakeup(handle);
105 int perf_output_begin(struct perf_output_handle *handle,
106 struct perf_event *event, unsigned int size)
108 struct ring_buffer *rb;
109 unsigned long tail, offset, head;
110 int have_lost, page_shift;
112 struct perf_event_header header;
119 * For inherited events we send all the output towards the parent.
122 event = event->parent;
124 rb = rcu_dereference(event->rb);
128 if (unlikely(!rb->nr_pages))
132 handle->event = event;
134 have_lost = local_read(&rb->lost);
135 if (unlikely(have_lost)) {
136 size += sizeof(lost_event);
137 if (event->attr.sample_id_all)
138 size += event->id_header_size;
141 perf_output_get_handle(handle);
144 tail = READ_ONCE(rb->user_page->data_tail);
145 offset = head = local_read(&rb->head);
146 if (!rb->overwrite &&
147 unlikely(CIRC_SPACE(head, tail, perf_data_size(rb)) < size))
151 * The above forms a control dependency barrier separating the
152 * @tail load above from the data stores below. Since the @tail
153 * load is required to compute the branch to fail below.
155 * A, matches D; the full memory barrier userspace SHOULD issue
156 * after reading the data and before storing the new tail
159 * See perf_output_put_handle().
163 } while (local_cmpxchg(&rb->head, offset, head) != offset);
166 * We rely on the implied barrier() by local_cmpxchg() to ensure
167 * none of the data stores below can be lifted up by the compiler.
170 if (unlikely(head - local_read(&rb->wakeup) > rb->watermark))
171 local_add(rb->watermark, &rb->wakeup);
173 page_shift = PAGE_SHIFT + page_order(rb);
175 handle->page = (offset >> page_shift) & (rb->nr_pages - 1);
176 offset &= (1UL << page_shift) - 1;
177 handle->addr = rb->data_pages[handle->page] + offset;
178 handle->size = (1UL << page_shift) - offset;
180 if (unlikely(have_lost)) {
181 struct perf_sample_data sample_data;
183 lost_event.header.size = sizeof(lost_event);
184 lost_event.header.type = PERF_RECORD_LOST;
185 lost_event.header.misc = 0;
186 lost_event.id = event->id;
187 lost_event.lost = local_xchg(&rb->lost, 0);
189 perf_event_header__init_id(&lost_event.header,
190 &sample_data, event);
191 perf_output_put(handle, lost_event);
192 perf_event__output_id_sample(event, handle, &sample_data);
198 local_inc(&rb->lost);
199 perf_output_put_handle(handle);
206 unsigned int perf_output_copy(struct perf_output_handle *handle,
207 const void *buf, unsigned int len)
209 return __output_copy(handle, buf, len);
212 unsigned int perf_output_skip(struct perf_output_handle *handle,
215 return __output_skip(handle, NULL, len);
218 void perf_output_end(struct perf_output_handle *handle)
220 perf_output_put_handle(handle);
225 ring_buffer_init(struct ring_buffer *rb, long watermark, int flags)
227 long max_size = perf_data_size(rb);
230 rb->watermark = min(max_size, watermark);
233 rb->watermark = max_size / 2;
235 if (flags & RING_BUFFER_WRITABLE)
240 atomic_set(&rb->refcount, 1);
242 INIT_LIST_HEAD(&rb->event_list);
243 spin_lock_init(&rb->event_lock);
247 * This is called before hardware starts writing to the AUX area to
248 * obtain an output handle and make sure there's room in the buffer.
249 * When the capture completes, call perf_aux_output_end() to commit
250 * the recorded data to the buffer.
252 * The ordering is similar to that of perf_output_{begin,end}, with
253 * the exception of (B), which should be taken care of by the pmu
254 * driver, since ordering rules will differ depending on hardware.
256 * Call this from pmu::start(); see the comment in perf_aux_output_end()
257 * about its use in pmu callbacks. Both can also be called from the PMI
260 void *perf_aux_output_begin(struct perf_output_handle *handle,
261 struct perf_event *event)
263 struct perf_event *output_event = event;
264 unsigned long aux_head, aux_tail;
265 struct ring_buffer *rb;
267 if (output_event->parent)
268 output_event = output_event->parent;
271 * Since this will typically be open across pmu::add/pmu::del, we
272 * grab ring_buffer's refcount instead of holding rcu read lock
273 * to make sure it doesn't disappear under us.
275 rb = ring_buffer_get(output_event);
279 if (!rb_has_aux(rb) || !atomic_inc_not_zero(&rb->aux_refcount))
283 * If rb::aux_mmap_count is zero (and rb_has_aux() above went through),
284 * the aux buffer is in perf_mmap_close(), about to get freed.
286 if (!atomic_read(&rb->aux_mmap_count))
290 * Nesting is not supported for AUX area, make sure nested
291 * writers are caught early
293 if (WARN_ON_ONCE(local_xchg(&rb->aux_nest, 1)))
296 aux_head = local_read(&rb->aux_head);
299 handle->event = event;
300 handle->head = aux_head;
304 * In overwrite mode, AUX data stores do not depend on aux_tail,
305 * therefore (A) control dependency barrier does not exist. The
306 * (B) <-> (C) ordering is still observed by the pmu driver.
308 if (!rb->aux_overwrite) {
309 aux_tail = ACCESS_ONCE(rb->user_page->aux_tail);
310 handle->wakeup = local_read(&rb->aux_wakeup) + rb->aux_watermark;
311 if (aux_head - aux_tail < perf_aux_size(rb))
312 handle->size = CIRC_SPACE(aux_head, aux_tail, perf_aux_size(rb));
315 * handle->size computation depends on aux_tail load; this forms a
316 * control dependency barrier separating aux_tail load from aux data
317 * store that will be enabled on successful return
319 if (!handle->size) { /* A, matches D */
320 event->pending_disable = 1;
321 perf_output_wakeup(handle);
322 local_set(&rb->aux_nest, 0);
327 return handle->rb->aux_priv;
335 handle->event = NULL;
341 * Commit the data written by hardware into the ring buffer by adjusting
342 * aux_head and posting a PERF_RECORD_AUX into the perf buffer. It is the
343 * pmu driver's responsibility to observe ordering rules of the hardware,
344 * so that all the data is externally visible before this is called.
346 * Note: this has to be called from pmu::stop() callback, as the assumption
347 * of the AUX buffer management code is that after pmu::stop(), the AUX
348 * transaction must be stopped and therefore drop the AUX reference count.
350 void perf_aux_output_end(struct perf_output_handle *handle, unsigned long size,
353 struct ring_buffer *rb = handle->rb;
354 bool wakeup = truncated;
355 unsigned long aux_head;
359 flags |= PERF_AUX_FLAG_TRUNCATED;
361 /* in overwrite mode, driver provides aux_head via handle */
362 if (rb->aux_overwrite) {
363 flags |= PERF_AUX_FLAG_OVERWRITE;
365 aux_head = handle->head;
366 local_set(&rb->aux_head, aux_head);
368 aux_head = local_read(&rb->aux_head);
369 local_add(size, &rb->aux_head);
374 * Only send RECORD_AUX if we have something useful to communicate
377 perf_event_aux_event(handle->event, aux_head, size, flags);
380 aux_head = rb->user_page->aux_head = local_read(&rb->aux_head);
382 if (aux_head - local_read(&rb->aux_wakeup) >= rb->aux_watermark) {
384 local_add(rb->aux_watermark, &rb->aux_wakeup);
389 handle->event->pending_disable = 1;
390 perf_output_wakeup(handle);
393 handle->event = NULL;
395 local_set(&rb->aux_nest, 0);
402 * Skip over a given number of bytes in the AUX buffer, due to, for example,
403 * hardware's alignment constraints.
405 int perf_aux_output_skip(struct perf_output_handle *handle, unsigned long size)
407 struct ring_buffer *rb = handle->rb;
408 unsigned long aux_head;
410 if (size > handle->size)
413 local_add(size, &rb->aux_head);
415 aux_head = rb->user_page->aux_head = local_read(&rb->aux_head);
416 if (aux_head - local_read(&rb->aux_wakeup) >= rb->aux_watermark) {
417 perf_output_wakeup(handle);
418 local_add(rb->aux_watermark, &rb->aux_wakeup);
419 handle->wakeup = local_read(&rb->aux_wakeup) +
423 handle->head = aux_head;
424 handle->size -= size;
429 void *perf_get_aux(struct perf_output_handle *handle)
431 /* this is only valid between perf_aux_output_begin and *_end */
435 return handle->rb->aux_priv;
438 #define PERF_AUX_GFP (GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_NORETRY)
440 static struct page *rb_alloc_aux_page(int node, int order)
444 if (order > MAX_ORDER)
448 page = alloc_pages_node(node, PERF_AUX_GFP, order);
449 } while (!page && order--);
453 * Communicate the allocation size to the driver:
454 * if we managed to secure a high-order allocation,
455 * set its first page's private to this order;
456 * !PagePrivate(page) means it's just a normal page.
458 split_page(page, order);
459 SetPagePrivate(page);
460 set_page_private(page, order);
466 static void rb_free_aux_page(struct ring_buffer *rb, int idx)
468 struct page *page = virt_to_page(rb->aux_pages[idx]);
470 ClearPagePrivate(page);
471 page->mapping = NULL;
475 static void __rb_free_aux(struct ring_buffer *rb)
480 * Should never happen, the last reference should be dropped from
481 * perf_mmap_close() path, which first stops aux transactions (which
482 * in turn are the atomic holders of aux_refcount) and then does the
483 * last rb_free_aux().
485 WARN_ON_ONCE(in_atomic());
488 rb->free_aux(rb->aux_priv);
493 if (rb->aux_nr_pages) {
494 for (pg = 0; pg < rb->aux_nr_pages; pg++)
495 rb_free_aux_page(rb, pg);
497 kfree(rb->aux_pages);
498 rb->aux_nr_pages = 0;
502 int rb_alloc_aux(struct ring_buffer *rb, struct perf_event *event,
503 pgoff_t pgoff, int nr_pages, long watermark, int flags)
505 bool overwrite = !(flags & RING_BUFFER_WRITABLE);
506 int node = (event->cpu == -1) ? -1 : cpu_to_node(event->cpu);
507 int ret = -ENOMEM, max_order = 0;
512 if (event->pmu->capabilities & PERF_PMU_CAP_AUX_NO_SG) {
514 * We need to start with the max_order that fits in nr_pages,
515 * not the other way around, hence ilog2() and not get_order.
517 max_order = ilog2(nr_pages);
520 * PMU requests more than one contiguous chunks of memory
521 * for SW double buffering
523 if ((event->pmu->capabilities & PERF_PMU_CAP_AUX_SW_DOUBLEBUF) &&
532 rb->aux_pages = kzalloc_node(nr_pages * sizeof(void *), GFP_KERNEL, node);
536 rb->free_aux = event->pmu->free_aux;
537 for (rb->aux_nr_pages = 0; rb->aux_nr_pages < nr_pages;) {
541 order = min(max_order, ilog2(nr_pages - rb->aux_nr_pages));
542 page = rb_alloc_aux_page(node, order);
546 for (last = rb->aux_nr_pages + (1 << page_private(page));
547 last > rb->aux_nr_pages; rb->aux_nr_pages++)
548 rb->aux_pages[rb->aux_nr_pages] = page_address(page++);
552 * In overwrite mode, PMUs that don't support SG may not handle more
553 * than one contiguous allocation, since they rely on PMI to do double
554 * buffering. In this case, the entire buffer has to be one contiguous
557 if ((event->pmu->capabilities & PERF_PMU_CAP_AUX_NO_SG) &&
559 struct page *page = virt_to_page(rb->aux_pages[0]);
561 if (page_private(page) != max_order)
565 rb->aux_priv = event->pmu->setup_aux(event, rb->aux_pages, nr_pages,
573 * aux_pages (and pmu driver's private data, aux_priv) will be
574 * referenced in both producer's and consumer's contexts, thus
575 * we keep a refcount here to make sure either of the two can
576 * reference them safely.
578 atomic_set(&rb->aux_refcount, 1);
580 rb->aux_overwrite = overwrite;
581 rb->aux_watermark = watermark;
583 if (!rb->aux_watermark && !rb->aux_overwrite)
584 rb->aux_watermark = nr_pages << (PAGE_SHIFT - 1);
588 rb->aux_pgoff = pgoff;
595 void rb_free_aux(struct ring_buffer *rb)
597 if (atomic_dec_and_test(&rb->aux_refcount))
601 #ifndef CONFIG_PERF_USE_VMALLOC
604 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
608 __perf_mmap_to_page(struct ring_buffer *rb, unsigned long pgoff)
610 if (pgoff > rb->nr_pages)
614 return virt_to_page(rb->user_page);
616 return virt_to_page(rb->data_pages[pgoff - 1]);
619 static void *perf_mmap_alloc_page(int cpu)
624 node = (cpu == -1) ? cpu : cpu_to_node(cpu);
625 page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
629 return page_address(page);
632 struct ring_buffer *rb_alloc(int nr_pages, long watermark, int cpu, int flags)
634 struct ring_buffer *rb;
638 size = sizeof(struct ring_buffer);
639 size += nr_pages * sizeof(void *);
641 rb = kzalloc(size, GFP_KERNEL);
645 rb->user_page = perf_mmap_alloc_page(cpu);
649 for (i = 0; i < nr_pages; i++) {
650 rb->data_pages[i] = perf_mmap_alloc_page(cpu);
651 if (!rb->data_pages[i])
652 goto fail_data_pages;
655 rb->nr_pages = nr_pages;
657 ring_buffer_init(rb, watermark, flags);
662 for (i--; i >= 0; i--)
663 free_page((unsigned long)rb->data_pages[i]);
665 free_page((unsigned long)rb->user_page);
674 static void perf_mmap_free_page(unsigned long addr)
676 struct page *page = virt_to_page((void *)addr);
678 page->mapping = NULL;
682 void rb_free(struct ring_buffer *rb)
686 perf_mmap_free_page((unsigned long)rb->user_page);
687 for (i = 0; i < rb->nr_pages; i++)
688 perf_mmap_free_page((unsigned long)rb->data_pages[i]);
693 static int data_page_nr(struct ring_buffer *rb)
695 return rb->nr_pages << page_order(rb);
699 __perf_mmap_to_page(struct ring_buffer *rb, unsigned long pgoff)
701 /* The '>' counts in the user page. */
702 if (pgoff > data_page_nr(rb))
705 return vmalloc_to_page((void *)rb->user_page + pgoff * PAGE_SIZE);
708 static void perf_mmap_unmark_page(void *addr)
710 struct page *page = vmalloc_to_page(addr);
712 page->mapping = NULL;
715 static void rb_free_work(struct work_struct *work)
717 struct ring_buffer *rb;
721 rb = container_of(work, struct ring_buffer, work);
722 nr = data_page_nr(rb);
724 base = rb->user_page;
725 /* The '<=' counts in the user page. */
726 for (i = 0; i <= nr; i++)
727 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
733 void rb_free(struct ring_buffer *rb)
735 schedule_work(&rb->work);
738 struct ring_buffer *rb_alloc(int nr_pages, long watermark, int cpu, int flags)
740 struct ring_buffer *rb;
744 size = sizeof(struct ring_buffer);
745 size += sizeof(void *);
747 rb = kzalloc(size, GFP_KERNEL);
751 INIT_WORK(&rb->work, rb_free_work);
753 all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
757 rb->user_page = all_buf;
758 rb->data_pages[0] = all_buf + PAGE_SIZE;
759 rb->page_order = ilog2(nr_pages);
760 rb->nr_pages = !!nr_pages;
762 ring_buffer_init(rb, watermark, flags);
776 perf_mmap_to_page(struct ring_buffer *rb, unsigned long pgoff)
778 if (rb->aux_nr_pages) {
779 /* above AUX space */
780 if (pgoff > rb->aux_pgoff + rb->aux_nr_pages)
784 if (pgoff >= rb->aux_pgoff)
785 return virt_to_page(rb->aux_pages[pgoff - rb->aux_pgoff]);
788 return __perf_mmap_to_page(rb, pgoff);