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// SPDX-License-Identifier: GPL-2.0
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
7 * Copyright (C) 2008-2011 Red Hat, Inc., Peter Zijlstra
8 * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
9 */
10
11#include <linux/perf_event.h>
12#include <linux/vmalloc.h>
13#include <linux/slab.h>
14#include <linux/circ_buf.h>
15#include <linux/poll.h>
16#include <linux/nospec.h>
17
18#include "internal.h"
19
20static void perf_output_wakeup(struct perf_output_handle *handle)
21{
22 atomic_set(&handle->rb->poll, EPOLLIN);
23
24 handle->event->pending_wakeup = 1;
25 irq_work_queue(&handle->event->pending);
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 */
36static void perf_output_get_handle(struct perf_output_handle *handle)
37{
38 struct perf_buffer *rb = handle->rb;
39
40 preempt_disable();
41
42 /*
43 * Avoid an explicit LOAD/STORE such that architectures with memops
44 * can use them.
45 */
46 (*(volatile unsigned int *)&rb->nest)++;
47 handle->wakeup = local_read(&rb->wakeup);
48}
49
50static void perf_output_put_handle(struct perf_output_handle *handle)
51{
52 struct perf_buffer *rb = handle->rb;
53 unsigned long head;
54 unsigned int nest;
55
56 /*
57 * If this isn't the outermost nesting, we don't have to update
58 * @rb->user_page->data_head.
59 */
60 nest = READ_ONCE(rb->nest);
61 if (nest > 1) {
62 WRITE_ONCE(rb->nest, nest - 1);
63 goto out;
64 }
65
66again:
67 /*
68 * In order to avoid publishing a head value that goes backwards,
69 * we must ensure the load of @rb->head happens after we've
70 * incremented @rb->nest.
71 *
72 * Otherwise we can observe a @rb->head value before one published
73 * by an IRQ/NMI happening between the load and the increment.
74 */
75 barrier();
76 head = local_read(&rb->head);
77
78 /*
79 * IRQ/NMI can happen here and advance @rb->head, causing our
80 * load above to be stale.
81 */
82
83 /*
84 * Since the mmap() consumer (userspace) can run on a different CPU:
85 *
86 * kernel user
87 *
88 * if (LOAD ->data_tail) { LOAD ->data_head
89 * (A) smp_rmb() (C)
90 * STORE $data LOAD $data
91 * smp_wmb() (B) smp_mb() (D)
92 * STORE ->data_head STORE ->data_tail
93 * }
94 *
95 * Where A pairs with D, and B pairs with C.
96 *
97 * In our case (A) is a control dependency that separates the load of
98 * the ->data_tail and the stores of $data. In case ->data_tail
99 * indicates there is no room in the buffer to store $data we do not.
100 *
101 * D needs to be a full barrier since it separates the data READ
102 * from the tail WRITE.
103 *
104 * For B a WMB is sufficient since it separates two WRITEs, and for C
105 * an RMB is sufficient since it separates two READs.
106 *
107 * See perf_output_begin().
108 */
109 smp_wmb(); /* B, matches C */
110 WRITE_ONCE(rb->user_page->data_head, head);
111
112 /*
113 * We must publish the head before decrementing the nest count,
114 * otherwise an IRQ/NMI can publish a more recent head value and our
115 * write will (temporarily) publish a stale value.
116 */
117 barrier();
118 WRITE_ONCE(rb->nest, 0);
119
120 /*
121 * Ensure we decrement @rb->nest before we validate the @rb->head.
122 * Otherwise we cannot be sure we caught the 'last' nested update.
123 */
124 barrier();
125 if (unlikely(head != local_read(&rb->head))) {
126 WRITE_ONCE(rb->nest, 1);
127 goto again;
128 }
129
130 if (handle->wakeup != local_read(&rb->wakeup))
131 perf_output_wakeup(handle);
132
133out:
134 preempt_enable();
135}
136
137static __always_inline bool
138ring_buffer_has_space(unsigned long head, unsigned long tail,
139 unsigned long data_size, unsigned int size,
140 bool backward)
141{
142 if (!backward)
143 return CIRC_SPACE(head, tail, data_size) >= size;
144 else
145 return CIRC_SPACE(tail, head, data_size) >= size;
146}
147
148static __always_inline int
149__perf_output_begin(struct perf_output_handle *handle,
150 struct perf_event *event, unsigned int size,
151 bool backward)
152{
153 struct perf_buffer *rb;
154 unsigned long tail, offset, head;
155 int have_lost, page_shift;
156 struct {
157 struct perf_event_header header;
158 u64 id;
159 u64 lost;
160 } lost_event;
161
162 rcu_read_lock();
163 /*
164 * For inherited events we send all the output towards the parent.
165 */
166 if (event->parent)
167 event = event->parent;
168
169 rb = rcu_dereference(event->rb);
170 if (unlikely(!rb))
171 goto out;
172
173 if (unlikely(rb->paused)) {
174 if (rb->nr_pages)
175 local_inc(&rb->lost);
176 goto out;
177 }
178
179 handle->rb = rb;
180 handle->event = event;
181
182 have_lost = local_read(&rb->lost);
183 if (unlikely(have_lost)) {
184 size += sizeof(lost_event);
185 if (event->attr.sample_id_all)
186 size += event->id_header_size;
187 }
188
189 perf_output_get_handle(handle);
190
191 do {
192 tail = READ_ONCE(rb->user_page->data_tail);
193 offset = head = local_read(&rb->head);
194 if (!rb->overwrite) {
195 if (unlikely(!ring_buffer_has_space(head, tail,
196 perf_data_size(rb),
197 size, backward)))
198 goto fail;
199 }
200
201 /*
202 * The above forms a control dependency barrier separating the
203 * @tail load above from the data stores below. Since the @tail
204 * load is required to compute the branch to fail below.
205 *
206 * A, matches D; the full memory barrier userspace SHOULD issue
207 * after reading the data and before storing the new tail
208 * position.
209 *
210 * See perf_output_put_handle().
211 */
212
213 if (!backward)
214 head += size;
215 else
216 head -= size;
217 } while (local_cmpxchg(&rb->head, offset, head) != offset);
218
219 if (backward) {
220 offset = head;
221 head = (u64)(-head);
222 }
223
224 /*
225 * We rely on the implied barrier() by local_cmpxchg() to ensure
226 * none of the data stores below can be lifted up by the compiler.
227 */
228
229 if (unlikely(head - local_read(&rb->wakeup) > rb->watermark))
230 local_add(rb->watermark, &rb->wakeup);
231
232 page_shift = PAGE_SHIFT + page_order(rb);
233
234 handle->page = (offset >> page_shift) & (rb->nr_pages - 1);
235 offset &= (1UL << page_shift) - 1;
236 handle->addr = rb->data_pages[handle->page] + offset;
237 handle->size = (1UL << page_shift) - offset;
238
239 if (unlikely(have_lost)) {
240 struct perf_sample_data sample_data;
241
242 lost_event.header.size = sizeof(lost_event);
243 lost_event.header.type = PERF_RECORD_LOST;
244 lost_event.header.misc = 0;
245 lost_event.id = event->id;
246 lost_event.lost = local_xchg(&rb->lost, 0);
247
248 perf_event_header__init_id(&lost_event.header,
249 &sample_data, event);
250 perf_output_put(handle, lost_event);
251 perf_event__output_id_sample(event, handle, &sample_data);
252 }
253
254 return 0;
255
256fail:
257 local_inc(&rb->lost);
258 perf_output_put_handle(handle);
259out:
260 rcu_read_unlock();
261
262 return -ENOSPC;
263}
264
265int perf_output_begin_forward(struct perf_output_handle *handle,
266 struct perf_event *event, unsigned int size)
267{
268 return __perf_output_begin(handle, event, size, false);
269}
270
271int perf_output_begin_backward(struct perf_output_handle *handle,
272 struct perf_event *event, unsigned int size)
273{
274 return __perf_output_begin(handle, event, size, true);
275}
276
277int perf_output_begin(struct perf_output_handle *handle,
278 struct perf_event *event, unsigned int size)
279{
280
281 return __perf_output_begin(handle, event, size,
282 unlikely(is_write_backward(event)));
283}
284
285unsigned int perf_output_copy(struct perf_output_handle *handle,
286 const void *buf, unsigned int len)
287{
288 return __output_copy(handle, buf, len);
289}
290
291unsigned int perf_output_skip(struct perf_output_handle *handle,
292 unsigned int len)
293{
294 return __output_skip(handle, NULL, len);
295}
296
297void perf_output_end(struct perf_output_handle *handle)
298{
299 perf_output_put_handle(handle);
300 rcu_read_unlock();
301}
302
303static void
304ring_buffer_init(struct perf_buffer *rb, long watermark, int flags)
305{
306 long max_size = perf_data_size(rb);
307
308 if (watermark)
309 rb->watermark = min(max_size, watermark);
310
311 if (!rb->watermark)
312 rb->watermark = max_size / 2;
313
314 if (flags & RING_BUFFER_WRITABLE)
315 rb->overwrite = 0;
316 else
317 rb->overwrite = 1;
318
319 refcount_set(&rb->refcount, 1);
320
321 INIT_LIST_HEAD(&rb->event_list);
322 spin_lock_init(&rb->event_lock);
323
324 /*
325 * perf_output_begin() only checks rb->paused, therefore
326 * rb->paused must be true if we have no pages for output.
327 */
328 if (!rb->nr_pages)
329 rb->paused = 1;
330}
331
332void perf_aux_output_flag(struct perf_output_handle *handle, u64 flags)
333{
334 /*
335 * OVERWRITE is determined by perf_aux_output_end() and can't
336 * be passed in directly.
337 */
338 if (WARN_ON_ONCE(flags & PERF_AUX_FLAG_OVERWRITE))
339 return;
340
341 handle->aux_flags |= flags;
342}
343EXPORT_SYMBOL_GPL(perf_aux_output_flag);
344
345/*
346 * This is called before hardware starts writing to the AUX area to
347 * obtain an output handle and make sure there's room in the buffer.
348 * When the capture completes, call perf_aux_output_end() to commit
349 * the recorded data to the buffer.
350 *
351 * The ordering is similar to that of perf_output_{begin,end}, with
352 * the exception of (B), which should be taken care of by the pmu
353 * driver, since ordering rules will differ depending on hardware.
354 *
355 * Call this from pmu::start(); see the comment in perf_aux_output_end()
356 * about its use in pmu callbacks. Both can also be called from the PMI
357 * handler if needed.
358 */
359void *perf_aux_output_begin(struct perf_output_handle *handle,
360 struct perf_event *event)
361{
362 struct perf_event *output_event = event;
363 unsigned long aux_head, aux_tail;
364 struct perf_buffer *rb;
365 unsigned int nest;
366
367 if (output_event->parent)
368 output_event = output_event->parent;
369
370 /*
371 * Since this will typically be open across pmu::add/pmu::del, we
372 * grab ring_buffer's refcount instead of holding rcu read lock
373 * to make sure it doesn't disappear under us.
374 */
375 rb = ring_buffer_get(output_event);
376 if (!rb)
377 return NULL;
378
379 if (!rb_has_aux(rb))
380 goto err;
381
382 /*
383 * If aux_mmap_count is zero, the aux buffer is in perf_mmap_close(),
384 * about to get freed, so we leave immediately.
385 *
386 * Checking rb::aux_mmap_count and rb::refcount has to be done in
387 * the same order, see perf_mmap_close. Otherwise we end up freeing
388 * aux pages in this path, which is a bug, because in_atomic().
389 */
390 if (!atomic_read(&rb->aux_mmap_count))
391 goto err;
392
393 if (!refcount_inc_not_zero(&rb->aux_refcount))
394 goto err;
395
396 nest = READ_ONCE(rb->aux_nest);
397 /*
398 * Nesting is not supported for AUX area, make sure nested
399 * writers are caught early
400 */
401 if (WARN_ON_ONCE(nest))
402 goto err_put;
403
404 WRITE_ONCE(rb->aux_nest, nest + 1);
405
406 aux_head = rb->aux_head;
407
408 handle->rb = rb;
409 handle->event = event;
410 handle->head = aux_head;
411 handle->size = 0;
412 handle->aux_flags = 0;
413
414 /*
415 * In overwrite mode, AUX data stores do not depend on aux_tail,
416 * therefore (A) control dependency barrier does not exist. The
417 * (B) <-> (C) ordering is still observed by the pmu driver.
418 */
419 if (!rb->aux_overwrite) {
420 aux_tail = READ_ONCE(rb->user_page->aux_tail);
421 handle->wakeup = rb->aux_wakeup + rb->aux_watermark;
422 if (aux_head - aux_tail < perf_aux_size(rb))
423 handle->size = CIRC_SPACE(aux_head, aux_tail, perf_aux_size(rb));
424
425 /*
426 * handle->size computation depends on aux_tail load; this forms a
427 * control dependency barrier separating aux_tail load from aux data
428 * store that will be enabled on successful return
429 */
430 if (!handle->size) { /* A, matches D */
431 event->pending_disable = smp_processor_id();
432 perf_output_wakeup(handle);
433 WRITE_ONCE(rb->aux_nest, 0);
434 goto err_put;
435 }
436 }
437
438 return handle->rb->aux_priv;
439
440err_put:
441 /* can't be last */
442 rb_free_aux(rb);
443
444err:
445 ring_buffer_put(rb);
446 handle->event = NULL;
447
448 return NULL;
449}
450EXPORT_SYMBOL_GPL(perf_aux_output_begin);
451
452static __always_inline bool rb_need_aux_wakeup(struct perf_buffer *rb)
453{
454 if (rb->aux_overwrite)
455 return false;
456
457 if (rb->aux_head - rb->aux_wakeup >= rb->aux_watermark) {
458 rb->aux_wakeup = rounddown(rb->aux_head, rb->aux_watermark);
459 return true;
460 }
461
462 return false;
463}
464
465/*
466 * Commit the data written by hardware into the ring buffer by adjusting
467 * aux_head and posting a PERF_RECORD_AUX into the perf buffer. It is the
468 * pmu driver's responsibility to observe ordering rules of the hardware,
469 * so that all the data is externally visible before this is called.
470 *
471 * Note: this has to be called from pmu::stop() callback, as the assumption
472 * of the AUX buffer management code is that after pmu::stop(), the AUX
473 * transaction must be stopped and therefore drop the AUX reference count.
474 */
475void perf_aux_output_end(struct perf_output_handle *handle, unsigned long size)
476{
477 bool wakeup = !!(handle->aux_flags & PERF_AUX_FLAG_TRUNCATED);
478 struct perf_buffer *rb = handle->rb;
479 unsigned long aux_head;
480
481 /* in overwrite mode, driver provides aux_head via handle */
482 if (rb->aux_overwrite) {
483 handle->aux_flags |= PERF_AUX_FLAG_OVERWRITE;
484
485 aux_head = handle->head;
486 rb->aux_head = aux_head;
487 } else {
488 handle->aux_flags &= ~PERF_AUX_FLAG_OVERWRITE;
489
490 aux_head = rb->aux_head;
491 rb->aux_head += size;
492 }
493
494 /*
495 * Only send RECORD_AUX if we have something useful to communicate
496 *
497 * Note: the OVERWRITE records by themselves are not considered
498 * useful, as they don't communicate any *new* information,
499 * aside from the short-lived offset, that becomes history at
500 * the next event sched-in and therefore isn't useful.
501 * The userspace that needs to copy out AUX data in overwrite
502 * mode should know to use user_page::aux_head for the actual
503 * offset. So, from now on we don't output AUX records that
504 * have *only* OVERWRITE flag set.
505 */
506 if (size || (handle->aux_flags & ~(u64)PERF_AUX_FLAG_OVERWRITE))
507 perf_event_aux_event(handle->event, aux_head, size,
508 handle->aux_flags);
509
510 WRITE_ONCE(rb->user_page->aux_head, rb->aux_head);
511 if (rb_need_aux_wakeup(rb))
512 wakeup = true;
513
514 if (wakeup) {
515 if (handle->aux_flags & PERF_AUX_FLAG_TRUNCATED)
516 handle->event->pending_disable = smp_processor_id();
517 perf_output_wakeup(handle);
518 }
519
520 handle->event = NULL;
521
522 WRITE_ONCE(rb->aux_nest, 0);
523 /* can't be last */
524 rb_free_aux(rb);
525 ring_buffer_put(rb);
526}
527EXPORT_SYMBOL_GPL(perf_aux_output_end);
528
529/*
530 * Skip over a given number of bytes in the AUX buffer, due to, for example,
531 * hardware's alignment constraints.
532 */
533int perf_aux_output_skip(struct perf_output_handle *handle, unsigned long size)
534{
535 struct perf_buffer *rb = handle->rb;
536
537 if (size > handle->size)
538 return -ENOSPC;
539
540 rb->aux_head += size;
541
542 WRITE_ONCE(rb->user_page->aux_head, rb->aux_head);
543 if (rb_need_aux_wakeup(rb)) {
544 perf_output_wakeup(handle);
545 handle->wakeup = rb->aux_wakeup + rb->aux_watermark;
546 }
547
548 handle->head = rb->aux_head;
549 handle->size -= size;
550
551 return 0;
552}
553EXPORT_SYMBOL_GPL(perf_aux_output_skip);
554
555void *perf_get_aux(struct perf_output_handle *handle)
556{
557 /* this is only valid between perf_aux_output_begin and *_end */
558 if (!handle->event)
559 return NULL;
560
561 return handle->rb->aux_priv;
562}
563EXPORT_SYMBOL_GPL(perf_get_aux);
564
565/*
566 * Copy out AUX data from an AUX handle.
567 */
568long perf_output_copy_aux(struct perf_output_handle *aux_handle,
569 struct perf_output_handle *handle,
570 unsigned long from, unsigned long to)
571{
572 struct perf_buffer *rb = aux_handle->rb;
573 unsigned long tocopy, remainder, len = 0;
574 void *addr;
575
576 from &= (rb->aux_nr_pages << PAGE_SHIFT) - 1;
577 to &= (rb->aux_nr_pages << PAGE_SHIFT) - 1;
578
579 do {
580 tocopy = PAGE_SIZE - offset_in_page(from);
581 if (to > from)
582 tocopy = min(tocopy, to - from);
583 if (!tocopy)
584 break;
585
586 addr = rb->aux_pages[from >> PAGE_SHIFT];
587 addr += offset_in_page(from);
588
589 remainder = perf_output_copy(handle, addr, tocopy);
590 if (remainder)
591 return -EFAULT;
592
593 len += tocopy;
594 from += tocopy;
595 from &= (rb->aux_nr_pages << PAGE_SHIFT) - 1;
596 } while (to != from);
597
598 return len;
599}
600
601#define PERF_AUX_GFP (GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_NORETRY)
602
603static struct page *rb_alloc_aux_page(int node, int order)
604{
605 struct page *page;
606
607 if (order > MAX_ORDER)
608 order = MAX_ORDER;
609
610 do {
611 page = alloc_pages_node(node, PERF_AUX_GFP, order);
612 } while (!page && order--);
613
614 if (page && order) {
615 /*
616 * Communicate the allocation size to the driver:
617 * if we managed to secure a high-order allocation,
618 * set its first page's private to this order;
619 * !PagePrivate(page) means it's just a normal page.
620 */
621 split_page(page, order);
622 SetPagePrivate(page);
623 set_page_private(page, order);
624 }
625
626 return page;
627}
628
629static void rb_free_aux_page(struct perf_buffer *rb, int idx)
630{
631 struct page *page = virt_to_page(rb->aux_pages[idx]);
632
633 ClearPagePrivate(page);
634 page->mapping = NULL;
635 __free_page(page);
636}
637
638static void __rb_free_aux(struct perf_buffer *rb)
639{
640 int pg;
641
642 /*
643 * Should never happen, the last reference should be dropped from
644 * perf_mmap_close() path, which first stops aux transactions (which
645 * in turn are the atomic holders of aux_refcount) and then does the
646 * last rb_free_aux().
647 */
648 WARN_ON_ONCE(in_atomic());
649
650 if (rb->aux_priv) {
651 rb->free_aux(rb->aux_priv);
652 rb->free_aux = NULL;
653 rb->aux_priv = NULL;
654 }
655
656 if (rb->aux_nr_pages) {
657 for (pg = 0; pg < rb->aux_nr_pages; pg++)
658 rb_free_aux_page(rb, pg);
659
660 kfree(rb->aux_pages);
661 rb->aux_nr_pages = 0;
662 }
663}
664
665int rb_alloc_aux(struct perf_buffer *rb, struct perf_event *event,
666 pgoff_t pgoff, int nr_pages, long watermark, int flags)
667{
668 bool overwrite = !(flags & RING_BUFFER_WRITABLE);
669 int node = (event->cpu == -1) ? -1 : cpu_to_node(event->cpu);
670 int ret = -ENOMEM, max_order;
671
672 if (!has_aux(event))
673 return -EOPNOTSUPP;
674
675 /*
676 * We need to start with the max_order that fits in nr_pages,
677 * not the other way around, hence ilog2() and not get_order.
678 */
679 max_order = ilog2(nr_pages);
680
681 /*
682 * PMU requests more than one contiguous chunks of memory
683 * for SW double buffering
684 */
685 if (!overwrite) {
686 if (!max_order)
687 return -EINVAL;
688
689 max_order--;
690 }
691
692 rb->aux_pages = kcalloc_node(nr_pages, sizeof(void *), GFP_KERNEL,
693 node);
694 if (!rb->aux_pages)
695 return -ENOMEM;
696
697 rb->free_aux = event->pmu->free_aux;
698 for (rb->aux_nr_pages = 0; rb->aux_nr_pages < nr_pages;) {
699 struct page *page;
700 int last, order;
701
702 order = min(max_order, ilog2(nr_pages - rb->aux_nr_pages));
703 page = rb_alloc_aux_page(node, order);
704 if (!page)
705 goto out;
706
707 for (last = rb->aux_nr_pages + (1 << page_private(page));
708 last > rb->aux_nr_pages; rb->aux_nr_pages++)
709 rb->aux_pages[rb->aux_nr_pages] = page_address(page++);
710 }
711
712 /*
713 * In overwrite mode, PMUs that don't support SG may not handle more
714 * than one contiguous allocation, since they rely on PMI to do double
715 * buffering. In this case, the entire buffer has to be one contiguous
716 * chunk.
717 */
718 if ((event->pmu->capabilities & PERF_PMU_CAP_AUX_NO_SG) &&
719 overwrite) {
720 struct page *page = virt_to_page(rb->aux_pages[0]);
721
722 if (page_private(page) != max_order)
723 goto out;
724 }
725
726 rb->aux_priv = event->pmu->setup_aux(event, rb->aux_pages, nr_pages,
727 overwrite);
728 if (!rb->aux_priv)
729 goto out;
730
731 ret = 0;
732
733 /*
734 * aux_pages (and pmu driver's private data, aux_priv) will be
735 * referenced in both producer's and consumer's contexts, thus
736 * we keep a refcount here to make sure either of the two can
737 * reference them safely.
738 */
739 refcount_set(&rb->aux_refcount, 1);
740
741 rb->aux_overwrite = overwrite;
742 rb->aux_watermark = watermark;
743
744 if (!rb->aux_watermark && !rb->aux_overwrite)
745 rb->aux_watermark = nr_pages << (PAGE_SHIFT - 1);
746
747out:
748 if (!ret)
749 rb->aux_pgoff = pgoff;
750 else
751 __rb_free_aux(rb);
752
753 return ret;
754}
755
756void rb_free_aux(struct perf_buffer *rb)
757{
758 if (refcount_dec_and_test(&rb->aux_refcount))
759 __rb_free_aux(rb);
760}
761
762#ifndef CONFIG_PERF_USE_VMALLOC
763
764/*
765 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
766 */
767
768static struct page *
769__perf_mmap_to_page(struct perf_buffer *rb, unsigned long pgoff)
770{
771 if (pgoff > rb->nr_pages)
772 return NULL;
773
774 if (pgoff == 0)
775 return virt_to_page(rb->user_page);
776
777 return virt_to_page(rb->data_pages[pgoff - 1]);
778}
779
780static void *perf_mmap_alloc_page(int cpu)
781{
782 struct page *page;
783 int node;
784
785 node = (cpu == -1) ? cpu : cpu_to_node(cpu);
786 page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
787 if (!page)
788 return NULL;
789
790 return page_address(page);
791}
792
793static void perf_mmap_free_page(void *addr)
794{
795 struct page *page = virt_to_page(addr);
796
797 page->mapping = NULL;
798 __free_page(page);
799}
800
801struct perf_buffer *rb_alloc(int nr_pages, long watermark, int cpu, int flags)
802{
803 struct perf_buffer *rb;
804 unsigned long size;
805 int i;
806
807 size = sizeof(struct perf_buffer);
808 size += nr_pages * sizeof(void *);
809
810 if (order_base_2(size) >= PAGE_SHIFT+MAX_ORDER)
811 goto fail;
812
813 rb = kzalloc(size, GFP_KERNEL);
814 if (!rb)
815 goto fail;
816
817 rb->user_page = perf_mmap_alloc_page(cpu);
818 if (!rb->user_page)
819 goto fail_user_page;
820
821 for (i = 0; i < nr_pages; i++) {
822 rb->data_pages[i] = perf_mmap_alloc_page(cpu);
823 if (!rb->data_pages[i])
824 goto fail_data_pages;
825 }
826
827 rb->nr_pages = nr_pages;
828
829 ring_buffer_init(rb, watermark, flags);
830
831 return rb;
832
833fail_data_pages:
834 for (i--; i >= 0; i--)
835 perf_mmap_free_page(rb->data_pages[i]);
836
837 perf_mmap_free_page(rb->user_page);
838
839fail_user_page:
840 kfree(rb);
841
842fail:
843 return NULL;
844}
845
846void rb_free(struct perf_buffer *rb)
847{
848 int i;
849
850 perf_mmap_free_page(rb->user_page);
851 for (i = 0; i < rb->nr_pages; i++)
852 perf_mmap_free_page(rb->data_pages[i]);
853 kfree(rb);
854}
855
856#else
857static int data_page_nr(struct perf_buffer *rb)
858{
859 return rb->nr_pages << page_order(rb);
860}
861
862static struct page *
863__perf_mmap_to_page(struct perf_buffer *rb, unsigned long pgoff)
864{
865 /* The '>' counts in the user page. */
866 if (pgoff > data_page_nr(rb))
867 return NULL;
868
869 return vmalloc_to_page((void *)rb->user_page + pgoff * PAGE_SIZE);
870}
871
872static void perf_mmap_unmark_page(void *addr)
873{
874 struct page *page = vmalloc_to_page(addr);
875
876 page->mapping = NULL;
877}
878
879static void rb_free_work(struct work_struct *work)
880{
881 struct perf_buffer *rb;
882 void *base;
883 int i, nr;
884
885 rb = container_of(work, struct perf_buffer, work);
886 nr = data_page_nr(rb);
887
888 base = rb->user_page;
889 /* The '<=' counts in the user page. */
890 for (i = 0; i <= nr; i++)
891 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
892
893 vfree(base);
894 kfree(rb);
895}
896
897void rb_free(struct perf_buffer *rb)
898{
899 schedule_work(&rb->work);
900}
901
902struct perf_buffer *rb_alloc(int nr_pages, long watermark, int cpu, int flags)
903{
904 struct perf_buffer *rb;
905 unsigned long size;
906 void *all_buf;
907
908 size = sizeof(struct perf_buffer);
909 size += sizeof(void *);
910
911 rb = kzalloc(size, GFP_KERNEL);
912 if (!rb)
913 goto fail;
914
915 INIT_WORK(&rb->work, rb_free_work);
916
917 all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
918 if (!all_buf)
919 goto fail_all_buf;
920
921 rb->user_page = all_buf;
922 rb->data_pages[0] = all_buf + PAGE_SIZE;
923 if (nr_pages) {
924 rb->nr_pages = 1;
925 rb->page_order = ilog2(nr_pages);
926 }
927
928 ring_buffer_init(rb, watermark, flags);
929
930 return rb;
931
932fail_all_buf:
933 kfree(rb);
934
935fail:
936 return NULL;
937}
938
939#endif
940
941struct page *
942perf_mmap_to_page(struct perf_buffer *rb, unsigned long pgoff)
943{
944 if (rb->aux_nr_pages) {
945 /* above AUX space */
946 if (pgoff > rb->aux_pgoff + rb->aux_nr_pages)
947 return NULL;
948
949 /* AUX space */
950 if (pgoff >= rb->aux_pgoff) {
951 int aux_pgoff = array_index_nospec(pgoff - rb->aux_pgoff, rb->aux_nr_pages);
952 return virt_to_page(rb->aux_pages[aux_pgoff]);
953 }
954 }
955
956 return __perf_mmap_to_page(rb, pgoff);
957}