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 <pzijlstr@redhat.com>
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>
18 static bool perf_output_space(struct ring_buffer *rb, unsigned long tail,
19 unsigned long offset, unsigned long head)
26 mask = perf_data_size(rb) - 1;
28 offset = (offset - tail) & mask;
29 head = (head - tail) & mask;
31 if ((int)(head - offset) < 0)
37 static void perf_output_wakeup(struct perf_output_handle *handle)
39 atomic_set(&handle->rb->poll, POLL_IN);
41 handle->event->pending_wakeup = 1;
42 irq_work_queue(&handle->event->pending);
46 * We need to ensure a later event_id doesn't publish a head when a former
47 * event isn't done writing. However since we need to deal with NMIs we
48 * cannot fully serialize things.
50 * We only publish the head (and generate a wakeup) when the outer-most
53 static void perf_output_get_handle(struct perf_output_handle *handle)
55 struct ring_buffer *rb = handle->rb;
59 handle->wakeup = local_read(&rb->wakeup);
62 static void perf_output_put_handle(struct perf_output_handle *handle)
64 struct ring_buffer *rb = handle->rb;
68 head = local_read(&rb->head);
71 * IRQ/NMI can happen here, which means we can miss a head update.
74 if (!local_dec_and_test(&rb->nest))
78 * Publish the known good head. Rely on the full barrier implied
79 * by atomic_dec_and_test() order the rb->head read and this
82 rb->user_page->data_head = head;
85 * Now check if we missed an update, rely on the (compiler)
86 * barrier in atomic_dec_and_test() to re-read rb->head.
88 if (unlikely(head != local_read(&rb->head))) {
93 if (handle->wakeup != local_read(&rb->wakeup))
94 perf_output_wakeup(handle);
100 int perf_output_begin(struct perf_output_handle *handle,
101 struct perf_event *event, unsigned int size,
104 struct ring_buffer *rb;
105 unsigned long tail, offset, head;
107 struct perf_sample_data sample_data;
109 struct perf_event_header header;
116 * For inherited events we send all the output towards the parent.
119 event = event->parent;
121 rb = rcu_dereference(event->rb);
126 handle->event = event;
127 handle->sample = sample;
132 have_lost = local_read(&rb->lost);
134 lost_event.header.size = sizeof(lost_event);
135 perf_event_header__init_id(&lost_event.header, &sample_data,
137 size += lost_event.header.size;
140 perf_output_get_handle(handle);
144 * Userspace could choose to issue a mb() before updating the
145 * tail pointer. So that all reads will be completed before the
148 tail = ACCESS_ONCE(rb->user_page->data_tail);
150 offset = head = local_read(&rb->head);
152 if (unlikely(!perf_output_space(rb, tail, offset, head)))
154 } while (local_cmpxchg(&rb->head, offset, head) != offset);
156 if (head - local_read(&rb->wakeup) > rb->watermark)
157 local_add(rb->watermark, &rb->wakeup);
159 handle->page = offset >> (PAGE_SHIFT + page_order(rb));
160 handle->page &= rb->nr_pages - 1;
161 handle->size = offset & ((PAGE_SIZE << page_order(rb)) - 1);
162 handle->addr = rb->data_pages[handle->page];
163 handle->addr += handle->size;
164 handle->size = (PAGE_SIZE << page_order(rb)) - handle->size;
167 lost_event.header.type = PERF_RECORD_LOST;
168 lost_event.header.misc = 0;
169 lost_event.id = event->id;
170 lost_event.lost = local_xchg(&rb->lost, 0);
172 perf_output_put(handle, lost_event);
173 perf_event__output_id_sample(event, handle, &sample_data);
179 local_inc(&rb->lost);
180 perf_output_put_handle(handle);
187 void perf_output_copy(struct perf_output_handle *handle,
188 const void *buf, unsigned int len)
190 __output_copy(handle, buf, len);
193 void perf_output_end(struct perf_output_handle *handle)
195 struct perf_event *event = handle->event;
196 struct ring_buffer *rb = handle->rb;
198 if (handle->sample && !event->attr.watermark) {
199 int wakeup_events = event->attr.wakeup_events;
202 int events = local_inc_return(&rb->events);
203 if (events >= wakeup_events) {
204 local_sub(wakeup_events, &rb->events);
205 local_inc(&rb->wakeup);
210 perf_output_put_handle(handle);
215 ring_buffer_init(struct ring_buffer *rb, long watermark, int flags)
217 long max_size = perf_data_size(rb);
220 rb->watermark = min(max_size, watermark);
223 rb->watermark = max_size / 2;
225 if (flags & RING_BUFFER_WRITABLE)
228 atomic_set(&rb->refcount, 1);
231 #ifndef CONFIG_PERF_USE_VMALLOC
234 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
238 perf_mmap_to_page(struct ring_buffer *rb, unsigned long pgoff)
240 if (pgoff > rb->nr_pages)
244 return virt_to_page(rb->user_page);
246 return virt_to_page(rb->data_pages[pgoff - 1]);
249 static void *perf_mmap_alloc_page(int cpu)
254 node = (cpu == -1) ? cpu : cpu_to_node(cpu);
255 page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
259 return page_address(page);
262 struct ring_buffer *rb_alloc(int nr_pages, long watermark, int cpu, int flags)
264 struct ring_buffer *rb;
268 size = sizeof(struct ring_buffer);
269 size += nr_pages * sizeof(void *);
271 rb = kzalloc(size, GFP_KERNEL);
275 rb->user_page = perf_mmap_alloc_page(cpu);
279 for (i = 0; i < nr_pages; i++) {
280 rb->data_pages[i] = perf_mmap_alloc_page(cpu);
281 if (!rb->data_pages[i])
282 goto fail_data_pages;
285 rb->nr_pages = nr_pages;
287 ring_buffer_init(rb, watermark, flags);
292 for (i--; i >= 0; i--)
293 free_page((unsigned long)rb->data_pages[i]);
295 free_page((unsigned long)rb->user_page);
304 static void perf_mmap_free_page(unsigned long addr)
306 struct page *page = virt_to_page((void *)addr);
308 page->mapping = NULL;
312 void rb_free(struct ring_buffer *rb)
316 perf_mmap_free_page((unsigned long)rb->user_page);
317 for (i = 0; i < rb->nr_pages; i++)
318 perf_mmap_free_page((unsigned long)rb->data_pages[i]);
325 perf_mmap_to_page(struct ring_buffer *rb, unsigned long pgoff)
327 if (pgoff > (1UL << page_order(rb)))
330 return vmalloc_to_page((void *)rb->user_page + pgoff * PAGE_SIZE);
333 static void perf_mmap_unmark_page(void *addr)
335 struct page *page = vmalloc_to_page(addr);
337 page->mapping = NULL;
340 static void rb_free_work(struct work_struct *work)
342 struct ring_buffer *rb;
346 rb = container_of(work, struct ring_buffer, work);
347 nr = 1 << page_order(rb);
349 base = rb->user_page;
350 for (i = 0; i < nr + 1; i++)
351 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
357 void rb_free(struct ring_buffer *rb)
359 schedule_work(&rb->work);
362 struct ring_buffer *rb_alloc(int nr_pages, long watermark, int cpu, int flags)
364 struct ring_buffer *rb;
368 size = sizeof(struct ring_buffer);
369 size += sizeof(void *);
371 rb = kzalloc(size, GFP_KERNEL);
375 INIT_WORK(&rb->work, rb_free_work);
377 all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
381 rb->user_page = all_buf;
382 rb->data_pages[0] = all_buf + PAGE_SIZE;
383 rb->page_order = ilog2(nr_pages);
386 ring_buffer_init(rb, watermark, flags);