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 | ||
105 | int perf_output_begin(struct perf_output_handle *handle, | |
a7ac67ea | 106 | struct perf_event *event, unsigned int size) |
76369139 FW |
107 | { |
108 | struct ring_buffer *rb; | |
109 | unsigned long tail, offset, head; | |
524feca5 | 110 | int have_lost, page_shift; |
76369139 FW |
111 | struct { |
112 | struct perf_event_header header; | |
113 | u64 id; | |
114 | u64 lost; | |
115 | } lost_event; | |
116 | ||
117 | rcu_read_lock(); | |
118 | /* | |
119 | * For inherited events we send all the output towards the parent. | |
120 | */ | |
121 | if (event->parent) | |
122 | event = event->parent; | |
123 | ||
124 | rb = rcu_dereference(event->rb); | |
c72b42a3 | 125 | if (unlikely(!rb)) |
76369139 FW |
126 | goto out; |
127 | ||
c72b42a3 | 128 | if (unlikely(!rb->nr_pages)) |
76369139 FW |
129 | goto out; |
130 | ||
c72b42a3 PZ |
131 | handle->rb = rb; |
132 | handle->event = event; | |
133 | ||
76369139 | 134 | have_lost = local_read(&rb->lost); |
c72b42a3 | 135 | if (unlikely(have_lost)) { |
d20a973f PZ |
136 | size += sizeof(lost_event); |
137 | if (event->attr.sample_id_all) | |
138 | size += event->id_header_size; | |
76369139 FW |
139 | } |
140 | ||
141 | perf_output_get_handle(handle); | |
142 | ||
143 | do { | |
105ff3cb | 144 | tail = READ_ONCE(rb->user_page->data_tail); |
76369139 | 145 | offset = head = local_read(&rb->head); |
26c86da8 PZ |
146 | if (!rb->overwrite && |
147 | unlikely(CIRC_SPACE(head, tail, perf_data_size(rb)) < size)) | |
76369139 | 148 | goto fail; |
c7f2e3cd PZ |
149 | |
150 | /* | |
151 | * The above forms a control dependency barrier separating the | |
152 | * @tail load above from the data stores below. Since the @tail | |
153 | * load is required to compute the branch to fail below. | |
154 | * | |
155 | * A, matches D; the full memory barrier userspace SHOULD issue | |
156 | * after reading the data and before storing the new tail | |
157 | * position. | |
158 | * | |
159 | * See perf_output_put_handle(). | |
160 | */ | |
161 | ||
26c86da8 | 162 | head += size; |
76369139 FW |
163 | } while (local_cmpxchg(&rb->head, offset, head) != offset); |
164 | ||
85f59edf | 165 | /* |
c7f2e3cd PZ |
166 | * We rely on the implied barrier() by local_cmpxchg() to ensure |
167 | * none of the data stores below can be lifted up by the compiler. | |
85f59edf | 168 | */ |
85f59edf | 169 | |
c72b42a3 | 170 | if (unlikely(head - local_read(&rb->wakeup) > rb->watermark)) |
76369139 FW |
171 | local_add(rb->watermark, &rb->wakeup); |
172 | ||
524feca5 PZ |
173 | page_shift = PAGE_SHIFT + page_order(rb); |
174 | ||
175 | handle->page = (offset >> page_shift) & (rb->nr_pages - 1); | |
176 | offset &= (1UL << page_shift) - 1; | |
177 | handle->addr = rb->data_pages[handle->page] + offset; | |
178 | handle->size = (1UL << page_shift) - offset; | |
76369139 | 179 | |
c72b42a3 | 180 | if (unlikely(have_lost)) { |
d20a973f PZ |
181 | struct perf_sample_data sample_data; |
182 | ||
183 | lost_event.header.size = sizeof(lost_event); | |
76369139 FW |
184 | lost_event.header.type = PERF_RECORD_LOST; |
185 | lost_event.header.misc = 0; | |
186 | lost_event.id = event->id; | |
187 | lost_event.lost = local_xchg(&rb->lost, 0); | |
188 | ||
d20a973f PZ |
189 | perf_event_header__init_id(&lost_event.header, |
190 | &sample_data, event); | |
76369139 FW |
191 | perf_output_put(handle, lost_event); |
192 | perf_event__output_id_sample(event, handle, &sample_data); | |
193 | } | |
194 | ||
195 | return 0; | |
196 | ||
197 | fail: | |
198 | local_inc(&rb->lost); | |
199 | perf_output_put_handle(handle); | |
200 | out: | |
201 | rcu_read_unlock(); | |
202 | ||
203 | return -ENOSPC; | |
204 | } | |
205 | ||
91d7753a | 206 | unsigned int perf_output_copy(struct perf_output_handle *handle, |
76369139 FW |
207 | const void *buf, unsigned int len) |
208 | { | |
91d7753a | 209 | return __output_copy(handle, buf, len); |
76369139 FW |
210 | } |
211 | ||
5685e0ff JO |
212 | unsigned int perf_output_skip(struct perf_output_handle *handle, |
213 | unsigned int len) | |
214 | { | |
215 | return __output_skip(handle, NULL, len); | |
216 | } | |
217 | ||
76369139 FW |
218 | void perf_output_end(struct perf_output_handle *handle) |
219 | { | |
76369139 FW |
220 | perf_output_put_handle(handle); |
221 | rcu_read_unlock(); | |
222 | } | |
223 | ||
224 | static void | |
225 | ring_buffer_init(struct ring_buffer *rb, long watermark, int flags) | |
226 | { | |
227 | long max_size = perf_data_size(rb); | |
228 | ||
229 | if (watermark) | |
230 | rb->watermark = min(max_size, watermark); | |
231 | ||
232 | if (!rb->watermark) | |
233 | rb->watermark = max_size / 2; | |
234 | ||
235 | if (flags & RING_BUFFER_WRITABLE) | |
dd9c086d SE |
236 | rb->overwrite = 0; |
237 | else | |
238 | rb->overwrite = 1; | |
76369139 FW |
239 | |
240 | atomic_set(&rb->refcount, 1); | |
10c6db11 PZ |
241 | |
242 | INIT_LIST_HEAD(&rb->event_list); | |
243 | spin_lock_init(&rb->event_lock); | |
76369139 FW |
244 | } |
245 | ||
fdc26706 AS |
246 | /* |
247 | * This is called before hardware starts writing to the AUX area to | |
248 | * obtain an output handle and make sure there's room in the buffer. | |
249 | * When the capture completes, call perf_aux_output_end() to commit | |
250 | * the recorded data to the buffer. | |
251 | * | |
252 | * The ordering is similar to that of perf_output_{begin,end}, with | |
253 | * the exception of (B), which should be taken care of by the pmu | |
254 | * driver, since ordering rules will differ depending on hardware. | |
255 | */ | |
256 | void *perf_aux_output_begin(struct perf_output_handle *handle, | |
257 | struct perf_event *event) | |
258 | { | |
259 | struct perf_event *output_event = event; | |
260 | unsigned long aux_head, aux_tail; | |
261 | struct ring_buffer *rb; | |
262 | ||
263 | if (output_event->parent) | |
264 | output_event = output_event->parent; | |
265 | ||
266 | /* | |
267 | * Since this will typically be open across pmu::add/pmu::del, we | |
268 | * grab ring_buffer's refcount instead of holding rcu read lock | |
269 | * to make sure it doesn't disappear under us. | |
270 | */ | |
271 | rb = ring_buffer_get(output_event); | |
272 | if (!rb) | |
273 | return NULL; | |
274 | ||
275 | if (!rb_has_aux(rb) || !atomic_inc_not_zero(&rb->aux_refcount)) | |
276 | goto err; | |
277 | ||
dcb10a96 AS |
278 | /* |
279 | * If rb::aux_mmap_count is zero (and rb_has_aux() above went through), | |
280 | * the aux buffer is in perf_mmap_close(), about to get freed. | |
281 | */ | |
282 | if (!atomic_read(&rb->aux_mmap_count)) | |
95ff4ca2 | 283 | goto err_put; |
dcb10a96 | 284 | |
fdc26706 AS |
285 | /* |
286 | * Nesting is not supported for AUX area, make sure nested | |
287 | * writers are caught early | |
288 | */ | |
289 | if (WARN_ON_ONCE(local_xchg(&rb->aux_nest, 1))) | |
290 | goto err_put; | |
291 | ||
292 | aux_head = local_read(&rb->aux_head); | |
fdc26706 AS |
293 | |
294 | handle->rb = rb; | |
295 | handle->event = event; | |
296 | handle->head = aux_head; | |
2023a0d2 | 297 | handle->size = 0; |
fdc26706 AS |
298 | |
299 | /* | |
2023a0d2 AS |
300 | * In overwrite mode, AUX data stores do not depend on aux_tail, |
301 | * therefore (A) control dependency barrier does not exist. The | |
302 | * (B) <-> (C) ordering is still observed by the pmu driver. | |
fdc26706 | 303 | */ |
2023a0d2 AS |
304 | if (!rb->aux_overwrite) { |
305 | aux_tail = ACCESS_ONCE(rb->user_page->aux_tail); | |
1a594131 | 306 | handle->wakeup = local_read(&rb->aux_wakeup) + rb->aux_watermark; |
2023a0d2 AS |
307 | if (aux_head - aux_tail < perf_aux_size(rb)) |
308 | handle->size = CIRC_SPACE(aux_head, aux_tail, perf_aux_size(rb)); | |
309 | ||
310 | /* | |
311 | * handle->size computation depends on aux_tail load; this forms a | |
312 | * control dependency barrier separating aux_tail load from aux data | |
313 | * store that will be enabled on successful return | |
314 | */ | |
315 | if (!handle->size) { /* A, matches D */ | |
316 | event->pending_disable = 1; | |
317 | perf_output_wakeup(handle); | |
318 | local_set(&rb->aux_nest, 0); | |
319 | goto err_put; | |
320 | } | |
fdc26706 AS |
321 | } |
322 | ||
323 | return handle->rb->aux_priv; | |
324 | ||
325 | err_put: | |
326 | rb_free_aux(rb); | |
327 | ||
328 | err: | |
95ff4ca2 | 329 | ring_buffer_put(rb); |
fdc26706 AS |
330 | handle->event = NULL; |
331 | ||
332 | return NULL; | |
333 | } | |
334 | ||
335 | /* | |
336 | * Commit the data written by hardware into the ring buffer by adjusting | |
337 | * aux_head and posting a PERF_RECORD_AUX into the perf buffer. It is the | |
338 | * pmu driver's responsibility to observe ordering rules of the hardware, | |
339 | * so that all the data is externally visible before this is called. | |
340 | */ | |
341 | void perf_aux_output_end(struct perf_output_handle *handle, unsigned long size, | |
342 | bool truncated) | |
343 | { | |
344 | struct ring_buffer *rb = handle->rb; | |
2023a0d2 | 345 | unsigned long aux_head; |
fdc26706 AS |
346 | u64 flags = 0; |
347 | ||
348 | if (truncated) | |
349 | flags |= PERF_AUX_FLAG_TRUNCATED; | |
350 | ||
2023a0d2 AS |
351 | /* in overwrite mode, driver provides aux_head via handle */ |
352 | if (rb->aux_overwrite) { | |
353 | flags |= PERF_AUX_FLAG_OVERWRITE; | |
354 | ||
355 | aux_head = handle->head; | |
356 | local_set(&rb->aux_head, aux_head); | |
357 | } else { | |
358 | aux_head = local_read(&rb->aux_head); | |
359 | local_add(size, &rb->aux_head); | |
360 | } | |
fdc26706 AS |
361 | |
362 | if (size || flags) { | |
363 | /* | |
364 | * Only send RECORD_AUX if we have something useful to communicate | |
365 | */ | |
366 | ||
367 | perf_event_aux_event(handle->event, aux_head, size, flags); | |
368 | } | |
369 | ||
1a594131 | 370 | aux_head = rb->user_page->aux_head = local_read(&rb->aux_head); |
fdc26706 | 371 | |
1a594131 AS |
372 | if (aux_head - local_read(&rb->aux_wakeup) >= rb->aux_watermark) { |
373 | perf_output_wakeup(handle); | |
374 | local_add(rb->aux_watermark, &rb->aux_wakeup); | |
375 | } | |
fdc26706 AS |
376 | handle->event = NULL; |
377 | ||
378 | local_set(&rb->aux_nest, 0); | |
379 | rb_free_aux(rb); | |
95ff4ca2 | 380 | ring_buffer_put(rb); |
fdc26706 AS |
381 | } |
382 | ||
383 | /* | |
384 | * Skip over a given number of bytes in the AUX buffer, due to, for example, | |
385 | * hardware's alignment constraints. | |
386 | */ | |
387 | int perf_aux_output_skip(struct perf_output_handle *handle, unsigned long size) | |
388 | { | |
389 | struct ring_buffer *rb = handle->rb; | |
390 | unsigned long aux_head; | |
391 | ||
392 | if (size > handle->size) | |
393 | return -ENOSPC; | |
394 | ||
395 | local_add(size, &rb->aux_head); | |
396 | ||
1a594131 AS |
397 | aux_head = rb->user_page->aux_head = local_read(&rb->aux_head); |
398 | if (aux_head - local_read(&rb->aux_wakeup) >= rb->aux_watermark) { | |
399 | perf_output_wakeup(handle); | |
400 | local_add(rb->aux_watermark, &rb->aux_wakeup); | |
401 | handle->wakeup = local_read(&rb->aux_wakeup) + | |
402 | rb->aux_watermark; | |
403 | } | |
404 | ||
fdc26706 AS |
405 | handle->head = aux_head; |
406 | handle->size -= size; | |
407 | ||
408 | return 0; | |
409 | } | |
410 | ||
411 | void *perf_get_aux(struct perf_output_handle *handle) | |
412 | { | |
413 | /* this is only valid between perf_aux_output_begin and *_end */ | |
414 | if (!handle->event) | |
415 | return NULL; | |
416 | ||
417 | return handle->rb->aux_priv; | |
418 | } | |
419 | ||
0a4e38e6 AS |
420 | #define PERF_AUX_GFP (GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_NORETRY) |
421 | ||
422 | static struct page *rb_alloc_aux_page(int node, int order) | |
423 | { | |
424 | struct page *page; | |
425 | ||
426 | if (order > MAX_ORDER) | |
427 | order = MAX_ORDER; | |
428 | ||
429 | do { | |
430 | page = alloc_pages_node(node, PERF_AUX_GFP, order); | |
431 | } while (!page && order--); | |
432 | ||
433 | if (page && order) { | |
434 | /* | |
c2ad6b51 AS |
435 | * Communicate the allocation size to the driver: |
436 | * if we managed to secure a high-order allocation, | |
437 | * set its first page's private to this order; | |
438 | * !PagePrivate(page) means it's just a normal page. | |
0a4e38e6 AS |
439 | */ |
440 | split_page(page, order); | |
441 | SetPagePrivate(page); | |
442 | set_page_private(page, order); | |
443 | } | |
444 | ||
445 | return page; | |
446 | } | |
447 | ||
448 | static void rb_free_aux_page(struct ring_buffer *rb, int idx) | |
449 | { | |
450 | struct page *page = virt_to_page(rb->aux_pages[idx]); | |
451 | ||
452 | ClearPagePrivate(page); | |
453 | page->mapping = NULL; | |
454 | __free_page(page); | |
455 | } | |
456 | ||
45c815f0 AS |
457 | static void __rb_free_aux(struct ring_buffer *rb) |
458 | { | |
459 | int pg; | |
460 | ||
95ff4ca2 AS |
461 | /* |
462 | * Should never happen, the last reference should be dropped from | |
463 | * perf_mmap_close() path, which first stops aux transactions (which | |
464 | * in turn are the atomic holders of aux_refcount) and then does the | |
465 | * last rb_free_aux(). | |
466 | */ | |
467 | WARN_ON_ONCE(in_atomic()); | |
468 | ||
45c815f0 AS |
469 | if (rb->aux_priv) { |
470 | rb->free_aux(rb->aux_priv); | |
471 | rb->free_aux = NULL; | |
472 | rb->aux_priv = NULL; | |
473 | } | |
474 | ||
475 | if (rb->aux_nr_pages) { | |
476 | for (pg = 0; pg < rb->aux_nr_pages; pg++) | |
477 | rb_free_aux_page(rb, pg); | |
478 | ||
479 | kfree(rb->aux_pages); | |
480 | rb->aux_nr_pages = 0; | |
481 | } | |
482 | } | |
483 | ||
45bfb2e5 | 484 | int rb_alloc_aux(struct ring_buffer *rb, struct perf_event *event, |
1a594131 | 485 | pgoff_t pgoff, int nr_pages, long watermark, int flags) |
45bfb2e5 PZ |
486 | { |
487 | bool overwrite = !(flags & RING_BUFFER_WRITABLE); | |
488 | int node = (event->cpu == -1) ? -1 : cpu_to_node(event->cpu); | |
0a4e38e6 | 489 | int ret = -ENOMEM, max_order = 0; |
45bfb2e5 PZ |
490 | |
491 | if (!has_aux(event)) | |
492 | return -ENOTSUPP; | |
493 | ||
6a279230 | 494 | if (event->pmu->capabilities & PERF_PMU_CAP_AUX_NO_SG) { |
0a4e38e6 AS |
495 | /* |
496 | * We need to start with the max_order that fits in nr_pages, | |
497 | * not the other way around, hence ilog2() and not get_order. | |
498 | */ | |
499 | max_order = ilog2(nr_pages); | |
500 | ||
6a279230 AS |
501 | /* |
502 | * PMU requests more than one contiguous chunks of memory | |
503 | * for SW double buffering | |
504 | */ | |
505 | if ((event->pmu->capabilities & PERF_PMU_CAP_AUX_SW_DOUBLEBUF) && | |
506 | !overwrite) { | |
507 | if (!max_order) | |
508 | return -EINVAL; | |
509 | ||
510 | max_order--; | |
511 | } | |
512 | } | |
513 | ||
45bfb2e5 PZ |
514 | rb->aux_pages = kzalloc_node(nr_pages * sizeof(void *), GFP_KERNEL, node); |
515 | if (!rb->aux_pages) | |
516 | return -ENOMEM; | |
517 | ||
518 | rb->free_aux = event->pmu->free_aux; | |
0a4e38e6 | 519 | for (rb->aux_nr_pages = 0; rb->aux_nr_pages < nr_pages;) { |
45bfb2e5 | 520 | struct page *page; |
0a4e38e6 | 521 | int last, order; |
45bfb2e5 | 522 | |
0a4e38e6 AS |
523 | order = min(max_order, ilog2(nr_pages - rb->aux_nr_pages)); |
524 | page = rb_alloc_aux_page(node, order); | |
45bfb2e5 PZ |
525 | if (!page) |
526 | goto out; | |
527 | ||
0a4e38e6 AS |
528 | for (last = rb->aux_nr_pages + (1 << page_private(page)); |
529 | last > rb->aux_nr_pages; rb->aux_nr_pages++) | |
530 | rb->aux_pages[rb->aux_nr_pages] = page_address(page++); | |
45bfb2e5 PZ |
531 | } |
532 | ||
aa319bcd AS |
533 | /* |
534 | * In overwrite mode, PMUs that don't support SG may not handle more | |
535 | * than one contiguous allocation, since they rely on PMI to do double | |
536 | * buffering. In this case, the entire buffer has to be one contiguous | |
537 | * chunk. | |
538 | */ | |
539 | if ((event->pmu->capabilities & PERF_PMU_CAP_AUX_NO_SG) && | |
540 | overwrite) { | |
541 | struct page *page = virt_to_page(rb->aux_pages[0]); | |
542 | ||
543 | if (page_private(page) != max_order) | |
544 | goto out; | |
545 | } | |
546 | ||
45bfb2e5 PZ |
547 | rb->aux_priv = event->pmu->setup_aux(event->cpu, rb->aux_pages, nr_pages, |
548 | overwrite); | |
549 | if (!rb->aux_priv) | |
550 | goto out; | |
551 | ||
552 | ret = 0; | |
553 | ||
554 | /* | |
555 | * aux_pages (and pmu driver's private data, aux_priv) will be | |
556 | * referenced in both producer's and consumer's contexts, thus | |
557 | * we keep a refcount here to make sure either of the two can | |
558 | * reference them safely. | |
559 | */ | |
560 | atomic_set(&rb->aux_refcount, 1); | |
561 | ||
2023a0d2 | 562 | rb->aux_overwrite = overwrite; |
1a594131 AS |
563 | rb->aux_watermark = watermark; |
564 | ||
565 | if (!rb->aux_watermark && !rb->aux_overwrite) | |
566 | rb->aux_watermark = nr_pages << (PAGE_SHIFT - 1); | |
2023a0d2 | 567 | |
45bfb2e5 PZ |
568 | out: |
569 | if (!ret) | |
570 | rb->aux_pgoff = pgoff; | |
571 | else | |
45c815f0 | 572 | __rb_free_aux(rb); |
45bfb2e5 PZ |
573 | |
574 | return ret; | |
575 | } | |
576 | ||
45bfb2e5 PZ |
577 | void rb_free_aux(struct ring_buffer *rb) |
578 | { | |
579 | if (atomic_dec_and_test(&rb->aux_refcount)) | |
580 | __rb_free_aux(rb); | |
581 | } | |
582 | ||
76369139 FW |
583 | #ifndef CONFIG_PERF_USE_VMALLOC |
584 | ||
585 | /* | |
586 | * Back perf_mmap() with regular GFP_KERNEL-0 pages. | |
587 | */ | |
588 | ||
45bfb2e5 PZ |
589 | static struct page * |
590 | __perf_mmap_to_page(struct ring_buffer *rb, unsigned long pgoff) | |
76369139 FW |
591 | { |
592 | if (pgoff > rb->nr_pages) | |
593 | return NULL; | |
594 | ||
595 | if (pgoff == 0) | |
596 | return virt_to_page(rb->user_page); | |
597 | ||
598 | return virt_to_page(rb->data_pages[pgoff - 1]); | |
599 | } | |
600 | ||
601 | static void *perf_mmap_alloc_page(int cpu) | |
602 | { | |
603 | struct page *page; | |
604 | int node; | |
605 | ||
606 | node = (cpu == -1) ? cpu : cpu_to_node(cpu); | |
607 | page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0); | |
608 | if (!page) | |
609 | return NULL; | |
610 | ||
611 | return page_address(page); | |
612 | } | |
613 | ||
614 | struct ring_buffer *rb_alloc(int nr_pages, long watermark, int cpu, int flags) | |
615 | { | |
616 | struct ring_buffer *rb; | |
617 | unsigned long size; | |
618 | int i; | |
619 | ||
620 | size = sizeof(struct ring_buffer); | |
621 | size += nr_pages * sizeof(void *); | |
622 | ||
623 | rb = kzalloc(size, GFP_KERNEL); | |
624 | if (!rb) | |
625 | goto fail; | |
626 | ||
627 | rb->user_page = perf_mmap_alloc_page(cpu); | |
628 | if (!rb->user_page) | |
629 | goto fail_user_page; | |
630 | ||
631 | for (i = 0; i < nr_pages; i++) { | |
632 | rb->data_pages[i] = perf_mmap_alloc_page(cpu); | |
633 | if (!rb->data_pages[i]) | |
634 | goto fail_data_pages; | |
635 | } | |
636 | ||
637 | rb->nr_pages = nr_pages; | |
638 | ||
639 | ring_buffer_init(rb, watermark, flags); | |
640 | ||
641 | return rb; | |
642 | ||
643 | fail_data_pages: | |
644 | for (i--; i >= 0; i--) | |
645 | free_page((unsigned long)rb->data_pages[i]); | |
646 | ||
647 | free_page((unsigned long)rb->user_page); | |
648 | ||
649 | fail_user_page: | |
650 | kfree(rb); | |
651 | ||
652 | fail: | |
653 | return NULL; | |
654 | } | |
655 | ||
656 | static void perf_mmap_free_page(unsigned long addr) | |
657 | { | |
658 | struct page *page = virt_to_page((void *)addr); | |
659 | ||
660 | page->mapping = NULL; | |
661 | __free_page(page); | |
662 | } | |
663 | ||
664 | void rb_free(struct ring_buffer *rb) | |
665 | { | |
666 | int i; | |
667 | ||
668 | perf_mmap_free_page((unsigned long)rb->user_page); | |
669 | for (i = 0; i < rb->nr_pages; i++) | |
670 | perf_mmap_free_page((unsigned long)rb->data_pages[i]); | |
671 | kfree(rb); | |
672 | } | |
673 | ||
674 | #else | |
5919b309 JO |
675 | static int data_page_nr(struct ring_buffer *rb) |
676 | { | |
677 | return rb->nr_pages << page_order(rb); | |
678 | } | |
76369139 | 679 | |
45bfb2e5 PZ |
680 | static struct page * |
681 | __perf_mmap_to_page(struct ring_buffer *rb, unsigned long pgoff) | |
76369139 | 682 | { |
5919b309 JO |
683 | /* The '>' counts in the user page. */ |
684 | if (pgoff > data_page_nr(rb)) | |
76369139 FW |
685 | return NULL; |
686 | ||
687 | return vmalloc_to_page((void *)rb->user_page + pgoff * PAGE_SIZE); | |
688 | } | |
689 | ||
690 | static void perf_mmap_unmark_page(void *addr) | |
691 | { | |
692 | struct page *page = vmalloc_to_page(addr); | |
693 | ||
694 | page->mapping = NULL; | |
695 | } | |
696 | ||
697 | static void rb_free_work(struct work_struct *work) | |
698 | { | |
699 | struct ring_buffer *rb; | |
700 | void *base; | |
701 | int i, nr; | |
702 | ||
703 | rb = container_of(work, struct ring_buffer, work); | |
5919b309 | 704 | nr = data_page_nr(rb); |
76369139 FW |
705 | |
706 | base = rb->user_page; | |
5919b309 JO |
707 | /* The '<=' counts in the user page. */ |
708 | for (i = 0; i <= nr; i++) | |
76369139 FW |
709 | perf_mmap_unmark_page(base + (i * PAGE_SIZE)); |
710 | ||
711 | vfree(base); | |
712 | kfree(rb); | |
713 | } | |
714 | ||
715 | void rb_free(struct ring_buffer *rb) | |
716 | { | |
717 | schedule_work(&rb->work); | |
718 | } | |
719 | ||
720 | struct ring_buffer *rb_alloc(int nr_pages, long watermark, int cpu, int flags) | |
721 | { | |
722 | struct ring_buffer *rb; | |
723 | unsigned long size; | |
724 | void *all_buf; | |
725 | ||
726 | size = sizeof(struct ring_buffer); | |
727 | size += sizeof(void *); | |
728 | ||
729 | rb = kzalloc(size, GFP_KERNEL); | |
730 | if (!rb) | |
731 | goto fail; | |
732 | ||
733 | INIT_WORK(&rb->work, rb_free_work); | |
734 | ||
735 | all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE); | |
736 | if (!all_buf) | |
737 | goto fail_all_buf; | |
738 | ||
739 | rb->user_page = all_buf; | |
740 | rb->data_pages[0] = all_buf + PAGE_SIZE; | |
8184059e PZ |
741 | if (nr_pages) { |
742 | rb->nr_pages = 1; | |
743 | rb->page_order = ilog2(nr_pages); | |
744 | } | |
76369139 FW |
745 | |
746 | ring_buffer_init(rb, watermark, flags); | |
747 | ||
748 | return rb; | |
749 | ||
750 | fail_all_buf: | |
751 | kfree(rb); | |
752 | ||
753 | fail: | |
754 | return NULL; | |
755 | } | |
756 | ||
757 | #endif | |
45bfb2e5 PZ |
758 | |
759 | struct page * | |
760 | perf_mmap_to_page(struct ring_buffer *rb, unsigned long pgoff) | |
761 | { | |
762 | if (rb->aux_nr_pages) { | |
763 | /* above AUX space */ | |
764 | if (pgoff > rb->aux_pgoff + rb->aux_nr_pages) | |
765 | return NULL; | |
766 | ||
767 | /* AUX space */ | |
768 | if (pgoff >= rb->aux_pgoff) | |
769 | return virt_to_page(rb->aux_pages[pgoff - rb->aux_pgoff]); | |
770 | } | |
771 | ||
772 | return __perf_mmap_to_page(rb, pgoff); | |
773 | } |