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>
15 #include <linux/circ_buf.h>
19 static void perf_output_wakeup(struct perf_output_handle *handle)
21 atomic_set(&handle->rb->poll, POLL_IN);
23 handle->event->pending_wakeup = 1;
24 irq_work_queue(&handle->event->pending);
28 * We need to ensure a later event_id doesn't publish a head when a former
29 * event isn't done writing. However since we need to deal with NMIs we
30 * cannot fully serialize things.
32 * We only publish the head (and generate a wakeup) when the outer-most
35 static void perf_output_get_handle(struct perf_output_handle *handle)
37 struct ring_buffer *rb = handle->rb;
41 handle->wakeup = local_read(&rb->wakeup);
44 static void perf_output_put_handle(struct perf_output_handle *handle)
46 struct ring_buffer *rb = handle->rb;
50 head = local_read(&rb->head);
53 * IRQ/NMI can happen here, which means we can miss a head update.
56 if (!local_dec_and_test(&rb->nest))
60 * Since the mmap() consumer (userspace) can run on a different CPU:
64 * READ ->data_tail READ ->data_head
65 * smp_mb() (A) smp_rmb() (C)
66 * WRITE $data READ $data
67 * smp_wmb() (B) smp_mb() (D)
68 * STORE ->data_head WRITE ->data_tail
70 * Where A pairs with D, and B pairs with C.
72 * I don't think A needs to be a full barrier because we won't in fact
73 * write data until we see the store from userspace. So we simply don't
74 * issue the data WRITE until we observe it. Be conservative for now.
76 * OTOH, D needs to be a full barrier since it separates the data READ
77 * from the tail WRITE.
79 * For B a WMB is sufficient since it separates two WRITEs, and for C
80 * an RMB is sufficient since it separates two READs.
82 * See perf_output_begin().
85 rb->user_page->data_head = head;
88 * Now check if we missed an update, rely on the (compiler)
89 * barrier in atomic_dec_and_test() to re-read rb->head.
91 if (unlikely(head != local_read(&rb->head))) {
96 if (handle->wakeup != local_read(&rb->wakeup))
97 perf_output_wakeup(handle);
103 int perf_output_begin(struct perf_output_handle *handle,
104 struct perf_event *event, unsigned int size)
106 struct ring_buffer *rb;
107 unsigned long tail, offset, head;
109 struct perf_sample_data sample_data;
111 struct perf_event_header header;
118 * For inherited events we send all the output towards the parent.
121 event = event->parent;
123 rb = rcu_dereference(event->rb);
128 handle->event = event;
133 have_lost = local_read(&rb->lost);
135 lost_event.header.size = sizeof(lost_event);
136 perf_event_header__init_id(&lost_event.header, &sample_data,
138 size += lost_event.header.size;
141 perf_output_get_handle(handle);
145 * Userspace could choose to issue a mb() before updating the
146 * tail pointer. So that all reads will be completed before the
149 * See perf_output_put_handle().
151 tail = ACCESS_ONCE(rb->user_page->data_tail);
153 offset = head = local_read(&rb->head);
154 if (!rb->overwrite &&
155 unlikely(CIRC_SPACE(head, tail, perf_data_size(rb)) < size))
158 } while (local_cmpxchg(&rb->head, offset, head) != offset);
160 if (head - local_read(&rb->wakeup) > rb->watermark)
161 local_add(rb->watermark, &rb->wakeup);
163 handle->page = offset >> (PAGE_SHIFT + page_order(rb));
164 handle->page &= rb->nr_pages - 1;
165 handle->size = offset & ((PAGE_SIZE << page_order(rb)) - 1);
166 handle->addr = rb->data_pages[handle->page];
167 handle->addr += handle->size;
168 handle->size = (PAGE_SIZE << page_order(rb)) - handle->size;
171 lost_event.header.type = PERF_RECORD_LOST;
172 lost_event.header.misc = 0;
173 lost_event.id = event->id;
174 lost_event.lost = local_xchg(&rb->lost, 0);
176 perf_output_put(handle, lost_event);
177 perf_event__output_id_sample(event, handle, &sample_data);
183 local_inc(&rb->lost);
184 perf_output_put_handle(handle);
191 unsigned int perf_output_copy(struct perf_output_handle *handle,
192 const void *buf, unsigned int len)
194 return __output_copy(handle, buf, len);
197 unsigned int perf_output_skip(struct perf_output_handle *handle,
200 return __output_skip(handle, NULL, len);
203 void perf_output_end(struct perf_output_handle *handle)
205 perf_output_put_handle(handle);
210 ring_buffer_init(struct ring_buffer *rb, long watermark, int flags)
212 long max_size = perf_data_size(rb);
215 rb->watermark = min(max_size, watermark);
218 rb->watermark = max_size / 2;
220 if (flags & RING_BUFFER_WRITABLE)
225 atomic_set(&rb->refcount, 1);
227 INIT_LIST_HEAD(&rb->event_list);
228 spin_lock_init(&rb->event_lock);
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]);
323 static int data_page_nr(struct ring_buffer *rb)
325 return rb->nr_pages << page_order(rb);
329 perf_mmap_to_page(struct ring_buffer *rb, unsigned long pgoff)
331 /* The '>' counts in the user page. */
332 if (pgoff > data_page_nr(rb))
335 return vmalloc_to_page((void *)rb->user_page + pgoff * PAGE_SIZE);
338 static void perf_mmap_unmark_page(void *addr)
340 struct page *page = vmalloc_to_page(addr);
342 page->mapping = NULL;
345 static void rb_free_work(struct work_struct *work)
347 struct ring_buffer *rb;
351 rb = container_of(work, struct ring_buffer, work);
352 nr = data_page_nr(rb);
354 base = rb->user_page;
355 /* The '<=' counts in the user page. */
356 for (i = 0; i <= nr; i++)
357 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
363 void rb_free(struct ring_buffer *rb)
365 schedule_work(&rb->work);
368 struct ring_buffer *rb_alloc(int nr_pages, long watermark, int cpu, int flags)
370 struct ring_buffer *rb;
374 size = sizeof(struct ring_buffer);
375 size += sizeof(void *);
377 rb = kzalloc(size, GFP_KERNEL);
381 INIT_WORK(&rb->work, rb_free_work);
383 all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
387 rb->user_page = all_buf;
388 rb->data_pages[0] = all_buf + PAGE_SIZE;
389 rb->page_order = ilog2(nr_pages);
390 rb->nr_pages = !!nr_pages;
392 ring_buffer_init(rb, watermark, flags);