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)
21 unsigned long sz = perf_data_size(rb);
22 unsigned long mask = sz - 1;
25 * check if user-writable
26 * overwrite : over-write its own tail
27 * !overwrite: buffer possibly drops events.
33 * verify that payload is not bigger than buffer
34 * otherwise masking logic may fail to detect
35 * the "not enough space" condition
37 if ((head - offset) > sz)
40 offset = (offset - tail) & mask;
41 head = (head - tail) & mask;
43 if ((int)(head - offset) < 0)
49 static void perf_output_wakeup(struct perf_output_handle *handle)
51 atomic_set(&handle->rb->poll, POLL_IN);
53 handle->event->pending_wakeup = 1;
54 irq_work_queue(&handle->event->pending);
58 * We need to ensure a later event_id doesn't publish a head when a former
59 * event isn't done writing. However since we need to deal with NMIs we
60 * cannot fully serialize things.
62 * We only publish the head (and generate a wakeup) when the outer-most
65 static void perf_output_get_handle(struct perf_output_handle *handle)
67 struct ring_buffer *rb = handle->rb;
71 handle->wakeup = local_read(&rb->wakeup);
74 static void perf_output_put_handle(struct perf_output_handle *handle)
76 struct ring_buffer *rb = handle->rb;
80 head = local_read(&rb->head);
83 * IRQ/NMI can happen here, which means we can miss a head update.
86 if (!local_dec_and_test(&rb->nest))
90 * Since the mmap() consumer (userspace) can run on a different CPU:
94 * READ ->data_tail READ ->data_head
95 * smp_mb() (A) smp_rmb() (C)
96 * WRITE $data READ $data
97 * smp_wmb() (B) smp_mb() (D)
98 * STORE ->data_head WRITE ->data_tail
100 * Where A pairs with D, and B pairs with C.
102 * I don't think A needs to be a full barrier because we won't in fact
103 * write data until we see the store from userspace. So we simply don't
104 * issue the data WRITE until we observe it. Be conservative for now.
106 * OTOH, D needs to be a full barrier since it separates the data READ
107 * from the tail WRITE.
109 * For B a WMB is sufficient since it separates two WRITEs, and for C
110 * an RMB is sufficient since it separates two READs.
112 * See perf_output_begin().
115 rb->user_page->data_head = head;
118 * Now check if we missed an update, rely on the (compiler)
119 * barrier in atomic_dec_and_test() to re-read rb->head.
121 if (unlikely(head != local_read(&rb->head))) {
122 local_inc(&rb->nest);
126 if (handle->wakeup != local_read(&rb->wakeup))
127 perf_output_wakeup(handle);
133 int perf_output_begin(struct perf_output_handle *handle,
134 struct perf_event *event, unsigned int size)
136 struct ring_buffer *rb;
137 unsigned long tail, offset, head;
139 struct perf_sample_data sample_data;
141 struct perf_event_header header;
148 * For inherited events we send all the output towards the parent.
151 event = event->parent;
153 rb = rcu_dereference(event->rb);
158 handle->event = event;
163 have_lost = local_read(&rb->lost);
165 lost_event.header.size = sizeof(lost_event);
166 perf_event_header__init_id(&lost_event.header, &sample_data,
168 size += lost_event.header.size;
171 perf_output_get_handle(handle);
175 * Userspace could choose to issue a mb() before updating the
176 * tail pointer. So that all reads will be completed before the
179 * See perf_output_put_handle().
181 tail = ACCESS_ONCE(rb->user_page->data_tail);
183 offset = head = local_read(&rb->head);
185 if (unlikely(!perf_output_space(rb, tail, offset, head)))
187 } while (local_cmpxchg(&rb->head, offset, head) != offset);
189 if (head - local_read(&rb->wakeup) > rb->watermark)
190 local_add(rb->watermark, &rb->wakeup);
192 handle->page = offset >> (PAGE_SHIFT + page_order(rb));
193 handle->page &= rb->nr_pages - 1;
194 handle->size = offset & ((PAGE_SIZE << page_order(rb)) - 1);
195 handle->addr = rb->data_pages[handle->page];
196 handle->addr += handle->size;
197 handle->size = (PAGE_SIZE << page_order(rb)) - handle->size;
200 lost_event.header.type = PERF_RECORD_LOST;
201 lost_event.header.misc = 0;
202 lost_event.id = event->id;
203 lost_event.lost = local_xchg(&rb->lost, 0);
205 perf_output_put(handle, lost_event);
206 perf_event__output_id_sample(event, handle, &sample_data);
212 local_inc(&rb->lost);
213 perf_output_put_handle(handle);
220 unsigned int perf_output_copy(struct perf_output_handle *handle,
221 const void *buf, unsigned int len)
223 return __output_copy(handle, buf, len);
226 unsigned int perf_output_skip(struct perf_output_handle *handle,
229 return __output_skip(handle, NULL, len);
232 void perf_output_end(struct perf_output_handle *handle)
234 perf_output_put_handle(handle);
239 ring_buffer_init(struct ring_buffer *rb, long watermark, int flags)
241 long max_size = perf_data_size(rb);
244 rb->watermark = min(max_size, watermark);
247 rb->watermark = max_size / 2;
249 if (flags & RING_BUFFER_WRITABLE)
254 atomic_set(&rb->refcount, 1);
256 INIT_LIST_HEAD(&rb->event_list);
257 spin_lock_init(&rb->event_lock);
260 #ifndef CONFIG_PERF_USE_VMALLOC
263 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
267 perf_mmap_to_page(struct ring_buffer *rb, unsigned long pgoff)
269 if (pgoff > rb->nr_pages)
273 return virt_to_page(rb->user_page);
275 return virt_to_page(rb->data_pages[pgoff - 1]);
278 static void *perf_mmap_alloc_page(int cpu)
283 node = (cpu == -1) ? cpu : cpu_to_node(cpu);
284 page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
288 return page_address(page);
291 struct ring_buffer *rb_alloc(int nr_pages, long watermark, int cpu, int flags)
293 struct ring_buffer *rb;
297 size = sizeof(struct ring_buffer);
298 size += nr_pages * sizeof(void *);
300 rb = kzalloc(size, GFP_KERNEL);
304 rb->user_page = perf_mmap_alloc_page(cpu);
308 for (i = 0; i < nr_pages; i++) {
309 rb->data_pages[i] = perf_mmap_alloc_page(cpu);
310 if (!rb->data_pages[i])
311 goto fail_data_pages;
314 rb->nr_pages = nr_pages;
316 ring_buffer_init(rb, watermark, flags);
321 for (i--; i >= 0; i--)
322 free_page((unsigned long)rb->data_pages[i]);
324 free_page((unsigned long)rb->user_page);
333 static void perf_mmap_free_page(unsigned long addr)
335 struct page *page = virt_to_page((void *)addr);
337 page->mapping = NULL;
341 void rb_free(struct ring_buffer *rb)
345 perf_mmap_free_page((unsigned long)rb->user_page);
346 for (i = 0; i < rb->nr_pages; i++)
347 perf_mmap_free_page((unsigned long)rb->data_pages[i]);
352 static int data_page_nr(struct ring_buffer *rb)
354 return rb->nr_pages << page_order(rb);
358 perf_mmap_to_page(struct ring_buffer *rb, unsigned long pgoff)
360 /* The '>' counts in the user page. */
361 if (pgoff > data_page_nr(rb))
364 return vmalloc_to_page((void *)rb->user_page + pgoff * PAGE_SIZE);
367 static void perf_mmap_unmark_page(void *addr)
369 struct page *page = vmalloc_to_page(addr);
371 page->mapping = NULL;
374 static void rb_free_work(struct work_struct *work)
376 struct ring_buffer *rb;
380 rb = container_of(work, struct ring_buffer, work);
381 nr = data_page_nr(rb);
383 base = rb->user_page;
384 /* The '<=' counts in the user page. */
385 for (i = 0; i <= nr; i++)
386 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
392 void rb_free(struct ring_buffer *rb)
394 schedule_work(&rb->work);
397 struct ring_buffer *rb_alloc(int nr_pages, long watermark, int cpu, int flags)
399 struct ring_buffer *rb;
403 size = sizeof(struct ring_buffer);
404 size += sizeof(void *);
406 rb = kzalloc(size, GFP_KERNEL);
410 INIT_WORK(&rb->work, rb_free_work);
412 all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
416 rb->user_page = all_buf;
417 rb->data_pages[0] = all_buf + PAGE_SIZE;
418 rb->page_order = ilog2(nr_pages);
419 rb->nr_pages = !!nr_pages;
421 ring_buffer_init(rb, watermark, flags);