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_irq);
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_sample_data *data,
151 struct perf_event *event, unsigned int size,
152 bool backward)
153{
154 struct perf_buffer *rb;
155 unsigned long tail, offset, head;
156 int have_lost, page_shift;
157 struct {
158 struct perf_event_header header;
159 u64 id;
160 u64 lost;
161 } lost_event;
162
163 rcu_read_lock();
164 /*
165 * For inherited events we send all the output towards the parent.
166 */
167 if (event->parent)
168 event = event->parent;
169
170 rb = rcu_dereference(event->rb);
171 if (unlikely(!rb))
172 goto out;
173
174 if (unlikely(rb->paused)) {
175 if (rb->nr_pages) {
176 local_inc(&rb->lost);
177 atomic64_inc(&event->lost_samples);
178 }
179 goto out;
180 }
181
182 handle->rb = rb;
183 handle->event = event;
184
185 have_lost = local_read(&rb->lost);
186 if (unlikely(have_lost)) {
187 size += sizeof(lost_event);
188 if (event->attr.sample_id_all)
189 size += event->id_header_size;
190 }
191
192 perf_output_get_handle(handle);
193
194 offset = local_read(&rb->head);
195 do {
196 head = offset;
197 tail = READ_ONCE(rb->user_page->data_tail);
198 if (!rb->overwrite) {
199 if (unlikely(!ring_buffer_has_space(head, tail,
200 perf_data_size(rb),
201 size, backward)))
202 goto fail;
203 }
204
205 /*
206 * The above forms a control dependency barrier separating the
207 * @tail load above from the data stores below. Since the @tail
208 * load is required to compute the branch to fail below.
209 *
210 * A, matches D; the full memory barrier userspace SHOULD issue
211 * after reading the data and before storing the new tail
212 * position.
213 *
214 * See perf_output_put_handle().
215 */
216
217 if (!backward)
218 head += size;
219 else
220 head -= size;
221 } while (!local_try_cmpxchg(&rb->head, &offset, head));
222
223 if (backward) {
224 offset = head;
225 head = (u64)(-head);
226 }
227
228 /*
229 * We rely on the implied barrier() by local_cmpxchg() to ensure
230 * none of the data stores below can be lifted up by the compiler.
231 */
232
233 if (unlikely(head - local_read(&rb->wakeup) > rb->watermark))
234 local_add(rb->watermark, &rb->wakeup);
235
236 page_shift = PAGE_SHIFT + page_order(rb);
237
238 handle->page = (offset >> page_shift) & (rb->nr_pages - 1);
239 offset &= (1UL << page_shift) - 1;
240 handle->addr = rb->data_pages[handle->page] + offset;
241 handle->size = (1UL << page_shift) - offset;
242
243 if (unlikely(have_lost)) {
244 lost_event.header.size = sizeof(lost_event);
245 lost_event.header.type = PERF_RECORD_LOST;
246 lost_event.header.misc = 0;
247 lost_event.id = event->id;
248 lost_event.lost = local_xchg(&rb->lost, 0);
249
250 /* XXX mostly redundant; @data is already fully initializes */
251 perf_event_header__init_id(&lost_event.header, data, event);
252 perf_output_put(handle, lost_event);
253 perf_event__output_id_sample(event, handle, data);
254 }
255
256 return 0;
257
258fail:
259 local_inc(&rb->lost);
260 atomic64_inc(&event->lost_samples);
261 perf_output_put_handle(handle);
262out:
263 rcu_read_unlock();
264
265 return -ENOSPC;
266}
267
268int perf_output_begin_forward(struct perf_output_handle *handle,
269 struct perf_sample_data *data,
270 struct perf_event *event, unsigned int size)
271{
272 return __perf_output_begin(handle, data, event, size, false);
273}
274
275int perf_output_begin_backward(struct perf_output_handle *handle,
276 struct perf_sample_data *data,
277 struct perf_event *event, unsigned int size)
278{
279 return __perf_output_begin(handle, data, event, size, true);
280}
281
282int perf_output_begin(struct perf_output_handle *handle,
283 struct perf_sample_data *data,
284 struct perf_event *event, unsigned int size)
285{
286
287 return __perf_output_begin(handle, data, event, size,
288 unlikely(is_write_backward(event)));
289}
290
291unsigned int perf_output_copy(struct perf_output_handle *handle,
292 const void *buf, unsigned int len)
293{
294 return __output_copy(handle, buf, len);
295}
296
297unsigned int perf_output_skip(struct perf_output_handle *handle,
298 unsigned int len)
299{
300 return __output_skip(handle, NULL, len);
301}
302
303void perf_output_end(struct perf_output_handle *handle)
304{
305 perf_output_put_handle(handle);
306 rcu_read_unlock();
307}
308
309static void
310ring_buffer_init(struct perf_buffer *rb, long watermark, int flags)
311{
312 long max_size = perf_data_size(rb);
313
314 if (watermark)
315 rb->watermark = min(max_size, watermark);
316
317 if (!rb->watermark)
318 rb->watermark = max_size / 2;
319
320 if (flags & RING_BUFFER_WRITABLE)
321 rb->overwrite = 0;
322 else
323 rb->overwrite = 1;
324
325 refcount_set(&rb->refcount, 1);
326
327 INIT_LIST_HEAD(&rb->event_list);
328 spin_lock_init(&rb->event_lock);
329
330 /*
331 * perf_output_begin() only checks rb->paused, therefore
332 * rb->paused must be true if we have no pages for output.
333 */
334 if (!rb->nr_pages)
335 rb->paused = 1;
336}
337
338void perf_aux_output_flag(struct perf_output_handle *handle, u64 flags)
339{
340 /*
341 * OVERWRITE is determined by perf_aux_output_end() and can't
342 * be passed in directly.
343 */
344 if (WARN_ON_ONCE(flags & PERF_AUX_FLAG_OVERWRITE))
345 return;
346
347 handle->aux_flags |= flags;
348}
349EXPORT_SYMBOL_GPL(perf_aux_output_flag);
350
351/*
352 * This is called before hardware starts writing to the AUX area to
353 * obtain an output handle and make sure there's room in the buffer.
354 * When the capture completes, call perf_aux_output_end() to commit
355 * the recorded data to the buffer.
356 *
357 * The ordering is similar to that of perf_output_{begin,end}, with
358 * the exception of (B), which should be taken care of by the pmu
359 * driver, since ordering rules will differ depending on hardware.
360 *
361 * Call this from pmu::start(); see the comment in perf_aux_output_end()
362 * about its use in pmu callbacks. Both can also be called from the PMI
363 * handler if needed.
364 */
365void *perf_aux_output_begin(struct perf_output_handle *handle,
366 struct perf_event *event)
367{
368 struct perf_event *output_event = event;
369 unsigned long aux_head, aux_tail;
370 struct perf_buffer *rb;
371 unsigned int nest;
372
373 if (output_event->parent)
374 output_event = output_event->parent;
375
376 /*
377 * Since this will typically be open across pmu::add/pmu::del, we
378 * grab ring_buffer's refcount instead of holding rcu read lock
379 * to make sure it doesn't disappear under us.
380 */
381 rb = ring_buffer_get(output_event);
382 if (!rb)
383 return NULL;
384
385 if (!rb_has_aux(rb))
386 goto err;
387
388 /*
389 * If aux_mmap_count is zero, the aux buffer is in perf_mmap_close(),
390 * about to get freed, so we leave immediately.
391 *
392 * Checking rb::aux_mmap_count and rb::refcount has to be done in
393 * the same order, see perf_mmap_close. Otherwise we end up freeing
394 * aux pages in this path, which is a bug, because in_atomic().
395 */
396 if (!atomic_read(&rb->aux_mmap_count))
397 goto err;
398
399 if (!refcount_inc_not_zero(&rb->aux_refcount))
400 goto err;
401
402 nest = READ_ONCE(rb->aux_nest);
403 /*
404 * Nesting is not supported for AUX area, make sure nested
405 * writers are caught early
406 */
407 if (WARN_ON_ONCE(nest))
408 goto err_put;
409
410 WRITE_ONCE(rb->aux_nest, nest + 1);
411
412 aux_head = rb->aux_head;
413
414 handle->rb = rb;
415 handle->event = event;
416 handle->head = aux_head;
417 handle->size = 0;
418 handle->aux_flags = 0;
419
420 /*
421 * In overwrite mode, AUX data stores do not depend on aux_tail,
422 * therefore (A) control dependency barrier does not exist. The
423 * (B) <-> (C) ordering is still observed by the pmu driver.
424 */
425 if (!rb->aux_overwrite) {
426 aux_tail = READ_ONCE(rb->user_page->aux_tail);
427 handle->wakeup = rb->aux_wakeup + rb->aux_watermark;
428 if (aux_head - aux_tail < perf_aux_size(rb))
429 handle->size = CIRC_SPACE(aux_head, aux_tail, perf_aux_size(rb));
430
431 /*
432 * handle->size computation depends on aux_tail load; this forms a
433 * control dependency barrier separating aux_tail load from aux data
434 * store that will be enabled on successful return
435 */
436 if (!handle->size) { /* A, matches D */
437 event->pending_disable = smp_processor_id();
438 perf_output_wakeup(handle);
439 WRITE_ONCE(rb->aux_nest, 0);
440 goto err_put;
441 }
442 }
443
444 return handle->rb->aux_priv;
445
446err_put:
447 /* can't be last */
448 rb_free_aux(rb);
449
450err:
451 ring_buffer_put(rb);
452 handle->event = NULL;
453
454 return NULL;
455}
456EXPORT_SYMBOL_GPL(perf_aux_output_begin);
457
458static __always_inline bool rb_need_aux_wakeup(struct perf_buffer *rb)
459{
460 if (rb->aux_overwrite)
461 return false;
462
463 if (rb->aux_head - rb->aux_wakeup >= rb->aux_watermark) {
464 rb->aux_wakeup = rounddown(rb->aux_head, rb->aux_watermark);
465 return true;
466 }
467
468 return false;
469}
470
471/*
472 * Commit the data written by hardware into the ring buffer by adjusting
473 * aux_head and posting a PERF_RECORD_AUX into the perf buffer. It is the
474 * pmu driver's responsibility to observe ordering rules of the hardware,
475 * so that all the data is externally visible before this is called.
476 *
477 * Note: this has to be called from pmu::stop() callback, as the assumption
478 * of the AUX buffer management code is that after pmu::stop(), the AUX
479 * transaction must be stopped and therefore drop the AUX reference count.
480 */
481void perf_aux_output_end(struct perf_output_handle *handle, unsigned long size)
482{
483 bool wakeup = !!(handle->aux_flags & PERF_AUX_FLAG_TRUNCATED);
484 struct perf_buffer *rb = handle->rb;
485 unsigned long aux_head;
486
487 /* in overwrite mode, driver provides aux_head via handle */
488 if (rb->aux_overwrite) {
489 handle->aux_flags |= PERF_AUX_FLAG_OVERWRITE;
490
491 aux_head = handle->head;
492 rb->aux_head = aux_head;
493 } else {
494 handle->aux_flags &= ~PERF_AUX_FLAG_OVERWRITE;
495
496 aux_head = rb->aux_head;
497 rb->aux_head += size;
498 }
499
500 /*
501 * Only send RECORD_AUX if we have something useful to communicate
502 *
503 * Note: the OVERWRITE records by themselves are not considered
504 * useful, as they don't communicate any *new* information,
505 * aside from the short-lived offset, that becomes history at
506 * the next event sched-in and therefore isn't useful.
507 * The userspace that needs to copy out AUX data in overwrite
508 * mode should know to use user_page::aux_head for the actual
509 * offset. So, from now on we don't output AUX records that
510 * have *only* OVERWRITE flag set.
511 */
512 if (size || (handle->aux_flags & ~(u64)PERF_AUX_FLAG_OVERWRITE))
513 perf_event_aux_event(handle->event, aux_head, size,
514 handle->aux_flags);
515
516 WRITE_ONCE(rb->user_page->aux_head, rb->aux_head);
517 if (rb_need_aux_wakeup(rb))
518 wakeup = true;
519
520 if (wakeup) {
521 if (handle->aux_flags & PERF_AUX_FLAG_TRUNCATED)
522 handle->event->pending_disable = smp_processor_id();
523 perf_output_wakeup(handle);
524 }
525
526 handle->event = NULL;
527
528 WRITE_ONCE(rb->aux_nest, 0);
529 /* can't be last */
530 rb_free_aux(rb);
531 ring_buffer_put(rb);
532}
533EXPORT_SYMBOL_GPL(perf_aux_output_end);
534
535/*
536 * Skip over a given number of bytes in the AUX buffer, due to, for example,
537 * hardware's alignment constraints.
538 */
539int perf_aux_output_skip(struct perf_output_handle *handle, unsigned long size)
540{
541 struct perf_buffer *rb = handle->rb;
542
543 if (size > handle->size)
544 return -ENOSPC;
545
546 rb->aux_head += size;
547
548 WRITE_ONCE(rb->user_page->aux_head, rb->aux_head);
549 if (rb_need_aux_wakeup(rb)) {
550 perf_output_wakeup(handle);
551 handle->wakeup = rb->aux_wakeup + rb->aux_watermark;
552 }
553
554 handle->head = rb->aux_head;
555 handle->size -= size;
556
557 return 0;
558}
559EXPORT_SYMBOL_GPL(perf_aux_output_skip);
560
561void *perf_get_aux(struct perf_output_handle *handle)
562{
563 /* this is only valid between perf_aux_output_begin and *_end */
564 if (!handle->event)
565 return NULL;
566
567 return handle->rb->aux_priv;
568}
569EXPORT_SYMBOL_GPL(perf_get_aux);
570
571/*
572 * Copy out AUX data from an AUX handle.
573 */
574long perf_output_copy_aux(struct perf_output_handle *aux_handle,
575 struct perf_output_handle *handle,
576 unsigned long from, unsigned long to)
577{
578 struct perf_buffer *rb = aux_handle->rb;
579 unsigned long tocopy, remainder, len = 0;
580 void *addr;
581
582 from &= (rb->aux_nr_pages << PAGE_SHIFT) - 1;
583 to &= (rb->aux_nr_pages << PAGE_SHIFT) - 1;
584
585 do {
586 tocopy = PAGE_SIZE - offset_in_page(from);
587 if (to > from)
588 tocopy = min(tocopy, to - from);
589 if (!tocopy)
590 break;
591
592 addr = rb->aux_pages[from >> PAGE_SHIFT];
593 addr += offset_in_page(from);
594
595 remainder = perf_output_copy(handle, addr, tocopy);
596 if (remainder)
597 return -EFAULT;
598
599 len += tocopy;
600 from += tocopy;
601 from &= (rb->aux_nr_pages << PAGE_SHIFT) - 1;
602 } while (to != from);
603
604 return len;
605}
606
607#define PERF_AUX_GFP (GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_NORETRY)
608
609static struct page *rb_alloc_aux_page(int node, int order)
610{
611 struct page *page;
612
613 if (order > MAX_PAGE_ORDER)
614 order = MAX_PAGE_ORDER;
615
616 do {
617 page = alloc_pages_node(node, PERF_AUX_GFP, order);
618 } while (!page && order--);
619
620 if (page && order) {
621 /*
622 * Communicate the allocation size to the driver:
623 * if we managed to secure a high-order allocation,
624 * set its first page's private to this order;
625 * !PagePrivate(page) means it's just a normal page.
626 */
627 split_page(page, order);
628 SetPagePrivate(page);
629 set_page_private(page, order);
630 }
631
632 return page;
633}
634
635static void rb_free_aux_page(struct perf_buffer *rb, int idx)
636{
637 struct page *page = virt_to_page(rb->aux_pages[idx]);
638
639 ClearPagePrivate(page);
640 page->mapping = NULL;
641 __free_page(page);
642}
643
644static void __rb_free_aux(struct perf_buffer *rb)
645{
646 int pg;
647
648 /*
649 * Should never happen, the last reference should be dropped from
650 * perf_mmap_close() path, which first stops aux transactions (which
651 * in turn are the atomic holders of aux_refcount) and then does the
652 * last rb_free_aux().
653 */
654 WARN_ON_ONCE(in_atomic());
655
656 if (rb->aux_priv) {
657 rb->free_aux(rb->aux_priv);
658 rb->free_aux = NULL;
659 rb->aux_priv = NULL;
660 }
661
662 if (rb->aux_nr_pages) {
663 for (pg = 0; pg < rb->aux_nr_pages; pg++)
664 rb_free_aux_page(rb, pg);
665
666 kfree(rb->aux_pages);
667 rb->aux_nr_pages = 0;
668 }
669}
670
671int rb_alloc_aux(struct perf_buffer *rb, struct perf_event *event,
672 pgoff_t pgoff, int nr_pages, long watermark, int flags)
673{
674 bool overwrite = !(flags & RING_BUFFER_WRITABLE);
675 int node = (event->cpu == -1) ? -1 : cpu_to_node(event->cpu);
676 int ret = -ENOMEM, max_order;
677
678 if (!has_aux(event))
679 return -EOPNOTSUPP;
680
681 if (!overwrite) {
682 /*
683 * Watermark defaults to half the buffer, and so does the
684 * max_order, to aid PMU drivers in double buffering.
685 */
686 if (!watermark)
687 watermark = nr_pages << (PAGE_SHIFT - 1);
688
689 /*
690 * Use aux_watermark as the basis for chunking to
691 * help PMU drivers honor the watermark.
692 */
693 max_order = get_order(watermark);
694 } else {
695 /*
696 * We need to start with the max_order that fits in nr_pages,
697 * not the other way around, hence ilog2() and not get_order.
698 */
699 max_order = ilog2(nr_pages);
700 watermark = 0;
701 }
702
703 /*
704 * kcalloc_node() is unable to allocate buffer if the size is larger
705 * than: PAGE_SIZE << MAX_PAGE_ORDER; directly bail out in this case.
706 */
707 if (get_order((unsigned long)nr_pages * sizeof(void *)) > MAX_PAGE_ORDER)
708 return -ENOMEM;
709 rb->aux_pages = kcalloc_node(nr_pages, sizeof(void *), GFP_KERNEL,
710 node);
711 if (!rb->aux_pages)
712 return -ENOMEM;
713
714 rb->free_aux = event->pmu->free_aux;
715 for (rb->aux_nr_pages = 0; rb->aux_nr_pages < nr_pages;) {
716 struct page *page;
717 int last, order;
718
719 order = min(max_order, ilog2(nr_pages - rb->aux_nr_pages));
720 page = rb_alloc_aux_page(node, order);
721 if (!page)
722 goto out;
723
724 for (last = rb->aux_nr_pages + (1 << page_private(page));
725 last > rb->aux_nr_pages; rb->aux_nr_pages++)
726 rb->aux_pages[rb->aux_nr_pages] = page_address(page++);
727 }
728
729 /*
730 * In overwrite mode, PMUs that don't support SG may not handle more
731 * than one contiguous allocation, since they rely on PMI to do double
732 * buffering. In this case, the entire buffer has to be one contiguous
733 * chunk.
734 */
735 if ((event->pmu->capabilities & PERF_PMU_CAP_AUX_NO_SG) &&
736 overwrite) {
737 struct page *page = virt_to_page(rb->aux_pages[0]);
738
739 if (page_private(page) != max_order)
740 goto out;
741 }
742
743 rb->aux_priv = event->pmu->setup_aux(event, rb->aux_pages, nr_pages,
744 overwrite);
745 if (!rb->aux_priv)
746 goto out;
747
748 ret = 0;
749
750 /*
751 * aux_pages (and pmu driver's private data, aux_priv) will be
752 * referenced in both producer's and consumer's contexts, thus
753 * we keep a refcount here to make sure either of the two can
754 * reference them safely.
755 */
756 refcount_set(&rb->aux_refcount, 1);
757
758 rb->aux_overwrite = overwrite;
759 rb->aux_watermark = watermark;
760
761out:
762 if (!ret)
763 rb->aux_pgoff = pgoff;
764 else
765 __rb_free_aux(rb);
766
767 return ret;
768}
769
770void rb_free_aux(struct perf_buffer *rb)
771{
772 if (refcount_dec_and_test(&rb->aux_refcount))
773 __rb_free_aux(rb);
774}
775
776#ifndef CONFIG_PERF_USE_VMALLOC
777
778/*
779 * Back perf_mmap() with regular GFP_KERNEL-0 pages.
780 */
781
782static struct page *
783__perf_mmap_to_page(struct perf_buffer *rb, unsigned long pgoff)
784{
785 if (pgoff > rb->nr_pages)
786 return NULL;
787
788 if (pgoff == 0)
789 return virt_to_page(rb->user_page);
790
791 return virt_to_page(rb->data_pages[pgoff - 1]);
792}
793
794static void *perf_mmap_alloc_page(int cpu)
795{
796 struct page *page;
797 int node;
798
799 node = (cpu == -1) ? cpu : cpu_to_node(cpu);
800 page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
801 if (!page)
802 return NULL;
803
804 return page_address(page);
805}
806
807static void perf_mmap_free_page(void *addr)
808{
809 struct page *page = virt_to_page(addr);
810
811 page->mapping = NULL;
812 __free_page(page);
813}
814
815struct perf_buffer *rb_alloc(int nr_pages, long watermark, int cpu, int flags)
816{
817 struct perf_buffer *rb;
818 unsigned long size;
819 int i, node;
820
821 size = sizeof(struct perf_buffer);
822 size += nr_pages * sizeof(void *);
823
824 if (order_base_2(size) > PAGE_SHIFT+MAX_PAGE_ORDER)
825 goto fail;
826
827 node = (cpu == -1) ? cpu : cpu_to_node(cpu);
828 rb = kzalloc_node(size, GFP_KERNEL, node);
829 if (!rb)
830 goto fail;
831
832 rb->user_page = perf_mmap_alloc_page(cpu);
833 if (!rb->user_page)
834 goto fail_user_page;
835
836 for (i = 0; i < nr_pages; i++) {
837 rb->data_pages[i] = perf_mmap_alloc_page(cpu);
838 if (!rb->data_pages[i])
839 goto fail_data_pages;
840 }
841
842 rb->nr_pages = nr_pages;
843
844 ring_buffer_init(rb, watermark, flags);
845
846 return rb;
847
848fail_data_pages:
849 for (i--; i >= 0; i--)
850 perf_mmap_free_page(rb->data_pages[i]);
851
852 perf_mmap_free_page(rb->user_page);
853
854fail_user_page:
855 kfree(rb);
856
857fail:
858 return NULL;
859}
860
861void rb_free(struct perf_buffer *rb)
862{
863 int i;
864
865 perf_mmap_free_page(rb->user_page);
866 for (i = 0; i < rb->nr_pages; i++)
867 perf_mmap_free_page(rb->data_pages[i]);
868 kfree(rb);
869}
870
871#else
872static struct page *
873__perf_mmap_to_page(struct perf_buffer *rb, unsigned long pgoff)
874{
875 /* The '>' counts in the user page. */
876 if (pgoff > data_page_nr(rb))
877 return NULL;
878
879 return vmalloc_to_page((void *)rb->user_page + pgoff * PAGE_SIZE);
880}
881
882static void perf_mmap_unmark_page(void *addr)
883{
884 struct page *page = vmalloc_to_page(addr);
885
886 page->mapping = NULL;
887}
888
889static void rb_free_work(struct work_struct *work)
890{
891 struct perf_buffer *rb;
892 void *base;
893 int i, nr;
894
895 rb = container_of(work, struct perf_buffer, work);
896 nr = data_page_nr(rb);
897
898 base = rb->user_page;
899 /* The '<=' counts in the user page. */
900 for (i = 0; i <= nr; i++)
901 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
902
903 vfree(base);
904 kfree(rb);
905}
906
907void rb_free(struct perf_buffer *rb)
908{
909 schedule_work(&rb->work);
910}
911
912struct perf_buffer *rb_alloc(int nr_pages, long watermark, int cpu, int flags)
913{
914 struct perf_buffer *rb;
915 unsigned long size;
916 void *all_buf;
917 int node;
918
919 size = sizeof(struct perf_buffer);
920 size += sizeof(void *);
921
922 node = (cpu == -1) ? cpu : cpu_to_node(cpu);
923 rb = kzalloc_node(size, GFP_KERNEL, node);
924 if (!rb)
925 goto fail;
926
927 INIT_WORK(&rb->work, rb_free_work);
928
929 all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
930 if (!all_buf)
931 goto fail_all_buf;
932
933 rb->user_page = all_buf;
934 rb->data_pages[0] = all_buf + PAGE_SIZE;
935 if (nr_pages) {
936 rb->nr_pages = 1;
937 rb->page_order = ilog2(nr_pages);
938 }
939
940 ring_buffer_init(rb, watermark, flags);
941
942 return rb;
943
944fail_all_buf:
945 kfree(rb);
946
947fail:
948 return NULL;
949}
950
951#endif
952
953struct page *
954perf_mmap_to_page(struct perf_buffer *rb, unsigned long pgoff)
955{
956 if (rb->aux_nr_pages) {
957 /* above AUX space */
958 if (pgoff > rb->aux_pgoff + rb->aux_nr_pages)
959 return NULL;
960
961 /* AUX space */
962 if (pgoff >= rb->aux_pgoff) {
963 int aux_pgoff = array_index_nospec(pgoff - rb->aux_pgoff, rb->aux_nr_pages);
964 return virt_to_page(rb->aux_pages[aux_pgoff]);
965 }
966 }
967
968 return __perf_mmap_to_page(rb, pgoff);
969}