]>
Commit | Line | Data |
---|---|---|
76369139 FW |
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 | |
90eec103 | 6 | * Copyright (C) 2008-2011 Red Hat, Inc., Peter Zijlstra |
d36b6910 | 7 | * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com> |
76369139 FW |
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> | |
26c86da8 | 15 | #include <linux/circ_buf.h> |
7c60fc0e | 16 | #include <linux/poll.h> |
76369139 FW |
17 | |
18 | #include "internal.h" | |
19 | ||
76369139 FW |
20 | static void perf_output_wakeup(struct perf_output_handle *handle) |
21 | { | |
7c60fc0e | 22 | atomic_set(&handle->rb->poll, POLLIN); |
76369139 | 23 | |
a8b0ca17 PZ |
24 | handle->event->pending_wakeup = 1; |
25 | irq_work_queue(&handle->event->pending); | |
76369139 FW |
26 | } |
27 | ||
28 | /* | |
29 | * We need to ensure a later event_id doesn't publish a head when a former | |
30 | * event isn't done writing. However since we need to deal with NMIs we | |
31 | * cannot fully serialize things. | |
32 | * | |
33 | * We only publish the head (and generate a wakeup) when the outer-most | |
34 | * event completes. | |
35 | */ | |
36 | static void perf_output_get_handle(struct perf_output_handle *handle) | |
37 | { | |
38 | struct ring_buffer *rb = handle->rb; | |
39 | ||
40 | preempt_disable(); | |
41 | local_inc(&rb->nest); | |
42 | handle->wakeup = local_read(&rb->wakeup); | |
43 | } | |
44 | ||
45 | static void perf_output_put_handle(struct perf_output_handle *handle) | |
46 | { | |
47 | struct ring_buffer *rb = handle->rb; | |
48 | unsigned long head; | |
49 | ||
50 | again: | |
51 | head = local_read(&rb->head); | |
52 | ||
53 | /* | |
54 | * IRQ/NMI can happen here, which means we can miss a head update. | |
55 | */ | |
56 | ||
57 | if (!local_dec_and_test(&rb->nest)) | |
58 | goto out; | |
59 | ||
60 | /* | |
bf378d34 PZ |
61 | * Since the mmap() consumer (userspace) can run on a different CPU: |
62 | * | |
63 | * kernel user | |
64 | * | |
c7f2e3cd PZ |
65 | * if (LOAD ->data_tail) { LOAD ->data_head |
66 | * (A) smp_rmb() (C) | |
67 | * STORE $data LOAD $data | |
68 | * smp_wmb() (B) smp_mb() (D) | |
69 | * STORE ->data_head STORE ->data_tail | |
70 | * } | |
bf378d34 PZ |
71 | * |
72 | * Where A pairs with D, and B pairs with C. | |
73 | * | |
c7f2e3cd PZ |
74 | * In our case (A) is a control dependency that separates the load of |
75 | * the ->data_tail and the stores of $data. In case ->data_tail | |
76 | * indicates there is no room in the buffer to store $data we do not. | |
bf378d34 | 77 | * |
c7f2e3cd | 78 | * D needs to be a full barrier since it separates the data READ |
bf378d34 PZ |
79 | * from the tail WRITE. |
80 | * | |
81 | * For B a WMB is sufficient since it separates two WRITEs, and for C | |
82 | * an RMB is sufficient since it separates two READs. | |
83 | * | |
84 | * See perf_output_begin(). | |
76369139 | 85 | */ |
c7f2e3cd | 86 | smp_wmb(); /* B, matches C */ |
76369139 FW |
87 | rb->user_page->data_head = head; |
88 | ||
89 | /* | |
394570b7 PZ |
90 | * Now check if we missed an update -- rely on previous implied |
91 | * compiler barriers to force a re-read. | |
76369139 FW |
92 | */ |
93 | if (unlikely(head != local_read(&rb->head))) { | |
94 | local_inc(&rb->nest); | |
95 | goto again; | |
96 | } | |
97 | ||
98 | if (handle->wakeup != local_read(&rb->wakeup)) | |
99 | perf_output_wakeup(handle); | |
100 | ||
101 | out: | |
102 | preempt_enable(); | |
103 | } | |
104 | ||
d1b26c70 WN |
105 | static bool __always_inline |
106 | ring_buffer_has_space(unsigned long head, unsigned long tail, | |
107 | unsigned long data_size, unsigned int size, | |
108 | bool backward) | |
109 | { | |
110 | if (!backward) | |
111 | return CIRC_SPACE(head, tail, data_size) >= size; | |
112 | else | |
113 | return CIRC_SPACE(tail, head, data_size) >= size; | |
114 | } | |
115 | ||
116 | static int __always_inline | |
117 | __perf_output_begin(struct perf_output_handle *handle, | |
118 | struct perf_event *event, unsigned int size, | |
119 | bool backward) | |
76369139 FW |
120 | { |
121 | struct ring_buffer *rb; | |
122 | unsigned long tail, offset, head; | |
524feca5 | 123 | int have_lost, page_shift; |
76369139 FW |
124 | struct { |
125 | struct perf_event_header header; | |
126 | u64 id; | |
127 | u64 lost; | |
128 | } lost_event; | |
129 | ||
130 | rcu_read_lock(); | |
131 | /* | |
132 | * For inherited events we send all the output towards the parent. | |
133 | */ | |
134 | if (event->parent) | |
135 | event = event->parent; | |
136 | ||
137 | rb = rcu_dereference(event->rb); | |
c72b42a3 | 138 | if (unlikely(!rb)) |
76369139 FW |
139 | goto out; |
140 | ||
86e7972f WN |
141 | if (unlikely(rb->paused)) { |
142 | if (rb->nr_pages) | |
143 | local_inc(&rb->lost); | |
76369139 | 144 | goto out; |
86e7972f | 145 | } |
76369139 | 146 | |
c72b42a3 PZ |
147 | handle->rb = rb; |
148 | handle->event = event; | |
149 | ||
76369139 | 150 | have_lost = local_read(&rb->lost); |
c72b42a3 | 151 | if (unlikely(have_lost)) { |
d20a973f PZ |
152 | size += sizeof(lost_event); |
153 | if (event->attr.sample_id_all) | |
154 | size += event->id_header_size; | |
76369139 FW |
155 | } |
156 | ||
157 | perf_output_get_handle(handle); | |
158 | ||
159 | do { | |
105ff3cb | 160 | tail = READ_ONCE(rb->user_page->data_tail); |
76369139 | 161 | offset = head = local_read(&rb->head); |
d1b26c70 WN |
162 | if (!rb->overwrite) { |
163 | if (unlikely(!ring_buffer_has_space(head, tail, | |
164 | perf_data_size(rb), | |
165 | size, backward))) | |
166 | goto fail; | |
167 | } | |
c7f2e3cd PZ |
168 | |
169 | /* | |
170 | * The above forms a control dependency barrier separating the | |
171 | * @tail load above from the data stores below. Since the @tail | |
172 | * load is required to compute the branch to fail below. | |
173 | * | |
174 | * A, matches D; the full memory barrier userspace SHOULD issue | |
175 | * after reading the data and before storing the new tail | |
176 | * position. | |
177 | * | |
178 | * See perf_output_put_handle(). | |
179 | */ | |
180 | ||
d1b26c70 WN |
181 | if (!backward) |
182 | head += size; | |
183 | else | |
184 | head -= size; | |
76369139 FW |
185 | } while (local_cmpxchg(&rb->head, offset, head) != offset); |
186 | ||
d1b26c70 WN |
187 | if (backward) { |
188 | offset = head; | |
189 | head = (u64)(-head); | |
190 | } | |
191 | ||
85f59edf | 192 | /* |
c7f2e3cd PZ |
193 | * We rely on the implied barrier() by local_cmpxchg() to ensure |
194 | * none of the data stores below can be lifted up by the compiler. | |
85f59edf | 195 | */ |
85f59edf | 196 | |
c72b42a3 | 197 | if (unlikely(head - local_read(&rb->wakeup) > rb->watermark)) |
76369139 FW |
198 | local_add(rb->watermark, &rb->wakeup); |
199 | ||
524feca5 PZ |
200 | page_shift = PAGE_SHIFT + page_order(rb); |
201 | ||
202 | handle->page = (offset >> page_shift) & (rb->nr_pages - 1); | |
203 | offset &= (1UL << page_shift) - 1; | |
204 | handle->addr = rb->data_pages[handle->page] + offset; | |
205 | handle->size = (1UL << page_shift) - offset; | |
76369139 | 206 | |
c72b42a3 | 207 | if (unlikely(have_lost)) { |
d20a973f PZ |
208 | struct perf_sample_data sample_data; |
209 | ||
210 | lost_event.header.size = sizeof(lost_event); | |
76369139 FW |
211 | lost_event.header.type = PERF_RECORD_LOST; |
212 | lost_event.header.misc = 0; | |
213 | lost_event.id = event->id; | |
214 | lost_event.lost = local_xchg(&rb->lost, 0); | |
215 | ||
d20a973f PZ |
216 | perf_event_header__init_id(&lost_event.header, |
217 | &sample_data, event); | |
76369139 FW |
218 | perf_output_put(handle, lost_event); |
219 | perf_event__output_id_sample(event, handle, &sample_data); | |
220 | } | |
221 | ||
222 | return 0; | |
223 | ||
224 | fail: | |
225 | local_inc(&rb->lost); | |
226 | perf_output_put_handle(handle); | |
227 | out: | |
228 | rcu_read_unlock(); | |
229 | ||
230 | return -ENOSPC; | |
231 | } | |
232 | ||
9ecda41a WN |
233 | int perf_output_begin_forward(struct perf_output_handle *handle, |
234 | struct perf_event *event, unsigned int size) | |
235 | { | |
236 | return __perf_output_begin(handle, event, size, false); | |
237 | } | |
238 | ||
239 | int perf_output_begin_backward(struct perf_output_handle *handle, | |
240 | struct perf_event *event, unsigned int size) | |
241 | { | |
242 | return __perf_output_begin(handle, event, size, true); | |
243 | } | |
244 | ||
d1b26c70 WN |
245 | int perf_output_begin(struct perf_output_handle *handle, |
246 | struct perf_event *event, unsigned int size) | |
247 | { | |
9ecda41a WN |
248 | |
249 | return __perf_output_begin(handle, event, size, | |
250 | unlikely(is_write_backward(event))); | |
d1b26c70 WN |
251 | } |
252 | ||
91d7753a | 253 | unsigned int perf_output_copy(struct perf_output_handle *handle, |
76369139 FW |
254 | const void *buf, unsigned int len) |
255 | { | |
91d7753a | 256 | return __output_copy(handle, buf, len); |
76369139 FW |
257 | } |
258 | ||
5685e0ff JO |
259 | unsigned int perf_output_skip(struct perf_output_handle *handle, |
260 | unsigned int len) | |
261 | { | |
262 | return __output_skip(handle, NULL, len); | |
263 | } | |
264 | ||
76369139 FW |
265 | void perf_output_end(struct perf_output_handle *handle) |
266 | { | |
76369139 FW |
267 | perf_output_put_handle(handle); |
268 | rcu_read_unlock(); | |
269 | } | |
270 | ||
271 | static void | |
272 | ring_buffer_init(struct ring_buffer *rb, long watermark, int flags) | |
273 | { | |
274 | long max_size = perf_data_size(rb); | |
275 | ||
276 | if (watermark) | |
277 | rb->watermark = min(max_size, watermark); | |
278 | ||
279 | if (!rb->watermark) | |
280 | rb->watermark = max_size / 2; | |
281 | ||
282 | if (flags & RING_BUFFER_WRITABLE) | |
dd9c086d SE |
283 | rb->overwrite = 0; |
284 | else | |
285 | rb->overwrite = 1; | |
76369139 FW |
286 | |
287 | atomic_set(&rb->refcount, 1); | |
10c6db11 PZ |
288 | |
289 | INIT_LIST_HEAD(&rb->event_list); | |
290 | spin_lock_init(&rb->event_lock); | |
86e7972f WN |
291 | |
292 | /* | |
293 | * perf_output_begin() only checks rb->paused, therefore | |
294 | * rb->paused must be true if we have no pages for output. | |
295 | */ | |
296 | if (!rb->nr_pages) | |
297 | rb->paused = 1; | |
76369139 FW |
298 | } |
299 | ||
fdc26706 AS |
300 | /* |
301 | * This is called before hardware starts writing to the AUX area to | |
302 | * obtain an output handle and make sure there's room in the buffer. | |
303 | * When the capture completes, call perf_aux_output_end() to commit | |
304 | * the recorded data to the buffer. | |
305 | * | |
306 | * The ordering is similar to that of perf_output_{begin,end}, with | |
307 | * the exception of (B), which should be taken care of by the pmu | |
308 | * driver, since ordering rules will differ depending on hardware. | |
af5bb4ed AS |
309 | * |
310 | * Call this from pmu::start(); see the comment in perf_aux_output_end() | |
311 | * about its use in pmu callbacks. Both can also be called from the PMI | |
312 | * handler if needed. | |
fdc26706 AS |
313 | */ |
314 | void *perf_aux_output_begin(struct perf_output_handle *handle, | |
315 | struct perf_event *event) | |
316 | { | |
317 | struct perf_event *output_event = event; | |
318 | unsigned long aux_head, aux_tail; | |
319 | struct ring_buffer *rb; | |
320 | ||
321 | if (output_event->parent) | |
322 | output_event = output_event->parent; | |
323 | ||
324 | /* | |
325 | * Since this will typically be open across pmu::add/pmu::del, we | |
326 | * grab ring_buffer's refcount instead of holding rcu read lock | |
327 | * to make sure it doesn't disappear under us. | |
328 | */ | |
329 | rb = ring_buffer_get(output_event); | |
330 | if (!rb) | |
331 | return NULL; | |
332 | ||
b79ccadd | 333 | if (!rb_has_aux(rb)) |
fdc26706 AS |
334 | goto err; |
335 | ||
dcb10a96 | 336 | /* |
b79ccadd AS |
337 | * If aux_mmap_count is zero, the aux buffer is in perf_mmap_close(), |
338 | * about to get freed, so we leave immediately. | |
339 | * | |
340 | * Checking rb::aux_mmap_count and rb::refcount has to be done in | |
341 | * the same order, see perf_mmap_close. Otherwise we end up freeing | |
342 | * aux pages in this path, which is a bug, because in_atomic(). | |
dcb10a96 AS |
343 | */ |
344 | if (!atomic_read(&rb->aux_mmap_count)) | |
b79ccadd AS |
345 | goto err; |
346 | ||
347 | if (!atomic_inc_not_zero(&rb->aux_refcount)) | |
348 | goto err; | |
dcb10a96 | 349 | |
fdc26706 AS |
350 | /* |
351 | * Nesting is not supported for AUX area, make sure nested | |
352 | * writers are caught early | |
353 | */ | |
354 | if (WARN_ON_ONCE(local_xchg(&rb->aux_nest, 1))) | |
355 | goto err_put; | |
356 | ||
357 | aux_head = local_read(&rb->aux_head); | |
fdc26706 AS |
358 | |
359 | handle->rb = rb; | |
360 | handle->event = event; | |
361 | handle->head = aux_head; | |
2023a0d2 | 362 | handle->size = 0; |
fdc26706 AS |
363 | |
364 | /* | |
2023a0d2 AS |
365 | * In overwrite mode, AUX data stores do not depend on aux_tail, |
366 | * therefore (A) control dependency barrier does not exist. The | |
367 | * (B) <-> (C) ordering is still observed by the pmu driver. | |
fdc26706 | 368 | */ |
2023a0d2 AS |
369 | if (!rb->aux_overwrite) { |
370 | aux_tail = ACCESS_ONCE(rb->user_page->aux_tail); | |
1a594131 | 371 | handle->wakeup = local_read(&rb->aux_wakeup) + rb->aux_watermark; |
2023a0d2 AS |
372 | if (aux_head - aux_tail < perf_aux_size(rb)) |
373 | handle->size = CIRC_SPACE(aux_head, aux_tail, perf_aux_size(rb)); | |
374 | ||
375 | /* | |
376 | * handle->size computation depends on aux_tail load; this forms a | |
377 | * control dependency barrier separating aux_tail load from aux data | |
378 | * store that will be enabled on successful return | |
379 | */ | |
380 | if (!handle->size) { /* A, matches D */ | |
381 | event->pending_disable = 1; | |
382 | perf_output_wakeup(handle); | |
383 | local_set(&rb->aux_nest, 0); | |
384 | goto err_put; | |
385 | } | |
fdc26706 AS |
386 | } |
387 | ||
388 | return handle->rb->aux_priv; | |
389 | ||
390 | err_put: | |
af5bb4ed | 391 | /* can't be last */ |
fdc26706 AS |
392 | rb_free_aux(rb); |
393 | ||
394 | err: | |
95ff4ca2 | 395 | ring_buffer_put(rb); |
fdc26706 AS |
396 | handle->event = NULL; |
397 | ||
398 | return NULL; | |
399 | } | |
400 | ||
401 | /* | |
402 | * Commit the data written by hardware into the ring buffer by adjusting | |
403 | * aux_head and posting a PERF_RECORD_AUX into the perf buffer. It is the | |
404 | * pmu driver's responsibility to observe ordering rules of the hardware, | |
405 | * so that all the data is externally visible before this is called. | |
af5bb4ed AS |
406 | * |
407 | * Note: this has to be called from pmu::stop() callback, as the assumption | |
408 | * of the AUX buffer management code is that after pmu::stop(), the AUX | |
409 | * transaction must be stopped and therefore drop the AUX reference count. | |
fdc26706 AS |
410 | */ |
411 | void perf_aux_output_end(struct perf_output_handle *handle, unsigned long size, | |
412 | bool truncated) | |
413 | { | |
414 | struct ring_buffer *rb = handle->rb; | |
3f56e687 | 415 | bool wakeup = truncated; |
2023a0d2 | 416 | unsigned long aux_head; |
fdc26706 AS |
417 | u64 flags = 0; |
418 | ||
419 | if (truncated) | |
420 | flags |= PERF_AUX_FLAG_TRUNCATED; | |
421 | ||
2023a0d2 AS |
422 | /* in overwrite mode, driver provides aux_head via handle */ |
423 | if (rb->aux_overwrite) { | |
424 | flags |= PERF_AUX_FLAG_OVERWRITE; | |
425 | ||
426 | aux_head = handle->head; | |
427 | local_set(&rb->aux_head, aux_head); | |
428 | } else { | |
429 | aux_head = local_read(&rb->aux_head); | |
430 | local_add(size, &rb->aux_head); | |
431 | } | |
fdc26706 AS |
432 | |
433 | if (size || flags) { | |
434 | /* | |
435 | * Only send RECORD_AUX if we have something useful to communicate | |
436 | */ | |
437 | ||
438 | perf_event_aux_event(handle->event, aux_head, size, flags); | |
439 | } | |
440 | ||
1a594131 | 441 | aux_head = rb->user_page->aux_head = local_read(&rb->aux_head); |
fdc26706 | 442 | |
1a594131 | 443 | if (aux_head - local_read(&rb->aux_wakeup) >= rb->aux_watermark) { |
3f56e687 | 444 | wakeup = true; |
1a594131 AS |
445 | local_add(rb->aux_watermark, &rb->aux_wakeup); |
446 | } | |
3f56e687 AS |
447 | |
448 | if (wakeup) { | |
449 | if (truncated) | |
450 | handle->event->pending_disable = 1; | |
451 | perf_output_wakeup(handle); | |
452 | } | |
453 | ||
fdc26706 AS |
454 | handle->event = NULL; |
455 | ||
456 | local_set(&rb->aux_nest, 0); | |
af5bb4ed | 457 | /* can't be last */ |
fdc26706 | 458 | rb_free_aux(rb); |
95ff4ca2 | 459 | ring_buffer_put(rb); |
fdc26706 AS |
460 | } |
461 | ||
462 | /* | |
463 | * Skip over a given number of bytes in the AUX buffer, due to, for example, | |
464 | * hardware's alignment constraints. | |
465 | */ | |
466 | int perf_aux_output_skip(struct perf_output_handle *handle, unsigned long size) | |
467 | { | |
468 | struct ring_buffer *rb = handle->rb; | |
469 | unsigned long aux_head; | |
470 | ||
471 | if (size > handle->size) | |
472 | return -ENOSPC; | |
473 | ||
474 | local_add(size, &rb->aux_head); | |
475 | ||
1a594131 AS |
476 | aux_head = rb->user_page->aux_head = local_read(&rb->aux_head); |
477 | if (aux_head - local_read(&rb->aux_wakeup) >= rb->aux_watermark) { | |
478 | perf_output_wakeup(handle); | |
479 | local_add(rb->aux_watermark, &rb->aux_wakeup); | |
480 | handle->wakeup = local_read(&rb->aux_wakeup) + | |
481 | rb->aux_watermark; | |
482 | } | |
483 | ||
fdc26706 AS |
484 | handle->head = aux_head; |
485 | handle->size -= size; | |
486 | ||
487 | return 0; | |
488 | } | |
489 | ||
490 | void *perf_get_aux(struct perf_output_handle *handle) | |
491 | { | |
492 | /* this is only valid between perf_aux_output_begin and *_end */ | |
493 | if (!handle->event) | |
494 | return NULL; | |
495 | ||
496 | return handle->rb->aux_priv; | |
497 | } | |
498 | ||
0a4e38e6 AS |
499 | #define PERF_AUX_GFP (GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_NORETRY) |
500 | ||
501 | static struct page *rb_alloc_aux_page(int node, int order) | |
502 | { | |
503 | struct page *page; | |
504 | ||
505 | if (order > MAX_ORDER) | |
506 | order = MAX_ORDER; | |
507 | ||
508 | do { | |
509 | page = alloc_pages_node(node, PERF_AUX_GFP, order); | |
510 | } while (!page && order--); | |
511 | ||
512 | if (page && order) { | |
513 | /* | |
c2ad6b51 AS |
514 | * Communicate the allocation size to the driver: |
515 | * if we managed to secure a high-order allocation, | |
516 | * set its first page's private to this order; | |
517 | * !PagePrivate(page) means it's just a normal page. | |
0a4e38e6 AS |
518 | */ |
519 | split_page(page, order); | |
520 | SetPagePrivate(page); | |
521 | set_page_private(page, order); | |
522 | } | |
523 | ||
524 | return page; | |
525 | } | |
526 | ||
527 | static void rb_free_aux_page(struct ring_buffer *rb, int idx) | |
528 | { | |
529 | struct page *page = virt_to_page(rb->aux_pages[idx]); | |
530 | ||
531 | ClearPagePrivate(page); | |
532 | page->mapping = NULL; | |
533 | __free_page(page); | |
534 | } | |
535 | ||
45c815f0 AS |
536 | static void __rb_free_aux(struct ring_buffer *rb) |
537 | { | |
538 | int pg; | |
539 | ||
95ff4ca2 AS |
540 | /* |
541 | * Should never happen, the last reference should be dropped from | |
542 | * perf_mmap_close() path, which first stops aux transactions (which | |
543 | * in turn are the atomic holders of aux_refcount) and then does the | |
544 | * last rb_free_aux(). | |
545 | */ | |
546 | WARN_ON_ONCE(in_atomic()); | |
547 | ||
45c815f0 AS |
548 | if (rb->aux_priv) { |
549 | rb->free_aux(rb->aux_priv); | |
550 | rb->free_aux = NULL; | |
551 | rb->aux_priv = NULL; | |
552 | } | |
553 | ||
554 | if (rb->aux_nr_pages) { | |
555 | for (pg = 0; pg < rb->aux_nr_pages; pg++) | |
556 | rb_free_aux_page(rb, pg); | |
557 | ||
558 | kfree(rb->aux_pages); | |
559 | rb->aux_nr_pages = 0; | |
560 | } | |
561 | } | |
562 | ||
45bfb2e5 | 563 | int rb_alloc_aux(struct ring_buffer *rb, struct perf_event *event, |
1a594131 | 564 | pgoff_t pgoff, int nr_pages, long watermark, int flags) |
45bfb2e5 PZ |
565 | { |
566 | bool overwrite = !(flags & RING_BUFFER_WRITABLE); | |
567 | int node = (event->cpu == -1) ? -1 : cpu_to_node(event->cpu); | |
0a4e38e6 | 568 | int ret = -ENOMEM, max_order = 0; |
45bfb2e5 PZ |
569 | |
570 | if (!has_aux(event)) | |
571 | return -ENOTSUPP; | |
572 | ||
6a279230 | 573 | if (event->pmu->capabilities & PERF_PMU_CAP_AUX_NO_SG) { |
0a4e38e6 AS |
574 | /* |
575 | * We need to start with the max_order that fits in nr_pages, | |
576 | * not the other way around, hence ilog2() and not get_order. | |
577 | */ | |
578 | max_order = ilog2(nr_pages); | |
579 | ||
6a279230 AS |
580 | /* |
581 | * PMU requests more than one contiguous chunks of memory | |
582 | * for SW double buffering | |
583 | */ | |
584 | if ((event->pmu->capabilities & PERF_PMU_CAP_AUX_SW_DOUBLEBUF) && | |
585 | !overwrite) { | |
586 | if (!max_order) | |
587 | return -EINVAL; | |
588 | ||
589 | max_order--; | |
590 | } | |
591 | } | |
592 | ||
45bfb2e5 PZ |
593 | rb->aux_pages = kzalloc_node(nr_pages * sizeof(void *), GFP_KERNEL, node); |
594 | if (!rb->aux_pages) | |
595 | return -ENOMEM; | |
596 | ||
597 | rb->free_aux = event->pmu->free_aux; | |
0a4e38e6 | 598 | for (rb->aux_nr_pages = 0; rb->aux_nr_pages < nr_pages;) { |
45bfb2e5 | 599 | struct page *page; |
0a4e38e6 | 600 | int last, order; |
45bfb2e5 | 601 | |
0a4e38e6 AS |
602 | order = min(max_order, ilog2(nr_pages - rb->aux_nr_pages)); |
603 | page = rb_alloc_aux_page(node, order); | |
45bfb2e5 PZ |
604 | if (!page) |
605 | goto out; | |
606 | ||
0a4e38e6 AS |
607 | for (last = rb->aux_nr_pages + (1 << page_private(page)); |
608 | last > rb->aux_nr_pages; rb->aux_nr_pages++) | |
609 | rb->aux_pages[rb->aux_nr_pages] = page_address(page++); | |
45bfb2e5 PZ |
610 | } |
611 | ||
aa319bcd AS |
612 | /* |
613 | * In overwrite mode, PMUs that don't support SG may not handle more | |
614 | * than one contiguous allocation, since they rely on PMI to do double | |
615 | * buffering. In this case, the entire buffer has to be one contiguous | |
616 | * chunk. | |
617 | */ | |
618 | if ((event->pmu->capabilities & PERF_PMU_CAP_AUX_NO_SG) && | |
619 | overwrite) { | |
620 | struct page *page = virt_to_page(rb->aux_pages[0]); | |
621 | ||
622 | if (page_private(page) != max_order) | |
623 | goto out; | |
624 | } | |
625 | ||
45bfb2e5 PZ |
626 | rb->aux_priv = event->pmu->setup_aux(event->cpu, rb->aux_pages, nr_pages, |
627 | overwrite); | |
628 | if (!rb->aux_priv) | |
629 | goto out; | |
630 | ||
631 | ret = 0; | |
632 | ||
633 | /* | |
634 | * aux_pages (and pmu driver's private data, aux_priv) will be | |
635 | * referenced in both producer's and consumer's contexts, thus | |
636 | * we keep a refcount here to make sure either of the two can | |
637 | * reference them safely. | |
638 | */ | |
639 | atomic_set(&rb->aux_refcount, 1); | |
640 | ||
2023a0d2 | 641 | rb->aux_overwrite = overwrite; |
1a594131 AS |
642 | rb->aux_watermark = watermark; |
643 | ||
644 | if (!rb->aux_watermark && !rb->aux_overwrite) | |
645 | rb->aux_watermark = nr_pages << (PAGE_SHIFT - 1); | |
2023a0d2 | 646 | |
45bfb2e5 PZ |
647 | out: |
648 | if (!ret) | |
649 | rb->aux_pgoff = pgoff; | |
650 | else | |
45c815f0 | 651 | __rb_free_aux(rb); |
45bfb2e5 PZ |
652 | |
653 | return ret; | |
654 | } | |
655 | ||
45bfb2e5 PZ |
656 | void rb_free_aux(struct ring_buffer *rb) |
657 | { | |
658 | if (atomic_dec_and_test(&rb->aux_refcount)) | |
659 | __rb_free_aux(rb); | |
660 | } | |
661 | ||
76369139 FW |
662 | #ifndef CONFIG_PERF_USE_VMALLOC |
663 | ||
664 | /* | |
665 | * Back perf_mmap() with regular GFP_KERNEL-0 pages. | |
666 | */ | |
667 | ||
45bfb2e5 PZ |
668 | static struct page * |
669 | __perf_mmap_to_page(struct ring_buffer *rb, unsigned long pgoff) | |
76369139 FW |
670 | { |
671 | if (pgoff > rb->nr_pages) | |
672 | return NULL; | |
673 | ||
674 | if (pgoff == 0) | |
675 | return virt_to_page(rb->user_page); | |
676 | ||
677 | return virt_to_page(rb->data_pages[pgoff - 1]); | |
678 | } | |
679 | ||
680 | static void *perf_mmap_alloc_page(int cpu) | |
681 | { | |
682 | struct page *page; | |
683 | int node; | |
684 | ||
685 | node = (cpu == -1) ? cpu : cpu_to_node(cpu); | |
686 | page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0); | |
687 | if (!page) | |
688 | return NULL; | |
689 | ||
690 | return page_address(page); | |
691 | } | |
692 | ||
693 | struct ring_buffer *rb_alloc(int nr_pages, long watermark, int cpu, int flags) | |
694 | { | |
695 | struct ring_buffer *rb; | |
696 | unsigned long size; | |
697 | int i; | |
698 | ||
699 | size = sizeof(struct ring_buffer); | |
700 | size += nr_pages * sizeof(void *); | |
701 | ||
702 | rb = kzalloc(size, GFP_KERNEL); | |
703 | if (!rb) | |
704 | goto fail; | |
705 | ||
706 | rb->user_page = perf_mmap_alloc_page(cpu); | |
707 | if (!rb->user_page) | |
708 | goto fail_user_page; | |
709 | ||
710 | for (i = 0; i < nr_pages; i++) { | |
711 | rb->data_pages[i] = perf_mmap_alloc_page(cpu); | |
712 | if (!rb->data_pages[i]) | |
713 | goto fail_data_pages; | |
714 | } | |
715 | ||
716 | rb->nr_pages = nr_pages; | |
717 | ||
718 | ring_buffer_init(rb, watermark, flags); | |
719 | ||
720 | return rb; | |
721 | ||
722 | fail_data_pages: | |
723 | for (i--; i >= 0; i--) | |
724 | free_page((unsigned long)rb->data_pages[i]); | |
725 | ||
726 | free_page((unsigned long)rb->user_page); | |
727 | ||
728 | fail_user_page: | |
729 | kfree(rb); | |
730 | ||
731 | fail: | |
732 | return NULL; | |
733 | } | |
734 | ||
735 | static void perf_mmap_free_page(unsigned long addr) | |
736 | { | |
737 | struct page *page = virt_to_page((void *)addr); | |
738 | ||
739 | page->mapping = NULL; | |
740 | __free_page(page); | |
741 | } | |
742 | ||
743 | void rb_free(struct ring_buffer *rb) | |
744 | { | |
745 | int i; | |
746 | ||
747 | perf_mmap_free_page((unsigned long)rb->user_page); | |
748 | for (i = 0; i < rb->nr_pages; i++) | |
749 | perf_mmap_free_page((unsigned long)rb->data_pages[i]); | |
750 | kfree(rb); | |
751 | } | |
752 | ||
753 | #else | |
5919b309 JO |
754 | static int data_page_nr(struct ring_buffer *rb) |
755 | { | |
756 | return rb->nr_pages << page_order(rb); | |
757 | } | |
76369139 | 758 | |
45bfb2e5 PZ |
759 | static struct page * |
760 | __perf_mmap_to_page(struct ring_buffer *rb, unsigned long pgoff) | |
76369139 | 761 | { |
5919b309 JO |
762 | /* The '>' counts in the user page. */ |
763 | if (pgoff > data_page_nr(rb)) | |
76369139 FW |
764 | return NULL; |
765 | ||
766 | return vmalloc_to_page((void *)rb->user_page + pgoff * PAGE_SIZE); | |
767 | } | |
768 | ||
769 | static void perf_mmap_unmark_page(void *addr) | |
770 | { | |
771 | struct page *page = vmalloc_to_page(addr); | |
772 | ||
773 | page->mapping = NULL; | |
774 | } | |
775 | ||
776 | static void rb_free_work(struct work_struct *work) | |
777 | { | |
778 | struct ring_buffer *rb; | |
779 | void *base; | |
780 | int i, nr; | |
781 | ||
782 | rb = container_of(work, struct ring_buffer, work); | |
5919b309 | 783 | nr = data_page_nr(rb); |
76369139 FW |
784 | |
785 | base = rb->user_page; | |
5919b309 JO |
786 | /* The '<=' counts in the user page. */ |
787 | for (i = 0; i <= nr; i++) | |
76369139 FW |
788 | perf_mmap_unmark_page(base + (i * PAGE_SIZE)); |
789 | ||
790 | vfree(base); | |
791 | kfree(rb); | |
792 | } | |
793 | ||
794 | void rb_free(struct ring_buffer *rb) | |
795 | { | |
796 | schedule_work(&rb->work); | |
797 | } | |
798 | ||
799 | struct ring_buffer *rb_alloc(int nr_pages, long watermark, int cpu, int flags) | |
800 | { | |
801 | struct ring_buffer *rb; | |
802 | unsigned long size; | |
803 | void *all_buf; | |
804 | ||
805 | size = sizeof(struct ring_buffer); | |
806 | size += sizeof(void *); | |
807 | ||
808 | rb = kzalloc(size, GFP_KERNEL); | |
809 | if (!rb) | |
810 | goto fail; | |
811 | ||
812 | INIT_WORK(&rb->work, rb_free_work); | |
813 | ||
814 | all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE); | |
815 | if (!all_buf) | |
816 | goto fail_all_buf; | |
817 | ||
818 | rb->user_page = all_buf; | |
819 | rb->data_pages[0] = all_buf + PAGE_SIZE; | |
8184059e PZ |
820 | if (nr_pages) { |
821 | rb->nr_pages = 1; | |
822 | rb->page_order = ilog2(nr_pages); | |
823 | } | |
76369139 FW |
824 | |
825 | ring_buffer_init(rb, watermark, flags); | |
826 | ||
827 | return rb; | |
828 | ||
829 | fail_all_buf: | |
830 | kfree(rb); | |
831 | ||
832 | fail: | |
833 | return NULL; | |
834 | } | |
835 | ||
836 | #endif | |
45bfb2e5 PZ |
837 | |
838 | struct page * | |
839 | perf_mmap_to_page(struct ring_buffer *rb, unsigned long pgoff) | |
840 | { | |
841 | if (rb->aux_nr_pages) { | |
842 | /* above AUX space */ | |
843 | if (pgoff > rb->aux_pgoff + rb->aux_nr_pages) | |
844 | return NULL; | |
845 | ||
846 | /* AUX space */ | |
847 | if (pgoff >= rb->aux_pgoff) | |
848 | return virt_to_page(rb->aux_pages[pgoff - rb->aux_pgoff]); | |
849 | } | |
850 | ||
851 | return __perf_mmap_to_page(rb, pgoff); | |
852 | } |