Loading...
1/*
2 * Performance events ring-buffer code:
3 *
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>
8 *
9 * For licensing details see kernel-base/COPYING
10 */
11
12#include <linux/perf_event.h>
13#include <linux/vmalloc.h>
14#include <linux/slab.h>
15#include <linux/circ_buf.h>
16
17#include "internal.h"
18
19static void perf_output_wakeup(struct perf_output_handle *handle)
20{
21 atomic_set(&handle->rb->poll, POLL_IN);
22
23 handle->event->pending_wakeup = 1;
24 irq_work_queue(&handle->event->pending);
25}
26
27/*
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.
31 *
32 * We only publish the head (and generate a wakeup) when the outer-most
33 * event completes.
34 */
35static void perf_output_get_handle(struct perf_output_handle *handle)
36{
37 struct ring_buffer *rb = handle->rb;
38
39 preempt_disable();
40 local_inc(&rb->nest);
41 handle->wakeup = local_read(&rb->wakeup);
42}
43
44static void perf_output_put_handle(struct perf_output_handle *handle)
45{
46 struct ring_buffer *rb = handle->rb;
47 unsigned long head;
48
49again:
50 head = local_read(&rb->head);
51
52 /*
53 * IRQ/NMI can happen here, which means we can miss a head update.
54 */
55
56 if (!local_dec_and_test(&rb->nest))
57 goto out;
58
59 /*
60 * Since the mmap() consumer (userspace) can run on a different CPU:
61 *
62 * kernel user
63 *
64 * if (LOAD ->data_tail) { LOAD ->data_head
65 * (A) smp_rmb() (C)
66 * STORE $data LOAD $data
67 * smp_wmb() (B) smp_mb() (D)
68 * STORE ->data_head STORE ->data_tail
69 * }
70 *
71 * Where A pairs with D, and B pairs with C.
72 *
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.
76 *
77 * D needs to be a full barrier since it separates the data READ
78 * from the tail WRITE.
79 *
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.
82 *
83 * See perf_output_begin().
84 */
85 smp_wmb(); /* B, matches C */
86 rb->user_page->data_head = head;
87
88 /*
89 * Now check if we missed an update -- rely on previous implied
90 * compiler barriers to force a re-read.
91 */
92 if (unlikely(head != local_read(&rb->head))) {
93 local_inc(&rb->nest);
94 goto again;
95 }
96
97 if (handle->wakeup != local_read(&rb->wakeup))
98 perf_output_wakeup(handle);
99
100out:
101 preempt_enable();
102}
103
104int perf_output_begin(struct perf_output_handle *handle,
105 struct perf_event *event, unsigned int size)
106{
107 struct ring_buffer *rb;
108 unsigned long tail, offset, head;
109 int have_lost, page_shift;
110 struct {
111 struct perf_event_header header;
112 u64 id;
113 u64 lost;
114 } lost_event;
115
116 rcu_read_lock();
117 /*
118 * For inherited events we send all the output towards the parent.
119 */
120 if (event->parent)
121 event = event->parent;
122
123 rb = rcu_dereference(event->rb);
124 if (unlikely(!rb))
125 goto out;
126
127 if (unlikely(!rb->nr_pages))
128 goto out;
129
130 handle->rb = rb;
131 handle->event = event;
132
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;
138 }
139
140 perf_output_get_handle(handle);
141
142 do {
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))
147 goto fail;
148
149 /*
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.
153 *
154 * A, matches D; the full memory barrier userspace SHOULD issue
155 * after reading the data and before storing the new tail
156 * position.
157 *
158 * See perf_output_put_handle().
159 */
160
161 head += size;
162 } while (local_cmpxchg(&rb->head, offset, head) != offset);
163
164 /*
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.
167 */
168
169 if (unlikely(head - local_read(&rb->wakeup) > rb->watermark))
170 local_add(rb->watermark, &rb->wakeup);
171
172 page_shift = PAGE_SHIFT + page_order(rb);
173
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;
178
179 if (unlikely(have_lost)) {
180 struct perf_sample_data sample_data;
181
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);
187
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);
192 }
193
194 return 0;
195
196fail:
197 local_inc(&rb->lost);
198 perf_output_put_handle(handle);
199out:
200 rcu_read_unlock();
201
202 return -ENOSPC;
203}
204
205unsigned int perf_output_copy(struct perf_output_handle *handle,
206 const void *buf, unsigned int len)
207{
208 return __output_copy(handle, buf, len);
209}
210
211unsigned int perf_output_skip(struct perf_output_handle *handle,
212 unsigned int len)
213{
214 return __output_skip(handle, NULL, len);
215}
216
217void perf_output_end(struct perf_output_handle *handle)
218{
219 perf_output_put_handle(handle);
220 rcu_read_unlock();
221}
222
223static void
224ring_buffer_init(struct ring_buffer *rb, long watermark, int flags)
225{
226 long max_size = perf_data_size(rb);
227
228 if (watermark)
229 rb->watermark = min(max_size, watermark);
230
231 if (!rb->watermark)
232 rb->watermark = max_size / 2;
233
234 if (flags & RING_BUFFER_WRITABLE)
235 rb->overwrite = 0;
236 else
237 rb->overwrite = 1;
238
239 atomic_set(&rb->refcount, 1);
240
241 INIT_LIST_HEAD(&rb->event_list);
242 spin_lock_init(&rb->event_lock);
243}
244
245#ifndef CONFIG_PERF_USE_VMALLOC
246
247/*
248 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
249 */
250
251struct page *
252perf_mmap_to_page(struct ring_buffer *rb, unsigned long pgoff)
253{
254 if (pgoff > rb->nr_pages)
255 return NULL;
256
257 if (pgoff == 0)
258 return virt_to_page(rb->user_page);
259
260 return virt_to_page(rb->data_pages[pgoff - 1]);
261}
262
263static void *perf_mmap_alloc_page(int cpu)
264{
265 struct page *page;
266 int node;
267
268 node = (cpu == -1) ? cpu : cpu_to_node(cpu);
269 page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
270 if (!page)
271 return NULL;
272
273 return page_address(page);
274}
275
276struct ring_buffer *rb_alloc(int nr_pages, long watermark, int cpu, int flags)
277{
278 struct ring_buffer *rb;
279 unsigned long size;
280 int i;
281
282 size = sizeof(struct ring_buffer);
283 size += nr_pages * sizeof(void *);
284
285 rb = kzalloc(size, GFP_KERNEL);
286 if (!rb)
287 goto fail;
288
289 rb->user_page = perf_mmap_alloc_page(cpu);
290 if (!rb->user_page)
291 goto fail_user_page;
292
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;
297 }
298
299 rb->nr_pages = nr_pages;
300
301 ring_buffer_init(rb, watermark, flags);
302
303 return rb;
304
305fail_data_pages:
306 for (i--; i >= 0; i--)
307 free_page((unsigned long)rb->data_pages[i]);
308
309 free_page((unsigned long)rb->user_page);
310
311fail_user_page:
312 kfree(rb);
313
314fail:
315 return NULL;
316}
317
318static void perf_mmap_free_page(unsigned long addr)
319{
320 struct page *page = virt_to_page((void *)addr);
321
322 page->mapping = NULL;
323 __free_page(page);
324}
325
326void rb_free(struct ring_buffer *rb)
327{
328 int i;
329
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]);
333 kfree(rb);
334}
335
336#else
337static int data_page_nr(struct ring_buffer *rb)
338{
339 return rb->nr_pages << page_order(rb);
340}
341
342struct page *
343perf_mmap_to_page(struct ring_buffer *rb, unsigned long pgoff)
344{
345 /* The '>' counts in the user page. */
346 if (pgoff > data_page_nr(rb))
347 return NULL;
348
349 return vmalloc_to_page((void *)rb->user_page + pgoff * PAGE_SIZE);
350}
351
352static void perf_mmap_unmark_page(void *addr)
353{
354 struct page *page = vmalloc_to_page(addr);
355
356 page->mapping = NULL;
357}
358
359static void rb_free_work(struct work_struct *work)
360{
361 struct ring_buffer *rb;
362 void *base;
363 int i, nr;
364
365 rb = container_of(work, struct ring_buffer, work);
366 nr = data_page_nr(rb);
367
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));
372
373 vfree(base);
374 kfree(rb);
375}
376
377void rb_free(struct ring_buffer *rb)
378{
379 schedule_work(&rb->work);
380}
381
382struct ring_buffer *rb_alloc(int nr_pages, long watermark, int cpu, int flags)
383{
384 struct ring_buffer *rb;
385 unsigned long size;
386 void *all_buf;
387
388 size = sizeof(struct ring_buffer);
389 size += sizeof(void *);
390
391 rb = kzalloc(size, GFP_KERNEL);
392 if (!rb)
393 goto fail;
394
395 INIT_WORK(&rb->work, rb_free_work);
396
397 all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
398 if (!all_buf)
399 goto fail_all_buf;
400
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;
405
406 ring_buffer_init(rb, watermark, flags);
407
408 return rb;
409
410fail_all_buf:
411 kfree(rb);
412
413fail:
414 return NULL;
415}
416
417#endif
1/*
2 * Performance events ring-buffer code:
3 *
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
7 * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
8 *
9 * For licensing details see kernel-base/COPYING
10 */
11
12#include <linux/perf_event.h>
13#include <linux/vmalloc.h>
14#include <linux/slab.h>
15#include <linux/circ_buf.h>
16#include <linux/poll.h>
17#include <linux/nospec.h>
18
19#include "internal.h"
20
21static void perf_output_wakeup(struct perf_output_handle *handle)
22{
23 atomic_set(&handle->rb->poll, EPOLLIN);
24
25 handle->event->pending_wakeup = 1;
26 irq_work_queue(&handle->event->pending);
27}
28
29/*
30 * We need to ensure a later event_id doesn't publish a head when a former
31 * event isn't done writing. However since we need to deal with NMIs we
32 * cannot fully serialize things.
33 *
34 * We only publish the head (and generate a wakeup) when the outer-most
35 * event completes.
36 */
37static void perf_output_get_handle(struct perf_output_handle *handle)
38{
39 struct ring_buffer *rb = handle->rb;
40
41 preempt_disable();
42 local_inc(&rb->nest);
43 handle->wakeup = local_read(&rb->wakeup);
44}
45
46static void perf_output_put_handle(struct perf_output_handle *handle)
47{
48 struct ring_buffer *rb = handle->rb;
49 unsigned long head;
50
51again:
52 head = local_read(&rb->head);
53
54 /*
55 * IRQ/NMI can happen here, which means we can miss a head update.
56 */
57
58 if (!local_dec_and_test(&rb->nest))
59 goto out;
60
61 /*
62 * Since the mmap() consumer (userspace) can run on a different CPU:
63 *
64 * kernel user
65 *
66 * if (LOAD ->data_tail) { LOAD ->data_head
67 * (A) smp_rmb() (C)
68 * STORE $data LOAD $data
69 * smp_wmb() (B) smp_mb() (D)
70 * STORE ->data_head STORE ->data_tail
71 * }
72 *
73 * Where A pairs with D, and B pairs with C.
74 *
75 * In our case (A) is a control dependency that separates the load of
76 * the ->data_tail and the stores of $data. In case ->data_tail
77 * indicates there is no room in the buffer to store $data we do not.
78 *
79 * D needs to be a full barrier since it separates the data READ
80 * from the tail WRITE.
81 *
82 * For B a WMB is sufficient since it separates two WRITEs, and for C
83 * an RMB is sufficient since it separates two READs.
84 *
85 * See perf_output_begin().
86 */
87 smp_wmb(); /* B, matches C */
88 rb->user_page->data_head = head;
89
90 /*
91 * Now check if we missed an update -- rely on previous implied
92 * compiler barriers to force a re-read.
93 */
94 if (unlikely(head != local_read(&rb->head))) {
95 local_inc(&rb->nest);
96 goto again;
97 }
98
99 if (handle->wakeup != local_read(&rb->wakeup))
100 perf_output_wakeup(handle);
101
102out:
103 preempt_enable();
104}
105
106static bool __always_inline
107ring_buffer_has_space(unsigned long head, unsigned long tail,
108 unsigned long data_size, unsigned int size,
109 bool backward)
110{
111 if (!backward)
112 return CIRC_SPACE(head, tail, data_size) >= size;
113 else
114 return CIRC_SPACE(tail, head, data_size) >= size;
115}
116
117static int __always_inline
118__perf_output_begin(struct perf_output_handle *handle,
119 struct perf_event *event, unsigned int size,
120 bool backward)
121{
122 struct ring_buffer *rb;
123 unsigned long tail, offset, head;
124 int have_lost, page_shift;
125 struct {
126 struct perf_event_header header;
127 u64 id;
128 u64 lost;
129 } lost_event;
130
131 rcu_read_lock();
132 /*
133 * For inherited events we send all the output towards the parent.
134 */
135 if (event->parent)
136 event = event->parent;
137
138 rb = rcu_dereference(event->rb);
139 if (unlikely(!rb))
140 goto out;
141
142 if (unlikely(rb->paused)) {
143 if (rb->nr_pages)
144 local_inc(&rb->lost);
145 goto out;
146 }
147
148 handle->rb = rb;
149 handle->event = event;
150
151 have_lost = local_read(&rb->lost);
152 if (unlikely(have_lost)) {
153 size += sizeof(lost_event);
154 if (event->attr.sample_id_all)
155 size += event->id_header_size;
156 }
157
158 perf_output_get_handle(handle);
159
160 do {
161 tail = READ_ONCE(rb->user_page->data_tail);
162 offset = head = local_read(&rb->head);
163 if (!rb->overwrite) {
164 if (unlikely(!ring_buffer_has_space(head, tail,
165 perf_data_size(rb),
166 size, backward)))
167 goto fail;
168 }
169
170 /*
171 * The above forms a control dependency barrier separating the
172 * @tail load above from the data stores below. Since the @tail
173 * load is required to compute the branch to fail below.
174 *
175 * A, matches D; the full memory barrier userspace SHOULD issue
176 * after reading the data and before storing the new tail
177 * position.
178 *
179 * See perf_output_put_handle().
180 */
181
182 if (!backward)
183 head += size;
184 else
185 head -= size;
186 } while (local_cmpxchg(&rb->head, offset, head) != offset);
187
188 if (backward) {
189 offset = head;
190 head = (u64)(-head);
191 }
192
193 /*
194 * We rely on the implied barrier() by local_cmpxchg() to ensure
195 * none of the data stores below can be lifted up by the compiler.
196 */
197
198 if (unlikely(head - local_read(&rb->wakeup) > rb->watermark))
199 local_add(rb->watermark, &rb->wakeup);
200
201 page_shift = PAGE_SHIFT + page_order(rb);
202
203 handle->page = (offset >> page_shift) & (rb->nr_pages - 1);
204 offset &= (1UL << page_shift) - 1;
205 handle->addr = rb->data_pages[handle->page] + offset;
206 handle->size = (1UL << page_shift) - offset;
207
208 if (unlikely(have_lost)) {
209 struct perf_sample_data sample_data;
210
211 lost_event.header.size = sizeof(lost_event);
212 lost_event.header.type = PERF_RECORD_LOST;
213 lost_event.header.misc = 0;
214 lost_event.id = event->id;
215 lost_event.lost = local_xchg(&rb->lost, 0);
216
217 perf_event_header__init_id(&lost_event.header,
218 &sample_data, event);
219 perf_output_put(handle, lost_event);
220 perf_event__output_id_sample(event, handle, &sample_data);
221 }
222
223 return 0;
224
225fail:
226 local_inc(&rb->lost);
227 perf_output_put_handle(handle);
228out:
229 rcu_read_unlock();
230
231 return -ENOSPC;
232}
233
234int perf_output_begin_forward(struct perf_output_handle *handle,
235 struct perf_event *event, unsigned int size)
236{
237 return __perf_output_begin(handle, event, size, false);
238}
239
240int perf_output_begin_backward(struct perf_output_handle *handle,
241 struct perf_event *event, unsigned int size)
242{
243 return __perf_output_begin(handle, event, size, true);
244}
245
246int perf_output_begin(struct perf_output_handle *handle,
247 struct perf_event *event, unsigned int size)
248{
249
250 return __perf_output_begin(handle, event, size,
251 unlikely(is_write_backward(event)));
252}
253
254unsigned int perf_output_copy(struct perf_output_handle *handle,
255 const void *buf, unsigned int len)
256{
257 return __output_copy(handle, buf, len);
258}
259
260unsigned int perf_output_skip(struct perf_output_handle *handle,
261 unsigned int len)
262{
263 return __output_skip(handle, NULL, len);
264}
265
266void perf_output_end(struct perf_output_handle *handle)
267{
268 perf_output_put_handle(handle);
269 rcu_read_unlock();
270}
271
272static void
273ring_buffer_init(struct ring_buffer *rb, long watermark, int flags)
274{
275 long max_size = perf_data_size(rb);
276
277 if (watermark)
278 rb->watermark = min(max_size, watermark);
279
280 if (!rb->watermark)
281 rb->watermark = max_size / 2;
282
283 if (flags & RING_BUFFER_WRITABLE)
284 rb->overwrite = 0;
285 else
286 rb->overwrite = 1;
287
288 atomic_set(&rb->refcount, 1);
289
290 INIT_LIST_HEAD(&rb->event_list);
291 spin_lock_init(&rb->event_lock);
292
293 /*
294 * perf_output_begin() only checks rb->paused, therefore
295 * rb->paused must be true if we have no pages for output.
296 */
297 if (!rb->nr_pages)
298 rb->paused = 1;
299}
300
301void perf_aux_output_flag(struct perf_output_handle *handle, u64 flags)
302{
303 /*
304 * OVERWRITE is determined by perf_aux_output_end() and can't
305 * be passed in directly.
306 */
307 if (WARN_ON_ONCE(flags & PERF_AUX_FLAG_OVERWRITE))
308 return;
309
310 handle->aux_flags |= flags;
311}
312EXPORT_SYMBOL_GPL(perf_aux_output_flag);
313
314/*
315 * This is called before hardware starts writing to the AUX area to
316 * obtain an output handle and make sure there's room in the buffer.
317 * When the capture completes, call perf_aux_output_end() to commit
318 * the recorded data to the buffer.
319 *
320 * The ordering is similar to that of perf_output_{begin,end}, with
321 * the exception of (B), which should be taken care of by the pmu
322 * driver, since ordering rules will differ depending on hardware.
323 *
324 * Call this from pmu::start(); see the comment in perf_aux_output_end()
325 * about its use in pmu callbacks. Both can also be called from the PMI
326 * handler if needed.
327 */
328void *perf_aux_output_begin(struct perf_output_handle *handle,
329 struct perf_event *event)
330{
331 struct perf_event *output_event = event;
332 unsigned long aux_head, aux_tail;
333 struct ring_buffer *rb;
334
335 if (output_event->parent)
336 output_event = output_event->parent;
337
338 /*
339 * Since this will typically be open across pmu::add/pmu::del, we
340 * grab ring_buffer's refcount instead of holding rcu read lock
341 * to make sure it doesn't disappear under us.
342 */
343 rb = ring_buffer_get(output_event);
344 if (!rb)
345 return NULL;
346
347 if (!rb_has_aux(rb))
348 goto err;
349
350 /*
351 * If aux_mmap_count is zero, the aux buffer is in perf_mmap_close(),
352 * about to get freed, so we leave immediately.
353 *
354 * Checking rb::aux_mmap_count and rb::refcount has to be done in
355 * the same order, see perf_mmap_close. Otherwise we end up freeing
356 * aux pages in this path, which is a bug, because in_atomic().
357 */
358 if (!atomic_read(&rb->aux_mmap_count))
359 goto err;
360
361 if (!atomic_inc_not_zero(&rb->aux_refcount))
362 goto err;
363
364 /*
365 * Nesting is not supported for AUX area, make sure nested
366 * writers are caught early
367 */
368 if (WARN_ON_ONCE(local_xchg(&rb->aux_nest, 1)))
369 goto err_put;
370
371 aux_head = rb->aux_head;
372
373 handle->rb = rb;
374 handle->event = event;
375 handle->head = aux_head;
376 handle->size = 0;
377 handle->aux_flags = 0;
378
379 /*
380 * In overwrite mode, AUX data stores do not depend on aux_tail,
381 * therefore (A) control dependency barrier does not exist. The
382 * (B) <-> (C) ordering is still observed by the pmu driver.
383 */
384 if (!rb->aux_overwrite) {
385 aux_tail = READ_ONCE(rb->user_page->aux_tail);
386 handle->wakeup = rb->aux_wakeup + rb->aux_watermark;
387 if (aux_head - aux_tail < perf_aux_size(rb))
388 handle->size = CIRC_SPACE(aux_head, aux_tail, perf_aux_size(rb));
389
390 /*
391 * handle->size computation depends on aux_tail load; this forms a
392 * control dependency barrier separating aux_tail load from aux data
393 * store that will be enabled on successful return
394 */
395 if (!handle->size) { /* A, matches D */
396 event->pending_disable = 1;
397 perf_output_wakeup(handle);
398 local_set(&rb->aux_nest, 0);
399 goto err_put;
400 }
401 }
402
403 return handle->rb->aux_priv;
404
405err_put:
406 /* can't be last */
407 rb_free_aux(rb);
408
409err:
410 ring_buffer_put(rb);
411 handle->event = NULL;
412
413 return NULL;
414}
415EXPORT_SYMBOL_GPL(perf_aux_output_begin);
416
417static bool __always_inline rb_need_aux_wakeup(struct ring_buffer *rb)
418{
419 if (rb->aux_overwrite)
420 return false;
421
422 if (rb->aux_head - rb->aux_wakeup >= rb->aux_watermark) {
423 rb->aux_wakeup = rounddown(rb->aux_head, rb->aux_watermark);
424 return true;
425 }
426
427 return false;
428}
429
430/*
431 * Commit the data written by hardware into the ring buffer by adjusting
432 * aux_head and posting a PERF_RECORD_AUX into the perf buffer. It is the
433 * pmu driver's responsibility to observe ordering rules of the hardware,
434 * so that all the data is externally visible before this is called.
435 *
436 * Note: this has to be called from pmu::stop() callback, as the assumption
437 * of the AUX buffer management code is that after pmu::stop(), the AUX
438 * transaction must be stopped and therefore drop the AUX reference count.
439 */
440void perf_aux_output_end(struct perf_output_handle *handle, unsigned long size)
441{
442 bool wakeup = !!(handle->aux_flags & PERF_AUX_FLAG_TRUNCATED);
443 struct ring_buffer *rb = handle->rb;
444 unsigned long aux_head;
445
446 /* in overwrite mode, driver provides aux_head via handle */
447 if (rb->aux_overwrite) {
448 handle->aux_flags |= PERF_AUX_FLAG_OVERWRITE;
449
450 aux_head = handle->head;
451 rb->aux_head = aux_head;
452 } else {
453 handle->aux_flags &= ~PERF_AUX_FLAG_OVERWRITE;
454
455 aux_head = rb->aux_head;
456 rb->aux_head += size;
457 }
458
459 if (size || handle->aux_flags) {
460 /*
461 * Only send RECORD_AUX if we have something useful to communicate
462 */
463
464 perf_event_aux_event(handle->event, aux_head, size,
465 handle->aux_flags);
466 }
467
468 rb->user_page->aux_head = rb->aux_head;
469 if (rb_need_aux_wakeup(rb))
470 wakeup = true;
471
472 if (wakeup) {
473 if (handle->aux_flags & PERF_AUX_FLAG_TRUNCATED)
474 handle->event->pending_disable = 1;
475 perf_output_wakeup(handle);
476 }
477
478 handle->event = NULL;
479
480 local_set(&rb->aux_nest, 0);
481 /* can't be last */
482 rb_free_aux(rb);
483 ring_buffer_put(rb);
484}
485EXPORT_SYMBOL_GPL(perf_aux_output_end);
486
487/*
488 * Skip over a given number of bytes in the AUX buffer, due to, for example,
489 * hardware's alignment constraints.
490 */
491int perf_aux_output_skip(struct perf_output_handle *handle, unsigned long size)
492{
493 struct ring_buffer *rb = handle->rb;
494
495 if (size > handle->size)
496 return -ENOSPC;
497
498 rb->aux_head += size;
499
500 rb->user_page->aux_head = rb->aux_head;
501 if (rb_need_aux_wakeup(rb)) {
502 perf_output_wakeup(handle);
503 handle->wakeup = rb->aux_wakeup + rb->aux_watermark;
504 }
505
506 handle->head = rb->aux_head;
507 handle->size -= size;
508
509 return 0;
510}
511EXPORT_SYMBOL_GPL(perf_aux_output_skip);
512
513void *perf_get_aux(struct perf_output_handle *handle)
514{
515 /* this is only valid between perf_aux_output_begin and *_end */
516 if (!handle->event)
517 return NULL;
518
519 return handle->rb->aux_priv;
520}
521EXPORT_SYMBOL_GPL(perf_get_aux);
522
523#define PERF_AUX_GFP (GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_NORETRY)
524
525static struct page *rb_alloc_aux_page(int node, int order)
526{
527 struct page *page;
528
529 if (order > MAX_ORDER)
530 order = MAX_ORDER;
531
532 do {
533 page = alloc_pages_node(node, PERF_AUX_GFP, order);
534 } while (!page && order--);
535
536 if (page && order) {
537 /*
538 * Communicate the allocation size to the driver:
539 * if we managed to secure a high-order allocation,
540 * set its first page's private to this order;
541 * !PagePrivate(page) means it's just a normal page.
542 */
543 split_page(page, order);
544 SetPagePrivate(page);
545 set_page_private(page, order);
546 }
547
548 return page;
549}
550
551static void rb_free_aux_page(struct ring_buffer *rb, int idx)
552{
553 struct page *page = virt_to_page(rb->aux_pages[idx]);
554
555 ClearPagePrivate(page);
556 page->mapping = NULL;
557 __free_page(page);
558}
559
560static void __rb_free_aux(struct ring_buffer *rb)
561{
562 int pg;
563
564 /*
565 * Should never happen, the last reference should be dropped from
566 * perf_mmap_close() path, which first stops aux transactions (which
567 * in turn are the atomic holders of aux_refcount) and then does the
568 * last rb_free_aux().
569 */
570 WARN_ON_ONCE(in_atomic());
571
572 if (rb->aux_priv) {
573 rb->free_aux(rb->aux_priv);
574 rb->free_aux = NULL;
575 rb->aux_priv = NULL;
576 }
577
578 if (rb->aux_nr_pages) {
579 for (pg = 0; pg < rb->aux_nr_pages; pg++)
580 rb_free_aux_page(rb, pg);
581
582 kfree(rb->aux_pages);
583 rb->aux_nr_pages = 0;
584 }
585}
586
587int rb_alloc_aux(struct ring_buffer *rb, struct perf_event *event,
588 pgoff_t pgoff, int nr_pages, long watermark, int flags)
589{
590 bool overwrite = !(flags & RING_BUFFER_WRITABLE);
591 int node = (event->cpu == -1) ? -1 : cpu_to_node(event->cpu);
592 int ret = -ENOMEM, max_order = 0;
593
594 if (!has_aux(event))
595 return -EOPNOTSUPP;
596
597 if (event->pmu->capabilities & PERF_PMU_CAP_AUX_NO_SG) {
598 /*
599 * We need to start with the max_order that fits in nr_pages,
600 * not the other way around, hence ilog2() and not get_order.
601 */
602 max_order = ilog2(nr_pages);
603
604 /*
605 * PMU requests more than one contiguous chunks of memory
606 * for SW double buffering
607 */
608 if ((event->pmu->capabilities & PERF_PMU_CAP_AUX_SW_DOUBLEBUF) &&
609 !overwrite) {
610 if (!max_order)
611 return -EINVAL;
612
613 max_order--;
614 }
615 }
616
617 rb->aux_pages = kzalloc_node(nr_pages * sizeof(void *), GFP_KERNEL, node);
618 if (!rb->aux_pages)
619 return -ENOMEM;
620
621 rb->free_aux = event->pmu->free_aux;
622 for (rb->aux_nr_pages = 0; rb->aux_nr_pages < nr_pages;) {
623 struct page *page;
624 int last, order;
625
626 order = min(max_order, ilog2(nr_pages - rb->aux_nr_pages));
627 page = rb_alloc_aux_page(node, order);
628 if (!page)
629 goto out;
630
631 for (last = rb->aux_nr_pages + (1 << page_private(page));
632 last > rb->aux_nr_pages; rb->aux_nr_pages++)
633 rb->aux_pages[rb->aux_nr_pages] = page_address(page++);
634 }
635
636 /*
637 * In overwrite mode, PMUs that don't support SG may not handle more
638 * than one contiguous allocation, since they rely on PMI to do double
639 * buffering. In this case, the entire buffer has to be one contiguous
640 * chunk.
641 */
642 if ((event->pmu->capabilities & PERF_PMU_CAP_AUX_NO_SG) &&
643 overwrite) {
644 struct page *page = virt_to_page(rb->aux_pages[0]);
645
646 if (page_private(page) != max_order)
647 goto out;
648 }
649
650 rb->aux_priv = event->pmu->setup_aux(event->cpu, rb->aux_pages, nr_pages,
651 overwrite);
652 if (!rb->aux_priv)
653 goto out;
654
655 ret = 0;
656
657 /*
658 * aux_pages (and pmu driver's private data, aux_priv) will be
659 * referenced in both producer's and consumer's contexts, thus
660 * we keep a refcount here to make sure either of the two can
661 * reference them safely.
662 */
663 atomic_set(&rb->aux_refcount, 1);
664
665 rb->aux_overwrite = overwrite;
666 rb->aux_watermark = watermark;
667
668 if (!rb->aux_watermark && !rb->aux_overwrite)
669 rb->aux_watermark = nr_pages << (PAGE_SHIFT - 1);
670
671out:
672 if (!ret)
673 rb->aux_pgoff = pgoff;
674 else
675 __rb_free_aux(rb);
676
677 return ret;
678}
679
680void rb_free_aux(struct ring_buffer *rb)
681{
682 if (atomic_dec_and_test(&rb->aux_refcount))
683 __rb_free_aux(rb);
684}
685
686#ifndef CONFIG_PERF_USE_VMALLOC
687
688/*
689 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
690 */
691
692static struct page *
693__perf_mmap_to_page(struct ring_buffer *rb, unsigned long pgoff)
694{
695 if (pgoff > rb->nr_pages)
696 return NULL;
697
698 if (pgoff == 0)
699 return virt_to_page(rb->user_page);
700
701 return virt_to_page(rb->data_pages[pgoff - 1]);
702}
703
704static void *perf_mmap_alloc_page(int cpu)
705{
706 struct page *page;
707 int node;
708
709 node = (cpu == -1) ? cpu : cpu_to_node(cpu);
710 page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
711 if (!page)
712 return NULL;
713
714 return page_address(page);
715}
716
717struct ring_buffer *rb_alloc(int nr_pages, long watermark, int cpu, int flags)
718{
719 struct ring_buffer *rb;
720 unsigned long size;
721 int i;
722
723 size = sizeof(struct ring_buffer);
724 size += nr_pages * sizeof(void *);
725
726 rb = kzalloc(size, GFP_KERNEL);
727 if (!rb)
728 goto fail;
729
730 rb->user_page = perf_mmap_alloc_page(cpu);
731 if (!rb->user_page)
732 goto fail_user_page;
733
734 for (i = 0; i < nr_pages; i++) {
735 rb->data_pages[i] = perf_mmap_alloc_page(cpu);
736 if (!rb->data_pages[i])
737 goto fail_data_pages;
738 }
739
740 rb->nr_pages = nr_pages;
741
742 ring_buffer_init(rb, watermark, flags);
743
744 return rb;
745
746fail_data_pages:
747 for (i--; i >= 0; i--)
748 free_page((unsigned long)rb->data_pages[i]);
749
750 free_page((unsigned long)rb->user_page);
751
752fail_user_page:
753 kfree(rb);
754
755fail:
756 return NULL;
757}
758
759static void perf_mmap_free_page(unsigned long addr)
760{
761 struct page *page = virt_to_page((void *)addr);
762
763 page->mapping = NULL;
764 __free_page(page);
765}
766
767void rb_free(struct ring_buffer *rb)
768{
769 int i;
770
771 perf_mmap_free_page((unsigned long)rb->user_page);
772 for (i = 0; i < rb->nr_pages; i++)
773 perf_mmap_free_page((unsigned long)rb->data_pages[i]);
774 kfree(rb);
775}
776
777#else
778static int data_page_nr(struct ring_buffer *rb)
779{
780 return rb->nr_pages << page_order(rb);
781}
782
783static struct page *
784__perf_mmap_to_page(struct ring_buffer *rb, unsigned long pgoff)
785{
786 /* The '>' counts in the user page. */
787 if (pgoff > data_page_nr(rb))
788 return NULL;
789
790 return vmalloc_to_page((void *)rb->user_page + pgoff * PAGE_SIZE);
791}
792
793static void perf_mmap_unmark_page(void *addr)
794{
795 struct page *page = vmalloc_to_page(addr);
796
797 page->mapping = NULL;
798}
799
800static void rb_free_work(struct work_struct *work)
801{
802 struct ring_buffer *rb;
803 void *base;
804 int i, nr;
805
806 rb = container_of(work, struct ring_buffer, work);
807 nr = data_page_nr(rb);
808
809 base = rb->user_page;
810 /* The '<=' counts in the user page. */
811 for (i = 0; i <= nr; i++)
812 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
813
814 vfree(base);
815 kfree(rb);
816}
817
818void rb_free(struct ring_buffer *rb)
819{
820 schedule_work(&rb->work);
821}
822
823struct ring_buffer *rb_alloc(int nr_pages, long watermark, int cpu, int flags)
824{
825 struct ring_buffer *rb;
826 unsigned long size;
827 void *all_buf;
828
829 size = sizeof(struct ring_buffer);
830 size += sizeof(void *);
831
832 rb = kzalloc(size, GFP_KERNEL);
833 if (!rb)
834 goto fail;
835
836 INIT_WORK(&rb->work, rb_free_work);
837
838 all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
839 if (!all_buf)
840 goto fail_all_buf;
841
842 rb->user_page = all_buf;
843 rb->data_pages[0] = all_buf + PAGE_SIZE;
844 if (nr_pages) {
845 rb->nr_pages = 1;
846 rb->page_order = ilog2(nr_pages);
847 }
848
849 ring_buffer_init(rb, watermark, flags);
850
851 return rb;
852
853fail_all_buf:
854 kfree(rb);
855
856fail:
857 return NULL;
858}
859
860#endif
861
862struct page *
863perf_mmap_to_page(struct ring_buffer *rb, unsigned long pgoff)
864{
865 if (rb->aux_nr_pages) {
866 /* above AUX space */
867 if (pgoff > rb->aux_pgoff + rb->aux_nr_pages)
868 return NULL;
869
870 /* AUX space */
871 if (pgoff >= rb->aux_pgoff) {
872 int aux_pgoff = array_index_nospec(pgoff - rb->aux_pgoff, rb->aux_nr_pages);
873 return virt_to_page(rb->aux_pages[aux_pgoff]);
874 }
875 }
876
877 return __perf_mmap_to_page(rb, pgoff);
878}