Commit | Line | Data |
---|---|---|
8e86e015 | 1 | // SPDX-License-Identifier: GPL-2.0 |
76369139 FW |
2 | /* |
3 | * Performance events ring-buffer code: | |
4 | * | |
5 | * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de> | |
6 | * Copyright (C) 2008-2011 Red Hat, Inc., Ingo Molnar | |
90eec103 | 7 | * Copyright (C) 2008-2011 Red Hat, Inc., Peter Zijlstra |
d36b6910 | 8 | * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com> |
76369139 FW |
9 | */ |
10 | ||
11 | #include <linux/perf_event.h> | |
12 | #include <linux/vmalloc.h> | |
13 | #include <linux/slab.h> | |
26c86da8 | 14 | #include <linux/circ_buf.h> |
7c60fc0e | 15 | #include <linux/poll.h> |
4411ec1d | 16 | #include <linux/nospec.h> |
76369139 FW |
17 | |
18 | #include "internal.h" | |
19 | ||
76369139 FW |
20 | static void perf_output_wakeup(struct perf_output_handle *handle) |
21 | { | |
a9a08845 | 22 | atomic_set(&handle->rb->poll, EPOLLIN); |
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 | ||
57d6a793 | 105 | static __always_inline bool |
d1b26c70 WN |
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 | ||
57d6a793 | 116 | static __always_inline int |
d1b26c70 WN |
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 | 286 | |
fecb8ed2 | 287 | refcount_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 | ||
f4c0b0aa WD |
300 | void perf_aux_output_flag(struct perf_output_handle *handle, u64 flags) |
301 | { | |
302 | /* | |
303 | * OVERWRITE is determined by perf_aux_output_end() and can't | |
304 | * be passed in directly. | |
305 | */ | |
306 | if (WARN_ON_ONCE(flags & PERF_AUX_FLAG_OVERWRITE)) | |
307 | return; | |
308 | ||
309 | handle->aux_flags |= flags; | |
310 | } | |
311 | EXPORT_SYMBOL_GPL(perf_aux_output_flag); | |
312 | ||
fdc26706 AS |
313 | /* |
314 | * This is called before hardware starts writing to the AUX area to | |
315 | * obtain an output handle and make sure there's room in the buffer. | |
316 | * When the capture completes, call perf_aux_output_end() to commit | |
317 | * the recorded data to the buffer. | |
318 | * | |
319 | * The ordering is similar to that of perf_output_{begin,end}, with | |
320 | * the exception of (B), which should be taken care of by the pmu | |
321 | * driver, since ordering rules will differ depending on hardware. | |
af5bb4ed AS |
322 | * |
323 | * Call this from pmu::start(); see the comment in perf_aux_output_end() | |
324 | * about its use in pmu callbacks. Both can also be called from the PMI | |
325 | * handler if needed. | |
fdc26706 AS |
326 | */ |
327 | void *perf_aux_output_begin(struct perf_output_handle *handle, | |
328 | struct perf_event *event) | |
329 | { | |
330 | struct perf_event *output_event = event; | |
331 | unsigned long aux_head, aux_tail; | |
332 | struct ring_buffer *rb; | |
333 | ||
334 | if (output_event->parent) | |
335 | output_event = output_event->parent; | |
336 | ||
337 | /* | |
338 | * Since this will typically be open across pmu::add/pmu::del, we | |
339 | * grab ring_buffer's refcount instead of holding rcu read lock | |
340 | * to make sure it doesn't disappear under us. | |
341 | */ | |
342 | rb = ring_buffer_get(output_event); | |
343 | if (!rb) | |
344 | return NULL; | |
345 | ||
b79ccadd | 346 | if (!rb_has_aux(rb)) |
fdc26706 AS |
347 | goto err; |
348 | ||
dcb10a96 | 349 | /* |
b79ccadd AS |
350 | * If aux_mmap_count is zero, the aux buffer is in perf_mmap_close(), |
351 | * about to get freed, so we leave immediately. | |
352 | * | |
353 | * Checking rb::aux_mmap_count and rb::refcount has to be done in | |
354 | * the same order, see perf_mmap_close. Otherwise we end up freeing | |
355 | * aux pages in this path, which is a bug, because in_atomic(). | |
dcb10a96 AS |
356 | */ |
357 | if (!atomic_read(&rb->aux_mmap_count)) | |
b79ccadd AS |
358 | goto err; |
359 | ||
ca3bb3d0 | 360 | if (!refcount_inc_not_zero(&rb->aux_refcount)) |
b79ccadd | 361 | goto err; |
dcb10a96 | 362 | |
fdc26706 AS |
363 | /* |
364 | * Nesting is not supported for AUX area, make sure nested | |
365 | * writers are caught early | |
366 | */ | |
367 | if (WARN_ON_ONCE(local_xchg(&rb->aux_nest, 1))) | |
368 | goto err_put; | |
369 | ||
2ab346cf | 370 | aux_head = rb->aux_head; |
fdc26706 AS |
371 | |
372 | handle->rb = rb; | |
373 | handle->event = event; | |
374 | handle->head = aux_head; | |
2023a0d2 | 375 | handle->size = 0; |
f4c0b0aa | 376 | handle->aux_flags = 0; |
fdc26706 AS |
377 | |
378 | /* | |
2023a0d2 AS |
379 | * In overwrite mode, AUX data stores do not depend on aux_tail, |
380 | * therefore (A) control dependency barrier does not exist. The | |
381 | * (B) <-> (C) ordering is still observed by the pmu driver. | |
fdc26706 | 382 | */ |
2023a0d2 | 383 | if (!rb->aux_overwrite) { |
6aa7de05 | 384 | aux_tail = READ_ONCE(rb->user_page->aux_tail); |
2ab346cf | 385 | handle->wakeup = rb->aux_wakeup + rb->aux_watermark; |
2023a0d2 AS |
386 | if (aux_head - aux_tail < perf_aux_size(rb)) |
387 | handle->size = CIRC_SPACE(aux_head, aux_tail, perf_aux_size(rb)); | |
388 | ||
389 | /* | |
390 | * handle->size computation depends on aux_tail load; this forms a | |
391 | * control dependency barrier separating aux_tail load from aux data | |
392 | * store that will be enabled on successful return | |
393 | */ | |
394 | if (!handle->size) { /* A, matches D */ | |
395 | event->pending_disable = 1; | |
396 | perf_output_wakeup(handle); | |
397 | local_set(&rb->aux_nest, 0); | |
398 | goto err_put; | |
399 | } | |
fdc26706 AS |
400 | } |
401 | ||
402 | return handle->rb->aux_priv; | |
403 | ||
404 | err_put: | |
af5bb4ed | 405 | /* can't be last */ |
fdc26706 AS |
406 | rb_free_aux(rb); |
407 | ||
408 | err: | |
95ff4ca2 | 409 | ring_buffer_put(rb); |
fdc26706 AS |
410 | handle->event = NULL; |
411 | ||
412 | return NULL; | |
413 | } | |
bc1d2020 | 414 | EXPORT_SYMBOL_GPL(perf_aux_output_begin); |
fdc26706 | 415 | |
57d6a793 | 416 | static __always_inline bool rb_need_aux_wakeup(struct ring_buffer *rb) |
441430eb AS |
417 | { |
418 | if (rb->aux_overwrite) | |
419 | return false; | |
420 | ||
421 | if (rb->aux_head - rb->aux_wakeup >= rb->aux_watermark) { | |
422 | rb->aux_wakeup = rounddown(rb->aux_head, rb->aux_watermark); | |
423 | return true; | |
424 | } | |
425 | ||
426 | return false; | |
427 | } | |
428 | ||
fdc26706 AS |
429 | /* |
430 | * Commit the data written by hardware into the ring buffer by adjusting | |
431 | * aux_head and posting a PERF_RECORD_AUX into the perf buffer. It is the | |
432 | * pmu driver's responsibility to observe ordering rules of the hardware, | |
433 | * so that all the data is externally visible before this is called. | |
af5bb4ed AS |
434 | * |
435 | * Note: this has to be called from pmu::stop() callback, as the assumption | |
436 | * of the AUX buffer management code is that after pmu::stop(), the AUX | |
437 | * transaction must be stopped and therefore drop the AUX reference count. | |
fdc26706 | 438 | */ |
f4c0b0aa | 439 | void perf_aux_output_end(struct perf_output_handle *handle, unsigned long size) |
fdc26706 | 440 | { |
ae0c2d99 | 441 | bool wakeup = !!(handle->aux_flags & PERF_AUX_FLAG_TRUNCATED); |
fdc26706 | 442 | struct ring_buffer *rb = handle->rb; |
2023a0d2 | 443 | unsigned long aux_head; |
fdc26706 | 444 | |
2023a0d2 AS |
445 | /* in overwrite mode, driver provides aux_head via handle */ |
446 | if (rb->aux_overwrite) { | |
f4c0b0aa | 447 | handle->aux_flags |= PERF_AUX_FLAG_OVERWRITE; |
2023a0d2 AS |
448 | |
449 | aux_head = handle->head; | |
2ab346cf | 450 | rb->aux_head = aux_head; |
2023a0d2 | 451 | } else { |
f4c0b0aa WD |
452 | handle->aux_flags &= ~PERF_AUX_FLAG_OVERWRITE; |
453 | ||
2ab346cf WD |
454 | aux_head = rb->aux_head; |
455 | rb->aux_head += size; | |
2023a0d2 | 456 | } |
fdc26706 | 457 | |
f4c0b0aa | 458 | if (size || handle->aux_flags) { |
fdc26706 AS |
459 | /* |
460 | * Only send RECORD_AUX if we have something useful to communicate | |
1627314f AS |
461 | * |
462 | * Note: the OVERWRITE records by themselves are not considered | |
463 | * useful, as they don't communicate any *new* information, | |
464 | * aside from the short-lived offset, that becomes history at | |
465 | * the next event sched-in and therefore isn't useful. | |
466 | * The userspace that needs to copy out AUX data in overwrite | |
467 | * mode should know to use user_page::aux_head for the actual | |
468 | * offset. So, from now on we don't output AUX records that | |
469 | * have *only* OVERWRITE flag set. | |
fdc26706 AS |
470 | */ |
471 | ||
1627314f AS |
472 | if (handle->aux_flags & ~(u64)PERF_AUX_FLAG_OVERWRITE) |
473 | perf_event_aux_event(handle->event, aux_head, size, | |
474 | handle->aux_flags); | |
fdc26706 AS |
475 | } |
476 | ||
2ab346cf | 477 | rb->user_page->aux_head = rb->aux_head; |
441430eb | 478 | if (rb_need_aux_wakeup(rb)) |
3f56e687 | 479 | wakeup = true; |
3f56e687 AS |
480 | |
481 | if (wakeup) { | |
f4c0b0aa | 482 | if (handle->aux_flags & PERF_AUX_FLAG_TRUNCATED) |
3f56e687 AS |
483 | handle->event->pending_disable = 1; |
484 | perf_output_wakeup(handle); | |
485 | } | |
486 | ||
fdc26706 AS |
487 | handle->event = NULL; |
488 | ||
489 | local_set(&rb->aux_nest, 0); | |
af5bb4ed | 490 | /* can't be last */ |
fdc26706 | 491 | rb_free_aux(rb); |
95ff4ca2 | 492 | ring_buffer_put(rb); |
fdc26706 | 493 | } |
bc1d2020 | 494 | EXPORT_SYMBOL_GPL(perf_aux_output_end); |
fdc26706 AS |
495 | |
496 | /* | |
497 | * Skip over a given number of bytes in the AUX buffer, due to, for example, | |
498 | * hardware's alignment constraints. | |
499 | */ | |
500 | int perf_aux_output_skip(struct perf_output_handle *handle, unsigned long size) | |
501 | { | |
502 | struct ring_buffer *rb = handle->rb; | |
fdc26706 AS |
503 | |
504 | if (size > handle->size) | |
505 | return -ENOSPC; | |
506 | ||
2ab346cf | 507 | rb->aux_head += size; |
fdc26706 | 508 | |
2ab346cf | 509 | rb->user_page->aux_head = rb->aux_head; |
441430eb | 510 | if (rb_need_aux_wakeup(rb)) { |
1a594131 | 511 | perf_output_wakeup(handle); |
2ab346cf | 512 | handle->wakeup = rb->aux_wakeup + rb->aux_watermark; |
1a594131 AS |
513 | } |
514 | ||
2ab346cf | 515 | handle->head = rb->aux_head; |
fdc26706 AS |
516 | handle->size -= size; |
517 | ||
518 | return 0; | |
519 | } | |
bc1d2020 | 520 | EXPORT_SYMBOL_GPL(perf_aux_output_skip); |
fdc26706 AS |
521 | |
522 | void *perf_get_aux(struct perf_output_handle *handle) | |
523 | { | |
524 | /* this is only valid between perf_aux_output_begin and *_end */ | |
525 | if (!handle->event) | |
526 | return NULL; | |
527 | ||
528 | return handle->rb->aux_priv; | |
529 | } | |
bc1d2020 | 530 | EXPORT_SYMBOL_GPL(perf_get_aux); |
fdc26706 | 531 | |
0a4e38e6 AS |
532 | #define PERF_AUX_GFP (GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_NORETRY) |
533 | ||
534 | static struct page *rb_alloc_aux_page(int node, int order) | |
535 | { | |
536 | struct page *page; | |
537 | ||
538 | if (order > MAX_ORDER) | |
539 | order = MAX_ORDER; | |
540 | ||
541 | do { | |
542 | page = alloc_pages_node(node, PERF_AUX_GFP, order); | |
543 | } while (!page && order--); | |
544 | ||
545 | if (page && order) { | |
546 | /* | |
c2ad6b51 AS |
547 | * Communicate the allocation size to the driver: |
548 | * if we managed to secure a high-order allocation, | |
549 | * set its first page's private to this order; | |
550 | * !PagePrivate(page) means it's just a normal page. | |
0a4e38e6 AS |
551 | */ |
552 | split_page(page, order); | |
553 | SetPagePrivate(page); | |
554 | set_page_private(page, order); | |
555 | } | |
556 | ||
557 | return page; | |
558 | } | |
559 | ||
560 | static void rb_free_aux_page(struct ring_buffer *rb, int idx) | |
561 | { | |
562 | struct page *page = virt_to_page(rb->aux_pages[idx]); | |
563 | ||
564 | ClearPagePrivate(page); | |
565 | page->mapping = NULL; | |
566 | __free_page(page); | |
567 | } | |
568 | ||
45c815f0 AS |
569 | static void __rb_free_aux(struct ring_buffer *rb) |
570 | { | |
571 | int pg; | |
572 | ||
95ff4ca2 AS |
573 | /* |
574 | * Should never happen, the last reference should be dropped from | |
575 | * perf_mmap_close() path, which first stops aux transactions (which | |
576 | * in turn are the atomic holders of aux_refcount) and then does the | |
577 | * last rb_free_aux(). | |
578 | */ | |
579 | WARN_ON_ONCE(in_atomic()); | |
580 | ||
45c815f0 AS |
581 | if (rb->aux_priv) { |
582 | rb->free_aux(rb->aux_priv); | |
583 | rb->free_aux = NULL; | |
584 | rb->aux_priv = NULL; | |
585 | } | |
586 | ||
587 | if (rb->aux_nr_pages) { | |
588 | for (pg = 0; pg < rb->aux_nr_pages; pg++) | |
589 | rb_free_aux_page(rb, pg); | |
590 | ||
591 | kfree(rb->aux_pages); | |
592 | rb->aux_nr_pages = 0; | |
593 | } | |
594 | } | |
595 | ||
45bfb2e5 | 596 | int rb_alloc_aux(struct ring_buffer *rb, struct perf_event *event, |
1a594131 | 597 | pgoff_t pgoff, int nr_pages, long watermark, int flags) |
45bfb2e5 PZ |
598 | { |
599 | bool overwrite = !(flags & RING_BUFFER_WRITABLE); | |
600 | int node = (event->cpu == -1) ? -1 : cpu_to_node(event->cpu); | |
0a4e38e6 | 601 | int ret = -ENOMEM, max_order = 0; |
45bfb2e5 PZ |
602 | |
603 | if (!has_aux(event)) | |
8a1898db | 604 | return -EOPNOTSUPP; |
45bfb2e5 | 605 | |
6a279230 | 606 | if (event->pmu->capabilities & PERF_PMU_CAP_AUX_NO_SG) { |
0a4e38e6 AS |
607 | /* |
608 | * We need to start with the max_order that fits in nr_pages, | |
609 | * not the other way around, hence ilog2() and not get_order. | |
610 | */ | |
611 | max_order = ilog2(nr_pages); | |
612 | ||
6a279230 AS |
613 | /* |
614 | * PMU requests more than one contiguous chunks of memory | |
615 | * for SW double buffering | |
616 | */ | |
617 | if ((event->pmu->capabilities & PERF_PMU_CAP_AUX_SW_DOUBLEBUF) && | |
618 | !overwrite) { | |
619 | if (!max_order) | |
620 | return -EINVAL; | |
621 | ||
622 | max_order--; | |
623 | } | |
624 | } | |
625 | ||
590b5b7d KC |
626 | rb->aux_pages = kcalloc_node(nr_pages, sizeof(void *), GFP_KERNEL, |
627 | node); | |
45bfb2e5 PZ |
628 | if (!rb->aux_pages) |
629 | return -ENOMEM; | |
630 | ||
631 | rb->free_aux = event->pmu->free_aux; | |
0a4e38e6 | 632 | for (rb->aux_nr_pages = 0; rb->aux_nr_pages < nr_pages;) { |
45bfb2e5 | 633 | struct page *page; |
0a4e38e6 | 634 | int last, order; |
45bfb2e5 | 635 | |
0a4e38e6 AS |
636 | order = min(max_order, ilog2(nr_pages - rb->aux_nr_pages)); |
637 | page = rb_alloc_aux_page(node, order); | |
45bfb2e5 PZ |
638 | if (!page) |
639 | goto out; | |
640 | ||
0a4e38e6 AS |
641 | for (last = rb->aux_nr_pages + (1 << page_private(page)); |
642 | last > rb->aux_nr_pages; rb->aux_nr_pages++) | |
643 | rb->aux_pages[rb->aux_nr_pages] = page_address(page++); | |
45bfb2e5 PZ |
644 | } |
645 | ||
aa319bcd AS |
646 | /* |
647 | * In overwrite mode, PMUs that don't support SG may not handle more | |
648 | * than one contiguous allocation, since they rely on PMI to do double | |
649 | * buffering. In this case, the entire buffer has to be one contiguous | |
650 | * chunk. | |
651 | */ | |
652 | if ((event->pmu->capabilities & PERF_PMU_CAP_AUX_NO_SG) && | |
653 | overwrite) { | |
654 | struct page *page = virt_to_page(rb->aux_pages[0]); | |
655 | ||
656 | if (page_private(page) != max_order) | |
657 | goto out; | |
658 | } | |
659 | ||
84001866 | 660 | rb->aux_priv = event->pmu->setup_aux(event, rb->aux_pages, nr_pages, |
45bfb2e5 PZ |
661 | overwrite); |
662 | if (!rb->aux_priv) | |
663 | goto out; | |
664 | ||
665 | ret = 0; | |
666 | ||
667 | /* | |
668 | * aux_pages (and pmu driver's private data, aux_priv) will be | |
669 | * referenced in both producer's and consumer's contexts, thus | |
670 | * we keep a refcount here to make sure either of the two can | |
671 | * reference them safely. | |
672 | */ | |
ca3bb3d0 | 673 | refcount_set(&rb->aux_refcount, 1); |
45bfb2e5 | 674 | |
2023a0d2 | 675 | rb->aux_overwrite = overwrite; |
1a594131 AS |
676 | rb->aux_watermark = watermark; |
677 | ||
678 | if (!rb->aux_watermark && !rb->aux_overwrite) | |
679 | rb->aux_watermark = nr_pages << (PAGE_SHIFT - 1); | |
2023a0d2 | 680 | |
45bfb2e5 PZ |
681 | out: |
682 | if (!ret) | |
683 | rb->aux_pgoff = pgoff; | |
684 | else | |
45c815f0 | 685 | __rb_free_aux(rb); |
45bfb2e5 PZ |
686 | |
687 | return ret; | |
688 | } | |
689 | ||
45bfb2e5 PZ |
690 | void rb_free_aux(struct ring_buffer *rb) |
691 | { | |
ca3bb3d0 | 692 | if (refcount_dec_and_test(&rb->aux_refcount)) |
45bfb2e5 PZ |
693 | __rb_free_aux(rb); |
694 | } | |
695 | ||
76369139 FW |
696 | #ifndef CONFIG_PERF_USE_VMALLOC |
697 | ||
698 | /* | |
699 | * Back perf_mmap() with regular GFP_KERNEL-0 pages. | |
700 | */ | |
701 | ||
45bfb2e5 PZ |
702 | static struct page * |
703 | __perf_mmap_to_page(struct ring_buffer *rb, unsigned long pgoff) | |
76369139 FW |
704 | { |
705 | if (pgoff > rb->nr_pages) | |
706 | return NULL; | |
707 | ||
708 | if (pgoff == 0) | |
709 | return virt_to_page(rb->user_page); | |
710 | ||
711 | return virt_to_page(rb->data_pages[pgoff - 1]); | |
712 | } | |
713 | ||
714 | static void *perf_mmap_alloc_page(int cpu) | |
715 | { | |
716 | struct page *page; | |
717 | int node; | |
718 | ||
719 | node = (cpu == -1) ? cpu : cpu_to_node(cpu); | |
720 | page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0); | |
721 | if (!page) | |
722 | return NULL; | |
723 | ||
724 | return page_address(page); | |
725 | } | |
726 | ||
727 | struct ring_buffer *rb_alloc(int nr_pages, long watermark, int cpu, int flags) | |
728 | { | |
729 | struct ring_buffer *rb; | |
730 | unsigned long size; | |
731 | int i; | |
732 | ||
733 | size = sizeof(struct ring_buffer); | |
734 | size += nr_pages * sizeof(void *); | |
735 | ||
528871b4 | 736 | if (order_base_2(size) >= PAGE_SHIFT+MAX_ORDER) |
9dff0aa9 MR |
737 | goto fail; |
738 | ||
76369139 FW |
739 | rb = kzalloc(size, GFP_KERNEL); |
740 | if (!rb) | |
741 | goto fail; | |
742 | ||
743 | rb->user_page = perf_mmap_alloc_page(cpu); | |
744 | if (!rb->user_page) | |
745 | goto fail_user_page; | |
746 | ||
747 | for (i = 0; i < nr_pages; i++) { | |
748 | rb->data_pages[i] = perf_mmap_alloc_page(cpu); | |
749 | if (!rb->data_pages[i]) | |
750 | goto fail_data_pages; | |
751 | } | |
752 | ||
753 | rb->nr_pages = nr_pages; | |
754 | ||
755 | ring_buffer_init(rb, watermark, flags); | |
756 | ||
757 | return rb; | |
758 | ||
759 | fail_data_pages: | |
760 | for (i--; i >= 0; i--) | |
761 | free_page((unsigned long)rb->data_pages[i]); | |
762 | ||
763 | free_page((unsigned long)rb->user_page); | |
764 | ||
765 | fail_user_page: | |
766 | kfree(rb); | |
767 | ||
768 | fail: | |
769 | return NULL; | |
770 | } | |
771 | ||
772 | static void perf_mmap_free_page(unsigned long addr) | |
773 | { | |
774 | struct page *page = virt_to_page((void *)addr); | |
775 | ||
776 | page->mapping = NULL; | |
777 | __free_page(page); | |
778 | } | |
779 | ||
780 | void rb_free(struct ring_buffer *rb) | |
781 | { | |
782 | int i; | |
783 | ||
784 | perf_mmap_free_page((unsigned long)rb->user_page); | |
785 | for (i = 0; i < rb->nr_pages; i++) | |
786 | perf_mmap_free_page((unsigned long)rb->data_pages[i]); | |
787 | kfree(rb); | |
788 | } | |
789 | ||
790 | #else | |
5919b309 JO |
791 | static int data_page_nr(struct ring_buffer *rb) |
792 | { | |
793 | return rb->nr_pages << page_order(rb); | |
794 | } | |
76369139 | 795 | |
45bfb2e5 PZ |
796 | static struct page * |
797 | __perf_mmap_to_page(struct ring_buffer *rb, unsigned long pgoff) | |
76369139 | 798 | { |
5919b309 JO |
799 | /* The '>' counts in the user page. */ |
800 | if (pgoff > data_page_nr(rb)) | |
76369139 FW |
801 | return NULL; |
802 | ||
803 | return vmalloc_to_page((void *)rb->user_page + pgoff * PAGE_SIZE); | |
804 | } | |
805 | ||
806 | static void perf_mmap_unmark_page(void *addr) | |
807 | { | |
808 | struct page *page = vmalloc_to_page(addr); | |
809 | ||
810 | page->mapping = NULL; | |
811 | } | |
812 | ||
813 | static void rb_free_work(struct work_struct *work) | |
814 | { | |
815 | struct ring_buffer *rb; | |
816 | void *base; | |
817 | int i, nr; | |
818 | ||
819 | rb = container_of(work, struct ring_buffer, work); | |
5919b309 | 820 | nr = data_page_nr(rb); |
76369139 FW |
821 | |
822 | base = rb->user_page; | |
5919b309 JO |
823 | /* The '<=' counts in the user page. */ |
824 | for (i = 0; i <= nr; i++) | |
76369139 FW |
825 | perf_mmap_unmark_page(base + (i * PAGE_SIZE)); |
826 | ||
827 | vfree(base); | |
828 | kfree(rb); | |
829 | } | |
830 | ||
831 | void rb_free(struct ring_buffer *rb) | |
832 | { | |
833 | schedule_work(&rb->work); | |
834 | } | |
835 | ||
836 | struct ring_buffer *rb_alloc(int nr_pages, long watermark, int cpu, int flags) | |
837 | { | |
838 | struct ring_buffer *rb; | |
839 | unsigned long size; | |
840 | void *all_buf; | |
841 | ||
842 | size = sizeof(struct ring_buffer); | |
843 | size += sizeof(void *); | |
844 | ||
845 | rb = kzalloc(size, GFP_KERNEL); | |
846 | if (!rb) | |
847 | goto fail; | |
848 | ||
849 | INIT_WORK(&rb->work, rb_free_work); | |
850 | ||
851 | all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE); | |
852 | if (!all_buf) | |
853 | goto fail_all_buf; | |
854 | ||
855 | rb->user_page = all_buf; | |
856 | rb->data_pages[0] = all_buf + PAGE_SIZE; | |
8184059e PZ |
857 | if (nr_pages) { |
858 | rb->nr_pages = 1; | |
859 | rb->page_order = ilog2(nr_pages); | |
860 | } | |
76369139 FW |
861 | |
862 | ring_buffer_init(rb, watermark, flags); | |
863 | ||
864 | return rb; | |
865 | ||
866 | fail_all_buf: | |
867 | kfree(rb); | |
868 | ||
869 | fail: | |
870 | return NULL; | |
871 | } | |
872 | ||
873 | #endif | |
45bfb2e5 PZ |
874 | |
875 | struct page * | |
876 | perf_mmap_to_page(struct ring_buffer *rb, unsigned long pgoff) | |
877 | { | |
878 | if (rb->aux_nr_pages) { | |
879 | /* above AUX space */ | |
880 | if (pgoff > rb->aux_pgoff + rb->aux_nr_pages) | |
881 | return NULL; | |
882 | ||
883 | /* AUX space */ | |
4411ec1d PZ |
884 | if (pgoff >= rb->aux_pgoff) { |
885 | int aux_pgoff = array_index_nospec(pgoff - rb->aux_pgoff, rb->aux_nr_pages); | |
886 | return virt_to_page(rb->aux_pages[aux_pgoff]); | |
887 | } | |
45bfb2e5 PZ |
888 | } |
889 | ||
890 | return __perf_mmap_to_page(rb, pgoff); | |
891 | } |