1 /*
2 * Generic ring buffer
3 *
4 * Copyright (C) 2008 Steven Rostedt <srostedt@redhat.com>
5 */
6 #include <linux/trace_events.h>
7 #include <linux/ring_buffer.h>
8 #include <linux/trace_clock.h>
9 #include <linux/trace_seq.h>
10 #include <linux/spinlock.h>
11 #include <linux/irq_work.h>
12 #include <linux/uaccess.h>
13 #include <linux/hardirq.h>
14 #include <linux/kthread.h> /* for self test */
15 #include <linux/kmemcheck.h>
16 #include <linux/module.h>
17 #include <linux/percpu.h>
18 #include <linux/mutex.h>
19 #include <linux/delay.h>
20 #include <linux/slab.h>
21 #include <linux/init.h>
22 #include <linux/hash.h>
23 #include <linux/list.h>
24 #include <linux/cpu.h>
25
26 #include <asm/local.h>
27
28 static void update_pages_handler(struct work_struct *work);
29
30 /*
31 * The ring buffer header is special. We must manually up keep it.
32 */
ring_buffer_print_entry_header(struct trace_seq * s)33 int ring_buffer_print_entry_header(struct trace_seq *s)
34 {
35 trace_seq_puts(s, "# compressed entry header\n");
36 trace_seq_puts(s, "\ttype_len : 5 bits\n");
37 trace_seq_puts(s, "\ttime_delta : 27 bits\n");
38 trace_seq_puts(s, "\tarray : 32 bits\n");
39 trace_seq_putc(s, '\n');
40 trace_seq_printf(s, "\tpadding : type == %d\n",
41 RINGBUF_TYPE_PADDING);
42 trace_seq_printf(s, "\ttime_extend : type == %d\n",
43 RINGBUF_TYPE_TIME_EXTEND);
44 trace_seq_printf(s, "\tdata max type_len == %d\n",
45 RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
46
47 return !trace_seq_has_overflowed(s);
48 }
49
50 /*
51 * The ring buffer is made up of a list of pages. A separate list of pages is
52 * allocated for each CPU. A writer may only write to a buffer that is
53 * associated with the CPU it is currently executing on. A reader may read
54 * from any per cpu buffer.
55 *
56 * The reader is special. For each per cpu buffer, the reader has its own
57 * reader page. When a reader has read the entire reader page, this reader
58 * page is swapped with another page in the ring buffer.
59 *
60 * Now, as long as the writer is off the reader page, the reader can do what
61 * ever it wants with that page. The writer will never write to that page
62 * again (as long as it is out of the ring buffer).
63 *
64 * Here's some silly ASCII art.
65 *
66 * +------+
67 * |reader| RING BUFFER
68 * |page |
69 * +------+ +---+ +---+ +---+
70 * | |-->| |-->| |
71 * +---+ +---+ +---+
72 * ^ |
73 * | |
74 * +---------------+
75 *
76 *
77 * +------+
78 * |reader| RING BUFFER
79 * |page |------------------v
80 * +------+ +---+ +---+ +---+
81 * | |-->| |-->| |
82 * +---+ +---+ +---+
83 * ^ |
84 * | |
85 * +---------------+
86 *
87 *
88 * +------+
89 * |reader| RING BUFFER
90 * |page |------------------v
91 * +------+ +---+ +---+ +---+
92 * ^ | |-->| |-->| |
93 * | +---+ +---+ +---+
94 * | |
95 * | |
96 * +------------------------------+
97 *
98 *
99 * +------+
100 * |buffer| RING BUFFER
101 * |page |------------------v
102 * +------+ +---+ +---+ +---+
103 * ^ | | | |-->| |
104 * | New +---+ +---+ +---+
105 * | Reader------^ |
106 * | page |
107 * +------------------------------+
108 *
109 *
110 * After we make this swap, the reader can hand this page off to the splice
111 * code and be done with it. It can even allocate a new page if it needs to
112 * and swap that into the ring buffer.
113 *
114 * We will be using cmpxchg soon to make all this lockless.
115 *
116 */
117
118 /* Used for individual buffers (after the counter) */
119 #define RB_BUFFER_OFF (1 << 20)
120
121 #define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
122
123 #define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
124 #define RB_ALIGNMENT 4U
125 #define RB_MAX_SMALL_DATA (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
126 #define RB_EVNT_MIN_SIZE 8U /* two 32bit words */
127
128 #ifndef CONFIG_HAVE_64BIT_ALIGNED_ACCESS
129 # define RB_FORCE_8BYTE_ALIGNMENT 0
130 # define RB_ARCH_ALIGNMENT RB_ALIGNMENT
131 #else
132 # define RB_FORCE_8BYTE_ALIGNMENT 1
133 # define RB_ARCH_ALIGNMENT 8U
134 #endif
135
136 #define RB_ALIGN_DATA __aligned(RB_ARCH_ALIGNMENT)
137
138 /* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
139 #define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
140
141 enum {
142 RB_LEN_TIME_EXTEND = 8,
143 RB_LEN_TIME_STAMP = 16,
144 };
145
146 #define skip_time_extend(event) \
147 ((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
148
rb_null_event(struct ring_buffer_event * event)149 static inline int rb_null_event(struct ring_buffer_event *event)
150 {
151 return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
152 }
153
rb_event_set_padding(struct ring_buffer_event * event)154 static void rb_event_set_padding(struct ring_buffer_event *event)
155 {
156 /* padding has a NULL time_delta */
157 event->type_len = RINGBUF_TYPE_PADDING;
158 event->time_delta = 0;
159 }
160
161 static unsigned
rb_event_data_length(struct ring_buffer_event * event)162 rb_event_data_length(struct ring_buffer_event *event)
163 {
164 unsigned length;
165
166 if (event->type_len)
167 length = event->type_len * RB_ALIGNMENT;
168 else
169 length = event->array[0];
170 return length + RB_EVNT_HDR_SIZE;
171 }
172
173 /*
174 * Return the length of the given event. Will return
175 * the length of the time extend if the event is a
176 * time extend.
177 */
178 static inline unsigned
rb_event_length(struct ring_buffer_event * event)179 rb_event_length(struct ring_buffer_event *event)
180 {
181 switch (event->type_len) {
182 case RINGBUF_TYPE_PADDING:
183 if (rb_null_event(event))
184 /* undefined */
185 return -1;
186 return event->array[0] + RB_EVNT_HDR_SIZE;
187
188 case RINGBUF_TYPE_TIME_EXTEND:
189 return RB_LEN_TIME_EXTEND;
190
191 case RINGBUF_TYPE_TIME_STAMP:
192 return RB_LEN_TIME_STAMP;
193
194 case RINGBUF_TYPE_DATA:
195 return rb_event_data_length(event);
196 default:
197 BUG();
198 }
199 /* not hit */
200 return 0;
201 }
202
203 /*
204 * Return total length of time extend and data,
205 * or just the event length for all other events.
206 */
207 static inline unsigned
rb_event_ts_length(struct ring_buffer_event * event)208 rb_event_ts_length(struct ring_buffer_event *event)
209 {
210 unsigned len = 0;
211
212 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
213 /* time extends include the data event after it */
214 len = RB_LEN_TIME_EXTEND;
215 event = skip_time_extend(event);
216 }
217 return len + rb_event_length(event);
218 }
219
220 /**
221 * ring_buffer_event_length - return the length of the event
222 * @event: the event to get the length of
223 *
224 * Returns the size of the data load of a data event.
225 * If the event is something other than a data event, it
226 * returns the size of the event itself. With the exception
227 * of a TIME EXTEND, where it still returns the size of the
228 * data load of the data event after it.
229 */
ring_buffer_event_length(struct ring_buffer_event * event)230 unsigned ring_buffer_event_length(struct ring_buffer_event *event)
231 {
232 unsigned length;
233
234 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
235 event = skip_time_extend(event);
236
237 length = rb_event_length(event);
238 if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
239 return length;
240 length -= RB_EVNT_HDR_SIZE;
241 if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
242 length -= sizeof(event->array[0]);
243 return length;
244 }
245 EXPORT_SYMBOL_GPL(ring_buffer_event_length);
246
247 /* inline for ring buffer fast paths */
248 static void *
rb_event_data(struct ring_buffer_event * event)249 rb_event_data(struct ring_buffer_event *event)
250 {
251 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
252 event = skip_time_extend(event);
253 BUG_ON(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
254 /* If length is in len field, then array[0] has the data */
255 if (event->type_len)
256 return (void *)&event->array[0];
257 /* Otherwise length is in array[0] and array[1] has the data */
258 return (void *)&event->array[1];
259 }
260
261 /**
262 * ring_buffer_event_data - return the data of the event
263 * @event: the event to get the data from
264 */
ring_buffer_event_data(struct ring_buffer_event * event)265 void *ring_buffer_event_data(struct ring_buffer_event *event)
266 {
267 return rb_event_data(event);
268 }
269 EXPORT_SYMBOL_GPL(ring_buffer_event_data);
270
271 #define for_each_buffer_cpu(buffer, cpu) \
272 for_each_cpu(cpu, buffer->cpumask)
273
274 #define TS_SHIFT 27
275 #define TS_MASK ((1ULL << TS_SHIFT) - 1)
276 #define TS_DELTA_TEST (~TS_MASK)
277
278 /* Flag when events were overwritten */
279 #define RB_MISSED_EVENTS (1 << 31)
280 /* Missed count stored at end */
281 #define RB_MISSED_STORED (1 << 30)
282
283 #define RB_MISSED_FLAGS (RB_MISSED_EVENTS|RB_MISSED_STORED)
284
285 struct buffer_data_page {
286 u64 time_stamp; /* page time stamp */
287 local_t commit; /* write committed index */
288 unsigned char data[] RB_ALIGN_DATA; /* data of buffer page */
289 };
290
291 /*
292 * Note, the buffer_page list must be first. The buffer pages
293 * are allocated in cache lines, which means that each buffer
294 * page will be at the beginning of a cache line, and thus
295 * the least significant bits will be zero. We use this to
296 * add flags in the list struct pointers, to make the ring buffer
297 * lockless.
298 */
299 struct buffer_page {
300 struct list_head list; /* list of buffer pages */
301 local_t write; /* index for next write */
302 unsigned read; /* index for next read */
303 local_t entries; /* entries on this page */
304 unsigned long real_end; /* real end of data */
305 struct buffer_data_page *page; /* Actual data page */
306 };
307
308 /*
309 * The buffer page counters, write and entries, must be reset
310 * atomically when crossing page boundaries. To synchronize this
311 * update, two counters are inserted into the number. One is
312 * the actual counter for the write position or count on the page.
313 *
314 * The other is a counter of updaters. Before an update happens
315 * the update partition of the counter is incremented. This will
316 * allow the updater to update the counter atomically.
317 *
318 * The counter is 20 bits, and the state data is 12.
319 */
320 #define RB_WRITE_MASK 0xfffff
321 #define RB_WRITE_INTCNT (1 << 20)
322
rb_init_page(struct buffer_data_page * bpage)323 static void rb_init_page(struct buffer_data_page *bpage)
324 {
325 local_set(&bpage->commit, 0);
326 }
327
328 /**
329 * ring_buffer_page_len - the size of data on the page.
330 * @page: The page to read
331 *
332 * Returns the amount of data on the page, including buffer page header.
333 */
ring_buffer_page_len(void * page)334 size_t ring_buffer_page_len(void *page)
335 {
336 struct buffer_data_page *bpage = page;
337
338 return (local_read(&bpage->commit) & ~RB_MISSED_FLAGS)
339 + BUF_PAGE_HDR_SIZE;
340 }
341
342 /*
343 * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
344 * this issue out.
345 */
free_buffer_page(struct buffer_page * bpage)346 static void free_buffer_page(struct buffer_page *bpage)
347 {
348 free_page((unsigned long)bpage->page);
349 kfree(bpage);
350 }
351
352 /*
353 * We need to fit the time_stamp delta into 27 bits.
354 */
test_time_stamp(u64 delta)355 static inline int test_time_stamp(u64 delta)
356 {
357 if (delta & TS_DELTA_TEST)
358 return 1;
359 return 0;
360 }
361
362 #define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
363
364 /* Max payload is BUF_PAGE_SIZE - header (8bytes) */
365 #define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
366
ring_buffer_print_page_header(struct trace_seq * s)367 int ring_buffer_print_page_header(struct trace_seq *s)
368 {
369 struct buffer_data_page field;
370
371 trace_seq_printf(s, "\tfield: u64 timestamp;\t"
372 "offset:0;\tsize:%u;\tsigned:%u;\n",
373 (unsigned int)sizeof(field.time_stamp),
374 (unsigned int)is_signed_type(u64));
375
376 trace_seq_printf(s, "\tfield: local_t commit;\t"
377 "offset:%u;\tsize:%u;\tsigned:%u;\n",
378 (unsigned int)offsetof(typeof(field), commit),
379 (unsigned int)sizeof(field.commit),
380 (unsigned int)is_signed_type(long));
381
382 trace_seq_printf(s, "\tfield: int overwrite;\t"
383 "offset:%u;\tsize:%u;\tsigned:%u;\n",
384 (unsigned int)offsetof(typeof(field), commit),
385 1,
386 (unsigned int)is_signed_type(long));
387
388 trace_seq_printf(s, "\tfield: char data;\t"
389 "offset:%u;\tsize:%u;\tsigned:%u;\n",
390 (unsigned int)offsetof(typeof(field), data),
391 (unsigned int)BUF_PAGE_SIZE,
392 (unsigned int)is_signed_type(char));
393
394 return !trace_seq_has_overflowed(s);
395 }
396
397 struct rb_irq_work {
398 struct irq_work work;
399 wait_queue_head_t waiters;
400 wait_queue_head_t full_waiters;
401 bool waiters_pending;
402 bool full_waiters_pending;
403 bool wakeup_full;
404 };
405
406 /*
407 * Structure to hold event state and handle nested events.
408 */
409 struct rb_event_info {
410 u64 ts;
411 u64 delta;
412 unsigned long length;
413 struct buffer_page *tail_page;
414 int add_timestamp;
415 };
416
417 /*
418 * Used for which event context the event is in.
419 * TRANSITION = 0
420 * NMI = 1
421 * IRQ = 2
422 * SOFTIRQ = 3
423 * NORMAL = 4
424 *
425 * See trace_recursive_lock() comment below for more details.
426 */
427 enum {
428 RB_CTX_TRANSITION,
429 RB_CTX_NMI,
430 RB_CTX_IRQ,
431 RB_CTX_SOFTIRQ,
432 RB_CTX_NORMAL,
433 RB_CTX_MAX
434 };
435
436 /*
437 * head_page == tail_page && head == tail then buffer is empty.
438 */
439 struct ring_buffer_per_cpu {
440 int cpu;
441 atomic_t record_disabled;
442 struct ring_buffer *buffer;
443 raw_spinlock_t reader_lock; /* serialize readers */
444 arch_spinlock_t lock;
445 struct lock_class_key lock_key;
446 unsigned long nr_pages;
447 unsigned int current_context;
448 struct list_head *pages;
449 struct buffer_page *head_page; /* read from head */
450 struct buffer_page *tail_page; /* write to tail */
451 struct buffer_page *commit_page; /* committed pages */
452 struct buffer_page *reader_page;
453 unsigned long lost_events;
454 unsigned long last_overrun;
455 local_t entries_bytes;
456 local_t entries;
457 local_t overrun;
458 local_t commit_overrun;
459 local_t dropped_events;
460 local_t committing;
461 local_t commits;
462 unsigned long read;
463 unsigned long read_bytes;
464 u64 write_stamp;
465 u64 read_stamp;
466 /* ring buffer pages to update, > 0 to add, < 0 to remove */
467 long nr_pages_to_update;
468 struct list_head new_pages; /* new pages to add */
469 struct work_struct update_pages_work;
470 struct completion update_done;
471
472 struct rb_irq_work irq_work;
473 };
474
475 struct ring_buffer {
476 unsigned flags;
477 int cpus;
478 atomic_t record_disabled;
479 atomic_t resize_disabled;
480 cpumask_var_t cpumask;
481
482 struct lock_class_key *reader_lock_key;
483
484 struct mutex mutex;
485
486 struct ring_buffer_per_cpu **buffers;
487
488 #ifdef CONFIG_HOTPLUG_CPU
489 struct notifier_block cpu_notify;
490 #endif
491 u64 (*clock)(void);
492
493 struct rb_irq_work irq_work;
494 };
495
496 struct ring_buffer_iter {
497 struct ring_buffer_per_cpu *cpu_buffer;
498 unsigned long head;
499 struct buffer_page *head_page;
500 struct buffer_page *cache_reader_page;
501 unsigned long cache_read;
502 u64 read_stamp;
503 };
504
505 /*
506 * rb_wake_up_waiters - wake up tasks waiting for ring buffer input
507 *
508 * Schedules a delayed work to wake up any task that is blocked on the
509 * ring buffer waiters queue.
510 */
rb_wake_up_waiters(struct irq_work * work)511 static void rb_wake_up_waiters(struct irq_work *work)
512 {
513 struct rb_irq_work *rbwork = container_of(work, struct rb_irq_work, work);
514
515 wake_up_all(&rbwork->waiters);
516 if (rbwork->wakeup_full) {
517 rbwork->wakeup_full = false;
518 wake_up_all(&rbwork->full_waiters);
519 }
520 }
521
522 /**
523 * ring_buffer_wait - wait for input to the ring buffer
524 * @buffer: buffer to wait on
525 * @cpu: the cpu buffer to wait on
526 * @full: wait until a full page is available, if @cpu != RING_BUFFER_ALL_CPUS
527 *
528 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
529 * as data is added to any of the @buffer's cpu buffers. Otherwise
530 * it will wait for data to be added to a specific cpu buffer.
531 */
ring_buffer_wait(struct ring_buffer * buffer,int cpu,bool full)532 int ring_buffer_wait(struct ring_buffer *buffer, int cpu, bool full)
533 {
534 struct ring_buffer_per_cpu *uninitialized_var(cpu_buffer);
535 DEFINE_WAIT(wait);
536 struct rb_irq_work *work;
537 int ret = 0;
538
539 /*
540 * Depending on what the caller is waiting for, either any
541 * data in any cpu buffer, or a specific buffer, put the
542 * caller on the appropriate wait queue.
543 */
544 if (cpu == RING_BUFFER_ALL_CPUS) {
545 work = &buffer->irq_work;
546 /* Full only makes sense on per cpu reads */
547 full = false;
548 } else {
549 if (!cpumask_test_cpu(cpu, buffer->cpumask))
550 return -ENODEV;
551 cpu_buffer = buffer->buffers[cpu];
552 work = &cpu_buffer->irq_work;
553 }
554
555
556 while (true) {
557 if (full)
558 prepare_to_wait(&work->full_waiters, &wait, TASK_INTERRUPTIBLE);
559 else
560 prepare_to_wait(&work->waiters, &wait, TASK_INTERRUPTIBLE);
561
562 /*
563 * The events can happen in critical sections where
564 * checking a work queue can cause deadlocks.
565 * After adding a task to the queue, this flag is set
566 * only to notify events to try to wake up the queue
567 * using irq_work.
568 *
569 * We don't clear it even if the buffer is no longer
570 * empty. The flag only causes the next event to run
571 * irq_work to do the work queue wake up. The worse
572 * that can happen if we race with !trace_empty() is that
573 * an event will cause an irq_work to try to wake up
574 * an empty queue.
575 *
576 * There's no reason to protect this flag either, as
577 * the work queue and irq_work logic will do the necessary
578 * synchronization for the wake ups. The only thing
579 * that is necessary is that the wake up happens after
580 * a task has been queued. It's OK for spurious wake ups.
581 */
582 if (full)
583 work->full_waiters_pending = true;
584 else
585 work->waiters_pending = true;
586
587 if (signal_pending(current)) {
588 ret = -EINTR;
589 break;
590 }
591
592 if (cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer))
593 break;
594
595 if (cpu != RING_BUFFER_ALL_CPUS &&
596 !ring_buffer_empty_cpu(buffer, cpu)) {
597 unsigned long flags;
598 bool pagebusy;
599
600 if (!full)
601 break;
602
603 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
604 pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
605 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
606
607 if (!pagebusy)
608 break;
609 }
610
611 schedule();
612 }
613
614 if (full)
615 finish_wait(&work->full_waiters, &wait);
616 else
617 finish_wait(&work->waiters, &wait);
618
619 return ret;
620 }
621
622 /**
623 * ring_buffer_poll_wait - poll on buffer input
624 * @buffer: buffer to wait on
625 * @cpu: the cpu buffer to wait on
626 * @filp: the file descriptor
627 * @poll_table: The poll descriptor
628 *
629 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
630 * as data is added to any of the @buffer's cpu buffers. Otherwise
631 * it will wait for data to be added to a specific cpu buffer.
632 *
633 * Returns POLLIN | POLLRDNORM if data exists in the buffers,
634 * zero otherwise.
635 */
ring_buffer_poll_wait(struct ring_buffer * buffer,int cpu,struct file * filp,poll_table * poll_table)636 int ring_buffer_poll_wait(struct ring_buffer *buffer, int cpu,
637 struct file *filp, poll_table *poll_table)
638 {
639 struct ring_buffer_per_cpu *cpu_buffer;
640 struct rb_irq_work *work;
641
642 if (cpu == RING_BUFFER_ALL_CPUS)
643 work = &buffer->irq_work;
644 else {
645 if (!cpumask_test_cpu(cpu, buffer->cpumask))
646 return -EINVAL;
647
648 cpu_buffer = buffer->buffers[cpu];
649 work = &cpu_buffer->irq_work;
650 }
651
652 poll_wait(filp, &work->waiters, poll_table);
653 work->waiters_pending = true;
654 /*
655 * There's a tight race between setting the waiters_pending and
656 * checking if the ring buffer is empty. Once the waiters_pending bit
657 * is set, the next event will wake the task up, but we can get stuck
658 * if there's only a single event in.
659 *
660 * FIXME: Ideally, we need a memory barrier on the writer side as well,
661 * but adding a memory barrier to all events will cause too much of a
662 * performance hit in the fast path. We only need a memory barrier when
663 * the buffer goes from empty to having content. But as this race is
664 * extremely small, and it's not a problem if another event comes in, we
665 * will fix it later.
666 */
667 smp_mb();
668
669 if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
670 (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
671 return POLLIN | POLLRDNORM;
672 return 0;
673 }
674
675 /* buffer may be either ring_buffer or ring_buffer_per_cpu */
676 #define RB_WARN_ON(b, cond) \
677 ({ \
678 int _____ret = unlikely(cond); \
679 if (_____ret) { \
680 if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
681 struct ring_buffer_per_cpu *__b = \
682 (void *)b; \
683 atomic_inc(&__b->buffer->record_disabled); \
684 } else \
685 atomic_inc(&b->record_disabled); \
686 WARN_ON(1); \
687 } \
688 _____ret; \
689 })
690
691 /* Up this if you want to test the TIME_EXTENTS and normalization */
692 #define DEBUG_SHIFT 0
693
rb_time_stamp(struct ring_buffer * buffer)694 static inline u64 rb_time_stamp(struct ring_buffer *buffer)
695 {
696 /* shift to debug/test normalization and TIME_EXTENTS */
697 return buffer->clock() << DEBUG_SHIFT;
698 }
699
ring_buffer_time_stamp(struct ring_buffer * buffer,int cpu)700 u64 ring_buffer_time_stamp(struct ring_buffer *buffer, int cpu)
701 {
702 u64 time;
703
704 preempt_disable_notrace();
705 time = rb_time_stamp(buffer);
706 preempt_enable_notrace();
707
708 return time;
709 }
710 EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
711
ring_buffer_normalize_time_stamp(struct ring_buffer * buffer,int cpu,u64 * ts)712 void ring_buffer_normalize_time_stamp(struct ring_buffer *buffer,
713 int cpu, u64 *ts)
714 {
715 /* Just stupid testing the normalize function and deltas */
716 *ts >>= DEBUG_SHIFT;
717 }
718 EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
719
720 /*
721 * Making the ring buffer lockless makes things tricky.
722 * Although writes only happen on the CPU that they are on,
723 * and they only need to worry about interrupts. Reads can
724 * happen on any CPU.
725 *
726 * The reader page is always off the ring buffer, but when the
727 * reader finishes with a page, it needs to swap its page with
728 * a new one from the buffer. The reader needs to take from
729 * the head (writes go to the tail). But if a writer is in overwrite
730 * mode and wraps, it must push the head page forward.
731 *
732 * Here lies the problem.
733 *
734 * The reader must be careful to replace only the head page, and
735 * not another one. As described at the top of the file in the
736 * ASCII art, the reader sets its old page to point to the next
737 * page after head. It then sets the page after head to point to
738 * the old reader page. But if the writer moves the head page
739 * during this operation, the reader could end up with the tail.
740 *
741 * We use cmpxchg to help prevent this race. We also do something
742 * special with the page before head. We set the LSB to 1.
743 *
744 * When the writer must push the page forward, it will clear the
745 * bit that points to the head page, move the head, and then set
746 * the bit that points to the new head page.
747 *
748 * We also don't want an interrupt coming in and moving the head
749 * page on another writer. Thus we use the second LSB to catch
750 * that too. Thus:
751 *
752 * head->list->prev->next bit 1 bit 0
753 * ------- -------
754 * Normal page 0 0
755 * Points to head page 0 1
756 * New head page 1 0
757 *
758 * Note we can not trust the prev pointer of the head page, because:
759 *
760 * +----+ +-----+ +-----+
761 * | |------>| T |---X--->| N |
762 * | |<------| | | |
763 * +----+ +-----+ +-----+
764 * ^ ^ |
765 * | +-----+ | |
766 * +----------| R |----------+ |
767 * | |<-----------+
768 * +-----+
769 *
770 * Key: ---X--> HEAD flag set in pointer
771 * T Tail page
772 * R Reader page
773 * N Next page
774 *
775 * (see __rb_reserve_next() to see where this happens)
776 *
777 * What the above shows is that the reader just swapped out
778 * the reader page with a page in the buffer, but before it
779 * could make the new header point back to the new page added
780 * it was preempted by a writer. The writer moved forward onto
781 * the new page added by the reader and is about to move forward
782 * again.
783 *
784 * You can see, it is legitimate for the previous pointer of
785 * the head (or any page) not to point back to itself. But only
786 * temporarially.
787 */
788
789 #define RB_PAGE_NORMAL 0UL
790 #define RB_PAGE_HEAD 1UL
791 #define RB_PAGE_UPDATE 2UL
792
793
794 #define RB_FLAG_MASK 3UL
795
796 /* PAGE_MOVED is not part of the mask */
797 #define RB_PAGE_MOVED 4UL
798
799 /*
800 * rb_list_head - remove any bit
801 */
rb_list_head(struct list_head * list)802 static struct list_head *rb_list_head(struct list_head *list)
803 {
804 unsigned long val = (unsigned long)list;
805
806 return (struct list_head *)(val & ~RB_FLAG_MASK);
807 }
808
809 /*
810 * rb_is_head_page - test if the given page is the head page
811 *
812 * Because the reader may move the head_page pointer, we can
813 * not trust what the head page is (it may be pointing to
814 * the reader page). But if the next page is a header page,
815 * its flags will be non zero.
816 */
817 static inline int
rb_is_head_page(struct ring_buffer_per_cpu * cpu_buffer,struct buffer_page * page,struct list_head * list)818 rb_is_head_page(struct ring_buffer_per_cpu *cpu_buffer,
819 struct buffer_page *page, struct list_head *list)
820 {
821 unsigned long val;
822
823 val = (unsigned long)list->next;
824
825 if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
826 return RB_PAGE_MOVED;
827
828 return val & RB_FLAG_MASK;
829 }
830
831 /*
832 * rb_is_reader_page
833 *
834 * The unique thing about the reader page, is that, if the
835 * writer is ever on it, the previous pointer never points
836 * back to the reader page.
837 */
rb_is_reader_page(struct buffer_page * page)838 static bool rb_is_reader_page(struct buffer_page *page)
839 {
840 struct list_head *list = page->list.prev;
841
842 return rb_list_head(list->next) != &page->list;
843 }
844
845 /*
846 * rb_set_list_to_head - set a list_head to be pointing to head.
847 */
rb_set_list_to_head(struct ring_buffer_per_cpu * cpu_buffer,struct list_head * list)848 static void rb_set_list_to_head(struct ring_buffer_per_cpu *cpu_buffer,
849 struct list_head *list)
850 {
851 unsigned long *ptr;
852
853 ptr = (unsigned long *)&list->next;
854 *ptr |= RB_PAGE_HEAD;
855 *ptr &= ~RB_PAGE_UPDATE;
856 }
857
858 /*
859 * rb_head_page_activate - sets up head page
860 */
rb_head_page_activate(struct ring_buffer_per_cpu * cpu_buffer)861 static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
862 {
863 struct buffer_page *head;
864
865 head = cpu_buffer->head_page;
866 if (!head)
867 return;
868
869 /*
870 * Set the previous list pointer to have the HEAD flag.
871 */
872 rb_set_list_to_head(cpu_buffer, head->list.prev);
873 }
874
rb_list_head_clear(struct list_head * list)875 static void rb_list_head_clear(struct list_head *list)
876 {
877 unsigned long *ptr = (unsigned long *)&list->next;
878
879 *ptr &= ~RB_FLAG_MASK;
880 }
881
882 /*
883 * rb_head_page_dactivate - clears head page ptr (for free list)
884 */
885 static void
rb_head_page_deactivate(struct ring_buffer_per_cpu * cpu_buffer)886 rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
887 {
888 struct list_head *hd;
889
890 /* Go through the whole list and clear any pointers found. */
891 rb_list_head_clear(cpu_buffer->pages);
892
893 list_for_each(hd, cpu_buffer->pages)
894 rb_list_head_clear(hd);
895 }
896
rb_head_page_set(struct ring_buffer_per_cpu * cpu_buffer,struct buffer_page * head,struct buffer_page * prev,int old_flag,int new_flag)897 static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
898 struct buffer_page *head,
899 struct buffer_page *prev,
900 int old_flag, int new_flag)
901 {
902 struct list_head *list;
903 unsigned long val = (unsigned long)&head->list;
904 unsigned long ret;
905
906 list = &prev->list;
907
908 val &= ~RB_FLAG_MASK;
909
910 ret = cmpxchg((unsigned long *)&list->next,
911 val | old_flag, val | new_flag);
912
913 /* check if the reader took the page */
914 if ((ret & ~RB_FLAG_MASK) != val)
915 return RB_PAGE_MOVED;
916
917 return ret & RB_FLAG_MASK;
918 }
919
rb_head_page_set_update(struct ring_buffer_per_cpu * cpu_buffer,struct buffer_page * head,struct buffer_page * prev,int old_flag)920 static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
921 struct buffer_page *head,
922 struct buffer_page *prev,
923 int old_flag)
924 {
925 return rb_head_page_set(cpu_buffer, head, prev,
926 old_flag, RB_PAGE_UPDATE);
927 }
928
rb_head_page_set_head(struct ring_buffer_per_cpu * cpu_buffer,struct buffer_page * head,struct buffer_page * prev,int old_flag)929 static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
930 struct buffer_page *head,
931 struct buffer_page *prev,
932 int old_flag)
933 {
934 return rb_head_page_set(cpu_buffer, head, prev,
935 old_flag, RB_PAGE_HEAD);
936 }
937
rb_head_page_set_normal(struct ring_buffer_per_cpu * cpu_buffer,struct buffer_page * head,struct buffer_page * prev,int old_flag)938 static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
939 struct buffer_page *head,
940 struct buffer_page *prev,
941 int old_flag)
942 {
943 return rb_head_page_set(cpu_buffer, head, prev,
944 old_flag, RB_PAGE_NORMAL);
945 }
946
rb_inc_page(struct ring_buffer_per_cpu * cpu_buffer,struct buffer_page ** bpage)947 static inline void rb_inc_page(struct ring_buffer_per_cpu *cpu_buffer,
948 struct buffer_page **bpage)
949 {
950 struct list_head *p = rb_list_head((*bpage)->list.next);
951
952 *bpage = list_entry(p, struct buffer_page, list);
953 }
954
955 static struct buffer_page *
rb_set_head_page(struct ring_buffer_per_cpu * cpu_buffer)956 rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
957 {
958 struct buffer_page *head;
959 struct buffer_page *page;
960 struct list_head *list;
961 int i;
962
963 if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
964 return NULL;
965
966 /* sanity check */
967 list = cpu_buffer->pages;
968 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
969 return NULL;
970
971 page = head = cpu_buffer->head_page;
972 /*
973 * It is possible that the writer moves the header behind
974 * where we started, and we miss in one loop.
975 * A second loop should grab the header, but we'll do
976 * three loops just because I'm paranoid.
977 */
978 for (i = 0; i < 3; i++) {
979 do {
980 if (rb_is_head_page(cpu_buffer, page, page->list.prev)) {
981 cpu_buffer->head_page = page;
982 return page;
983 }
984 rb_inc_page(cpu_buffer, &page);
985 } while (page != head);
986 }
987
988 RB_WARN_ON(cpu_buffer, 1);
989
990 return NULL;
991 }
992
rb_head_page_replace(struct buffer_page * old,struct buffer_page * new)993 static int rb_head_page_replace(struct buffer_page *old,
994 struct buffer_page *new)
995 {
996 unsigned long *ptr = (unsigned long *)&old->list.prev->next;
997 unsigned long val;
998 unsigned long ret;
999
1000 val = *ptr & ~RB_FLAG_MASK;
1001 val |= RB_PAGE_HEAD;
1002
1003 ret = cmpxchg(ptr, val, (unsigned long)&new->list);
1004
1005 return ret == val;
1006 }
1007
1008 /*
1009 * rb_tail_page_update - move the tail page forward
1010 *
1011 * Returns 1 if moved tail page, 0 if someone else did.
1012 */
rb_tail_page_update(struct ring_buffer_per_cpu * cpu_buffer,struct buffer_page * tail_page,struct buffer_page * next_page)1013 static int rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
1014 struct buffer_page *tail_page,
1015 struct buffer_page *next_page)
1016 {
1017 struct buffer_page *old_tail;
1018 unsigned long old_entries;
1019 unsigned long old_write;
1020 int ret = 0;
1021
1022 /*
1023 * The tail page now needs to be moved forward.
1024 *
1025 * We need to reset the tail page, but without messing
1026 * with possible erasing of data brought in by interrupts
1027 * that have moved the tail page and are currently on it.
1028 *
1029 * We add a counter to the write field to denote this.
1030 */
1031 old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
1032 old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
1033
1034 /*
1035 * Just make sure we have seen our old_write and synchronize
1036 * with any interrupts that come in.
1037 */
1038 barrier();
1039
1040 /*
1041 * If the tail page is still the same as what we think
1042 * it is, then it is up to us to update the tail
1043 * pointer.
1044 */
1045 if (tail_page == cpu_buffer->tail_page) {
1046 /* Zero the write counter */
1047 unsigned long val = old_write & ~RB_WRITE_MASK;
1048 unsigned long eval = old_entries & ~RB_WRITE_MASK;
1049
1050 /*
1051 * This will only succeed if an interrupt did
1052 * not come in and change it. In which case, we
1053 * do not want to modify it.
1054 *
1055 * We add (void) to let the compiler know that we do not care
1056 * about the return value of these functions. We use the
1057 * cmpxchg to only update if an interrupt did not already
1058 * do it for us. If the cmpxchg fails, we don't care.
1059 */
1060 (void)local_cmpxchg(&next_page->write, old_write, val);
1061 (void)local_cmpxchg(&next_page->entries, old_entries, eval);
1062
1063 /*
1064 * No need to worry about races with clearing out the commit.
1065 * it only can increment when a commit takes place. But that
1066 * only happens in the outer most nested commit.
1067 */
1068 local_set(&next_page->page->commit, 0);
1069
1070 old_tail = cmpxchg(&cpu_buffer->tail_page,
1071 tail_page, next_page);
1072
1073 if (old_tail == tail_page)
1074 ret = 1;
1075 }
1076
1077 return ret;
1078 }
1079
rb_check_bpage(struct ring_buffer_per_cpu * cpu_buffer,struct buffer_page * bpage)1080 static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
1081 struct buffer_page *bpage)
1082 {
1083 unsigned long val = (unsigned long)bpage;
1084
1085 if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK))
1086 return 1;
1087
1088 return 0;
1089 }
1090
1091 /**
1092 * rb_check_list - make sure a pointer to a list has the last bits zero
1093 */
rb_check_list(struct ring_buffer_per_cpu * cpu_buffer,struct list_head * list)1094 static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer,
1095 struct list_head *list)
1096 {
1097 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev))
1098 return 1;
1099 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next))
1100 return 1;
1101 return 0;
1102 }
1103
1104 /**
1105 * rb_check_pages - integrity check of buffer pages
1106 * @cpu_buffer: CPU buffer with pages to test
1107 *
1108 * As a safety measure we check to make sure the data pages have not
1109 * been corrupted.
1110 */
rb_check_pages(struct ring_buffer_per_cpu * cpu_buffer)1111 static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
1112 {
1113 struct list_head *head = cpu_buffer->pages;
1114 struct buffer_page *bpage, *tmp;
1115
1116 /* Reset the head page if it exists */
1117 if (cpu_buffer->head_page)
1118 rb_set_head_page(cpu_buffer);
1119
1120 rb_head_page_deactivate(cpu_buffer);
1121
1122 if (RB_WARN_ON(cpu_buffer, head->next->prev != head))
1123 return -1;
1124 if (RB_WARN_ON(cpu_buffer, head->prev->next != head))
1125 return -1;
1126
1127 if (rb_check_list(cpu_buffer, head))
1128 return -1;
1129
1130 list_for_each_entry_safe(bpage, tmp, head, list) {
1131 if (RB_WARN_ON(cpu_buffer,
1132 bpage->list.next->prev != &bpage->list))
1133 return -1;
1134 if (RB_WARN_ON(cpu_buffer,
1135 bpage->list.prev->next != &bpage->list))
1136 return -1;
1137 if (rb_check_list(cpu_buffer, &bpage->list))
1138 return -1;
1139 }
1140
1141 rb_head_page_activate(cpu_buffer);
1142
1143 return 0;
1144 }
1145
__rb_allocate_pages(long nr_pages,struct list_head * pages,int cpu)1146 static int __rb_allocate_pages(long nr_pages, struct list_head *pages, int cpu)
1147 {
1148 struct buffer_page *bpage, *tmp;
1149 long i;
1150
1151 for (i = 0; i < nr_pages; i++) {
1152 struct page *page;
1153 /*
1154 * __GFP_NORETRY flag makes sure that the allocation fails
1155 * gracefully without invoking oom-killer and the system is
1156 * not destabilized.
1157 */
1158 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1159 GFP_KERNEL | __GFP_NORETRY,
1160 cpu_to_node(cpu));
1161 if (!bpage)
1162 goto free_pages;
1163
1164 list_add(&bpage->list, pages);
1165
1166 page = alloc_pages_node(cpu_to_node(cpu),
1167 GFP_KERNEL | __GFP_NORETRY, 0);
1168 if (!page)
1169 goto free_pages;
1170 bpage->page = page_address(page);
1171 rb_init_page(bpage->page);
1172 }
1173
1174 return 0;
1175
1176 free_pages:
1177 list_for_each_entry_safe(bpage, tmp, pages, list) {
1178 list_del_init(&bpage->list);
1179 free_buffer_page(bpage);
1180 }
1181
1182 return -ENOMEM;
1183 }
1184
rb_allocate_pages(struct ring_buffer_per_cpu * cpu_buffer,unsigned long nr_pages)1185 static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1186 unsigned long nr_pages)
1187 {
1188 LIST_HEAD(pages);
1189
1190 WARN_ON(!nr_pages);
1191
1192 if (__rb_allocate_pages(nr_pages, &pages, cpu_buffer->cpu))
1193 return -ENOMEM;
1194
1195 /*
1196 * The ring buffer page list is a circular list that does not
1197 * start and end with a list head. All page list items point to
1198 * other pages.
1199 */
1200 cpu_buffer->pages = pages.next;
1201 list_del(&pages);
1202
1203 cpu_buffer->nr_pages = nr_pages;
1204
1205 rb_check_pages(cpu_buffer);
1206
1207 return 0;
1208 }
1209
1210 static struct ring_buffer_per_cpu *
rb_allocate_cpu_buffer(struct ring_buffer * buffer,long nr_pages,int cpu)1211 rb_allocate_cpu_buffer(struct ring_buffer *buffer, long nr_pages, int cpu)
1212 {
1213 struct ring_buffer_per_cpu *cpu_buffer;
1214 struct buffer_page *bpage;
1215 struct page *page;
1216 int ret;
1217
1218 cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1219 GFP_KERNEL, cpu_to_node(cpu));
1220 if (!cpu_buffer)
1221 return NULL;
1222
1223 cpu_buffer->cpu = cpu;
1224 cpu_buffer->buffer = buffer;
1225 raw_spin_lock_init(&cpu_buffer->reader_lock);
1226 lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1227 cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1228 INIT_WORK(&cpu_buffer->update_pages_work, update_pages_handler);
1229 init_completion(&cpu_buffer->update_done);
1230 init_irq_work(&cpu_buffer->irq_work.work, rb_wake_up_waiters);
1231 init_waitqueue_head(&cpu_buffer->irq_work.waiters);
1232 init_waitqueue_head(&cpu_buffer->irq_work.full_waiters);
1233
1234 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1235 GFP_KERNEL, cpu_to_node(cpu));
1236 if (!bpage)
1237 goto fail_free_buffer;
1238
1239 rb_check_bpage(cpu_buffer, bpage);
1240
1241 cpu_buffer->reader_page = bpage;
1242 page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0);
1243 if (!page)
1244 goto fail_free_reader;
1245 bpage->page = page_address(page);
1246 rb_init_page(bpage->page);
1247
1248 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1249 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1250
1251 ret = rb_allocate_pages(cpu_buffer, nr_pages);
1252 if (ret < 0)
1253 goto fail_free_reader;
1254
1255 cpu_buffer->head_page
1256 = list_entry(cpu_buffer->pages, struct buffer_page, list);
1257 cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1258
1259 rb_head_page_activate(cpu_buffer);
1260
1261 return cpu_buffer;
1262
1263 fail_free_reader:
1264 free_buffer_page(cpu_buffer->reader_page);
1265
1266 fail_free_buffer:
1267 kfree(cpu_buffer);
1268 return NULL;
1269 }
1270
rb_free_cpu_buffer(struct ring_buffer_per_cpu * cpu_buffer)1271 static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1272 {
1273 struct list_head *head = cpu_buffer->pages;
1274 struct buffer_page *bpage, *tmp;
1275
1276 free_buffer_page(cpu_buffer->reader_page);
1277
1278 rb_head_page_deactivate(cpu_buffer);
1279
1280 if (head) {
1281 list_for_each_entry_safe(bpage, tmp, head, list) {
1282 list_del_init(&bpage->list);
1283 free_buffer_page(bpage);
1284 }
1285 bpage = list_entry(head, struct buffer_page, list);
1286 free_buffer_page(bpage);
1287 }
1288
1289 kfree(cpu_buffer);
1290 }
1291
1292 #ifdef CONFIG_HOTPLUG_CPU
1293 static int rb_cpu_notify(struct notifier_block *self,
1294 unsigned long action, void *hcpu);
1295 #endif
1296
1297 /**
1298 * __ring_buffer_alloc - allocate a new ring_buffer
1299 * @size: the size in bytes per cpu that is needed.
1300 * @flags: attributes to set for the ring buffer.
1301 *
1302 * Currently the only flag that is available is the RB_FL_OVERWRITE
1303 * flag. This flag means that the buffer will overwrite old data
1304 * when the buffer wraps. If this flag is not set, the buffer will
1305 * drop data when the tail hits the head.
1306 */
__ring_buffer_alloc(unsigned long size,unsigned flags,struct lock_class_key * key)1307 struct ring_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1308 struct lock_class_key *key)
1309 {
1310 struct ring_buffer *buffer;
1311 long nr_pages;
1312 int bsize;
1313 int cpu;
1314
1315 /* keep it in its own cache line */
1316 buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1317 GFP_KERNEL);
1318 if (!buffer)
1319 return NULL;
1320
1321 if (!alloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1322 goto fail_free_buffer;
1323
1324 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1325 buffer->flags = flags;
1326 buffer->clock = trace_clock_local;
1327 buffer->reader_lock_key = key;
1328
1329 init_irq_work(&buffer->irq_work.work, rb_wake_up_waiters);
1330 init_waitqueue_head(&buffer->irq_work.waiters);
1331
1332 /* need at least two pages */
1333 if (nr_pages < 2)
1334 nr_pages = 2;
1335
1336 /*
1337 * In case of non-hotplug cpu, if the ring-buffer is allocated
1338 * in early initcall, it will not be notified of secondary cpus.
1339 * In that off case, we need to allocate for all possible cpus.
1340 */
1341 #ifdef CONFIG_HOTPLUG_CPU
1342 cpu_notifier_register_begin();
1343 cpumask_copy(buffer->cpumask, cpu_online_mask);
1344 #else
1345 cpumask_copy(buffer->cpumask, cpu_possible_mask);
1346 #endif
1347 buffer->cpus = nr_cpu_ids;
1348
1349 bsize = sizeof(void *) * nr_cpu_ids;
1350 buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1351 GFP_KERNEL);
1352 if (!buffer->buffers)
1353 goto fail_free_cpumask;
1354
1355 for_each_buffer_cpu(buffer, cpu) {
1356 buffer->buffers[cpu] =
1357 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
1358 if (!buffer->buffers[cpu])
1359 goto fail_free_buffers;
1360 }
1361
1362 #ifdef CONFIG_HOTPLUG_CPU
1363 buffer->cpu_notify.notifier_call = rb_cpu_notify;
1364 buffer->cpu_notify.priority = 0;
1365 __register_cpu_notifier(&buffer->cpu_notify);
1366 cpu_notifier_register_done();
1367 #endif
1368
1369 mutex_init(&buffer->mutex);
1370
1371 return buffer;
1372
1373 fail_free_buffers:
1374 for_each_buffer_cpu(buffer, cpu) {
1375 if (buffer->buffers[cpu])
1376 rb_free_cpu_buffer(buffer->buffers[cpu]);
1377 }
1378 kfree(buffer->buffers);
1379
1380 fail_free_cpumask:
1381 free_cpumask_var(buffer->cpumask);
1382 #ifdef CONFIG_HOTPLUG_CPU
1383 cpu_notifier_register_done();
1384 #endif
1385
1386 fail_free_buffer:
1387 kfree(buffer);
1388 return NULL;
1389 }
1390 EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1391
1392 /**
1393 * ring_buffer_free - free a ring buffer.
1394 * @buffer: the buffer to free.
1395 */
1396 void
ring_buffer_free(struct ring_buffer * buffer)1397 ring_buffer_free(struct ring_buffer *buffer)
1398 {
1399 int cpu;
1400
1401 #ifdef CONFIG_HOTPLUG_CPU
1402 cpu_notifier_register_begin();
1403 __unregister_cpu_notifier(&buffer->cpu_notify);
1404 #endif
1405
1406 for_each_buffer_cpu(buffer, cpu)
1407 rb_free_cpu_buffer(buffer->buffers[cpu]);
1408
1409 #ifdef CONFIG_HOTPLUG_CPU
1410 cpu_notifier_register_done();
1411 #endif
1412
1413 kfree(buffer->buffers);
1414 free_cpumask_var(buffer->cpumask);
1415
1416 kfree(buffer);
1417 }
1418 EXPORT_SYMBOL_GPL(ring_buffer_free);
1419
ring_buffer_set_clock(struct ring_buffer * buffer,u64 (* clock)(void))1420 void ring_buffer_set_clock(struct ring_buffer *buffer,
1421 u64 (*clock)(void))
1422 {
1423 buffer->clock = clock;
1424 }
1425
1426 static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1427
rb_page_entries(struct buffer_page * bpage)1428 static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1429 {
1430 return local_read(&bpage->entries) & RB_WRITE_MASK;
1431 }
1432
rb_page_write(struct buffer_page * bpage)1433 static inline unsigned long rb_page_write(struct buffer_page *bpage)
1434 {
1435 return local_read(&bpage->write) & RB_WRITE_MASK;
1436 }
1437
1438 static int
rb_remove_pages(struct ring_buffer_per_cpu * cpu_buffer,unsigned long nr_pages)1439 rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned long nr_pages)
1440 {
1441 struct list_head *tail_page, *to_remove, *next_page;
1442 struct buffer_page *to_remove_page, *tmp_iter_page;
1443 struct buffer_page *last_page, *first_page;
1444 unsigned long nr_removed;
1445 unsigned long head_bit;
1446 int page_entries;
1447
1448 head_bit = 0;
1449
1450 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1451 atomic_inc(&cpu_buffer->record_disabled);
1452 /*
1453 * We don't race with the readers since we have acquired the reader
1454 * lock. We also don't race with writers after disabling recording.
1455 * This makes it easy to figure out the first and the last page to be
1456 * removed from the list. We unlink all the pages in between including
1457 * the first and last pages. This is done in a busy loop so that we
1458 * lose the least number of traces.
1459 * The pages are freed after we restart recording and unlock readers.
1460 */
1461 tail_page = &cpu_buffer->tail_page->list;
1462
1463 /*
1464 * tail page might be on reader page, we remove the next page
1465 * from the ring buffer
1466 */
1467 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
1468 tail_page = rb_list_head(tail_page->next);
1469 to_remove = tail_page;
1470
1471 /* start of pages to remove */
1472 first_page = list_entry(rb_list_head(to_remove->next),
1473 struct buffer_page, list);
1474
1475 for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
1476 to_remove = rb_list_head(to_remove)->next;
1477 head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
1478 }
1479
1480 next_page = rb_list_head(to_remove)->next;
1481
1482 /*
1483 * Now we remove all pages between tail_page and next_page.
1484 * Make sure that we have head_bit value preserved for the
1485 * next page
1486 */
1487 tail_page->next = (struct list_head *)((unsigned long)next_page |
1488 head_bit);
1489 next_page = rb_list_head(next_page);
1490 next_page->prev = tail_page;
1491
1492 /* make sure pages points to a valid page in the ring buffer */
1493 cpu_buffer->pages = next_page;
1494
1495 /* update head page */
1496 if (head_bit)
1497 cpu_buffer->head_page = list_entry(next_page,
1498 struct buffer_page, list);
1499
1500 /*
1501 * change read pointer to make sure any read iterators reset
1502 * themselves
1503 */
1504 cpu_buffer->read = 0;
1505
1506 /* pages are removed, resume tracing and then free the pages */
1507 atomic_dec(&cpu_buffer->record_disabled);
1508 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1509
1510 RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
1511
1512 /* last buffer page to remove */
1513 last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
1514 list);
1515 tmp_iter_page = first_page;
1516
1517 do {
1518 cond_resched();
1519
1520 to_remove_page = tmp_iter_page;
1521 rb_inc_page(cpu_buffer, &tmp_iter_page);
1522
1523 /* update the counters */
1524 page_entries = rb_page_entries(to_remove_page);
1525 if (page_entries) {
1526 /*
1527 * If something was added to this page, it was full
1528 * since it is not the tail page. So we deduct the
1529 * bytes consumed in ring buffer from here.
1530 * Increment overrun to account for the lost events.
1531 */
1532 local_add(page_entries, &cpu_buffer->overrun);
1533 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1534 }
1535
1536 /*
1537 * We have already removed references to this list item, just
1538 * free up the buffer_page and its page
1539 */
1540 free_buffer_page(to_remove_page);
1541 nr_removed--;
1542
1543 } while (to_remove_page != last_page);
1544
1545 RB_WARN_ON(cpu_buffer, nr_removed);
1546
1547 return nr_removed == 0;
1548 }
1549
1550 static int
rb_insert_pages(struct ring_buffer_per_cpu * cpu_buffer)1551 rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
1552 {
1553 struct list_head *pages = &cpu_buffer->new_pages;
1554 int retries, success;
1555
1556 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1557 /*
1558 * We are holding the reader lock, so the reader page won't be swapped
1559 * in the ring buffer. Now we are racing with the writer trying to
1560 * move head page and the tail page.
1561 * We are going to adapt the reader page update process where:
1562 * 1. We first splice the start and end of list of new pages between
1563 * the head page and its previous page.
1564 * 2. We cmpxchg the prev_page->next to point from head page to the
1565 * start of new pages list.
1566 * 3. Finally, we update the head->prev to the end of new list.
1567 *
1568 * We will try this process 10 times, to make sure that we don't keep
1569 * spinning.
1570 */
1571 retries = 10;
1572 success = 0;
1573 while (retries--) {
1574 struct list_head *head_page, *prev_page, *r;
1575 struct list_head *last_page, *first_page;
1576 struct list_head *head_page_with_bit;
1577
1578 head_page = &rb_set_head_page(cpu_buffer)->list;
1579 if (!head_page)
1580 break;
1581 prev_page = head_page->prev;
1582
1583 first_page = pages->next;
1584 last_page = pages->prev;
1585
1586 head_page_with_bit = (struct list_head *)
1587 ((unsigned long)head_page | RB_PAGE_HEAD);
1588
1589 last_page->next = head_page_with_bit;
1590 first_page->prev = prev_page;
1591
1592 r = cmpxchg(&prev_page->next, head_page_with_bit, first_page);
1593
1594 if (r == head_page_with_bit) {
1595 /*
1596 * yay, we replaced the page pointer to our new list,
1597 * now, we just have to update to head page's prev
1598 * pointer to point to end of list
1599 */
1600 head_page->prev = last_page;
1601 success = 1;
1602 break;
1603 }
1604 }
1605
1606 if (success)
1607 INIT_LIST_HEAD(pages);
1608 /*
1609 * If we weren't successful in adding in new pages, warn and stop
1610 * tracing
1611 */
1612 RB_WARN_ON(cpu_buffer, !success);
1613 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1614
1615 /* free pages if they weren't inserted */
1616 if (!success) {
1617 struct buffer_page *bpage, *tmp;
1618 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1619 list) {
1620 list_del_init(&bpage->list);
1621 free_buffer_page(bpage);
1622 }
1623 }
1624 return success;
1625 }
1626
rb_update_pages(struct ring_buffer_per_cpu * cpu_buffer)1627 static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
1628 {
1629 int success;
1630
1631 if (cpu_buffer->nr_pages_to_update > 0)
1632 success = rb_insert_pages(cpu_buffer);
1633 else
1634 success = rb_remove_pages(cpu_buffer,
1635 -cpu_buffer->nr_pages_to_update);
1636
1637 if (success)
1638 cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
1639 }
1640
update_pages_handler(struct work_struct * work)1641 static void update_pages_handler(struct work_struct *work)
1642 {
1643 struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
1644 struct ring_buffer_per_cpu, update_pages_work);
1645 rb_update_pages(cpu_buffer);
1646 complete(&cpu_buffer->update_done);
1647 }
1648
1649 /**
1650 * ring_buffer_resize - resize the ring buffer
1651 * @buffer: the buffer to resize.
1652 * @size: the new size.
1653 * @cpu_id: the cpu buffer to resize
1654 *
1655 * Minimum size is 2 * BUF_PAGE_SIZE.
1656 *
1657 * Returns 0 on success and < 0 on failure.
1658 */
ring_buffer_resize(struct ring_buffer * buffer,unsigned long size,int cpu_id)1659 int ring_buffer_resize(struct ring_buffer *buffer, unsigned long size,
1660 int cpu_id)
1661 {
1662 struct ring_buffer_per_cpu *cpu_buffer;
1663 unsigned long nr_pages;
1664 int cpu, err;
1665
1666 /*
1667 * Always succeed at resizing a non-existent buffer:
1668 */
1669 if (!buffer)
1670 return 0;
1671
1672 /* Make sure the requested buffer exists */
1673 if (cpu_id != RING_BUFFER_ALL_CPUS &&
1674 !cpumask_test_cpu(cpu_id, buffer->cpumask))
1675 return 0;
1676
1677 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1678
1679 /* we need a minimum of two pages */
1680 if (nr_pages < 2)
1681 nr_pages = 2;
1682
1683 size = nr_pages * BUF_PAGE_SIZE;
1684
1685 /*
1686 * Don't succeed if resizing is disabled, as a reader might be
1687 * manipulating the ring buffer and is expecting a sane state while
1688 * this is true.
1689 */
1690 if (atomic_read(&buffer->resize_disabled))
1691 return -EBUSY;
1692
1693 /* prevent another thread from changing buffer sizes */
1694 mutex_lock(&buffer->mutex);
1695
1696 if (cpu_id == RING_BUFFER_ALL_CPUS) {
1697 /* calculate the pages to update */
1698 for_each_buffer_cpu(buffer, cpu) {
1699 cpu_buffer = buffer->buffers[cpu];
1700
1701 cpu_buffer->nr_pages_to_update = nr_pages -
1702 cpu_buffer->nr_pages;
1703 /*
1704 * nothing more to do for removing pages or no update
1705 */
1706 if (cpu_buffer->nr_pages_to_update <= 0)
1707 continue;
1708 /*
1709 * to add pages, make sure all new pages can be
1710 * allocated without receiving ENOMEM
1711 */
1712 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1713 if (__rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1714 &cpu_buffer->new_pages, cpu)) {
1715 /* not enough memory for new pages */
1716 err = -ENOMEM;
1717 goto out_err;
1718 }
1719 }
1720
1721 get_online_cpus();
1722 /*
1723 * Fire off all the required work handlers
1724 * We can't schedule on offline CPUs, but it's not necessary
1725 * since we can change their buffer sizes without any race.
1726 */
1727 for_each_buffer_cpu(buffer, cpu) {
1728 cpu_buffer = buffer->buffers[cpu];
1729 if (!cpu_buffer->nr_pages_to_update)
1730 continue;
1731
1732 /* Can't run something on an offline CPU. */
1733 if (!cpu_online(cpu)) {
1734 rb_update_pages(cpu_buffer);
1735 cpu_buffer->nr_pages_to_update = 0;
1736 } else {
1737 schedule_work_on(cpu,
1738 &cpu_buffer->update_pages_work);
1739 }
1740 }
1741
1742 /* wait for all the updates to complete */
1743 for_each_buffer_cpu(buffer, cpu) {
1744 cpu_buffer = buffer->buffers[cpu];
1745 if (!cpu_buffer->nr_pages_to_update)
1746 continue;
1747
1748 if (cpu_online(cpu))
1749 wait_for_completion(&cpu_buffer->update_done);
1750 cpu_buffer->nr_pages_to_update = 0;
1751 }
1752
1753 put_online_cpus();
1754 } else {
1755 /* Make sure this CPU has been intitialized */
1756 if (!cpumask_test_cpu(cpu_id, buffer->cpumask))
1757 goto out;
1758
1759 cpu_buffer = buffer->buffers[cpu_id];
1760
1761 if (nr_pages == cpu_buffer->nr_pages)
1762 goto out;
1763
1764 cpu_buffer->nr_pages_to_update = nr_pages -
1765 cpu_buffer->nr_pages;
1766
1767 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1768 if (cpu_buffer->nr_pages_to_update > 0 &&
1769 __rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1770 &cpu_buffer->new_pages, cpu_id)) {
1771 err = -ENOMEM;
1772 goto out_err;
1773 }
1774
1775 get_online_cpus();
1776
1777 /* Can't run something on an offline CPU. */
1778 if (!cpu_online(cpu_id))
1779 rb_update_pages(cpu_buffer);
1780 else {
1781 schedule_work_on(cpu_id,
1782 &cpu_buffer->update_pages_work);
1783 wait_for_completion(&cpu_buffer->update_done);
1784 }
1785
1786 cpu_buffer->nr_pages_to_update = 0;
1787 put_online_cpus();
1788 }
1789
1790 out:
1791 /*
1792 * The ring buffer resize can happen with the ring buffer
1793 * enabled, so that the update disturbs the tracing as little
1794 * as possible. But if the buffer is disabled, we do not need
1795 * to worry about that, and we can take the time to verify
1796 * that the buffer is not corrupt.
1797 */
1798 if (atomic_read(&buffer->record_disabled)) {
1799 atomic_inc(&buffer->record_disabled);
1800 /*
1801 * Even though the buffer was disabled, we must make sure
1802 * that it is truly disabled before calling rb_check_pages.
1803 * There could have been a race between checking
1804 * record_disable and incrementing it.
1805 */
1806 synchronize_sched();
1807 for_each_buffer_cpu(buffer, cpu) {
1808 cpu_buffer = buffer->buffers[cpu];
1809 rb_check_pages(cpu_buffer);
1810 }
1811 atomic_dec(&buffer->record_disabled);
1812 }
1813
1814 mutex_unlock(&buffer->mutex);
1815 return 0;
1816
1817 out_err:
1818 for_each_buffer_cpu(buffer, cpu) {
1819 struct buffer_page *bpage, *tmp;
1820
1821 cpu_buffer = buffer->buffers[cpu];
1822 cpu_buffer->nr_pages_to_update = 0;
1823
1824 if (list_empty(&cpu_buffer->new_pages))
1825 continue;
1826
1827 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1828 list) {
1829 list_del_init(&bpage->list);
1830 free_buffer_page(bpage);
1831 }
1832 }
1833 mutex_unlock(&buffer->mutex);
1834 return err;
1835 }
1836 EXPORT_SYMBOL_GPL(ring_buffer_resize);
1837
ring_buffer_change_overwrite(struct ring_buffer * buffer,int val)1838 void ring_buffer_change_overwrite(struct ring_buffer *buffer, int val)
1839 {
1840 mutex_lock(&buffer->mutex);
1841 if (val)
1842 buffer->flags |= RB_FL_OVERWRITE;
1843 else
1844 buffer->flags &= ~RB_FL_OVERWRITE;
1845 mutex_unlock(&buffer->mutex);
1846 }
1847 EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
1848
1849 static inline void *
__rb_data_page_index(struct buffer_data_page * bpage,unsigned index)1850 __rb_data_page_index(struct buffer_data_page *bpage, unsigned index)
1851 {
1852 return bpage->data + index;
1853 }
1854
__rb_page_index(struct buffer_page * bpage,unsigned index)1855 static inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
1856 {
1857 return bpage->page->data + index;
1858 }
1859
1860 static inline struct ring_buffer_event *
rb_reader_event(struct ring_buffer_per_cpu * cpu_buffer)1861 rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
1862 {
1863 return __rb_page_index(cpu_buffer->reader_page,
1864 cpu_buffer->reader_page->read);
1865 }
1866
1867 static inline struct ring_buffer_event *
rb_iter_head_event(struct ring_buffer_iter * iter)1868 rb_iter_head_event(struct ring_buffer_iter *iter)
1869 {
1870 return __rb_page_index(iter->head_page, iter->head);
1871 }
1872
rb_page_commit(struct buffer_page * bpage)1873 static inline unsigned rb_page_commit(struct buffer_page *bpage)
1874 {
1875 return local_read(&bpage->page->commit);
1876 }
1877
1878 /* Size is determined by what has been committed */
rb_page_size(struct buffer_page * bpage)1879 static inline unsigned rb_page_size(struct buffer_page *bpage)
1880 {
1881 return rb_page_commit(bpage);
1882 }
1883
1884 static inline unsigned
rb_commit_index(struct ring_buffer_per_cpu * cpu_buffer)1885 rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
1886 {
1887 return rb_page_commit(cpu_buffer->commit_page);
1888 }
1889
1890 static inline unsigned
rb_event_index(struct ring_buffer_event * event)1891 rb_event_index(struct ring_buffer_event *event)
1892 {
1893 unsigned long addr = (unsigned long)event;
1894
1895 return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
1896 }
1897
rb_inc_iter(struct ring_buffer_iter * iter)1898 static void rb_inc_iter(struct ring_buffer_iter *iter)
1899 {
1900 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
1901
1902 /*
1903 * The iterator could be on the reader page (it starts there).
1904 * But the head could have moved, since the reader was
1905 * found. Check for this case and assign the iterator
1906 * to the head page instead of next.
1907 */
1908 if (iter->head_page == cpu_buffer->reader_page)
1909 iter->head_page = rb_set_head_page(cpu_buffer);
1910 else
1911 rb_inc_page(cpu_buffer, &iter->head_page);
1912
1913 iter->read_stamp = iter->head_page->page->time_stamp;
1914 iter->head = 0;
1915 }
1916
1917 /*
1918 * rb_handle_head_page - writer hit the head page
1919 *
1920 * Returns: +1 to retry page
1921 * 0 to continue
1922 * -1 on error
1923 */
1924 static int
rb_handle_head_page(struct ring_buffer_per_cpu * cpu_buffer,struct buffer_page * tail_page,struct buffer_page * next_page)1925 rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
1926 struct buffer_page *tail_page,
1927 struct buffer_page *next_page)
1928 {
1929 struct buffer_page *new_head;
1930 int entries;
1931 int type;
1932 int ret;
1933
1934 entries = rb_page_entries(next_page);
1935
1936 /*
1937 * The hard part is here. We need to move the head
1938 * forward, and protect against both readers on
1939 * other CPUs and writers coming in via interrupts.
1940 */
1941 type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
1942 RB_PAGE_HEAD);
1943
1944 /*
1945 * type can be one of four:
1946 * NORMAL - an interrupt already moved it for us
1947 * HEAD - we are the first to get here.
1948 * UPDATE - we are the interrupt interrupting
1949 * a current move.
1950 * MOVED - a reader on another CPU moved the next
1951 * pointer to its reader page. Give up
1952 * and try again.
1953 */
1954
1955 switch (type) {
1956 case RB_PAGE_HEAD:
1957 /*
1958 * We changed the head to UPDATE, thus
1959 * it is our responsibility to update
1960 * the counters.
1961 */
1962 local_add(entries, &cpu_buffer->overrun);
1963 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1964
1965 /*
1966 * The entries will be zeroed out when we move the
1967 * tail page.
1968 */
1969
1970 /* still more to do */
1971 break;
1972
1973 case RB_PAGE_UPDATE:
1974 /*
1975 * This is an interrupt that interrupt the
1976 * previous update. Still more to do.
1977 */
1978 break;
1979 case RB_PAGE_NORMAL:
1980 /*
1981 * An interrupt came in before the update
1982 * and processed this for us.
1983 * Nothing left to do.
1984 */
1985 return 1;
1986 case RB_PAGE_MOVED:
1987 /*
1988 * The reader is on another CPU and just did
1989 * a swap with our next_page.
1990 * Try again.
1991 */
1992 return 1;
1993 default:
1994 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
1995 return -1;
1996 }
1997
1998 /*
1999 * Now that we are here, the old head pointer is
2000 * set to UPDATE. This will keep the reader from
2001 * swapping the head page with the reader page.
2002 * The reader (on another CPU) will spin till
2003 * we are finished.
2004 *
2005 * We just need to protect against interrupts
2006 * doing the job. We will set the next pointer
2007 * to HEAD. After that, we set the old pointer
2008 * to NORMAL, but only if it was HEAD before.
2009 * otherwise we are an interrupt, and only
2010 * want the outer most commit to reset it.
2011 */
2012 new_head = next_page;
2013 rb_inc_page(cpu_buffer, &new_head);
2014
2015 ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
2016 RB_PAGE_NORMAL);
2017
2018 /*
2019 * Valid returns are:
2020 * HEAD - an interrupt came in and already set it.
2021 * NORMAL - One of two things:
2022 * 1) We really set it.
2023 * 2) A bunch of interrupts came in and moved
2024 * the page forward again.
2025 */
2026 switch (ret) {
2027 case RB_PAGE_HEAD:
2028 case RB_PAGE_NORMAL:
2029 /* OK */
2030 break;
2031 default:
2032 RB_WARN_ON(cpu_buffer, 1);
2033 return -1;
2034 }
2035
2036 /*
2037 * It is possible that an interrupt came in,
2038 * set the head up, then more interrupts came in
2039 * and moved it again. When we get back here,
2040 * the page would have been set to NORMAL but we
2041 * just set it back to HEAD.
2042 *
2043 * How do you detect this? Well, if that happened
2044 * the tail page would have moved.
2045 */
2046 if (ret == RB_PAGE_NORMAL) {
2047 /*
2048 * If the tail had moved passed next, then we need
2049 * to reset the pointer.
2050 */
2051 if (cpu_buffer->tail_page != tail_page &&
2052 cpu_buffer->tail_page != next_page)
2053 rb_head_page_set_normal(cpu_buffer, new_head,
2054 next_page,
2055 RB_PAGE_HEAD);
2056 }
2057
2058 /*
2059 * If this was the outer most commit (the one that
2060 * changed the original pointer from HEAD to UPDATE),
2061 * then it is up to us to reset it to NORMAL.
2062 */
2063 if (type == RB_PAGE_HEAD) {
2064 ret = rb_head_page_set_normal(cpu_buffer, next_page,
2065 tail_page,
2066 RB_PAGE_UPDATE);
2067 if (RB_WARN_ON(cpu_buffer,
2068 ret != RB_PAGE_UPDATE))
2069 return -1;
2070 }
2071
2072 return 0;
2073 }
2074
2075 static inline void
rb_reset_tail(struct ring_buffer_per_cpu * cpu_buffer,unsigned long tail,struct rb_event_info * info)2076 rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2077 unsigned long tail, struct rb_event_info *info)
2078 {
2079 struct buffer_page *tail_page = info->tail_page;
2080 struct ring_buffer_event *event;
2081 unsigned long length = info->length;
2082
2083 /*
2084 * Only the event that crossed the page boundary
2085 * must fill the old tail_page with padding.
2086 */
2087 if (tail >= BUF_PAGE_SIZE) {
2088 /*
2089 * If the page was filled, then we still need
2090 * to update the real_end. Reset it to zero
2091 * and the reader will ignore it.
2092 */
2093 if (tail == BUF_PAGE_SIZE)
2094 tail_page->real_end = 0;
2095
2096 local_sub(length, &tail_page->write);
2097 return;
2098 }
2099
2100 event = __rb_page_index(tail_page, tail);
2101 kmemcheck_annotate_bitfield(event, bitfield);
2102
2103 /* account for padding bytes */
2104 local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes);
2105
2106 /*
2107 * Save the original length to the meta data.
2108 * This will be used by the reader to add lost event
2109 * counter.
2110 */
2111 tail_page->real_end = tail;
2112
2113 /*
2114 * If this event is bigger than the minimum size, then
2115 * we need to be careful that we don't subtract the
2116 * write counter enough to allow another writer to slip
2117 * in on this page.
2118 * We put in a discarded commit instead, to make sure
2119 * that this space is not used again.
2120 *
2121 * If we are less than the minimum size, we don't need to
2122 * worry about it.
2123 */
2124 if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
2125 /* No room for any events */
2126
2127 /* Mark the rest of the page with padding */
2128 rb_event_set_padding(event);
2129
2130 /* Set the write back to the previous setting */
2131 local_sub(length, &tail_page->write);
2132 return;
2133 }
2134
2135 /* Put in a discarded event */
2136 event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
2137 event->type_len = RINGBUF_TYPE_PADDING;
2138 /* time delta must be non zero */
2139 event->time_delta = 1;
2140
2141 /* Set write to end of buffer */
2142 length = (tail + length) - BUF_PAGE_SIZE;
2143 local_sub(length, &tail_page->write);
2144 }
2145
2146 /*
2147 * This is the slow path, force gcc not to inline it.
2148 */
2149 static noinline struct ring_buffer_event *
rb_move_tail(struct ring_buffer_per_cpu * cpu_buffer,unsigned long tail,struct rb_event_info * info)2150 rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2151 unsigned long tail, struct rb_event_info *info)
2152 {
2153 struct buffer_page *tail_page = info->tail_page;
2154 struct buffer_page *commit_page = cpu_buffer->commit_page;
2155 struct ring_buffer *buffer = cpu_buffer->buffer;
2156 struct buffer_page *next_page;
2157 int ret;
2158 u64 ts;
2159
2160 next_page = tail_page;
2161
2162 rb_inc_page(cpu_buffer, &next_page);
2163
2164 /*
2165 * If for some reason, we had an interrupt storm that made
2166 * it all the way around the buffer, bail, and warn
2167 * about it.
2168 */
2169 if (unlikely(next_page == commit_page)) {
2170 local_inc(&cpu_buffer->commit_overrun);
2171 goto out_reset;
2172 }
2173
2174 /*
2175 * This is where the fun begins!
2176 *
2177 * We are fighting against races between a reader that
2178 * could be on another CPU trying to swap its reader
2179 * page with the buffer head.
2180 *
2181 * We are also fighting against interrupts coming in and
2182 * moving the head or tail on us as well.
2183 *
2184 * If the next page is the head page then we have filled
2185 * the buffer, unless the commit page is still on the
2186 * reader page.
2187 */
2188 if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) {
2189
2190 /*
2191 * If the commit is not on the reader page, then
2192 * move the header page.
2193 */
2194 if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2195 /*
2196 * If we are not in overwrite mode,
2197 * this is easy, just stop here.
2198 */
2199 if (!(buffer->flags & RB_FL_OVERWRITE)) {
2200 local_inc(&cpu_buffer->dropped_events);
2201 goto out_reset;
2202 }
2203
2204 ret = rb_handle_head_page(cpu_buffer,
2205 tail_page,
2206 next_page);
2207 if (ret < 0)
2208 goto out_reset;
2209 if (ret)
2210 goto out_again;
2211 } else {
2212 /*
2213 * We need to be careful here too. The
2214 * commit page could still be on the reader
2215 * page. We could have a small buffer, and
2216 * have filled up the buffer with events
2217 * from interrupts and such, and wrapped.
2218 *
2219 * Note, if the tail page is also the on the
2220 * reader_page, we let it move out.
2221 */
2222 if (unlikely((cpu_buffer->commit_page !=
2223 cpu_buffer->tail_page) &&
2224 (cpu_buffer->commit_page ==
2225 cpu_buffer->reader_page))) {
2226 local_inc(&cpu_buffer->commit_overrun);
2227 goto out_reset;
2228 }
2229 }
2230 }
2231
2232 ret = rb_tail_page_update(cpu_buffer, tail_page, next_page);
2233 if (ret) {
2234 /*
2235 * Nested commits always have zero deltas, so
2236 * just reread the time stamp
2237 */
2238 ts = rb_time_stamp(buffer);
2239 next_page->page->time_stamp = ts;
2240 }
2241
2242 out_again:
2243
2244 rb_reset_tail(cpu_buffer, tail, info);
2245
2246 /* fail and let the caller try again */
2247 return ERR_PTR(-EAGAIN);
2248
2249 out_reset:
2250 /* reset write */
2251 rb_reset_tail(cpu_buffer, tail, info);
2252
2253 return NULL;
2254 }
2255
2256 /* Slow path, do not inline */
2257 static noinline struct ring_buffer_event *
rb_add_time_stamp(struct ring_buffer_event * event,u64 delta)2258 rb_add_time_stamp(struct ring_buffer_event *event, u64 delta)
2259 {
2260 event->type_len = RINGBUF_TYPE_TIME_EXTEND;
2261
2262 /* Not the first event on the page? */
2263 if (rb_event_index(event)) {
2264 event->time_delta = delta & TS_MASK;
2265 event->array[0] = delta >> TS_SHIFT;
2266 } else {
2267 /* nope, just zero it */
2268 event->time_delta = 0;
2269 event->array[0] = 0;
2270 }
2271
2272 return skip_time_extend(event);
2273 }
2274
2275 static inline bool rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2276 struct ring_buffer_event *event);
2277
2278 /**
2279 * rb_update_event - update event type and data
2280 * @event: the event to update
2281 * @type: the type of event
2282 * @length: the size of the event field in the ring buffer
2283 *
2284 * Update the type and data fields of the event. The length
2285 * is the actual size that is written to the ring buffer,
2286 * and with this, we can determine what to place into the
2287 * data field.
2288 */
2289 static void
rb_update_event(struct ring_buffer_per_cpu * cpu_buffer,struct ring_buffer_event * event,struct rb_event_info * info)2290 rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
2291 struct ring_buffer_event *event,
2292 struct rb_event_info *info)
2293 {
2294 unsigned length = info->length;
2295 u64 delta = info->delta;
2296
2297 /* Only a commit updates the timestamp */
2298 if (unlikely(!rb_event_is_commit(cpu_buffer, event)))
2299 delta = 0;
2300
2301 /*
2302 * If we need to add a timestamp, then we
2303 * add it to the start of the resevered space.
2304 */
2305 if (unlikely(info->add_timestamp)) {
2306 event = rb_add_time_stamp(event, delta);
2307 length -= RB_LEN_TIME_EXTEND;
2308 delta = 0;
2309 }
2310
2311 event->time_delta = delta;
2312 length -= RB_EVNT_HDR_SIZE;
2313 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
2314 event->type_len = 0;
2315 event->array[0] = length;
2316 } else
2317 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
2318 }
2319
rb_calculate_event_length(unsigned length)2320 static unsigned rb_calculate_event_length(unsigned length)
2321 {
2322 struct ring_buffer_event event; /* Used only for sizeof array */
2323
2324 /* zero length can cause confusions */
2325 if (!length)
2326 length++;
2327
2328 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
2329 length += sizeof(event.array[0]);
2330
2331 length += RB_EVNT_HDR_SIZE;
2332 length = ALIGN(length, RB_ARCH_ALIGNMENT);
2333
2334 /*
2335 * In case the time delta is larger than the 27 bits for it
2336 * in the header, we need to add a timestamp. If another
2337 * event comes in when trying to discard this one to increase
2338 * the length, then the timestamp will be added in the allocated
2339 * space of this event. If length is bigger than the size needed
2340 * for the TIME_EXTEND, then padding has to be used. The events
2341 * length must be either RB_LEN_TIME_EXTEND, or greater than or equal
2342 * to RB_LEN_TIME_EXTEND + 8, as 8 is the minimum size for padding.
2343 * As length is a multiple of 4, we only need to worry if it
2344 * is 12 (RB_LEN_TIME_EXTEND + 4).
2345 */
2346 if (length == RB_LEN_TIME_EXTEND + RB_ALIGNMENT)
2347 length += RB_ALIGNMENT;
2348
2349 return length;
2350 }
2351
2352 #ifndef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
sched_clock_stable(void)2353 static inline bool sched_clock_stable(void)
2354 {
2355 return true;
2356 }
2357 #endif
2358
2359 static inline int
rb_try_to_discard(struct ring_buffer_per_cpu * cpu_buffer,struct ring_buffer_event * event)2360 rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2361 struct ring_buffer_event *event)
2362 {
2363 unsigned long new_index, old_index;
2364 struct buffer_page *bpage;
2365 unsigned long index;
2366 unsigned long addr;
2367
2368 new_index = rb_event_index(event);
2369 old_index = new_index + rb_event_ts_length(event);
2370 addr = (unsigned long)event;
2371 addr &= PAGE_MASK;
2372
2373 bpage = cpu_buffer->tail_page;
2374
2375 if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2376 unsigned long write_mask =
2377 local_read(&bpage->write) & ~RB_WRITE_MASK;
2378 unsigned long event_length = rb_event_length(event);
2379 /*
2380 * This is on the tail page. It is possible that
2381 * a write could come in and move the tail page
2382 * and write to the next page. That is fine
2383 * because we just shorten what is on this page.
2384 */
2385 old_index += write_mask;
2386 new_index += write_mask;
2387 index = local_cmpxchg(&bpage->write, old_index, new_index);
2388 if (index == old_index) {
2389 /* update counters */
2390 local_sub(event_length, &cpu_buffer->entries_bytes);
2391 return 1;
2392 }
2393 }
2394
2395 /* could not discard */
2396 return 0;
2397 }
2398
rb_start_commit(struct ring_buffer_per_cpu * cpu_buffer)2399 static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2400 {
2401 local_inc(&cpu_buffer->committing);
2402 local_inc(&cpu_buffer->commits);
2403 }
2404
2405 static void
rb_set_commit_to_write(struct ring_buffer_per_cpu * cpu_buffer)2406 rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
2407 {
2408 unsigned long max_count;
2409
2410 /*
2411 * We only race with interrupts and NMIs on this CPU.
2412 * If we own the commit event, then we can commit
2413 * all others that interrupted us, since the interruptions
2414 * are in stack format (they finish before they come
2415 * back to us). This allows us to do a simple loop to
2416 * assign the commit to the tail.
2417 */
2418 again:
2419 max_count = cpu_buffer->nr_pages * 100;
2420
2421 while (cpu_buffer->commit_page != cpu_buffer->tail_page) {
2422 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
2423 return;
2424 if (RB_WARN_ON(cpu_buffer,
2425 rb_is_reader_page(cpu_buffer->tail_page)))
2426 return;
2427 local_set(&cpu_buffer->commit_page->page->commit,
2428 rb_page_write(cpu_buffer->commit_page));
2429 rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
2430 cpu_buffer->write_stamp =
2431 cpu_buffer->commit_page->page->time_stamp;
2432 /* add barrier to keep gcc from optimizing too much */
2433 barrier();
2434 }
2435 while (rb_commit_index(cpu_buffer) !=
2436 rb_page_write(cpu_buffer->commit_page)) {
2437
2438 local_set(&cpu_buffer->commit_page->page->commit,
2439 rb_page_write(cpu_buffer->commit_page));
2440 RB_WARN_ON(cpu_buffer,
2441 local_read(&cpu_buffer->commit_page->page->commit) &
2442 ~RB_WRITE_MASK);
2443 barrier();
2444 }
2445
2446 /* again, keep gcc from optimizing */
2447 barrier();
2448
2449 /*
2450 * If an interrupt came in just after the first while loop
2451 * and pushed the tail page forward, we will be left with
2452 * a dangling commit that will never go forward.
2453 */
2454 if (unlikely(cpu_buffer->commit_page != cpu_buffer->tail_page))
2455 goto again;
2456 }
2457
rb_end_commit(struct ring_buffer_per_cpu * cpu_buffer)2458 static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
2459 {
2460 unsigned long commits;
2461
2462 if (RB_WARN_ON(cpu_buffer,
2463 !local_read(&cpu_buffer->committing)))
2464 return;
2465
2466 again:
2467 commits = local_read(&cpu_buffer->commits);
2468 /* synchronize with interrupts */
2469 barrier();
2470 if (local_read(&cpu_buffer->committing) == 1)
2471 rb_set_commit_to_write(cpu_buffer);
2472
2473 local_dec(&cpu_buffer->committing);
2474
2475 /* synchronize with interrupts */
2476 barrier();
2477
2478 /*
2479 * Need to account for interrupts coming in between the
2480 * updating of the commit page and the clearing of the
2481 * committing counter.
2482 */
2483 if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
2484 !local_read(&cpu_buffer->committing)) {
2485 local_inc(&cpu_buffer->committing);
2486 goto again;
2487 }
2488 }
2489
rb_event_discard(struct ring_buffer_event * event)2490 static inline void rb_event_discard(struct ring_buffer_event *event)
2491 {
2492 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
2493 event = skip_time_extend(event);
2494
2495 /* array[0] holds the actual length for the discarded event */
2496 event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
2497 event->type_len = RINGBUF_TYPE_PADDING;
2498 /* time delta must be non zero */
2499 if (!event->time_delta)
2500 event->time_delta = 1;
2501 }
2502
2503 static inline bool
rb_event_is_commit(struct ring_buffer_per_cpu * cpu_buffer,struct ring_buffer_event * event)2504 rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2505 struct ring_buffer_event *event)
2506 {
2507 unsigned long addr = (unsigned long)event;
2508 unsigned long index;
2509
2510 index = rb_event_index(event);
2511 addr &= PAGE_MASK;
2512
2513 return cpu_buffer->commit_page->page == (void *)addr &&
2514 rb_commit_index(cpu_buffer) == index;
2515 }
2516
2517 static void
rb_update_write_stamp(struct ring_buffer_per_cpu * cpu_buffer,struct ring_buffer_event * event)2518 rb_update_write_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2519 struct ring_buffer_event *event)
2520 {
2521 u64 delta;
2522
2523 /*
2524 * The event first in the commit queue updates the
2525 * time stamp.
2526 */
2527 if (rb_event_is_commit(cpu_buffer, event)) {
2528 /*
2529 * A commit event that is first on a page
2530 * updates the write timestamp with the page stamp
2531 */
2532 if (!rb_event_index(event))
2533 cpu_buffer->write_stamp =
2534 cpu_buffer->commit_page->page->time_stamp;
2535 else if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
2536 delta = event->array[0];
2537 delta <<= TS_SHIFT;
2538 delta += event->time_delta;
2539 cpu_buffer->write_stamp += delta;
2540 } else
2541 cpu_buffer->write_stamp += event->time_delta;
2542 }
2543 }
2544
rb_commit(struct ring_buffer_per_cpu * cpu_buffer,struct ring_buffer_event * event)2545 static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
2546 struct ring_buffer_event *event)
2547 {
2548 local_inc(&cpu_buffer->entries);
2549 rb_update_write_stamp(cpu_buffer, event);
2550 rb_end_commit(cpu_buffer);
2551 }
2552
2553 static __always_inline void
rb_wakeups(struct ring_buffer * buffer,struct ring_buffer_per_cpu * cpu_buffer)2554 rb_wakeups(struct ring_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer)
2555 {
2556 bool pagebusy;
2557
2558 if (buffer->irq_work.waiters_pending) {
2559 buffer->irq_work.waiters_pending = false;
2560 /* irq_work_queue() supplies it's own memory barriers */
2561 irq_work_queue(&buffer->irq_work.work);
2562 }
2563
2564 if (cpu_buffer->irq_work.waiters_pending) {
2565 cpu_buffer->irq_work.waiters_pending = false;
2566 /* irq_work_queue() supplies it's own memory barriers */
2567 irq_work_queue(&cpu_buffer->irq_work.work);
2568 }
2569
2570 pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
2571
2572 if (!pagebusy && cpu_buffer->irq_work.full_waiters_pending) {
2573 cpu_buffer->irq_work.wakeup_full = true;
2574 cpu_buffer->irq_work.full_waiters_pending = false;
2575 /* irq_work_queue() supplies it's own memory barriers */
2576 irq_work_queue(&cpu_buffer->irq_work.work);
2577 }
2578 }
2579
2580 /*
2581 * The lock and unlock are done within a preempt disable section.
2582 * The current_context per_cpu variable can only be modified
2583 * by the current task between lock and unlock. But it can
2584 * be modified more than once via an interrupt. To pass this
2585 * information from the lock to the unlock without having to
2586 * access the 'in_interrupt()' functions again (which do show
2587 * a bit of overhead in something as critical as function tracing,
2588 * we use a bitmask trick.
2589 *
2590 * bit 1 = NMI context
2591 * bit 2 = IRQ context
2592 * bit 3 = SoftIRQ context
2593 * bit 4 = normal context.
2594 *
2595 * This works because this is the order of contexts that can
2596 * preempt other contexts. A SoftIRQ never preempts an IRQ
2597 * context.
2598 *
2599 * When the context is determined, the corresponding bit is
2600 * checked and set (if it was set, then a recursion of that context
2601 * happened).
2602 *
2603 * On unlock, we need to clear this bit. To do so, just subtract
2604 * 1 from the current_context and AND it to itself.
2605 *
2606 * (binary)
2607 * 101 - 1 = 100
2608 * 101 & 100 = 100 (clearing bit zero)
2609 *
2610 * 1010 - 1 = 1001
2611 * 1010 & 1001 = 1000 (clearing bit 1)
2612 *
2613 * The least significant bit can be cleared this way, and it
2614 * just so happens that it is the same bit corresponding to
2615 * the current context.
2616 *
2617 * Now the TRANSITION bit breaks the above slightly. The TRANSITION bit
2618 * is set when a recursion is detected at the current context, and if
2619 * the TRANSITION bit is already set, it will fail the recursion.
2620 * This is needed because there's a lag between the changing of
2621 * interrupt context and updating the preempt count. In this case,
2622 * a false positive will be found. To handle this, one extra recursion
2623 * is allowed, and this is done by the TRANSITION bit. If the TRANSITION
2624 * bit is already set, then it is considered a recursion and the function
2625 * ends. Otherwise, the TRANSITION bit is set, and that bit is returned.
2626 *
2627 * On the trace_recursive_unlock(), the TRANSITION bit will be the first
2628 * to be cleared. Even if it wasn't the context that set it. That is,
2629 * if an interrupt comes in while NORMAL bit is set and the ring buffer
2630 * is called before preempt_count() is updated, since the check will
2631 * be on the NORMAL bit, the TRANSITION bit will then be set. If an
2632 * NMI then comes in, it will set the NMI bit, but when the NMI code
2633 * does the trace_recursive_unlock() it will clear the TRANSTION bit
2634 * and leave the NMI bit set. But this is fine, because the interrupt
2635 * code that set the TRANSITION bit will then clear the NMI bit when it
2636 * calls trace_recursive_unlock(). If another NMI comes in, it will
2637 * set the TRANSITION bit and continue.
2638 *
2639 * Note: The TRANSITION bit only handles a single transition between context.
2640 */
2641
2642 static __always_inline int
trace_recursive_lock(struct ring_buffer_per_cpu * cpu_buffer)2643 trace_recursive_lock(struct ring_buffer_per_cpu *cpu_buffer)
2644 {
2645 unsigned int val = cpu_buffer->current_context;
2646 int bit;
2647
2648 if (in_interrupt()) {
2649 if (in_nmi())
2650 bit = RB_CTX_NMI;
2651 else if (in_irq())
2652 bit = RB_CTX_IRQ;
2653 else
2654 bit = RB_CTX_SOFTIRQ;
2655 } else
2656 bit = RB_CTX_NORMAL;
2657
2658 if (unlikely(val & (1 << bit))) {
2659 /*
2660 * It is possible that this was called by transitioning
2661 * between interrupt context, and preempt_count() has not
2662 * been updated yet. In this case, use the TRANSITION bit.
2663 */
2664 bit = RB_CTX_TRANSITION;
2665 if (val & (1 << bit))
2666 return 1;
2667 }
2668
2669 val |= (1 << bit);
2670 cpu_buffer->current_context = val;
2671
2672 return 0;
2673 }
2674
2675 static __always_inline void
trace_recursive_unlock(struct ring_buffer_per_cpu * cpu_buffer)2676 trace_recursive_unlock(struct ring_buffer_per_cpu *cpu_buffer)
2677 {
2678 cpu_buffer->current_context &= cpu_buffer->current_context - 1;
2679 }
2680
2681 /**
2682 * ring_buffer_unlock_commit - commit a reserved
2683 * @buffer: The buffer to commit to
2684 * @event: The event pointer to commit.
2685 *
2686 * This commits the data to the ring buffer, and releases any locks held.
2687 *
2688 * Must be paired with ring_buffer_lock_reserve.
2689 */
ring_buffer_unlock_commit(struct ring_buffer * buffer,struct ring_buffer_event * event)2690 int ring_buffer_unlock_commit(struct ring_buffer *buffer,
2691 struct ring_buffer_event *event)
2692 {
2693 struct ring_buffer_per_cpu *cpu_buffer;
2694 int cpu = raw_smp_processor_id();
2695
2696 cpu_buffer = buffer->buffers[cpu];
2697
2698 rb_commit(cpu_buffer, event);
2699
2700 rb_wakeups(buffer, cpu_buffer);
2701
2702 trace_recursive_unlock(cpu_buffer);
2703
2704 preempt_enable_notrace();
2705
2706 return 0;
2707 }
2708 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
2709
2710 static noinline void
rb_handle_timestamp(struct ring_buffer_per_cpu * cpu_buffer,struct rb_event_info * info)2711 rb_handle_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2712 struct rb_event_info *info)
2713 {
2714 WARN_ONCE(info->delta > (1ULL << 59),
2715 KERN_WARNING "Delta way too big! %llu ts=%llu write stamp = %llu\n%s",
2716 (unsigned long long)info->delta,
2717 (unsigned long long)info->ts,
2718 (unsigned long long)cpu_buffer->write_stamp,
2719 sched_clock_stable() ? "" :
2720 "If you just came from a suspend/resume,\n"
2721 "please switch to the trace global clock:\n"
2722 " echo global > /sys/kernel/debug/tracing/trace_clock\n");
2723 info->add_timestamp = 1;
2724 }
2725
2726 static struct ring_buffer_event *
__rb_reserve_next(struct ring_buffer_per_cpu * cpu_buffer,struct rb_event_info * info)2727 __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
2728 struct rb_event_info *info)
2729 {
2730 struct ring_buffer_event *event;
2731 struct buffer_page *tail_page;
2732 unsigned long tail, write;
2733
2734 /*
2735 * If the time delta since the last event is too big to
2736 * hold in the time field of the event, then we append a
2737 * TIME EXTEND event ahead of the data event.
2738 */
2739 if (unlikely(info->add_timestamp))
2740 info->length += RB_LEN_TIME_EXTEND;
2741
2742 tail_page = info->tail_page = cpu_buffer->tail_page;
2743 write = local_add_return(info->length, &tail_page->write);
2744
2745 /* set write to only the index of the write */
2746 write &= RB_WRITE_MASK;
2747 tail = write - info->length;
2748
2749 /*
2750 * If this is the first commit on the page, then it has the same
2751 * timestamp as the page itself.
2752 */
2753 if (!tail)
2754 info->delta = 0;
2755
2756 /* See if we shot pass the end of this buffer page */
2757 if (unlikely(write > BUF_PAGE_SIZE))
2758 return rb_move_tail(cpu_buffer, tail, info);
2759
2760 /* We reserved something on the buffer */
2761
2762 event = __rb_page_index(tail_page, tail);
2763 kmemcheck_annotate_bitfield(event, bitfield);
2764 rb_update_event(cpu_buffer, event, info);
2765
2766 local_inc(&tail_page->entries);
2767
2768 /*
2769 * If this is the first commit on the page, then update
2770 * its timestamp.
2771 */
2772 if (!tail)
2773 tail_page->page->time_stamp = info->ts;
2774
2775 /* account for these added bytes */
2776 local_add(info->length, &cpu_buffer->entries_bytes);
2777
2778 return event;
2779 }
2780
2781 static struct ring_buffer_event *
rb_reserve_next_event(struct ring_buffer * buffer,struct ring_buffer_per_cpu * cpu_buffer,unsigned long length)2782 rb_reserve_next_event(struct ring_buffer *buffer,
2783 struct ring_buffer_per_cpu *cpu_buffer,
2784 unsigned long length)
2785 {
2786 struct ring_buffer_event *event;
2787 struct rb_event_info info;
2788 int nr_loops = 0;
2789 u64 diff;
2790
2791 rb_start_commit(cpu_buffer);
2792
2793 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2794 /*
2795 * Due to the ability to swap a cpu buffer from a buffer
2796 * it is possible it was swapped before we committed.
2797 * (committing stops a swap). We check for it here and
2798 * if it happened, we have to fail the write.
2799 */
2800 barrier();
2801 if (unlikely(ACCESS_ONCE(cpu_buffer->buffer) != buffer)) {
2802 local_dec(&cpu_buffer->committing);
2803 local_dec(&cpu_buffer->commits);
2804 return NULL;
2805 }
2806 #endif
2807
2808 info.length = rb_calculate_event_length(length);
2809 again:
2810 info.add_timestamp = 0;
2811 info.delta = 0;
2812
2813 /*
2814 * We allow for interrupts to reenter here and do a trace.
2815 * If one does, it will cause this original code to loop
2816 * back here. Even with heavy interrupts happening, this
2817 * should only happen a few times in a row. If this happens
2818 * 1000 times in a row, there must be either an interrupt
2819 * storm or we have something buggy.
2820 * Bail!
2821 */
2822 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
2823 goto out_fail;
2824
2825 info.ts = rb_time_stamp(cpu_buffer->buffer);
2826 diff = info.ts - cpu_buffer->write_stamp;
2827
2828 /* make sure this diff is calculated here */
2829 barrier();
2830
2831 /* Did the write stamp get updated already? */
2832 if (likely(info.ts >= cpu_buffer->write_stamp)) {
2833 info.delta = diff;
2834 if (unlikely(test_time_stamp(info.delta)))
2835 rb_handle_timestamp(cpu_buffer, &info);
2836 }
2837
2838 event = __rb_reserve_next(cpu_buffer, &info);
2839
2840 if (unlikely(PTR_ERR(event) == -EAGAIN)) {
2841 if (info.add_timestamp)
2842 info.length -= RB_LEN_TIME_EXTEND;
2843 goto again;
2844 }
2845
2846 if (!event)
2847 goto out_fail;
2848
2849 return event;
2850
2851 out_fail:
2852 rb_end_commit(cpu_buffer);
2853 return NULL;
2854 }
2855
2856 /**
2857 * ring_buffer_lock_reserve - reserve a part of the buffer
2858 * @buffer: the ring buffer to reserve from
2859 * @length: the length of the data to reserve (excluding event header)
2860 *
2861 * Returns a reseverd event on the ring buffer to copy directly to.
2862 * The user of this interface will need to get the body to write into
2863 * and can use the ring_buffer_event_data() interface.
2864 *
2865 * The length is the length of the data needed, not the event length
2866 * which also includes the event header.
2867 *
2868 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2869 * If NULL is returned, then nothing has been allocated or locked.
2870 */
2871 struct ring_buffer_event *
ring_buffer_lock_reserve(struct ring_buffer * buffer,unsigned long length)2872 ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length)
2873 {
2874 struct ring_buffer_per_cpu *cpu_buffer;
2875 struct ring_buffer_event *event;
2876 int cpu;
2877
2878 /* If we are tracing schedule, we don't want to recurse */
2879 preempt_disable_notrace();
2880
2881 if (unlikely(atomic_read(&buffer->record_disabled)))
2882 goto out;
2883
2884 cpu = raw_smp_processor_id();
2885
2886 if (unlikely(!cpumask_test_cpu(cpu, buffer->cpumask)))
2887 goto out;
2888
2889 cpu_buffer = buffer->buffers[cpu];
2890
2891 if (unlikely(atomic_read(&cpu_buffer->record_disabled)))
2892 goto out;
2893
2894 if (unlikely(length > BUF_MAX_DATA_SIZE))
2895 goto out;
2896
2897 if (unlikely(trace_recursive_lock(cpu_buffer)))
2898 goto out;
2899
2900 event = rb_reserve_next_event(buffer, cpu_buffer, length);
2901 if (!event)
2902 goto out_unlock;
2903
2904 return event;
2905
2906 out_unlock:
2907 trace_recursive_unlock(cpu_buffer);
2908 out:
2909 preempt_enable_notrace();
2910 return NULL;
2911 }
2912 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
2913
2914 /*
2915 * Decrement the entries to the page that an event is on.
2916 * The event does not even need to exist, only the pointer
2917 * to the page it is on. This may only be called before the commit
2918 * takes place.
2919 */
2920 static inline void
rb_decrement_entry(struct ring_buffer_per_cpu * cpu_buffer,struct ring_buffer_event * event)2921 rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
2922 struct ring_buffer_event *event)
2923 {
2924 unsigned long addr = (unsigned long)event;
2925 struct buffer_page *bpage = cpu_buffer->commit_page;
2926 struct buffer_page *start;
2927
2928 addr &= PAGE_MASK;
2929
2930 /* Do the likely case first */
2931 if (likely(bpage->page == (void *)addr)) {
2932 local_dec(&bpage->entries);
2933 return;
2934 }
2935
2936 /*
2937 * Because the commit page may be on the reader page we
2938 * start with the next page and check the end loop there.
2939 */
2940 rb_inc_page(cpu_buffer, &bpage);
2941 start = bpage;
2942 do {
2943 if (bpage->page == (void *)addr) {
2944 local_dec(&bpage->entries);
2945 return;
2946 }
2947 rb_inc_page(cpu_buffer, &bpage);
2948 } while (bpage != start);
2949
2950 /* commit not part of this buffer?? */
2951 RB_WARN_ON(cpu_buffer, 1);
2952 }
2953
2954 /**
2955 * ring_buffer_commit_discard - discard an event that has not been committed
2956 * @buffer: the ring buffer
2957 * @event: non committed event to discard
2958 *
2959 * Sometimes an event that is in the ring buffer needs to be ignored.
2960 * This function lets the user discard an event in the ring buffer
2961 * and then that event will not be read later.
2962 *
2963 * This function only works if it is called before the the item has been
2964 * committed. It will try to free the event from the ring buffer
2965 * if another event has not been added behind it.
2966 *
2967 * If another event has been added behind it, it will set the event
2968 * up as discarded, and perform the commit.
2969 *
2970 * If this function is called, do not call ring_buffer_unlock_commit on
2971 * the event.
2972 */
ring_buffer_discard_commit(struct ring_buffer * buffer,struct ring_buffer_event * event)2973 void ring_buffer_discard_commit(struct ring_buffer *buffer,
2974 struct ring_buffer_event *event)
2975 {
2976 struct ring_buffer_per_cpu *cpu_buffer;
2977 int cpu;
2978
2979 /* The event is discarded regardless */
2980 rb_event_discard(event);
2981
2982 cpu = smp_processor_id();
2983 cpu_buffer = buffer->buffers[cpu];
2984
2985 /*
2986 * This must only be called if the event has not been
2987 * committed yet. Thus we can assume that preemption
2988 * is still disabled.
2989 */
2990 RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
2991
2992 rb_decrement_entry(cpu_buffer, event);
2993 if (rb_try_to_discard(cpu_buffer, event))
2994 goto out;
2995
2996 /*
2997 * The commit is still visible by the reader, so we
2998 * must still update the timestamp.
2999 */
3000 rb_update_write_stamp(cpu_buffer, event);
3001 out:
3002 rb_end_commit(cpu_buffer);
3003
3004 trace_recursive_unlock(cpu_buffer);
3005
3006 preempt_enable_notrace();
3007
3008 }
3009 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
3010
3011 /**
3012 * ring_buffer_write - write data to the buffer without reserving
3013 * @buffer: The ring buffer to write to.
3014 * @length: The length of the data being written (excluding the event header)
3015 * @data: The data to write to the buffer.
3016 *
3017 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
3018 * one function. If you already have the data to write to the buffer, it
3019 * may be easier to simply call this function.
3020 *
3021 * Note, like ring_buffer_lock_reserve, the length is the length of the data
3022 * and not the length of the event which would hold the header.
3023 */
ring_buffer_write(struct ring_buffer * buffer,unsigned long length,void * data)3024 int ring_buffer_write(struct ring_buffer *buffer,
3025 unsigned long length,
3026 void *data)
3027 {
3028 struct ring_buffer_per_cpu *cpu_buffer;
3029 struct ring_buffer_event *event;
3030 void *body;
3031 int ret = -EBUSY;
3032 int cpu;
3033
3034 preempt_disable_notrace();
3035
3036 if (atomic_read(&buffer->record_disabled))
3037 goto out;
3038
3039 cpu = raw_smp_processor_id();
3040
3041 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3042 goto out;
3043
3044 cpu_buffer = buffer->buffers[cpu];
3045
3046 if (atomic_read(&cpu_buffer->record_disabled))
3047 goto out;
3048
3049 if (length > BUF_MAX_DATA_SIZE)
3050 goto out;
3051
3052 if (unlikely(trace_recursive_lock(cpu_buffer)))
3053 goto out;
3054
3055 event = rb_reserve_next_event(buffer, cpu_buffer, length);
3056 if (!event)
3057 goto out_unlock;
3058
3059 body = rb_event_data(event);
3060
3061 memcpy(body, data, length);
3062
3063 rb_commit(cpu_buffer, event);
3064
3065 rb_wakeups(buffer, cpu_buffer);
3066
3067 ret = 0;
3068
3069 out_unlock:
3070 trace_recursive_unlock(cpu_buffer);
3071
3072 out:
3073 preempt_enable_notrace();
3074
3075 return ret;
3076 }
3077 EXPORT_SYMBOL_GPL(ring_buffer_write);
3078
rb_per_cpu_empty(struct ring_buffer_per_cpu * cpu_buffer)3079 static bool rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
3080 {
3081 struct buffer_page *reader = cpu_buffer->reader_page;
3082 struct buffer_page *head = rb_set_head_page(cpu_buffer);
3083 struct buffer_page *commit = cpu_buffer->commit_page;
3084
3085 /* In case of error, head will be NULL */
3086 if (unlikely(!head))
3087 return true;
3088
3089 /* Reader should exhaust content in reader page */
3090 if (reader->read != rb_page_commit(reader))
3091 return false;
3092
3093 /*
3094 * If writers are committing on the reader page, knowing all
3095 * committed content has been read, the ring buffer is empty.
3096 */
3097 if (commit == reader)
3098 return true;
3099
3100 /*
3101 * If writers are committing on a page other than reader page
3102 * and head page, there should always be content to read.
3103 */
3104 if (commit != head)
3105 return false;
3106
3107 /*
3108 * Writers are committing on the head page, we just need
3109 * to care about there're committed data, and the reader will
3110 * swap reader page with head page when it is to read data.
3111 */
3112 return rb_page_commit(commit) == 0;
3113 }
3114
3115 /**
3116 * ring_buffer_record_disable - stop all writes into the buffer
3117 * @buffer: The ring buffer to stop writes to.
3118 *
3119 * This prevents all writes to the buffer. Any attempt to write
3120 * to the buffer after this will fail and return NULL.
3121 *
3122 * The caller should call synchronize_sched() after this.
3123 */
ring_buffer_record_disable(struct ring_buffer * buffer)3124 void ring_buffer_record_disable(struct ring_buffer *buffer)
3125 {
3126 atomic_inc(&buffer->record_disabled);
3127 }
3128 EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
3129
3130 /**
3131 * ring_buffer_record_enable - enable writes to the buffer
3132 * @buffer: The ring buffer to enable writes
3133 *
3134 * Note, multiple disables will need the same number of enables
3135 * to truly enable the writing (much like preempt_disable).
3136 */
ring_buffer_record_enable(struct ring_buffer * buffer)3137 void ring_buffer_record_enable(struct ring_buffer *buffer)
3138 {
3139 atomic_dec(&buffer->record_disabled);
3140 }
3141 EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
3142
3143 /**
3144 * ring_buffer_record_off - stop all writes into the buffer
3145 * @buffer: The ring buffer to stop writes to.
3146 *
3147 * This prevents all writes to the buffer. Any attempt to write
3148 * to the buffer after this will fail and return NULL.
3149 *
3150 * This is different than ring_buffer_record_disable() as
3151 * it works like an on/off switch, where as the disable() version
3152 * must be paired with a enable().
3153 */
ring_buffer_record_off(struct ring_buffer * buffer)3154 void ring_buffer_record_off(struct ring_buffer *buffer)
3155 {
3156 unsigned int rd;
3157 unsigned int new_rd;
3158
3159 do {
3160 rd = atomic_read(&buffer->record_disabled);
3161 new_rd = rd | RB_BUFFER_OFF;
3162 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3163 }
3164 EXPORT_SYMBOL_GPL(ring_buffer_record_off);
3165
3166 /**
3167 * ring_buffer_record_on - restart writes into the buffer
3168 * @buffer: The ring buffer to start writes to.
3169 *
3170 * This enables all writes to the buffer that was disabled by
3171 * ring_buffer_record_off().
3172 *
3173 * This is different than ring_buffer_record_enable() as
3174 * it works like an on/off switch, where as the enable() version
3175 * must be paired with a disable().
3176 */
ring_buffer_record_on(struct ring_buffer * buffer)3177 void ring_buffer_record_on(struct ring_buffer *buffer)
3178 {
3179 unsigned int rd;
3180 unsigned int new_rd;
3181
3182 do {
3183 rd = atomic_read(&buffer->record_disabled);
3184 new_rd = rd & ~RB_BUFFER_OFF;
3185 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3186 }
3187 EXPORT_SYMBOL_GPL(ring_buffer_record_on);
3188
3189 /**
3190 * ring_buffer_record_is_on - return true if the ring buffer can write
3191 * @buffer: The ring buffer to see if write is enabled
3192 *
3193 * Returns true if the ring buffer is in a state that it accepts writes.
3194 */
ring_buffer_record_is_on(struct ring_buffer * buffer)3195 int ring_buffer_record_is_on(struct ring_buffer *buffer)
3196 {
3197 return !atomic_read(&buffer->record_disabled);
3198 }
3199
3200 /**
3201 * ring_buffer_record_is_set_on - return true if the ring buffer is set writable
3202 * @buffer: The ring buffer to see if write is set enabled
3203 *
3204 * Returns true if the ring buffer is set writable by ring_buffer_record_on().
3205 * Note that this does NOT mean it is in a writable state.
3206 *
3207 * It may return true when the ring buffer has been disabled by
3208 * ring_buffer_record_disable(), as that is a temporary disabling of
3209 * the ring buffer.
3210 */
ring_buffer_record_is_set_on(struct ring_buffer * buffer)3211 int ring_buffer_record_is_set_on(struct ring_buffer *buffer)
3212 {
3213 return !(atomic_read(&buffer->record_disabled) & RB_BUFFER_OFF);
3214 }
3215
3216 /**
3217 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
3218 * @buffer: The ring buffer to stop writes to.
3219 * @cpu: The CPU buffer to stop
3220 *
3221 * This prevents all writes to the buffer. Any attempt to write
3222 * to the buffer after this will fail and return NULL.
3223 *
3224 * The caller should call synchronize_sched() after this.
3225 */
ring_buffer_record_disable_cpu(struct ring_buffer * buffer,int cpu)3226 void ring_buffer_record_disable_cpu(struct ring_buffer *buffer, int cpu)
3227 {
3228 struct ring_buffer_per_cpu *cpu_buffer;
3229
3230 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3231 return;
3232
3233 cpu_buffer = buffer->buffers[cpu];
3234 atomic_inc(&cpu_buffer->record_disabled);
3235 }
3236 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
3237
3238 /**
3239 * ring_buffer_record_enable_cpu - enable writes to the buffer
3240 * @buffer: The ring buffer to enable writes
3241 * @cpu: The CPU to enable.
3242 *
3243 * Note, multiple disables will need the same number of enables
3244 * to truly enable the writing (much like preempt_disable).
3245 */
ring_buffer_record_enable_cpu(struct ring_buffer * buffer,int cpu)3246 void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu)
3247 {
3248 struct ring_buffer_per_cpu *cpu_buffer;
3249
3250 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3251 return;
3252
3253 cpu_buffer = buffer->buffers[cpu];
3254 atomic_dec(&cpu_buffer->record_disabled);
3255 }
3256 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
3257
3258 /*
3259 * The total entries in the ring buffer is the running counter
3260 * of entries entered into the ring buffer, minus the sum of
3261 * the entries read from the ring buffer and the number of
3262 * entries that were overwritten.
3263 */
3264 static inline unsigned long
rb_num_of_entries(struct ring_buffer_per_cpu * cpu_buffer)3265 rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
3266 {
3267 return local_read(&cpu_buffer->entries) -
3268 (local_read(&cpu_buffer->overrun) + cpu_buffer->read);
3269 }
3270
3271 /**
3272 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
3273 * @buffer: The ring buffer
3274 * @cpu: The per CPU buffer to read from.
3275 */
ring_buffer_oldest_event_ts(struct ring_buffer * buffer,int cpu)3276 u64 ring_buffer_oldest_event_ts(struct ring_buffer *buffer, int cpu)
3277 {
3278 unsigned long flags;
3279 struct ring_buffer_per_cpu *cpu_buffer;
3280 struct buffer_page *bpage;
3281 u64 ret = 0;
3282
3283 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3284 return 0;
3285
3286 cpu_buffer = buffer->buffers[cpu];
3287 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3288 /*
3289 * if the tail is on reader_page, oldest time stamp is on the reader
3290 * page
3291 */
3292 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
3293 bpage = cpu_buffer->reader_page;
3294 else
3295 bpage = rb_set_head_page(cpu_buffer);
3296 if (bpage)
3297 ret = bpage->page->time_stamp;
3298 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3299
3300 return ret;
3301 }
3302 EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
3303
3304 /**
3305 * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
3306 * @buffer: The ring buffer
3307 * @cpu: The per CPU buffer to read from.
3308 */
ring_buffer_bytes_cpu(struct ring_buffer * buffer,int cpu)3309 unsigned long ring_buffer_bytes_cpu(struct ring_buffer *buffer, int cpu)
3310 {
3311 struct ring_buffer_per_cpu *cpu_buffer;
3312 unsigned long ret;
3313
3314 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3315 return 0;
3316
3317 cpu_buffer = buffer->buffers[cpu];
3318 ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
3319
3320 return ret;
3321 }
3322 EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
3323
3324 /**
3325 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
3326 * @buffer: The ring buffer
3327 * @cpu: The per CPU buffer to get the entries from.
3328 */
ring_buffer_entries_cpu(struct ring_buffer * buffer,int cpu)3329 unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu)
3330 {
3331 struct ring_buffer_per_cpu *cpu_buffer;
3332
3333 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3334 return 0;
3335
3336 cpu_buffer = buffer->buffers[cpu];
3337
3338 return rb_num_of_entries(cpu_buffer);
3339 }
3340 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
3341
3342 /**
3343 * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
3344 * buffer wrapping around (only if RB_FL_OVERWRITE is on).
3345 * @buffer: The ring buffer
3346 * @cpu: The per CPU buffer to get the number of overruns from
3347 */
ring_buffer_overrun_cpu(struct ring_buffer * buffer,int cpu)3348 unsigned long ring_buffer_overrun_cpu(struct ring_buffer *buffer, int cpu)
3349 {
3350 struct ring_buffer_per_cpu *cpu_buffer;
3351 unsigned long ret;
3352
3353 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3354 return 0;
3355
3356 cpu_buffer = buffer->buffers[cpu];
3357 ret = local_read(&cpu_buffer->overrun);
3358
3359 return ret;
3360 }
3361 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
3362
3363 /**
3364 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
3365 * commits failing due to the buffer wrapping around while there are uncommitted
3366 * events, such as during an interrupt storm.
3367 * @buffer: The ring buffer
3368 * @cpu: The per CPU buffer to get the number of overruns from
3369 */
3370 unsigned long
ring_buffer_commit_overrun_cpu(struct ring_buffer * buffer,int cpu)3371 ring_buffer_commit_overrun_cpu(struct ring_buffer *buffer, int cpu)
3372 {
3373 struct ring_buffer_per_cpu *cpu_buffer;
3374 unsigned long ret;
3375
3376 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3377 return 0;
3378
3379 cpu_buffer = buffer->buffers[cpu];
3380 ret = local_read(&cpu_buffer->commit_overrun);
3381
3382 return ret;
3383 }
3384 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
3385
3386 /**
3387 * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
3388 * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
3389 * @buffer: The ring buffer
3390 * @cpu: The per CPU buffer to get the number of overruns from
3391 */
3392 unsigned long
ring_buffer_dropped_events_cpu(struct ring_buffer * buffer,int cpu)3393 ring_buffer_dropped_events_cpu(struct ring_buffer *buffer, int cpu)
3394 {
3395 struct ring_buffer_per_cpu *cpu_buffer;
3396 unsigned long ret;
3397
3398 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3399 return 0;
3400
3401 cpu_buffer = buffer->buffers[cpu];
3402 ret = local_read(&cpu_buffer->dropped_events);
3403
3404 return ret;
3405 }
3406 EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
3407
3408 /**
3409 * ring_buffer_read_events_cpu - get the number of events successfully read
3410 * @buffer: The ring buffer
3411 * @cpu: The per CPU buffer to get the number of events read
3412 */
3413 unsigned long
ring_buffer_read_events_cpu(struct ring_buffer * buffer,int cpu)3414 ring_buffer_read_events_cpu(struct ring_buffer *buffer, int cpu)
3415 {
3416 struct ring_buffer_per_cpu *cpu_buffer;
3417
3418 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3419 return 0;
3420
3421 cpu_buffer = buffer->buffers[cpu];
3422 return cpu_buffer->read;
3423 }
3424 EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu);
3425
3426 /**
3427 * ring_buffer_entries - get the number of entries in a buffer
3428 * @buffer: The ring buffer
3429 *
3430 * Returns the total number of entries in the ring buffer
3431 * (all CPU entries)
3432 */
ring_buffer_entries(struct ring_buffer * buffer)3433 unsigned long ring_buffer_entries(struct ring_buffer *buffer)
3434 {
3435 struct ring_buffer_per_cpu *cpu_buffer;
3436 unsigned long entries = 0;
3437 int cpu;
3438
3439 /* if you care about this being correct, lock the buffer */
3440 for_each_buffer_cpu(buffer, cpu) {
3441 cpu_buffer = buffer->buffers[cpu];
3442 entries += rb_num_of_entries(cpu_buffer);
3443 }
3444
3445 return entries;
3446 }
3447 EXPORT_SYMBOL_GPL(ring_buffer_entries);
3448
3449 /**
3450 * ring_buffer_overruns - get the number of overruns in buffer
3451 * @buffer: The ring buffer
3452 *
3453 * Returns the total number of overruns in the ring buffer
3454 * (all CPU entries)
3455 */
ring_buffer_overruns(struct ring_buffer * buffer)3456 unsigned long ring_buffer_overruns(struct ring_buffer *buffer)
3457 {
3458 struct ring_buffer_per_cpu *cpu_buffer;
3459 unsigned long overruns = 0;
3460 int cpu;
3461
3462 /* if you care about this being correct, lock the buffer */
3463 for_each_buffer_cpu(buffer, cpu) {
3464 cpu_buffer = buffer->buffers[cpu];
3465 overruns += local_read(&cpu_buffer->overrun);
3466 }
3467
3468 return overruns;
3469 }
3470 EXPORT_SYMBOL_GPL(ring_buffer_overruns);
3471
rb_iter_reset(struct ring_buffer_iter * iter)3472 static void rb_iter_reset(struct ring_buffer_iter *iter)
3473 {
3474 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3475
3476 /* Iterator usage is expected to have record disabled */
3477 iter->head_page = cpu_buffer->reader_page;
3478 iter->head = cpu_buffer->reader_page->read;
3479
3480 iter->cache_reader_page = iter->head_page;
3481 iter->cache_read = cpu_buffer->read;
3482
3483 if (iter->head)
3484 iter->read_stamp = cpu_buffer->read_stamp;
3485 else
3486 iter->read_stamp = iter->head_page->page->time_stamp;
3487 }
3488
3489 /**
3490 * ring_buffer_iter_reset - reset an iterator
3491 * @iter: The iterator to reset
3492 *
3493 * Resets the iterator, so that it will start from the beginning
3494 * again.
3495 */
ring_buffer_iter_reset(struct ring_buffer_iter * iter)3496 void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
3497 {
3498 struct ring_buffer_per_cpu *cpu_buffer;
3499 unsigned long flags;
3500
3501 if (!iter)
3502 return;
3503
3504 cpu_buffer = iter->cpu_buffer;
3505
3506 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3507 rb_iter_reset(iter);
3508 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3509 }
3510 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
3511
3512 /**
3513 * ring_buffer_iter_empty - check if an iterator has no more to read
3514 * @iter: The iterator to check
3515 */
ring_buffer_iter_empty(struct ring_buffer_iter * iter)3516 int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
3517 {
3518 struct ring_buffer_per_cpu *cpu_buffer;
3519 struct buffer_page *reader;
3520 struct buffer_page *head_page;
3521 struct buffer_page *commit_page;
3522 unsigned commit;
3523
3524 cpu_buffer = iter->cpu_buffer;
3525
3526 /* Remember, trace recording is off when iterator is in use */
3527 reader = cpu_buffer->reader_page;
3528 head_page = cpu_buffer->head_page;
3529 commit_page = cpu_buffer->commit_page;
3530 commit = rb_page_commit(commit_page);
3531
3532 return ((iter->head_page == commit_page && iter->head == commit) ||
3533 (iter->head_page == reader && commit_page == head_page &&
3534 head_page->read == commit &&
3535 iter->head == rb_page_commit(cpu_buffer->reader_page)));
3536 }
3537 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
3538
3539 static void
rb_update_read_stamp(struct ring_buffer_per_cpu * cpu_buffer,struct ring_buffer_event * event)3540 rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
3541 struct ring_buffer_event *event)
3542 {
3543 u64 delta;
3544
3545 switch (event->type_len) {
3546 case RINGBUF_TYPE_PADDING:
3547 return;
3548
3549 case RINGBUF_TYPE_TIME_EXTEND:
3550 delta = event->array[0];
3551 delta <<= TS_SHIFT;
3552 delta += event->time_delta;
3553 cpu_buffer->read_stamp += delta;
3554 return;
3555
3556 case RINGBUF_TYPE_TIME_STAMP:
3557 /* FIXME: not implemented */
3558 return;
3559
3560 case RINGBUF_TYPE_DATA:
3561 cpu_buffer->read_stamp += event->time_delta;
3562 return;
3563
3564 default:
3565 BUG();
3566 }
3567 return;
3568 }
3569
3570 static void
rb_update_iter_read_stamp(struct ring_buffer_iter * iter,struct ring_buffer_event * event)3571 rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
3572 struct ring_buffer_event *event)
3573 {
3574 u64 delta;
3575
3576 switch (event->type_len) {
3577 case RINGBUF_TYPE_PADDING:
3578 return;
3579
3580 case RINGBUF_TYPE_TIME_EXTEND:
3581 delta = event->array[0];
3582 delta <<= TS_SHIFT;
3583 delta += event->time_delta;
3584 iter->read_stamp += delta;
3585 return;
3586
3587 case RINGBUF_TYPE_TIME_STAMP:
3588 /* FIXME: not implemented */
3589 return;
3590
3591 case RINGBUF_TYPE_DATA:
3592 iter->read_stamp += event->time_delta;
3593 return;
3594
3595 default:
3596 BUG();
3597 }
3598 return;
3599 }
3600
3601 static struct buffer_page *
rb_get_reader_page(struct ring_buffer_per_cpu * cpu_buffer)3602 rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
3603 {
3604 struct buffer_page *reader = NULL;
3605 unsigned long overwrite;
3606 unsigned long flags;
3607 int nr_loops = 0;
3608 int ret;
3609
3610 local_irq_save(flags);
3611 arch_spin_lock(&cpu_buffer->lock);
3612
3613 again:
3614 /*
3615 * This should normally only loop twice. But because the
3616 * start of the reader inserts an empty page, it causes
3617 * a case where we will loop three times. There should be no
3618 * reason to loop four times (that I know of).
3619 */
3620 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
3621 reader = NULL;
3622 goto out;
3623 }
3624
3625 reader = cpu_buffer->reader_page;
3626
3627 /* If there's more to read, return this page */
3628 if (cpu_buffer->reader_page->read < rb_page_size(reader))
3629 goto out;
3630
3631 /* Never should we have an index greater than the size */
3632 if (RB_WARN_ON(cpu_buffer,
3633 cpu_buffer->reader_page->read > rb_page_size(reader)))
3634 goto out;
3635
3636 /* check if we caught up to the tail */
3637 reader = NULL;
3638 if (cpu_buffer->commit_page == cpu_buffer->reader_page)
3639 goto out;
3640
3641 /* Don't bother swapping if the ring buffer is empty */
3642 if (rb_num_of_entries(cpu_buffer) == 0)
3643 goto out;
3644
3645 /*
3646 * Reset the reader page to size zero.
3647 */
3648 local_set(&cpu_buffer->reader_page->write, 0);
3649 local_set(&cpu_buffer->reader_page->entries, 0);
3650 local_set(&cpu_buffer->reader_page->page->commit, 0);
3651 cpu_buffer->reader_page->real_end = 0;
3652
3653 spin:
3654 /*
3655 * Splice the empty reader page into the list around the head.
3656 */
3657 reader = rb_set_head_page(cpu_buffer);
3658 if (!reader)
3659 goto out;
3660 cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
3661 cpu_buffer->reader_page->list.prev = reader->list.prev;
3662
3663 /*
3664 * cpu_buffer->pages just needs to point to the buffer, it
3665 * has no specific buffer page to point to. Lets move it out
3666 * of our way so we don't accidentally swap it.
3667 */
3668 cpu_buffer->pages = reader->list.prev;
3669
3670 /* The reader page will be pointing to the new head */
3671 rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list);
3672
3673 /*
3674 * We want to make sure we read the overruns after we set up our
3675 * pointers to the next object. The writer side does a
3676 * cmpxchg to cross pages which acts as the mb on the writer
3677 * side. Note, the reader will constantly fail the swap
3678 * while the writer is updating the pointers, so this
3679 * guarantees that the overwrite recorded here is the one we
3680 * want to compare with the last_overrun.
3681 */
3682 smp_mb();
3683 overwrite = local_read(&(cpu_buffer->overrun));
3684
3685 /*
3686 * Here's the tricky part.
3687 *
3688 * We need to move the pointer past the header page.
3689 * But we can only do that if a writer is not currently
3690 * moving it. The page before the header page has the
3691 * flag bit '1' set if it is pointing to the page we want.
3692 * but if the writer is in the process of moving it
3693 * than it will be '2' or already moved '0'.
3694 */
3695
3696 ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
3697
3698 /*
3699 * If we did not convert it, then we must try again.
3700 */
3701 if (!ret)
3702 goto spin;
3703
3704 /*
3705 * Yeah! We succeeded in replacing the page.
3706 *
3707 * Now make the new head point back to the reader page.
3708 */
3709 rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
3710 rb_inc_page(cpu_buffer, &cpu_buffer->head_page);
3711
3712 /* Finally update the reader page to the new head */
3713 cpu_buffer->reader_page = reader;
3714 cpu_buffer->reader_page->read = 0;
3715
3716 if (overwrite != cpu_buffer->last_overrun) {
3717 cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
3718 cpu_buffer->last_overrun = overwrite;
3719 }
3720
3721 goto again;
3722
3723 out:
3724 /* Update the read_stamp on the first event */
3725 if (reader && reader->read == 0)
3726 cpu_buffer->read_stamp = reader->page->time_stamp;
3727
3728 arch_spin_unlock(&cpu_buffer->lock);
3729 local_irq_restore(flags);
3730
3731 return reader;
3732 }
3733
rb_advance_reader(struct ring_buffer_per_cpu * cpu_buffer)3734 static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
3735 {
3736 struct ring_buffer_event *event;
3737 struct buffer_page *reader;
3738 unsigned length;
3739
3740 reader = rb_get_reader_page(cpu_buffer);
3741
3742 /* This function should not be called when buffer is empty */
3743 if (RB_WARN_ON(cpu_buffer, !reader))
3744 return;
3745
3746 event = rb_reader_event(cpu_buffer);
3747
3748 if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
3749 cpu_buffer->read++;
3750
3751 rb_update_read_stamp(cpu_buffer, event);
3752
3753 length = rb_event_length(event);
3754 cpu_buffer->reader_page->read += length;
3755 }
3756
rb_advance_iter(struct ring_buffer_iter * iter)3757 static void rb_advance_iter(struct ring_buffer_iter *iter)
3758 {
3759 struct ring_buffer_per_cpu *cpu_buffer;
3760 struct ring_buffer_event *event;
3761 unsigned length;
3762
3763 cpu_buffer = iter->cpu_buffer;
3764
3765 /*
3766 * Check if we are at the end of the buffer.
3767 */
3768 if (iter->head >= rb_page_size(iter->head_page)) {
3769 /* discarded commits can make the page empty */
3770 if (iter->head_page == cpu_buffer->commit_page)
3771 return;
3772 rb_inc_iter(iter);
3773 return;
3774 }
3775
3776 event = rb_iter_head_event(iter);
3777
3778 length = rb_event_length(event);
3779
3780 /*
3781 * This should not be called to advance the header if we are
3782 * at the tail of the buffer.
3783 */
3784 if (RB_WARN_ON(cpu_buffer,
3785 (iter->head_page == cpu_buffer->commit_page) &&
3786 (iter->head + length > rb_commit_index(cpu_buffer))))
3787 return;
3788
3789 rb_update_iter_read_stamp(iter, event);
3790
3791 iter->head += length;
3792
3793 /* check for end of page padding */
3794 if ((iter->head >= rb_page_size(iter->head_page)) &&
3795 (iter->head_page != cpu_buffer->commit_page))
3796 rb_inc_iter(iter);
3797 }
3798
rb_lost_events(struct ring_buffer_per_cpu * cpu_buffer)3799 static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
3800 {
3801 return cpu_buffer->lost_events;
3802 }
3803
3804 static struct ring_buffer_event *
rb_buffer_peek(struct ring_buffer_per_cpu * cpu_buffer,u64 * ts,unsigned long * lost_events)3805 rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
3806 unsigned long *lost_events)
3807 {
3808 struct ring_buffer_event *event;
3809 struct buffer_page *reader;
3810 int nr_loops = 0;
3811
3812 again:
3813 /*
3814 * We repeat when a time extend is encountered.
3815 * Since the time extend is always attached to a data event,
3816 * we should never loop more than once.
3817 * (We never hit the following condition more than twice).
3818 */
3819 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3820 return NULL;
3821
3822 reader = rb_get_reader_page(cpu_buffer);
3823 if (!reader)
3824 return NULL;
3825
3826 event = rb_reader_event(cpu_buffer);
3827
3828 switch (event->type_len) {
3829 case RINGBUF_TYPE_PADDING:
3830 if (rb_null_event(event))
3831 RB_WARN_ON(cpu_buffer, 1);
3832 /*
3833 * Because the writer could be discarding every
3834 * event it creates (which would probably be bad)
3835 * if we were to go back to "again" then we may never
3836 * catch up, and will trigger the warn on, or lock
3837 * the box. Return the padding, and we will release
3838 * the current locks, and try again.
3839 */
3840 return event;
3841
3842 case RINGBUF_TYPE_TIME_EXTEND:
3843 /* Internal data, OK to advance */
3844 rb_advance_reader(cpu_buffer);
3845 goto again;
3846
3847 case RINGBUF_TYPE_TIME_STAMP:
3848 /* FIXME: not implemented */
3849 rb_advance_reader(cpu_buffer);
3850 goto again;
3851
3852 case RINGBUF_TYPE_DATA:
3853 if (ts) {
3854 *ts = cpu_buffer->read_stamp + event->time_delta;
3855 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3856 cpu_buffer->cpu, ts);
3857 }
3858 if (lost_events)
3859 *lost_events = rb_lost_events(cpu_buffer);
3860 return event;
3861
3862 default:
3863 BUG();
3864 }
3865
3866 return NULL;
3867 }
3868 EXPORT_SYMBOL_GPL(ring_buffer_peek);
3869
3870 static struct ring_buffer_event *
rb_iter_peek(struct ring_buffer_iter * iter,u64 * ts)3871 rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3872 {
3873 struct ring_buffer *buffer;
3874 struct ring_buffer_per_cpu *cpu_buffer;
3875 struct ring_buffer_event *event;
3876 int nr_loops = 0;
3877
3878 cpu_buffer = iter->cpu_buffer;
3879 buffer = cpu_buffer->buffer;
3880
3881 /*
3882 * Check if someone performed a consuming read to
3883 * the buffer. A consuming read invalidates the iterator
3884 * and we need to reset the iterator in this case.
3885 */
3886 if (unlikely(iter->cache_read != cpu_buffer->read ||
3887 iter->cache_reader_page != cpu_buffer->reader_page))
3888 rb_iter_reset(iter);
3889
3890 again:
3891 if (ring_buffer_iter_empty(iter))
3892 return NULL;
3893
3894 /*
3895 * We repeat when a time extend is encountered or we hit
3896 * the end of the page. Since the time extend is always attached
3897 * to a data event, we should never loop more than three times.
3898 * Once for going to next page, once on time extend, and
3899 * finally once to get the event.
3900 * (We never hit the following condition more than thrice).
3901 */
3902 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3))
3903 return NULL;
3904
3905 if (rb_per_cpu_empty(cpu_buffer))
3906 return NULL;
3907
3908 if (iter->head >= rb_page_size(iter->head_page)) {
3909 rb_inc_iter(iter);
3910 goto again;
3911 }
3912
3913 event = rb_iter_head_event(iter);
3914
3915 switch (event->type_len) {
3916 case RINGBUF_TYPE_PADDING:
3917 if (rb_null_event(event)) {
3918 rb_inc_iter(iter);
3919 goto again;
3920 }
3921 rb_advance_iter(iter);
3922 return event;
3923
3924 case RINGBUF_TYPE_TIME_EXTEND:
3925 /* Internal data, OK to advance */
3926 rb_advance_iter(iter);
3927 goto again;
3928
3929 case RINGBUF_TYPE_TIME_STAMP:
3930 /* FIXME: not implemented */
3931 rb_advance_iter(iter);
3932 goto again;
3933
3934 case RINGBUF_TYPE_DATA:
3935 if (ts) {
3936 *ts = iter->read_stamp + event->time_delta;
3937 ring_buffer_normalize_time_stamp(buffer,
3938 cpu_buffer->cpu, ts);
3939 }
3940 return event;
3941
3942 default:
3943 BUG();
3944 }
3945
3946 return NULL;
3947 }
3948 EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
3949
rb_reader_lock(struct ring_buffer_per_cpu * cpu_buffer)3950 static inline bool rb_reader_lock(struct ring_buffer_per_cpu *cpu_buffer)
3951 {
3952 if (likely(!in_nmi())) {
3953 raw_spin_lock(&cpu_buffer->reader_lock);
3954 return true;
3955 }
3956
3957 /*
3958 * If an NMI die dumps out the content of the ring buffer
3959 * trylock must be used to prevent a deadlock if the NMI
3960 * preempted a task that holds the ring buffer locks. If
3961 * we get the lock then all is fine, if not, then continue
3962 * to do the read, but this can corrupt the ring buffer,
3963 * so it must be permanently disabled from future writes.
3964 * Reading from NMI is a oneshot deal.
3965 */
3966 if (raw_spin_trylock(&cpu_buffer->reader_lock))
3967 return true;
3968
3969 /* Continue without locking, but disable the ring buffer */
3970 atomic_inc(&cpu_buffer->record_disabled);
3971 return false;
3972 }
3973
3974 static inline void
rb_reader_unlock(struct ring_buffer_per_cpu * cpu_buffer,bool locked)3975 rb_reader_unlock(struct ring_buffer_per_cpu *cpu_buffer, bool locked)
3976 {
3977 if (likely(locked))
3978 raw_spin_unlock(&cpu_buffer->reader_lock);
3979 return;
3980 }
3981
3982 /**
3983 * ring_buffer_peek - peek at the next event to be read
3984 * @buffer: The ring buffer to read
3985 * @cpu: The cpu to peak at
3986 * @ts: The timestamp counter of this event.
3987 * @lost_events: a variable to store if events were lost (may be NULL)
3988 *
3989 * This will return the event that will be read next, but does
3990 * not consume the data.
3991 */
3992 struct ring_buffer_event *
ring_buffer_peek(struct ring_buffer * buffer,int cpu,u64 * ts,unsigned long * lost_events)3993 ring_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts,
3994 unsigned long *lost_events)
3995 {
3996 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3997 struct ring_buffer_event *event;
3998 unsigned long flags;
3999 bool dolock;
4000
4001 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4002 return NULL;
4003
4004 again:
4005 local_irq_save(flags);
4006 dolock = rb_reader_lock(cpu_buffer);
4007 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
4008 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4009 rb_advance_reader(cpu_buffer);
4010 rb_reader_unlock(cpu_buffer, dolock);
4011 local_irq_restore(flags);
4012
4013 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4014 goto again;
4015
4016 return event;
4017 }
4018
4019 /**
4020 * ring_buffer_iter_peek - peek at the next event to be read
4021 * @iter: The ring buffer iterator
4022 * @ts: The timestamp counter of this event.
4023 *
4024 * This will return the event that will be read next, but does
4025 * not increment the iterator.
4026 */
4027 struct ring_buffer_event *
ring_buffer_iter_peek(struct ring_buffer_iter * iter,u64 * ts)4028 ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
4029 {
4030 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4031 struct ring_buffer_event *event;
4032 unsigned long flags;
4033
4034 again:
4035 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4036 event = rb_iter_peek(iter, ts);
4037 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4038
4039 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4040 goto again;
4041
4042 return event;
4043 }
4044
4045 /**
4046 * ring_buffer_consume - return an event and consume it
4047 * @buffer: The ring buffer to get the next event from
4048 * @cpu: the cpu to read the buffer from
4049 * @ts: a variable to store the timestamp (may be NULL)
4050 * @lost_events: a variable to store if events were lost (may be NULL)
4051 *
4052 * Returns the next event in the ring buffer, and that event is consumed.
4053 * Meaning, that sequential reads will keep returning a different event,
4054 * and eventually empty the ring buffer if the producer is slower.
4055 */
4056 struct ring_buffer_event *
ring_buffer_consume(struct ring_buffer * buffer,int cpu,u64 * ts,unsigned long * lost_events)4057 ring_buffer_consume(struct ring_buffer *buffer, int cpu, u64 *ts,
4058 unsigned long *lost_events)
4059 {
4060 struct ring_buffer_per_cpu *cpu_buffer;
4061 struct ring_buffer_event *event = NULL;
4062 unsigned long flags;
4063 bool dolock;
4064
4065 again:
4066 /* might be called in atomic */
4067 preempt_disable();
4068
4069 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4070 goto out;
4071
4072 cpu_buffer = buffer->buffers[cpu];
4073 local_irq_save(flags);
4074 dolock = rb_reader_lock(cpu_buffer);
4075
4076 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
4077 if (event) {
4078 cpu_buffer->lost_events = 0;
4079 rb_advance_reader(cpu_buffer);
4080 }
4081
4082 rb_reader_unlock(cpu_buffer, dolock);
4083 local_irq_restore(flags);
4084
4085 out:
4086 preempt_enable();
4087
4088 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4089 goto again;
4090
4091 return event;
4092 }
4093 EXPORT_SYMBOL_GPL(ring_buffer_consume);
4094
4095 /**
4096 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
4097 * @buffer: The ring buffer to read from
4098 * @cpu: The cpu buffer to iterate over
4099 * @flags: gfp flags to use for memory allocation
4100 *
4101 * This performs the initial preparations necessary to iterate
4102 * through the buffer. Memory is allocated, buffer recording
4103 * is disabled, and the iterator pointer is returned to the caller.
4104 *
4105 * Disabling buffer recordng prevents the reading from being
4106 * corrupted. This is not a consuming read, so a producer is not
4107 * expected.
4108 *
4109 * After a sequence of ring_buffer_read_prepare calls, the user is
4110 * expected to make at least one call to ring_buffer_read_prepare_sync.
4111 * Afterwards, ring_buffer_read_start is invoked to get things going
4112 * for real.
4113 *
4114 * This overall must be paired with ring_buffer_read_finish.
4115 */
4116 struct ring_buffer_iter *
ring_buffer_read_prepare(struct ring_buffer * buffer,int cpu,gfp_t flags)4117 ring_buffer_read_prepare(struct ring_buffer *buffer, int cpu, gfp_t flags)
4118 {
4119 struct ring_buffer_per_cpu *cpu_buffer;
4120 struct ring_buffer_iter *iter;
4121
4122 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4123 return NULL;
4124
4125 iter = kmalloc(sizeof(*iter), flags);
4126 if (!iter)
4127 return NULL;
4128
4129 cpu_buffer = buffer->buffers[cpu];
4130
4131 iter->cpu_buffer = cpu_buffer;
4132
4133 atomic_inc(&buffer->resize_disabled);
4134 atomic_inc(&cpu_buffer->record_disabled);
4135
4136 return iter;
4137 }
4138 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
4139
4140 /**
4141 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
4142 *
4143 * All previously invoked ring_buffer_read_prepare calls to prepare
4144 * iterators will be synchronized. Afterwards, read_buffer_read_start
4145 * calls on those iterators are allowed.
4146 */
4147 void
ring_buffer_read_prepare_sync(void)4148 ring_buffer_read_prepare_sync(void)
4149 {
4150 synchronize_sched();
4151 }
4152 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
4153
4154 /**
4155 * ring_buffer_read_start - start a non consuming read of the buffer
4156 * @iter: The iterator returned by ring_buffer_read_prepare
4157 *
4158 * This finalizes the startup of an iteration through the buffer.
4159 * The iterator comes from a call to ring_buffer_read_prepare and
4160 * an intervening ring_buffer_read_prepare_sync must have been
4161 * performed.
4162 *
4163 * Must be paired with ring_buffer_read_finish.
4164 */
4165 void
ring_buffer_read_start(struct ring_buffer_iter * iter)4166 ring_buffer_read_start(struct ring_buffer_iter *iter)
4167 {
4168 struct ring_buffer_per_cpu *cpu_buffer;
4169 unsigned long flags;
4170
4171 if (!iter)
4172 return;
4173
4174 cpu_buffer = iter->cpu_buffer;
4175
4176 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4177 arch_spin_lock(&cpu_buffer->lock);
4178 rb_iter_reset(iter);
4179 arch_spin_unlock(&cpu_buffer->lock);
4180 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4181 }
4182 EXPORT_SYMBOL_GPL(ring_buffer_read_start);
4183
4184 /**
4185 * ring_buffer_read_finish - finish reading the iterator of the buffer
4186 * @iter: The iterator retrieved by ring_buffer_start
4187 *
4188 * This re-enables the recording to the buffer, and frees the
4189 * iterator.
4190 */
4191 void
ring_buffer_read_finish(struct ring_buffer_iter * iter)4192 ring_buffer_read_finish(struct ring_buffer_iter *iter)
4193 {
4194 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4195 unsigned long flags;
4196
4197 /*
4198 * Ring buffer is disabled from recording, here's a good place
4199 * to check the integrity of the ring buffer.
4200 * Must prevent readers from trying to read, as the check
4201 * clears the HEAD page and readers require it.
4202 */
4203 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4204 rb_check_pages(cpu_buffer);
4205 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4206
4207 atomic_dec(&cpu_buffer->record_disabled);
4208 atomic_dec(&cpu_buffer->buffer->resize_disabled);
4209 kfree(iter);
4210 }
4211 EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
4212
4213 /**
4214 * ring_buffer_read - read the next item in the ring buffer by the iterator
4215 * @iter: The ring buffer iterator
4216 * @ts: The time stamp of the event read.
4217 *
4218 * This reads the next event in the ring buffer and increments the iterator.
4219 */
4220 struct ring_buffer_event *
ring_buffer_read(struct ring_buffer_iter * iter,u64 * ts)4221 ring_buffer_read(struct ring_buffer_iter *iter, u64 *ts)
4222 {
4223 struct ring_buffer_event *event;
4224 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4225 unsigned long flags;
4226
4227 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4228 again:
4229 event = rb_iter_peek(iter, ts);
4230 if (!event)
4231 goto out;
4232
4233 if (event->type_len == RINGBUF_TYPE_PADDING)
4234 goto again;
4235
4236 rb_advance_iter(iter);
4237 out:
4238 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4239
4240 return event;
4241 }
4242 EXPORT_SYMBOL_GPL(ring_buffer_read);
4243
4244 /**
4245 * ring_buffer_size - return the size of the ring buffer (in bytes)
4246 * @buffer: The ring buffer.
4247 */
ring_buffer_size(struct ring_buffer * buffer,int cpu)4248 unsigned long ring_buffer_size(struct ring_buffer *buffer, int cpu)
4249 {
4250 /*
4251 * Earlier, this method returned
4252 * BUF_PAGE_SIZE * buffer->nr_pages
4253 * Since the nr_pages field is now removed, we have converted this to
4254 * return the per cpu buffer value.
4255 */
4256 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4257 return 0;
4258
4259 return BUF_PAGE_SIZE * buffer->buffers[cpu]->nr_pages;
4260 }
4261 EXPORT_SYMBOL_GPL(ring_buffer_size);
4262
4263 static void
rb_reset_cpu(struct ring_buffer_per_cpu * cpu_buffer)4264 rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
4265 {
4266 rb_head_page_deactivate(cpu_buffer);
4267
4268 cpu_buffer->head_page
4269 = list_entry(cpu_buffer->pages, struct buffer_page, list);
4270 local_set(&cpu_buffer->head_page->write, 0);
4271 local_set(&cpu_buffer->head_page->entries, 0);
4272 local_set(&cpu_buffer->head_page->page->commit, 0);
4273
4274 cpu_buffer->head_page->read = 0;
4275
4276 cpu_buffer->tail_page = cpu_buffer->head_page;
4277 cpu_buffer->commit_page = cpu_buffer->head_page;
4278
4279 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
4280 INIT_LIST_HEAD(&cpu_buffer->new_pages);
4281 local_set(&cpu_buffer->reader_page->write, 0);
4282 local_set(&cpu_buffer->reader_page->entries, 0);
4283 local_set(&cpu_buffer->reader_page->page->commit, 0);
4284 cpu_buffer->reader_page->read = 0;
4285
4286 local_set(&cpu_buffer->entries_bytes, 0);
4287 local_set(&cpu_buffer->overrun, 0);
4288 local_set(&cpu_buffer->commit_overrun, 0);
4289 local_set(&cpu_buffer->dropped_events, 0);
4290 local_set(&cpu_buffer->entries, 0);
4291 local_set(&cpu_buffer->committing, 0);
4292 local_set(&cpu_buffer->commits, 0);
4293 cpu_buffer->read = 0;
4294 cpu_buffer->read_bytes = 0;
4295
4296 cpu_buffer->write_stamp = 0;
4297 cpu_buffer->read_stamp = 0;
4298
4299 cpu_buffer->lost_events = 0;
4300 cpu_buffer->last_overrun = 0;
4301
4302 rb_head_page_activate(cpu_buffer);
4303 }
4304
4305 /**
4306 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
4307 * @buffer: The ring buffer to reset a per cpu buffer of
4308 * @cpu: The CPU buffer to be reset
4309 */
ring_buffer_reset_cpu(struct ring_buffer * buffer,int cpu)4310 void ring_buffer_reset_cpu(struct ring_buffer *buffer, int cpu)
4311 {
4312 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4313 unsigned long flags;
4314
4315 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4316 return;
4317 /* prevent another thread from changing buffer sizes */
4318 mutex_lock(&buffer->mutex);
4319
4320 atomic_inc(&buffer->resize_disabled);
4321 atomic_inc(&cpu_buffer->record_disabled);
4322
4323 /* Make sure all commits have finished */
4324 synchronize_sched();
4325
4326 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4327
4328 if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
4329 goto out;
4330
4331 arch_spin_lock(&cpu_buffer->lock);
4332
4333 rb_reset_cpu(cpu_buffer);
4334
4335 arch_spin_unlock(&cpu_buffer->lock);
4336
4337 out:
4338 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4339
4340 atomic_dec(&cpu_buffer->record_disabled);
4341 atomic_dec(&buffer->resize_disabled);
4342
4343 mutex_unlock(&buffer->mutex);
4344 }
4345 EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
4346
4347 /**
4348 * ring_buffer_reset - reset a ring buffer
4349 * @buffer: The ring buffer to reset all cpu buffers
4350 */
ring_buffer_reset(struct ring_buffer * buffer)4351 void ring_buffer_reset(struct ring_buffer *buffer)
4352 {
4353 int cpu;
4354
4355 for_each_buffer_cpu(buffer, cpu)
4356 ring_buffer_reset_cpu(buffer, cpu);
4357 }
4358 EXPORT_SYMBOL_GPL(ring_buffer_reset);
4359
4360 /**
4361 * rind_buffer_empty - is the ring buffer empty?
4362 * @buffer: The ring buffer to test
4363 */
ring_buffer_empty(struct ring_buffer * buffer)4364 bool ring_buffer_empty(struct ring_buffer *buffer)
4365 {
4366 struct ring_buffer_per_cpu *cpu_buffer;
4367 unsigned long flags;
4368 bool dolock;
4369 int cpu;
4370 int ret;
4371
4372 /* yes this is racy, but if you don't like the race, lock the buffer */
4373 for_each_buffer_cpu(buffer, cpu) {
4374 cpu_buffer = buffer->buffers[cpu];
4375 local_irq_save(flags);
4376 dolock = rb_reader_lock(cpu_buffer);
4377 ret = rb_per_cpu_empty(cpu_buffer);
4378 rb_reader_unlock(cpu_buffer, dolock);
4379 local_irq_restore(flags);
4380
4381 if (!ret)
4382 return false;
4383 }
4384
4385 return true;
4386 }
4387 EXPORT_SYMBOL_GPL(ring_buffer_empty);
4388
4389 /**
4390 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
4391 * @buffer: The ring buffer
4392 * @cpu: The CPU buffer to test
4393 */
ring_buffer_empty_cpu(struct ring_buffer * buffer,int cpu)4394 bool ring_buffer_empty_cpu(struct ring_buffer *buffer, int cpu)
4395 {
4396 struct ring_buffer_per_cpu *cpu_buffer;
4397 unsigned long flags;
4398 bool dolock;
4399 int ret;
4400
4401 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4402 return true;
4403
4404 cpu_buffer = buffer->buffers[cpu];
4405 local_irq_save(flags);
4406 dolock = rb_reader_lock(cpu_buffer);
4407 ret = rb_per_cpu_empty(cpu_buffer);
4408 rb_reader_unlock(cpu_buffer, dolock);
4409 local_irq_restore(flags);
4410
4411 return ret;
4412 }
4413 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
4414
4415 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
4416 /**
4417 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
4418 * @buffer_a: One buffer to swap with
4419 * @buffer_b: The other buffer to swap with
4420 *
4421 * This function is useful for tracers that want to take a "snapshot"
4422 * of a CPU buffer and has another back up buffer lying around.
4423 * it is expected that the tracer handles the cpu buffer not being
4424 * used at the moment.
4425 */
ring_buffer_swap_cpu(struct ring_buffer * buffer_a,struct ring_buffer * buffer_b,int cpu)4426 int ring_buffer_swap_cpu(struct ring_buffer *buffer_a,
4427 struct ring_buffer *buffer_b, int cpu)
4428 {
4429 struct ring_buffer_per_cpu *cpu_buffer_a;
4430 struct ring_buffer_per_cpu *cpu_buffer_b;
4431 int ret = -EINVAL;
4432
4433 if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
4434 !cpumask_test_cpu(cpu, buffer_b->cpumask))
4435 goto out;
4436
4437 cpu_buffer_a = buffer_a->buffers[cpu];
4438 cpu_buffer_b = buffer_b->buffers[cpu];
4439
4440 /* At least make sure the two buffers are somewhat the same */
4441 if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
4442 goto out;
4443
4444 ret = -EAGAIN;
4445
4446 if (atomic_read(&buffer_a->record_disabled))
4447 goto out;
4448
4449 if (atomic_read(&buffer_b->record_disabled))
4450 goto out;
4451
4452 if (atomic_read(&cpu_buffer_a->record_disabled))
4453 goto out;
4454
4455 if (atomic_read(&cpu_buffer_b->record_disabled))
4456 goto out;
4457
4458 /*
4459 * We can't do a synchronize_sched here because this
4460 * function can be called in atomic context.
4461 * Normally this will be called from the same CPU as cpu.
4462 * If not it's up to the caller to protect this.
4463 */
4464 atomic_inc(&cpu_buffer_a->record_disabled);
4465 atomic_inc(&cpu_buffer_b->record_disabled);
4466
4467 ret = -EBUSY;
4468 if (local_read(&cpu_buffer_a->committing))
4469 goto out_dec;
4470 if (local_read(&cpu_buffer_b->committing))
4471 goto out_dec;
4472
4473 buffer_a->buffers[cpu] = cpu_buffer_b;
4474 buffer_b->buffers[cpu] = cpu_buffer_a;
4475
4476 cpu_buffer_b->buffer = buffer_a;
4477 cpu_buffer_a->buffer = buffer_b;
4478
4479 ret = 0;
4480
4481 out_dec:
4482 atomic_dec(&cpu_buffer_a->record_disabled);
4483 atomic_dec(&cpu_buffer_b->record_disabled);
4484 out:
4485 return ret;
4486 }
4487 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
4488 #endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
4489
4490 /**
4491 * ring_buffer_alloc_read_page - allocate a page to read from buffer
4492 * @buffer: the buffer to allocate for.
4493 * @cpu: the cpu buffer to allocate.
4494 *
4495 * This function is used in conjunction with ring_buffer_read_page.
4496 * When reading a full page from the ring buffer, these functions
4497 * can be used to speed up the process. The calling function should
4498 * allocate a few pages first with this function. Then when it
4499 * needs to get pages from the ring buffer, it passes the result
4500 * of this function into ring_buffer_read_page, which will swap
4501 * the page that was allocated, with the read page of the buffer.
4502 *
4503 * Returns:
4504 * The page allocated, or NULL on error.
4505 */
ring_buffer_alloc_read_page(struct ring_buffer * buffer,int cpu)4506 void *ring_buffer_alloc_read_page(struct ring_buffer *buffer, int cpu)
4507 {
4508 struct buffer_data_page *bpage;
4509 struct page *page;
4510
4511 page = alloc_pages_node(cpu_to_node(cpu),
4512 GFP_KERNEL | __GFP_NORETRY, 0);
4513 if (!page)
4514 return NULL;
4515
4516 bpage = page_address(page);
4517
4518 rb_init_page(bpage);
4519
4520 return bpage;
4521 }
4522 EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
4523
4524 /**
4525 * ring_buffer_free_read_page - free an allocated read page
4526 * @buffer: the buffer the page was allocate for
4527 * @data: the page to free
4528 *
4529 * Free a page allocated from ring_buffer_alloc_read_page.
4530 */
ring_buffer_free_read_page(struct ring_buffer * buffer,void * data)4531 void ring_buffer_free_read_page(struct ring_buffer *buffer, void *data)
4532 {
4533 free_page((unsigned long)data);
4534 }
4535 EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
4536
4537 /**
4538 * ring_buffer_read_page - extract a page from the ring buffer
4539 * @buffer: buffer to extract from
4540 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
4541 * @len: amount to extract
4542 * @cpu: the cpu of the buffer to extract
4543 * @full: should the extraction only happen when the page is full.
4544 *
4545 * This function will pull out a page from the ring buffer and consume it.
4546 * @data_page must be the address of the variable that was returned
4547 * from ring_buffer_alloc_read_page. This is because the page might be used
4548 * to swap with a page in the ring buffer.
4549 *
4550 * for example:
4551 * rpage = ring_buffer_alloc_read_page(buffer, cpu);
4552 * if (!rpage)
4553 * return error;
4554 * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
4555 * if (ret >= 0)
4556 * process_page(rpage, ret);
4557 *
4558 * When @full is set, the function will not return true unless
4559 * the writer is off the reader page.
4560 *
4561 * Note: it is up to the calling functions to handle sleeps and wakeups.
4562 * The ring buffer can be used anywhere in the kernel and can not
4563 * blindly call wake_up. The layer that uses the ring buffer must be
4564 * responsible for that.
4565 *
4566 * Returns:
4567 * >=0 if data has been transferred, returns the offset of consumed data.
4568 * <0 if no data has been transferred.
4569 */
ring_buffer_read_page(struct ring_buffer * buffer,void ** data_page,size_t len,int cpu,int full)4570 int ring_buffer_read_page(struct ring_buffer *buffer,
4571 void **data_page, size_t len, int cpu, int full)
4572 {
4573 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4574 struct ring_buffer_event *event;
4575 struct buffer_data_page *bpage;
4576 struct buffer_page *reader;
4577 unsigned long missed_events;
4578 unsigned long flags;
4579 unsigned int commit;
4580 unsigned int read;
4581 u64 save_timestamp;
4582 int ret = -1;
4583
4584 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4585 goto out;
4586
4587 /*
4588 * If len is not big enough to hold the page header, then
4589 * we can not copy anything.
4590 */
4591 if (len <= BUF_PAGE_HDR_SIZE)
4592 goto out;
4593
4594 len -= BUF_PAGE_HDR_SIZE;
4595
4596 if (!data_page)
4597 goto out;
4598
4599 bpage = *data_page;
4600 if (!bpage)
4601 goto out;
4602
4603 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4604
4605 reader = rb_get_reader_page(cpu_buffer);
4606 if (!reader)
4607 goto out_unlock;
4608
4609 event = rb_reader_event(cpu_buffer);
4610
4611 read = reader->read;
4612 commit = rb_page_commit(reader);
4613
4614 /* Check if any events were dropped */
4615 missed_events = cpu_buffer->lost_events;
4616
4617 /*
4618 * If this page has been partially read or
4619 * if len is not big enough to read the rest of the page or
4620 * a writer is still on the page, then
4621 * we must copy the data from the page to the buffer.
4622 * Otherwise, we can simply swap the page with the one passed in.
4623 */
4624 if (read || (len < (commit - read)) ||
4625 cpu_buffer->reader_page == cpu_buffer->commit_page) {
4626 struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
4627 unsigned int rpos = read;
4628 unsigned int pos = 0;
4629 unsigned int size;
4630
4631 if (full)
4632 goto out_unlock;
4633
4634 if (len > (commit - read))
4635 len = (commit - read);
4636
4637 /* Always keep the time extend and data together */
4638 size = rb_event_ts_length(event);
4639
4640 if (len < size)
4641 goto out_unlock;
4642
4643 /* save the current timestamp, since the user will need it */
4644 save_timestamp = cpu_buffer->read_stamp;
4645
4646 /* Need to copy one event at a time */
4647 do {
4648 /* We need the size of one event, because
4649 * rb_advance_reader only advances by one event,
4650 * whereas rb_event_ts_length may include the size of
4651 * one or two events.
4652 * We have already ensured there's enough space if this
4653 * is a time extend. */
4654 size = rb_event_length(event);
4655 memcpy(bpage->data + pos, rpage->data + rpos, size);
4656
4657 len -= size;
4658
4659 rb_advance_reader(cpu_buffer);
4660 rpos = reader->read;
4661 pos += size;
4662
4663 if (rpos >= commit)
4664 break;
4665
4666 event = rb_reader_event(cpu_buffer);
4667 /* Always keep the time extend and data together */
4668 size = rb_event_ts_length(event);
4669 } while (len >= size);
4670
4671 /* update bpage */
4672 local_set(&bpage->commit, pos);
4673 bpage->time_stamp = save_timestamp;
4674
4675 /* we copied everything to the beginning */
4676 read = 0;
4677 } else {
4678 /* update the entry counter */
4679 cpu_buffer->read += rb_page_entries(reader);
4680 cpu_buffer->read_bytes += BUF_PAGE_SIZE;
4681
4682 /* swap the pages */
4683 rb_init_page(bpage);
4684 bpage = reader->page;
4685 reader->page = *data_page;
4686 local_set(&reader->write, 0);
4687 local_set(&reader->entries, 0);
4688 reader->read = 0;
4689 *data_page = bpage;
4690
4691 /*
4692 * Use the real_end for the data size,
4693 * This gives us a chance to store the lost events
4694 * on the page.
4695 */
4696 if (reader->real_end)
4697 local_set(&bpage->commit, reader->real_end);
4698 }
4699 ret = read;
4700
4701 cpu_buffer->lost_events = 0;
4702
4703 commit = local_read(&bpage->commit);
4704 /*
4705 * Set a flag in the commit field if we lost events
4706 */
4707 if (missed_events) {
4708 /* If there is room at the end of the page to save the
4709 * missed events, then record it there.
4710 */
4711 if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
4712 memcpy(&bpage->data[commit], &missed_events,
4713 sizeof(missed_events));
4714 local_add(RB_MISSED_STORED, &bpage->commit);
4715 commit += sizeof(missed_events);
4716 }
4717 local_add(RB_MISSED_EVENTS, &bpage->commit);
4718 }
4719
4720 /*
4721 * This page may be off to user land. Zero it out here.
4722 */
4723 if (commit < BUF_PAGE_SIZE)
4724 memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
4725
4726 out_unlock:
4727 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4728
4729 out:
4730 return ret;
4731 }
4732 EXPORT_SYMBOL_GPL(ring_buffer_read_page);
4733
4734 #ifdef CONFIG_HOTPLUG_CPU
rb_cpu_notify(struct notifier_block * self,unsigned long action,void * hcpu)4735 static int rb_cpu_notify(struct notifier_block *self,
4736 unsigned long action, void *hcpu)
4737 {
4738 struct ring_buffer *buffer =
4739 container_of(self, struct ring_buffer, cpu_notify);
4740 long cpu = (long)hcpu;
4741 long nr_pages_same;
4742 int cpu_i;
4743 unsigned long nr_pages;
4744
4745 switch (action) {
4746 case CPU_UP_PREPARE:
4747 case CPU_UP_PREPARE_FROZEN:
4748 if (cpumask_test_cpu(cpu, buffer->cpumask))
4749 return NOTIFY_OK;
4750
4751 nr_pages = 0;
4752 nr_pages_same = 1;
4753 /* check if all cpu sizes are same */
4754 for_each_buffer_cpu(buffer, cpu_i) {
4755 /* fill in the size from first enabled cpu */
4756 if (nr_pages == 0)
4757 nr_pages = buffer->buffers[cpu_i]->nr_pages;
4758 if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
4759 nr_pages_same = 0;
4760 break;
4761 }
4762 }
4763 /* allocate minimum pages, user can later expand it */
4764 if (!nr_pages_same)
4765 nr_pages = 2;
4766 buffer->buffers[cpu] =
4767 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
4768 if (!buffer->buffers[cpu]) {
4769 WARN(1, "failed to allocate ring buffer on CPU %ld\n",
4770 cpu);
4771 return NOTIFY_OK;
4772 }
4773 smp_wmb();
4774 cpumask_set_cpu(cpu, buffer->cpumask);
4775 break;
4776 case CPU_DOWN_PREPARE:
4777 case CPU_DOWN_PREPARE_FROZEN:
4778 /*
4779 * Do nothing.
4780 * If we were to free the buffer, then the user would
4781 * lose any trace that was in the buffer.
4782 */
4783 break;
4784 default:
4785 break;
4786 }
4787 return NOTIFY_OK;
4788 }
4789 #endif
4790
4791 #ifdef CONFIG_RING_BUFFER_STARTUP_TEST
4792 /*
4793 * This is a basic integrity check of the ring buffer.
4794 * Late in the boot cycle this test will run when configured in.
4795 * It will kick off a thread per CPU that will go into a loop
4796 * writing to the per cpu ring buffer various sizes of data.
4797 * Some of the data will be large items, some small.
4798 *
4799 * Another thread is created that goes into a spin, sending out
4800 * IPIs to the other CPUs to also write into the ring buffer.
4801 * this is to test the nesting ability of the buffer.
4802 *
4803 * Basic stats are recorded and reported. If something in the
4804 * ring buffer should happen that's not expected, a big warning
4805 * is displayed and all ring buffers are disabled.
4806 */
4807 static struct task_struct *rb_threads[NR_CPUS] __initdata;
4808
4809 struct rb_test_data {
4810 struct ring_buffer *buffer;
4811 unsigned long events;
4812 unsigned long bytes_written;
4813 unsigned long bytes_alloc;
4814 unsigned long bytes_dropped;
4815 unsigned long events_nested;
4816 unsigned long bytes_written_nested;
4817 unsigned long bytes_alloc_nested;
4818 unsigned long bytes_dropped_nested;
4819 int min_size_nested;
4820 int max_size_nested;
4821 int max_size;
4822 int min_size;
4823 int cpu;
4824 int cnt;
4825 };
4826
4827 static struct rb_test_data rb_data[NR_CPUS] __initdata;
4828
4829 /* 1 meg per cpu */
4830 #define RB_TEST_BUFFER_SIZE 1048576
4831
4832 static char rb_string[] __initdata =
4833 "abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
4834 "?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
4835 "!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
4836
4837 static bool rb_test_started __initdata;
4838
4839 struct rb_item {
4840 int size;
4841 char str[];
4842 };
4843
rb_write_something(struct rb_test_data * data,bool nested)4844 static __init int rb_write_something(struct rb_test_data *data, bool nested)
4845 {
4846 struct ring_buffer_event *event;
4847 struct rb_item *item;
4848 bool started;
4849 int event_len;
4850 int size;
4851 int len;
4852 int cnt;
4853
4854 /* Have nested writes different that what is written */
4855 cnt = data->cnt + (nested ? 27 : 0);
4856
4857 /* Multiply cnt by ~e, to make some unique increment */
4858 size = (data->cnt * 68 / 25) % (sizeof(rb_string) - 1);
4859
4860 len = size + sizeof(struct rb_item);
4861
4862 started = rb_test_started;
4863 /* read rb_test_started before checking buffer enabled */
4864 smp_rmb();
4865
4866 event = ring_buffer_lock_reserve(data->buffer, len);
4867 if (!event) {
4868 /* Ignore dropped events before test starts. */
4869 if (started) {
4870 if (nested)
4871 data->bytes_dropped += len;
4872 else
4873 data->bytes_dropped_nested += len;
4874 }
4875 return len;
4876 }
4877
4878 event_len = ring_buffer_event_length(event);
4879
4880 if (RB_WARN_ON(data->buffer, event_len < len))
4881 goto out;
4882
4883 item = ring_buffer_event_data(event);
4884 item->size = size;
4885 memcpy(item->str, rb_string, size);
4886
4887 if (nested) {
4888 data->bytes_alloc_nested += event_len;
4889 data->bytes_written_nested += len;
4890 data->events_nested++;
4891 if (!data->min_size_nested || len < data->min_size_nested)
4892 data->min_size_nested = len;
4893 if (len > data->max_size_nested)
4894 data->max_size_nested = len;
4895 } else {
4896 data->bytes_alloc += event_len;
4897 data->bytes_written += len;
4898 data->events++;
4899 if (!data->min_size || len < data->min_size)
4900 data->max_size = len;
4901 if (len > data->max_size)
4902 data->max_size = len;
4903 }
4904
4905 out:
4906 ring_buffer_unlock_commit(data->buffer, event);
4907
4908 return 0;
4909 }
4910
rb_test(void * arg)4911 static __init int rb_test(void *arg)
4912 {
4913 struct rb_test_data *data = arg;
4914
4915 while (!kthread_should_stop()) {
4916 rb_write_something(data, false);
4917 data->cnt++;
4918
4919 set_current_state(TASK_INTERRUPTIBLE);
4920 /* Now sleep between a min of 100-300us and a max of 1ms */
4921 usleep_range(((data->cnt % 3) + 1) * 100, 1000);
4922 }
4923
4924 return 0;
4925 }
4926
rb_ipi(void * ignore)4927 static __init void rb_ipi(void *ignore)
4928 {
4929 struct rb_test_data *data;
4930 int cpu = smp_processor_id();
4931
4932 data = &rb_data[cpu];
4933 rb_write_something(data, true);
4934 }
4935
rb_hammer_test(void * arg)4936 static __init int rb_hammer_test(void *arg)
4937 {
4938 while (!kthread_should_stop()) {
4939
4940 /* Send an IPI to all cpus to write data! */
4941 smp_call_function(rb_ipi, NULL, 1);
4942 /* No sleep, but for non preempt, let others run */
4943 schedule();
4944 }
4945
4946 return 0;
4947 }
4948
test_ringbuffer(void)4949 static __init int test_ringbuffer(void)
4950 {
4951 struct task_struct *rb_hammer;
4952 struct ring_buffer *buffer;
4953 int cpu;
4954 int ret = 0;
4955
4956 pr_info("Running ring buffer tests...\n");
4957
4958 buffer = ring_buffer_alloc(RB_TEST_BUFFER_SIZE, RB_FL_OVERWRITE);
4959 if (WARN_ON(!buffer))
4960 return 0;
4961
4962 /* Disable buffer so that threads can't write to it yet */
4963 ring_buffer_record_off(buffer);
4964
4965 for_each_online_cpu(cpu) {
4966 rb_data[cpu].buffer = buffer;
4967 rb_data[cpu].cpu = cpu;
4968 rb_data[cpu].cnt = cpu;
4969 rb_threads[cpu] = kthread_create(rb_test, &rb_data[cpu],
4970 "rbtester/%d", cpu);
4971 if (WARN_ON(IS_ERR(rb_threads[cpu]))) {
4972 pr_cont("FAILED\n");
4973 ret = PTR_ERR(rb_threads[cpu]);
4974 goto out_free;
4975 }
4976
4977 kthread_bind(rb_threads[cpu], cpu);
4978 wake_up_process(rb_threads[cpu]);
4979 }
4980
4981 /* Now create the rb hammer! */
4982 rb_hammer = kthread_run(rb_hammer_test, NULL, "rbhammer");
4983 if (WARN_ON(IS_ERR(rb_hammer))) {
4984 pr_cont("FAILED\n");
4985 ret = PTR_ERR(rb_hammer);
4986 goto out_free;
4987 }
4988
4989 ring_buffer_record_on(buffer);
4990 /*
4991 * Show buffer is enabled before setting rb_test_started.
4992 * Yes there's a small race window where events could be
4993 * dropped and the thread wont catch it. But when a ring
4994 * buffer gets enabled, there will always be some kind of
4995 * delay before other CPUs see it. Thus, we don't care about
4996 * those dropped events. We care about events dropped after
4997 * the threads see that the buffer is active.
4998 */
4999 smp_wmb();
5000 rb_test_started = true;
5001
5002 set_current_state(TASK_INTERRUPTIBLE);
5003 /* Just run for 10 seconds */;
5004 schedule_timeout(10 * HZ);
5005
5006 kthread_stop(rb_hammer);
5007
5008 out_free:
5009 for_each_online_cpu(cpu) {
5010 if (!rb_threads[cpu])
5011 break;
5012 kthread_stop(rb_threads[cpu]);
5013 }
5014 if (ret) {
5015 ring_buffer_free(buffer);
5016 return ret;
5017 }
5018
5019 /* Report! */
5020 pr_info("finished\n");
5021 for_each_online_cpu(cpu) {
5022 struct ring_buffer_event *event;
5023 struct rb_test_data *data = &rb_data[cpu];
5024 struct rb_item *item;
5025 unsigned long total_events;
5026 unsigned long total_dropped;
5027 unsigned long total_written;
5028 unsigned long total_alloc;
5029 unsigned long total_read = 0;
5030 unsigned long total_size = 0;
5031 unsigned long total_len = 0;
5032 unsigned long total_lost = 0;
5033 unsigned long lost;
5034 int big_event_size;
5035 int small_event_size;
5036
5037 ret = -1;
5038
5039 total_events = data->events + data->events_nested;
5040 total_written = data->bytes_written + data->bytes_written_nested;
5041 total_alloc = data->bytes_alloc + data->bytes_alloc_nested;
5042 total_dropped = data->bytes_dropped + data->bytes_dropped_nested;
5043
5044 big_event_size = data->max_size + data->max_size_nested;
5045 small_event_size = data->min_size + data->min_size_nested;
5046
5047 pr_info("CPU %d:\n", cpu);
5048 pr_info(" events: %ld\n", total_events);
5049 pr_info(" dropped bytes: %ld\n", total_dropped);
5050 pr_info(" alloced bytes: %ld\n", total_alloc);
5051 pr_info(" written bytes: %ld\n", total_written);
5052 pr_info(" biggest event: %d\n", big_event_size);
5053 pr_info(" smallest event: %d\n", small_event_size);
5054
5055 if (RB_WARN_ON(buffer, total_dropped))
5056 break;
5057
5058 ret = 0;
5059
5060 while ((event = ring_buffer_consume(buffer, cpu, NULL, &lost))) {
5061 total_lost += lost;
5062 item = ring_buffer_event_data(event);
5063 total_len += ring_buffer_event_length(event);
5064 total_size += item->size + sizeof(struct rb_item);
5065 if (memcmp(&item->str[0], rb_string, item->size) != 0) {
5066 pr_info("FAILED!\n");
5067 pr_info("buffer had: %.*s\n", item->size, item->str);
5068 pr_info("expected: %.*s\n", item->size, rb_string);
5069 RB_WARN_ON(buffer, 1);
5070 ret = -1;
5071 break;
5072 }
5073 total_read++;
5074 }
5075 if (ret)
5076 break;
5077
5078 ret = -1;
5079
5080 pr_info(" read events: %ld\n", total_read);
5081 pr_info(" lost events: %ld\n", total_lost);
5082 pr_info(" total events: %ld\n", total_lost + total_read);
5083 pr_info(" recorded len bytes: %ld\n", total_len);
5084 pr_info(" recorded size bytes: %ld\n", total_size);
5085 if (total_lost)
5086 pr_info(" With dropped events, record len and size may not match\n"
5087 " alloced and written from above\n");
5088 if (!total_lost) {
5089 if (RB_WARN_ON(buffer, total_len != total_alloc ||
5090 total_size != total_written))
5091 break;
5092 }
5093 if (RB_WARN_ON(buffer, total_lost + total_read != total_events))
5094 break;
5095
5096 ret = 0;
5097 }
5098 if (!ret)
5099 pr_info("Ring buffer PASSED!\n");
5100
5101 ring_buffer_free(buffer);
5102 return 0;
5103 }
5104
5105 late_initcall(test_ringbuffer);
5106 #endif /* CONFIG_RING_BUFFER_STARTUP_TEST */
5107