1 // SPDX-License-Identifier: GPL-2.0
3 * Performance events ring-buffer code:
6 * Copyright (C) 2008-2011 Red Hat, Inc., Ingo Molnar
7 * Copyright (C) 2008-2011 Red Hat, Inc., Peter Zijlstra
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>
20 static void perf_output_wakeup(struct perf_output_handle *handle)
22 atomic_set(&handle->rb->poll, EPOLLIN);
24 handle->event->pending_wakeup = 1;
26 if (*perf_event_fasync(handle->event) && !handle->event->pending_kill)
27 handle->event->pending_kill = POLL_IN;
29 irq_work_queue(&handle->event->pending_irq);
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.
37 * We only publish the head (and generate a wakeup) when the outer-most
40 static void perf_output_get_handle(struct perf_output_handle *handle)
42 struct perf_buffer *rb = handle->rb;
47 * Avoid an explicit LOAD/STORE such that architectures with memops
50 (*(volatile unsigned int *)&rb->nest)++;
51 handle->wakeup = local_read(&rb->wakeup);
54 static void perf_output_put_handle(struct perf_output_handle *handle)
56 struct perf_buffer *rb = handle->rb;
61 * If this isn't the outermost nesting, we don't have to update
62 * @rb->user_page->data_head.
64 nest = READ_ONCE(rb->nest);
66 WRITE_ONCE(rb->nest, nest - 1);
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.
76 * Otherwise we can observe a @rb->head value before one published
77 * by an IRQ/NMI happening between the load and the increment.
80 head = local_read(&rb->head);
83 * IRQ/NMI can happen here and advance @rb->head, causing our
84 * load above to be stale.
88 * Since the mmap() consumer (userspace) can run on a different CPU:
92 * if (LOAD ->data_tail) { LOAD ->data_head
94 * STORE $data LOAD $data
95 * smp_wmb() (B) smp_mb() (D)
96 * STORE ->data_head STORE ->data_tail
99 * Where A pairs with D, and B pairs with C.
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.
105 * D needs to be a full barrier since it separates the data READ
106 * from the tail WRITE.
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.
111 * See perf_output_begin().
113 smp_wmb(); /* B, matches C */
114 WRITE_ONCE(rb->user_page->data_head, head);
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.
122 WRITE_ONCE(rb->nest, 0);
125 * Ensure we decrement @rb->nest before we validate the @rb->head.
126 * Otherwise we cannot be sure we caught the 'last' nested update.
129 if (unlikely(head != local_read(&rb->head))) {
130 WRITE_ONCE(rb->nest, 1);
134 if (handle->wakeup != local_read(&rb->wakeup))
135 perf_output_wakeup(handle);
141 static __always_inline bool
142 ring_buffer_has_space(unsigned long head, unsigned long tail,
143 unsigned long data_size, unsigned int size,
147 return CIRC_SPACE(head, tail, data_size) >= size;
149 return CIRC_SPACE(tail, head, data_size) >= size;
152 static __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,
158 struct perf_buffer *rb;
159 unsigned long tail, offset, head;
160 int have_lost, page_shift;
162 struct perf_event_header header;
169 * For inherited events we send all the output towards the parent.
172 event = event->parent;
174 rb = rcu_dereference(event->rb);
178 if (unlikely(rb->paused)) {
180 local_inc(&rb->lost);
181 atomic64_inc(&event->lost_samples);
187 handle->event = event;
189 have_lost = local_read(&rb->lost);
190 if (unlikely(have_lost)) {
191 size += sizeof(lost_event);
192 if (event->attr.sample_id_all)
193 size += event->id_header_size;
196 perf_output_get_handle(handle);
198 offset = local_read(&rb->head);
201 tail = READ_ONCE(rb->user_page->data_tail);
202 if (!rb->overwrite) {
203 if (unlikely(!ring_buffer_has_space(head, tail,
210 * The above forms a control dependency barrier separating the
211 * @tail load above from the data stores below. Since the @tail
212 * load is required to compute the branch to fail below.
214 * A, matches D; the full memory barrier userspace SHOULD issue
215 * after reading the data and before storing the new tail
218 * See perf_output_put_handle().
225 } while (!local_try_cmpxchg(&rb->head, &offset, head));
233 * We rely on the implied barrier() by local_cmpxchg() to ensure
234 * none of the data stores below can be lifted up by the compiler.
237 if (unlikely(head - local_read(&rb->wakeup) > rb->watermark))
238 local_add(rb->watermark, &rb->wakeup);
240 page_shift = PAGE_SHIFT + page_order(rb);
242 handle->page = (offset >> page_shift) & (rb->nr_pages - 1);
243 offset &= (1UL << page_shift) - 1;
244 handle->addr = rb->data_pages[handle->page] + offset;
245 handle->size = (1UL << page_shift) - offset;
247 if (unlikely(have_lost)) {
248 lost_event.header.size = sizeof(lost_event);
249 lost_event.header.type = PERF_RECORD_LOST;
250 lost_event.header.misc = 0;
251 lost_event.id = event->id;
252 lost_event.lost = local_xchg(&rb->lost, 0);
254 /* XXX mostly redundant; @data is already fully initializes */
255 perf_event_header__init_id(&lost_event.header, data, event);
256 perf_output_put(handle, lost_event);
257 perf_event__output_id_sample(event, handle, data);
263 local_inc(&rb->lost);
264 atomic64_inc(&event->lost_samples);
265 perf_output_put_handle(handle);
272 int perf_output_begin_forward(struct perf_output_handle *handle,
273 struct perf_sample_data *data,
274 struct perf_event *event, unsigned int size)
276 return __perf_output_begin(handle, data, event, size, false);
279 int perf_output_begin_backward(struct perf_output_handle *handle,
280 struct perf_sample_data *data,
281 struct perf_event *event, unsigned int size)
283 return __perf_output_begin(handle, data, event, size, true);
286 int perf_output_begin(struct perf_output_handle *handle,
287 struct perf_sample_data *data,
288 struct perf_event *event, unsigned int size)
291 return __perf_output_begin(handle, data, event, size,
292 unlikely(is_write_backward(event)));
295 unsigned int perf_output_copy(struct perf_output_handle *handle,
296 const void *buf, unsigned int len)
298 return __output_copy(handle, buf, len);
301 unsigned int perf_output_skip(struct perf_output_handle *handle,
304 return __output_skip(handle, NULL, len);
307 void perf_output_end(struct perf_output_handle *handle)
309 perf_output_put_handle(handle);
314 ring_buffer_init(struct perf_buffer *rb, long watermark, int flags)
316 long max_size = perf_data_size(rb);
319 rb->watermark = min(max_size, watermark);
322 rb->watermark = max_size / 2;
324 if (flags & RING_BUFFER_WRITABLE)
329 refcount_set(&rb->refcount, 1);
331 INIT_LIST_HEAD(&rb->event_list);
332 spin_lock_init(&rb->event_lock);
335 * perf_output_begin() only checks rb->paused, therefore
336 * rb->paused must be true if we have no pages for output.
342 void perf_aux_output_flag(struct perf_output_handle *handle, u64 flags)
345 * OVERWRITE is determined by perf_aux_output_end() and can't
346 * be passed in directly.
348 if (WARN_ON_ONCE(flags & PERF_AUX_FLAG_OVERWRITE))
351 handle->aux_flags |= flags;
353 EXPORT_SYMBOL_GPL(perf_aux_output_flag);
356 * This is called before hardware starts writing to the AUX area to
357 * obtain an output handle and make sure there's room in the buffer.
358 * When the capture completes, call perf_aux_output_end() to commit
359 * the recorded data to the buffer.
361 * The ordering is similar to that of perf_output_{begin,end}, with
362 * the exception of (B), which should be taken care of by the pmu
363 * driver, since ordering rules will differ depending on hardware.
365 * Call this from pmu::start(); see the comment in perf_aux_output_end()
366 * about its use in pmu callbacks. Both can also be called from the PMI
369 void *perf_aux_output_begin(struct perf_output_handle *handle,
370 struct perf_event *event)
372 struct perf_event *output_event = event;
373 unsigned long aux_head, aux_tail;
374 struct perf_buffer *rb;
377 if (output_event->parent)
378 output_event = output_event->parent;
381 * Since this will typically be open across pmu::add/pmu::del, we
382 * grab ring_buffer's refcount instead of holding rcu read lock
383 * to make sure it doesn't disappear under us.
385 rb = ring_buffer_get(output_event);
393 * If aux_mmap_count is zero, the aux buffer is in perf_mmap_close(),
394 * about to get freed, so we leave immediately.
396 * Checking rb::aux_mmap_count and rb::refcount has to be done in
397 * the same order, see perf_mmap_close. Otherwise we end up freeing
398 * aux pages in this path, which is a bug, because in_atomic().
400 if (!atomic_read(&rb->aux_mmap_count))
403 if (!refcount_inc_not_zero(&rb->aux_refcount))
406 nest = READ_ONCE(rb->aux_nest);
408 * Nesting is not supported for AUX area, make sure nested
409 * writers are caught early
411 if (WARN_ON_ONCE(nest))
414 WRITE_ONCE(rb->aux_nest, nest + 1);
416 aux_head = rb->aux_head;
419 handle->event = event;
420 handle->head = aux_head;
422 handle->aux_flags = 0;
425 * In overwrite mode, AUX data stores do not depend on aux_tail,
426 * therefore (A) control dependency barrier does not exist. The
427 * (B) <-> (C) ordering is still observed by the pmu driver.
429 if (!rb->aux_overwrite) {
430 aux_tail = READ_ONCE(rb->user_page->aux_tail);
431 handle->wakeup = rb->aux_wakeup + rb->aux_watermark;
432 if (aux_head - aux_tail < perf_aux_size(rb))
433 handle->size = CIRC_SPACE(aux_head, aux_tail, perf_aux_size(rb));
436 * handle->size computation depends on aux_tail load; this forms a
437 * control dependency barrier separating aux_tail load from aux data
438 * store that will be enabled on successful return
440 if (!handle->size) { /* A, matches D */
441 event->pending_disable = smp_processor_id();
442 perf_output_wakeup(handle);
443 WRITE_ONCE(rb->aux_nest, 0);
448 return handle->rb->aux_priv;
456 handle->event = NULL;
460 EXPORT_SYMBOL_GPL(perf_aux_output_begin);
462 static __always_inline bool rb_need_aux_wakeup(struct perf_buffer *rb)
464 if (rb->aux_overwrite)
467 if (rb->aux_head - rb->aux_wakeup >= rb->aux_watermark) {
468 rb->aux_wakeup = rounddown(rb->aux_head, rb->aux_watermark);
476 * Commit the data written by hardware into the ring buffer by adjusting
477 * aux_head and posting a PERF_RECORD_AUX into the perf buffer. It is the
478 * pmu driver's responsibility to observe ordering rules of the hardware,
479 * so that all the data is externally visible before this is called.
481 * Note: this has to be called from pmu::stop() callback, as the assumption
482 * of the AUX buffer management code is that after pmu::stop(), the AUX
483 * transaction must be stopped and therefore drop the AUX reference count.
485 void perf_aux_output_end(struct perf_output_handle *handle, unsigned long size)
487 bool wakeup = !!(handle->aux_flags & PERF_AUX_FLAG_TRUNCATED);
488 struct perf_buffer *rb = handle->rb;
489 unsigned long aux_head;
491 /* in overwrite mode, driver provides aux_head via handle */
492 if (rb->aux_overwrite) {
493 handle->aux_flags |= PERF_AUX_FLAG_OVERWRITE;
495 aux_head = handle->head;
496 rb->aux_head = aux_head;
498 handle->aux_flags &= ~PERF_AUX_FLAG_OVERWRITE;
500 aux_head = rb->aux_head;
501 rb->aux_head += size;
505 * Only send RECORD_AUX if we have something useful to communicate
507 * Note: the OVERWRITE records by themselves are not considered
508 * useful, as they don't communicate any *new* information,
509 * aside from the short-lived offset, that becomes history at
510 * the next event sched-in and therefore isn't useful.
511 * The userspace that needs to copy out AUX data in overwrite
512 * mode should know to use user_page::aux_head for the actual
513 * offset. So, from now on we don't output AUX records that
514 * have *only* OVERWRITE flag set.
516 if (size || (handle->aux_flags & ~(u64)PERF_AUX_FLAG_OVERWRITE))
517 perf_event_aux_event(handle->event, aux_head, size,
520 WRITE_ONCE(rb->user_page->aux_head, rb->aux_head);
521 if (rb_need_aux_wakeup(rb))
525 if (handle->aux_flags & PERF_AUX_FLAG_TRUNCATED)
526 handle->event->pending_disable = smp_processor_id();
527 perf_output_wakeup(handle);
530 handle->event = NULL;
532 WRITE_ONCE(rb->aux_nest, 0);
537 EXPORT_SYMBOL_GPL(perf_aux_output_end);
540 * Skip over a given number of bytes in the AUX buffer, due to, for example,
541 * hardware's alignment constraints.
543 int perf_aux_output_skip(struct perf_output_handle *handle, unsigned long size)
545 struct perf_buffer *rb = handle->rb;
547 if (size > handle->size)
550 rb->aux_head += size;
552 WRITE_ONCE(rb->user_page->aux_head, rb->aux_head);
553 if (rb_need_aux_wakeup(rb)) {
554 perf_output_wakeup(handle);
555 handle->wakeup = rb->aux_wakeup + rb->aux_watermark;
558 handle->head = rb->aux_head;
559 handle->size -= size;
563 EXPORT_SYMBOL_GPL(perf_aux_output_skip);
565 void *perf_get_aux(struct perf_output_handle *handle)
567 /* this is only valid between perf_aux_output_begin and *_end */
571 return handle->rb->aux_priv;
573 EXPORT_SYMBOL_GPL(perf_get_aux);
576 * Copy out AUX data from an AUX handle.
578 long perf_output_copy_aux(struct perf_output_handle *aux_handle,
579 struct perf_output_handle *handle,
580 unsigned long from, unsigned long to)
582 struct perf_buffer *rb = aux_handle->rb;
583 unsigned long tocopy, remainder, len = 0;
586 from &= (rb->aux_nr_pages << PAGE_SHIFT) - 1;
587 to &= (rb->aux_nr_pages << PAGE_SHIFT) - 1;
590 tocopy = PAGE_SIZE - offset_in_page(from);
592 tocopy = min(tocopy, to - from);
596 addr = rb->aux_pages[from >> PAGE_SHIFT];
597 addr += offset_in_page(from);
599 remainder = perf_output_copy(handle, addr, tocopy);
605 from &= (rb->aux_nr_pages << PAGE_SHIFT) - 1;
606 } while (to != from);
611 #define PERF_AUX_GFP (GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_NORETRY)
613 static struct page *rb_alloc_aux_page(int node, int order)
617 if (order > MAX_PAGE_ORDER)
618 order = MAX_PAGE_ORDER;
621 page = alloc_pages_node(node, PERF_AUX_GFP, order);
622 } while (!page && order--);
626 * Communicate the allocation size to the driver:
627 * if we managed to secure a high-order allocation,
628 * set its first page's private to this order;
629 * !PagePrivate(page) means it's just a normal page.
631 split_page(page, order);
632 SetPagePrivate(page);
633 set_page_private(page, order);
639 static void rb_free_aux_page(struct perf_buffer *rb, int idx)
641 struct page *page = virt_to_page(rb->aux_pages[idx]);
643 ClearPagePrivate(page);
644 page->mapping = NULL;
648 static void __rb_free_aux(struct perf_buffer *rb)
653 * Should never happen, the last reference should be dropped from
654 * perf_mmap_close() path, which first stops aux transactions (which
655 * in turn are the atomic holders of aux_refcount) and then does the
656 * last rb_free_aux().
658 WARN_ON_ONCE(in_atomic());
661 rb->free_aux(rb->aux_priv);
666 if (rb->aux_nr_pages) {
667 for (pg = 0; pg < rb->aux_nr_pages; pg++)
668 rb_free_aux_page(rb, pg);
670 kfree(rb->aux_pages);
671 rb->aux_nr_pages = 0;
675 int rb_alloc_aux(struct perf_buffer *rb, struct perf_event *event,
676 pgoff_t pgoff, int nr_pages, long watermark, int flags)
678 bool overwrite = !(flags & RING_BUFFER_WRITABLE);
679 int node = (event->cpu == -1) ? -1 : cpu_to_node(event->cpu);
680 int ret = -ENOMEM, max_order;
687 * Watermark defaults to half the buffer, and so does the
688 * max_order, to aid PMU drivers in double buffering.
691 watermark = nr_pages << (PAGE_SHIFT - 1);
694 * Use aux_watermark as the basis for chunking to
695 * help PMU drivers honor the watermark.
697 max_order = get_order(watermark);
700 * We need to start with the max_order that fits in nr_pages,
701 * not the other way around, hence ilog2() and not get_order.
703 max_order = ilog2(nr_pages);
708 * kcalloc_node() is unable to allocate buffer if the size is larger
709 * than: PAGE_SIZE << MAX_PAGE_ORDER; directly bail out in this case.
711 if (get_order((unsigned long)nr_pages * sizeof(void *)) > MAX_PAGE_ORDER)
713 rb->aux_pages = kcalloc_node(nr_pages, sizeof(void *), GFP_KERNEL,
718 rb->free_aux = event->pmu->free_aux;
719 for (rb->aux_nr_pages = 0; rb->aux_nr_pages < nr_pages;) {
723 order = min(max_order, ilog2(nr_pages - rb->aux_nr_pages));
724 page = rb_alloc_aux_page(node, order);
728 for (last = rb->aux_nr_pages + (1 << page_private(page));
729 last > rb->aux_nr_pages; rb->aux_nr_pages++)
730 rb->aux_pages[rb->aux_nr_pages] = page_address(page++);
734 * In overwrite mode, PMUs that don't support SG may not handle more
735 * than one contiguous allocation, since they rely on PMI to do double
736 * buffering. In this case, the entire buffer has to be one contiguous
739 if ((event->pmu->capabilities & PERF_PMU_CAP_AUX_NO_SG) &&
741 struct page *page = virt_to_page(rb->aux_pages[0]);
743 if (page_private(page) != max_order)
747 rb->aux_priv = event->pmu->setup_aux(event, rb->aux_pages, nr_pages,
755 * aux_pages (and pmu driver's private data, aux_priv) will be
756 * referenced in both producer's and consumer's contexts, thus
757 * we keep a refcount here to make sure either of the two can
758 * reference them safely.
760 refcount_set(&rb->aux_refcount, 1);
762 rb->aux_overwrite = overwrite;
763 rb->aux_watermark = watermark;
767 rb->aux_pgoff = pgoff;
774 void rb_free_aux(struct perf_buffer *rb)
776 if (refcount_dec_and_test(&rb->aux_refcount))
780 #ifndef CONFIG_PERF_USE_VMALLOC
783 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
787 __perf_mmap_to_page(struct perf_buffer *rb, unsigned long pgoff)
789 if (pgoff > rb->nr_pages)
793 return virt_to_page(rb->user_page);
795 return virt_to_page(rb->data_pages[pgoff - 1]);
798 static void *perf_mmap_alloc_page(int cpu)
803 node = (cpu == -1) ? cpu : cpu_to_node(cpu);
804 page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
808 return page_address(page);
811 static void perf_mmap_free_page(void *addr)
813 struct page *page = virt_to_page(addr);
815 page->mapping = NULL;
819 struct perf_buffer *rb_alloc(int nr_pages, long watermark, int cpu, int flags)
821 struct perf_buffer *rb;
825 size = sizeof(struct perf_buffer);
826 size += nr_pages * sizeof(void *);
828 if (order_base_2(size) > PAGE_SHIFT+MAX_PAGE_ORDER)
831 node = (cpu == -1) ? cpu : cpu_to_node(cpu);
832 rb = kzalloc_node(size, GFP_KERNEL, node);
836 rb->user_page = perf_mmap_alloc_page(cpu);
840 for (i = 0; i < nr_pages; i++) {
841 rb->data_pages[i] = perf_mmap_alloc_page(cpu);
842 if (!rb->data_pages[i])
843 goto fail_data_pages;
846 rb->nr_pages = nr_pages;
848 ring_buffer_init(rb, watermark, flags);
853 for (i--; i >= 0; i--)
854 perf_mmap_free_page(rb->data_pages[i]);
856 perf_mmap_free_page(rb->user_page);
865 void rb_free(struct perf_buffer *rb)
869 perf_mmap_free_page(rb->user_page);
870 for (i = 0; i < rb->nr_pages; i++)
871 perf_mmap_free_page(rb->data_pages[i]);
877 __perf_mmap_to_page(struct perf_buffer *rb, unsigned long pgoff)
879 /* The '>' counts in the user page. */
880 if (pgoff > data_page_nr(rb))
883 return vmalloc_to_page((void *)rb->user_page + pgoff * PAGE_SIZE);
886 static void perf_mmap_unmark_page(void *addr)
888 struct page *page = vmalloc_to_page(addr);
890 page->mapping = NULL;
893 static void rb_free_work(struct work_struct *work)
895 struct perf_buffer *rb;
899 rb = container_of(work, struct perf_buffer, work);
900 nr = data_page_nr(rb);
902 base = rb->user_page;
903 /* The '<=' counts in the user page. */
904 for (i = 0; i <= nr; i++)
905 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
911 void rb_free(struct perf_buffer *rb)
913 schedule_work(&rb->work);
916 struct perf_buffer *rb_alloc(int nr_pages, long watermark, int cpu, int flags)
918 struct perf_buffer *rb;
923 size = sizeof(struct perf_buffer);
924 size += sizeof(void *);
926 node = (cpu == -1) ? cpu : cpu_to_node(cpu);
927 rb = kzalloc_node(size, GFP_KERNEL, node);
931 INIT_WORK(&rb->work, rb_free_work);
933 all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
937 rb->user_page = all_buf;
938 rb->data_pages[0] = all_buf + PAGE_SIZE;
941 rb->page_order = ilog2(nr_pages);
944 ring_buffer_init(rb, watermark, flags);
958 perf_mmap_to_page(struct perf_buffer *rb, unsigned long pgoff)
960 if (rb->aux_nr_pages) {
961 /* above AUX space */
962 if (pgoff > rb->aux_pgoff + rb->aux_nr_pages)
966 if (pgoff >= rb->aux_pgoff) {
967 int aux_pgoff = array_index_nospec(pgoff - rb->aux_pgoff, rb->aux_nr_pages);
968 return virt_to_page(rb->aux_pages[aux_pgoff]);
972 return __perf_mmap_to_page(rb, pgoff);