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;
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 static __always_inline bool
106 ring_buffer_has_space(unsigned long head, unsigned long tail,
107 unsigned long data_size, unsigned int size,
111 return CIRC_SPACE(head, tail, data_size) >= size;
113 return CIRC_SPACE(tail, head, data_size) >= size;
116 static __always_inline int
117 __perf_output_begin(struct perf_output_handle *handle,
118 struct perf_event *event, unsigned int size,
121 struct ring_buffer *rb;
122 unsigned long tail, offset, head;
123 int have_lost, page_shift;
125 struct perf_event_header header;
132 * For inherited events we send all the output towards the parent.
135 event = event->parent;
137 rb = rcu_dereference(event->rb);
141 if (unlikely(rb->paused)) {
143 local_inc(&rb->lost);
148 handle->event = event;
150 have_lost = local_read(&rb->lost);
151 if (unlikely(have_lost)) {
152 size += sizeof(lost_event);
153 if (event->attr.sample_id_all)
154 size += event->id_header_size;
157 perf_output_get_handle(handle);
160 tail = READ_ONCE(rb->user_page->data_tail);
161 offset = head = local_read(&rb->head);
162 if (!rb->overwrite) {
163 if (unlikely(!ring_buffer_has_space(head, tail,
170 * The above forms a control dependency barrier separating the
171 * @tail load above from the data stores below. Since the @tail
172 * load is required to compute the branch to fail below.
174 * A, matches D; the full memory barrier userspace SHOULD issue
175 * after reading the data and before storing the new tail
178 * See perf_output_put_handle().
185 } while (local_cmpxchg(&rb->head, offset, head) != offset);
193 * We rely on the implied barrier() by local_cmpxchg() to ensure
194 * none of the data stores below can be lifted up by the compiler.
197 if (unlikely(head - local_read(&rb->wakeup) > rb->watermark))
198 local_add(rb->watermark, &rb->wakeup);
200 page_shift = PAGE_SHIFT + page_order(rb);
202 handle->page = (offset >> page_shift) & (rb->nr_pages - 1);
203 offset &= (1UL << page_shift) - 1;
204 handle->addr = rb->data_pages[handle->page] + offset;
205 handle->size = (1UL << page_shift) - offset;
207 if (unlikely(have_lost)) {
208 struct perf_sample_data sample_data;
210 lost_event.header.size = sizeof(lost_event);
211 lost_event.header.type = PERF_RECORD_LOST;
212 lost_event.header.misc = 0;
213 lost_event.id = event->id;
214 lost_event.lost = local_xchg(&rb->lost, 0);
216 perf_event_header__init_id(&lost_event.header,
217 &sample_data, event);
218 perf_output_put(handle, lost_event);
219 perf_event__output_id_sample(event, handle, &sample_data);
225 local_inc(&rb->lost);
226 perf_output_put_handle(handle);
233 int perf_output_begin_forward(struct perf_output_handle *handle,
234 struct perf_event *event, unsigned int size)
236 return __perf_output_begin(handle, event, size, false);
239 int perf_output_begin_backward(struct perf_output_handle *handle,
240 struct perf_event *event, unsigned int size)
242 return __perf_output_begin(handle, event, size, true);
245 int perf_output_begin(struct perf_output_handle *handle,
246 struct perf_event *event, unsigned int size)
249 return __perf_output_begin(handle, event, size,
250 unlikely(is_write_backward(event)));
253 unsigned int perf_output_copy(struct perf_output_handle *handle,
254 const void *buf, unsigned int len)
256 return __output_copy(handle, buf, len);
259 unsigned int perf_output_skip(struct perf_output_handle *handle,
262 return __output_skip(handle, NULL, len);
265 void perf_output_end(struct perf_output_handle *handle)
267 perf_output_put_handle(handle);
272 ring_buffer_init(struct ring_buffer *rb, long watermark, int flags)
274 long max_size = perf_data_size(rb);
277 rb->watermark = min(max_size, watermark);
280 rb->watermark = max_size / 2;
282 if (flags & RING_BUFFER_WRITABLE)
287 refcount_set(&rb->refcount, 1);
289 INIT_LIST_HEAD(&rb->event_list);
290 spin_lock_init(&rb->event_lock);
293 * perf_output_begin() only checks rb->paused, therefore
294 * rb->paused must be true if we have no pages for output.
300 void perf_aux_output_flag(struct perf_output_handle *handle, u64 flags)
303 * OVERWRITE is determined by perf_aux_output_end() and can't
304 * be passed in directly.
306 if (WARN_ON_ONCE(flags & PERF_AUX_FLAG_OVERWRITE))
309 handle->aux_flags |= flags;
311 EXPORT_SYMBOL_GPL(perf_aux_output_flag);
314 * This is called before hardware starts writing to the AUX area to
315 * obtain an output handle and make sure there's room in the buffer.
316 * When the capture completes, call perf_aux_output_end() to commit
317 * the recorded data to the buffer.
319 * The ordering is similar to that of perf_output_{begin,end}, with
320 * the exception of (B), which should be taken care of by the pmu
321 * driver, since ordering rules will differ depending on hardware.
323 * Call this from pmu::start(); see the comment in perf_aux_output_end()
324 * about its use in pmu callbacks. Both can also be called from the PMI
327 void *perf_aux_output_begin(struct perf_output_handle *handle,
328 struct perf_event *event)
330 struct perf_event *output_event = event;
331 unsigned long aux_head, aux_tail;
332 struct ring_buffer *rb;
334 if (output_event->parent)
335 output_event = output_event->parent;
338 * Since this will typically be open across pmu::add/pmu::del, we
339 * grab ring_buffer's refcount instead of holding rcu read lock
340 * to make sure it doesn't disappear under us.
342 rb = ring_buffer_get(output_event);
350 * If aux_mmap_count is zero, the aux buffer is in perf_mmap_close(),
351 * about to get freed, so we leave immediately.
353 * Checking rb::aux_mmap_count and rb::refcount has to be done in
354 * the same order, see perf_mmap_close. Otherwise we end up freeing
355 * aux pages in this path, which is a bug, because in_atomic().
357 if (!atomic_read(&rb->aux_mmap_count))
360 if (!refcount_inc_not_zero(&rb->aux_refcount))
364 * Nesting is not supported for AUX area, make sure nested
365 * writers are caught early
367 if (WARN_ON_ONCE(local_xchg(&rb->aux_nest, 1)))
370 aux_head = rb->aux_head;
373 handle->event = event;
374 handle->head = aux_head;
376 handle->aux_flags = 0;
379 * In overwrite mode, AUX data stores do not depend on aux_tail,
380 * therefore (A) control dependency barrier does not exist. The
381 * (B) <-> (C) ordering is still observed by the pmu driver.
383 if (!rb->aux_overwrite) {
384 aux_tail = READ_ONCE(rb->user_page->aux_tail);
385 handle->wakeup = rb->aux_wakeup + rb->aux_watermark;
386 if (aux_head - aux_tail < perf_aux_size(rb))
387 handle->size = CIRC_SPACE(aux_head, aux_tail, perf_aux_size(rb));
390 * handle->size computation depends on aux_tail load; this forms a
391 * control dependency barrier separating aux_tail load from aux data
392 * store that will be enabled on successful return
394 if (!handle->size) { /* A, matches D */
395 event->pending_disable = 1;
396 perf_output_wakeup(handle);
397 local_set(&rb->aux_nest, 0);
402 return handle->rb->aux_priv;
410 handle->event = NULL;
414 EXPORT_SYMBOL_GPL(perf_aux_output_begin);
416 static __always_inline bool rb_need_aux_wakeup(struct ring_buffer *rb)
418 if (rb->aux_overwrite)
421 if (rb->aux_head - rb->aux_wakeup >= rb->aux_watermark) {
422 rb->aux_wakeup = rounddown(rb->aux_head, rb->aux_watermark);
430 * Commit the data written by hardware into the ring buffer by adjusting
431 * aux_head and posting a PERF_RECORD_AUX into the perf buffer. It is the
432 * pmu driver's responsibility to observe ordering rules of the hardware,
433 * so that all the data is externally visible before this is called.
435 * Note: this has to be called from pmu::stop() callback, as the assumption
436 * of the AUX buffer management code is that after pmu::stop(), the AUX
437 * transaction must be stopped and therefore drop the AUX reference count.
439 void perf_aux_output_end(struct perf_output_handle *handle, unsigned long size)
441 bool wakeup = !!(handle->aux_flags & PERF_AUX_FLAG_TRUNCATED);
442 struct ring_buffer *rb = handle->rb;
443 unsigned long aux_head;
445 /* in overwrite mode, driver provides aux_head via handle */
446 if (rb->aux_overwrite) {
447 handle->aux_flags |= PERF_AUX_FLAG_OVERWRITE;
449 aux_head = handle->head;
450 rb->aux_head = aux_head;
452 handle->aux_flags &= ~PERF_AUX_FLAG_OVERWRITE;
454 aux_head = rb->aux_head;
455 rb->aux_head += size;
458 if (size || handle->aux_flags) {
460 * Only send RECORD_AUX if we have something useful to communicate
462 * Note: the OVERWRITE records by themselves are not considered
463 * useful, as they don't communicate any *new* information,
464 * aside from the short-lived offset, that becomes history at
465 * the next event sched-in and therefore isn't useful.
466 * The userspace that needs to copy out AUX data in overwrite
467 * mode should know to use user_page::aux_head for the actual
468 * offset. So, from now on we don't output AUX records that
469 * have *only* OVERWRITE flag set.
472 if (handle->aux_flags & ~(u64)PERF_AUX_FLAG_OVERWRITE)
473 perf_event_aux_event(handle->event, aux_head, size,
477 rb->user_page->aux_head = rb->aux_head;
478 if (rb_need_aux_wakeup(rb))
482 if (handle->aux_flags & PERF_AUX_FLAG_TRUNCATED)
483 handle->event->pending_disable = 1;
484 perf_output_wakeup(handle);
487 handle->event = NULL;
489 local_set(&rb->aux_nest, 0);
494 EXPORT_SYMBOL_GPL(perf_aux_output_end);
497 * Skip over a given number of bytes in the AUX buffer, due to, for example,
498 * hardware's alignment constraints.
500 int perf_aux_output_skip(struct perf_output_handle *handle, unsigned long size)
502 struct ring_buffer *rb = handle->rb;
504 if (size > handle->size)
507 rb->aux_head += size;
509 rb->user_page->aux_head = rb->aux_head;
510 if (rb_need_aux_wakeup(rb)) {
511 perf_output_wakeup(handle);
512 handle->wakeup = rb->aux_wakeup + rb->aux_watermark;
515 handle->head = rb->aux_head;
516 handle->size -= size;
520 EXPORT_SYMBOL_GPL(perf_aux_output_skip);
522 void *perf_get_aux(struct perf_output_handle *handle)
524 /* this is only valid between perf_aux_output_begin and *_end */
528 return handle->rb->aux_priv;
530 EXPORT_SYMBOL_GPL(perf_get_aux);
532 #define PERF_AUX_GFP (GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_NORETRY)
534 static struct page *rb_alloc_aux_page(int node, int order)
538 if (order > MAX_ORDER)
542 page = alloc_pages_node(node, PERF_AUX_GFP, order);
543 } while (!page && order--);
547 * Communicate the allocation size to the driver:
548 * if we managed to secure a high-order allocation,
549 * set its first page's private to this order;
550 * !PagePrivate(page) means it's just a normal page.
552 split_page(page, order);
553 SetPagePrivate(page);
554 set_page_private(page, order);
560 static void rb_free_aux_page(struct ring_buffer *rb, int idx)
562 struct page *page = virt_to_page(rb->aux_pages[idx]);
564 ClearPagePrivate(page);
565 page->mapping = NULL;
569 static void __rb_free_aux(struct ring_buffer *rb)
574 * Should never happen, the last reference should be dropped from
575 * perf_mmap_close() path, which first stops aux transactions (which
576 * in turn are the atomic holders of aux_refcount) and then does the
577 * last rb_free_aux().
579 WARN_ON_ONCE(in_atomic());
582 rb->free_aux(rb->aux_priv);
587 if (rb->aux_nr_pages) {
588 for (pg = 0; pg < rb->aux_nr_pages; pg++)
589 rb_free_aux_page(rb, pg);
591 kfree(rb->aux_pages);
592 rb->aux_nr_pages = 0;
596 int rb_alloc_aux(struct ring_buffer *rb, struct perf_event *event,
597 pgoff_t pgoff, int nr_pages, long watermark, int flags)
599 bool overwrite = !(flags & RING_BUFFER_WRITABLE);
600 int node = (event->cpu == -1) ? -1 : cpu_to_node(event->cpu);
601 int ret = -ENOMEM, max_order;
607 * We need to start with the max_order that fits in nr_pages,
608 * not the other way around, hence ilog2() and not get_order.
610 max_order = ilog2(nr_pages);
613 * PMU requests more than one contiguous chunks of memory
614 * for SW double buffering
616 if ((event->pmu->capabilities & PERF_PMU_CAP_AUX_SW_DOUBLEBUF) &&
624 rb->aux_pages = kcalloc_node(nr_pages, sizeof(void *), GFP_KERNEL,
629 rb->free_aux = event->pmu->free_aux;
630 for (rb->aux_nr_pages = 0; rb->aux_nr_pages < nr_pages;) {
634 order = min(max_order, ilog2(nr_pages - rb->aux_nr_pages));
635 page = rb_alloc_aux_page(node, order);
639 for (last = rb->aux_nr_pages + (1 << page_private(page));
640 last > rb->aux_nr_pages; rb->aux_nr_pages++)
641 rb->aux_pages[rb->aux_nr_pages] = page_address(page++);
645 * In overwrite mode, PMUs that don't support SG may not handle more
646 * than one contiguous allocation, since they rely on PMI to do double
647 * buffering. In this case, the entire buffer has to be one contiguous
650 if ((event->pmu->capabilities & PERF_PMU_CAP_AUX_NO_SG) &&
652 struct page *page = virt_to_page(rb->aux_pages[0]);
654 if (page_private(page) != max_order)
658 rb->aux_priv = event->pmu->setup_aux(event, rb->aux_pages, nr_pages,
666 * aux_pages (and pmu driver's private data, aux_priv) will be
667 * referenced in both producer's and consumer's contexts, thus
668 * we keep a refcount here to make sure either of the two can
669 * reference them safely.
671 refcount_set(&rb->aux_refcount, 1);
673 rb->aux_overwrite = overwrite;
674 rb->aux_watermark = watermark;
676 if (!rb->aux_watermark && !rb->aux_overwrite)
677 rb->aux_watermark = nr_pages << (PAGE_SHIFT - 1);
681 rb->aux_pgoff = pgoff;
688 void rb_free_aux(struct ring_buffer *rb)
690 if (refcount_dec_and_test(&rb->aux_refcount))
694 #ifndef CONFIG_PERF_USE_VMALLOC
697 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
701 __perf_mmap_to_page(struct ring_buffer *rb, unsigned long pgoff)
703 if (pgoff > rb->nr_pages)
707 return virt_to_page(rb->user_page);
709 return virt_to_page(rb->data_pages[pgoff - 1]);
712 static void *perf_mmap_alloc_page(int cpu)
717 node = (cpu == -1) ? cpu : cpu_to_node(cpu);
718 page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
722 return page_address(page);
725 struct ring_buffer *rb_alloc(int nr_pages, long watermark, int cpu, int flags)
727 struct ring_buffer *rb;
731 size = sizeof(struct ring_buffer);
732 size += nr_pages * sizeof(void *);
734 if (order_base_2(size) >= PAGE_SHIFT+MAX_ORDER)
737 rb = kzalloc(size, GFP_KERNEL);
741 rb->user_page = perf_mmap_alloc_page(cpu);
745 for (i = 0; i < nr_pages; i++) {
746 rb->data_pages[i] = perf_mmap_alloc_page(cpu);
747 if (!rb->data_pages[i])
748 goto fail_data_pages;
751 rb->nr_pages = nr_pages;
753 ring_buffer_init(rb, watermark, flags);
758 for (i--; i >= 0; i--)
759 free_page((unsigned long)rb->data_pages[i]);
761 free_page((unsigned long)rb->user_page);
770 static void perf_mmap_free_page(unsigned long addr)
772 struct page *page = virt_to_page((void *)addr);
774 page->mapping = NULL;
778 void rb_free(struct ring_buffer *rb)
782 perf_mmap_free_page((unsigned long)rb->user_page);
783 for (i = 0; i < rb->nr_pages; i++)
784 perf_mmap_free_page((unsigned long)rb->data_pages[i]);
789 static int data_page_nr(struct ring_buffer *rb)
791 return rb->nr_pages << page_order(rb);
795 __perf_mmap_to_page(struct ring_buffer *rb, unsigned long pgoff)
797 /* The '>' counts in the user page. */
798 if (pgoff > data_page_nr(rb))
801 return vmalloc_to_page((void *)rb->user_page + pgoff * PAGE_SIZE);
804 static void perf_mmap_unmark_page(void *addr)
806 struct page *page = vmalloc_to_page(addr);
808 page->mapping = NULL;
811 static void rb_free_work(struct work_struct *work)
813 struct ring_buffer *rb;
817 rb = container_of(work, struct ring_buffer, work);
818 nr = data_page_nr(rb);
820 base = rb->user_page;
821 /* The '<=' counts in the user page. */
822 for (i = 0; i <= nr; i++)
823 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
829 void rb_free(struct ring_buffer *rb)
831 schedule_work(&rb->work);
834 struct ring_buffer *rb_alloc(int nr_pages, long watermark, int cpu, int flags)
836 struct ring_buffer *rb;
840 size = sizeof(struct ring_buffer);
841 size += sizeof(void *);
843 rb = kzalloc(size, GFP_KERNEL);
847 INIT_WORK(&rb->work, rb_free_work);
849 all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
853 rb->user_page = all_buf;
854 rb->data_pages[0] = all_buf + PAGE_SIZE;
857 rb->page_order = ilog2(nr_pages);
860 ring_buffer_init(rb, watermark, flags);
874 perf_mmap_to_page(struct ring_buffer *rb, unsigned long pgoff)
876 if (rb->aux_nr_pages) {
877 /* above AUX space */
878 if (pgoff > rb->aux_pgoff + rb->aux_nr_pages)
882 if (pgoff >= rb->aux_pgoff) {
883 int aux_pgoff = array_index_nospec(pgoff - rb->aux_pgoff, rb->aux_nr_pages);
884 return virt_to_page(rb->aux_pages[aux_pgoff]);
888 return __perf_mmap_to_page(rb, pgoff);