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 * if (LOAD ->data_tail) { LOAD ->data_head
66 * STORE $data LOAD $data
67 * smp_wmb() (B) smp_mb() (D)
68 * STORE ->data_head STORE ->data_tail
71 * Where A pairs with D, and B pairs with C.
73 * In our case (A) is a control dependency that separates the load of
74 * the ->data_tail and the stores of $data. In case ->data_tail
75 * indicates there is no room in the buffer to store $data we do not.
77 * D needs to be a full barrier since it separates the data READ
78 * from the tail WRITE.
80 * For B a WMB is sufficient since it separates two WRITEs, and for C
81 * an RMB is sufficient since it separates two READs.
83 * See perf_output_begin().
85 smp_wmb(); /* B, matches C */
86 rb->user_page->data_head = head;
89 * Now check if we missed an update -- rely on previous implied
90 * compiler barriers to force a re-read.
92 if (unlikely(head != local_read(&rb->head))) {
97 if (handle->wakeup != local_read(&rb->wakeup))
98 perf_output_wakeup(handle);
104 int perf_output_begin(struct perf_output_handle *handle,
105 struct perf_event *event, unsigned int size)
107 struct ring_buffer *rb;
108 unsigned long tail, offset, head;
109 int have_lost, page_shift;
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);
127 if (unlikely(!rb->nr_pages))
131 handle->event = event;
133 have_lost = local_read(&rb->lost);
134 if (unlikely(have_lost)) {
135 size += sizeof(lost_event);
136 if (event->attr.sample_id_all)
137 size += event->id_header_size;
140 perf_output_get_handle(handle);
143 tail = ACCESS_ONCE(rb->user_page->data_tail);
144 offset = head = local_read(&rb->head);
145 if (!rb->overwrite &&
146 unlikely(CIRC_SPACE(head, tail, perf_data_size(rb)) < size))
150 * The above forms a control dependency barrier separating the
151 * @tail load above from the data stores below. Since the @tail
152 * load is required to compute the branch to fail below.
154 * A, matches D; the full memory barrier userspace SHOULD issue
155 * after reading the data and before storing the new tail
158 * See perf_output_put_handle().
162 } while (local_cmpxchg(&rb->head, offset, head) != offset);
165 * We rely on the implied barrier() by local_cmpxchg() to ensure
166 * none of the data stores below can be lifted up by the compiler.
169 if (unlikely(head - local_read(&rb->wakeup) > rb->watermark))
170 local_add(rb->watermark, &rb->wakeup);
172 page_shift = PAGE_SHIFT + page_order(rb);
174 handle->page = (offset >> page_shift) & (rb->nr_pages - 1);
175 offset &= (1UL << page_shift) - 1;
176 handle->addr = rb->data_pages[handle->page] + offset;
177 handle->size = (1UL << page_shift) - offset;
179 if (unlikely(have_lost)) {
180 struct perf_sample_data sample_data;
182 lost_event.header.size = sizeof(lost_event);
183 lost_event.header.type = PERF_RECORD_LOST;
184 lost_event.header.misc = 0;
185 lost_event.id = event->id;
186 lost_event.lost = local_xchg(&rb->lost, 0);
188 perf_event_header__init_id(&lost_event.header,
189 &sample_data, event);
190 perf_output_put(handle, lost_event);
191 perf_event__output_id_sample(event, handle, &sample_data);
197 local_inc(&rb->lost);
198 perf_output_put_handle(handle);
205 unsigned int perf_output_copy(struct perf_output_handle *handle,
206 const void *buf, unsigned int len)
208 return __output_copy(handle, buf, len);
211 unsigned int perf_output_skip(struct perf_output_handle *handle,
214 return __output_skip(handle, NULL, len);
217 void perf_output_end(struct perf_output_handle *handle)
219 perf_output_put_handle(handle);
224 ring_buffer_init(struct ring_buffer *rb, long watermark, int flags)
226 long max_size = perf_data_size(rb);
229 rb->watermark = min(max_size, watermark);
232 rb->watermark = max_size / 2;
234 if (flags & RING_BUFFER_WRITABLE)
239 atomic_set(&rb->refcount, 1);
241 INIT_LIST_HEAD(&rb->event_list);
242 spin_lock_init(&rb->event_lock);
245 #ifndef CONFIG_PERF_USE_VMALLOC
248 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
252 perf_mmap_to_page(struct ring_buffer *rb, unsigned long pgoff)
254 if (pgoff > rb->nr_pages)
258 return virt_to_page(rb->user_page);
260 return virt_to_page(rb->data_pages[pgoff - 1]);
263 static void *perf_mmap_alloc_page(int cpu)
268 node = (cpu == -1) ? cpu : cpu_to_node(cpu);
269 page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
273 return page_address(page);
276 struct ring_buffer *rb_alloc(int nr_pages, long watermark, int cpu, int flags)
278 struct ring_buffer *rb;
282 size = sizeof(struct ring_buffer);
283 size += nr_pages * sizeof(void *);
285 rb = kzalloc(size, GFP_KERNEL);
289 rb->user_page = perf_mmap_alloc_page(cpu);
293 for (i = 0; i < nr_pages; i++) {
294 rb->data_pages[i] = perf_mmap_alloc_page(cpu);
295 if (!rb->data_pages[i])
296 goto fail_data_pages;
299 rb->nr_pages = nr_pages;
301 ring_buffer_init(rb, watermark, flags);
306 for (i--; i >= 0; i--)
307 free_page((unsigned long)rb->data_pages[i]);
309 free_page((unsigned long)rb->user_page);
318 static void perf_mmap_free_page(unsigned long addr)
320 struct page *page = virt_to_page((void *)addr);
322 page->mapping = NULL;
326 void rb_free(struct ring_buffer *rb)
330 perf_mmap_free_page((unsigned long)rb->user_page);
331 for (i = 0; i < rb->nr_pages; i++)
332 perf_mmap_free_page((unsigned long)rb->data_pages[i]);
337 static int data_page_nr(struct ring_buffer *rb)
339 return rb->nr_pages << page_order(rb);
343 perf_mmap_to_page(struct ring_buffer *rb, unsigned long pgoff)
345 /* The '>' counts in the user page. */
346 if (pgoff > data_page_nr(rb))
349 return vmalloc_to_page((void *)rb->user_page + pgoff * PAGE_SIZE);
352 static void perf_mmap_unmark_page(void *addr)
354 struct page *page = vmalloc_to_page(addr);
356 page->mapping = NULL;
359 static void rb_free_work(struct work_struct *work)
361 struct ring_buffer *rb;
365 rb = container_of(work, struct ring_buffer, work);
366 nr = data_page_nr(rb);
368 base = rb->user_page;
369 /* The '<=' counts in the user page. */
370 for (i = 0; i <= nr; i++)
371 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
377 void rb_free(struct ring_buffer *rb)
379 schedule_work(&rb->work);
382 struct ring_buffer *rb_alloc(int nr_pages, long watermark, int cpu, int flags)
384 struct ring_buffer *rb;
388 size = sizeof(struct ring_buffer);
389 size += sizeof(void *);
391 rb = kzalloc(size, GFP_KERNEL);
395 INIT_WORK(&rb->work, rb_free_work);
397 all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
401 rb->user_page = all_buf;
402 rb->data_pages[0] = all_buf + PAGE_SIZE;
403 rb->page_order = ilog2(nr_pages);
404 rb->nr_pages = !!nr_pages;
406 ring_buffer_init(rb, watermark, flags);