xref: /openbmc/linux/kernel/trace/ring_buffer.c (revision d2999e1b)
1 /*
2  * Generic ring buffer
3  *
4  * Copyright (C) 2008 Steven Rostedt <srostedt@redhat.com>
5  */
6 #include <linux/ftrace_event.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/debugfs.h>
13 #include <linux/uaccess.h>
14 #include <linux/hardirq.h>
15 #include <linux/kthread.h>	/* for self test */
16 #include <linux/kmemcheck.h>
17 #include <linux/module.h>
18 #include <linux/percpu.h>
19 #include <linux/mutex.h>
20 #include <linux/delay.h>
21 #include <linux/slab.h>
22 #include <linux/init.h>
23 #include <linux/hash.h>
24 #include <linux/list.h>
25 #include <linux/cpu.h>
26 #include <linux/fs.h>
27 
28 #include <asm/local.h>
29 
30 static void update_pages_handler(struct work_struct *work);
31 
32 /*
33  * The ring buffer header is special. We must manually up keep it.
34  */
35 int ring_buffer_print_entry_header(struct trace_seq *s)
36 {
37 	int ret;
38 
39 	ret = trace_seq_puts(s, "# compressed entry header\n");
40 	ret = trace_seq_puts(s, "\ttype_len    :    5 bits\n");
41 	ret = trace_seq_puts(s, "\ttime_delta  :   27 bits\n");
42 	ret = trace_seq_puts(s, "\tarray       :   32 bits\n");
43 	ret = trace_seq_putc(s, '\n');
44 	ret = trace_seq_printf(s, "\tpadding     : type == %d\n",
45 			       RINGBUF_TYPE_PADDING);
46 	ret = trace_seq_printf(s, "\ttime_extend : type == %d\n",
47 			       RINGBUF_TYPE_TIME_EXTEND);
48 	ret = trace_seq_printf(s, "\tdata max type_len  == %d\n",
49 			       RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
50 
51 	return ret;
52 }
53 
54 /*
55  * The ring buffer is made up of a list of pages. A separate list of pages is
56  * allocated for each CPU. A writer may only write to a buffer that is
57  * associated with the CPU it is currently executing on.  A reader may read
58  * from any per cpu buffer.
59  *
60  * The reader is special. For each per cpu buffer, the reader has its own
61  * reader page. When a reader has read the entire reader page, this reader
62  * page is swapped with another page in the ring buffer.
63  *
64  * Now, as long as the writer is off the reader page, the reader can do what
65  * ever it wants with that page. The writer will never write to that page
66  * again (as long as it is out of the ring buffer).
67  *
68  * Here's some silly ASCII art.
69  *
70  *   +------+
71  *   |reader|          RING BUFFER
72  *   |page  |
73  *   +------+        +---+   +---+   +---+
74  *                   |   |-->|   |-->|   |
75  *                   +---+   +---+   +---+
76  *                     ^               |
77  *                     |               |
78  *                     +---------------+
79  *
80  *
81  *   +------+
82  *   |reader|          RING BUFFER
83  *   |page  |------------------v
84  *   +------+        +---+   +---+   +---+
85  *                   |   |-->|   |-->|   |
86  *                   +---+   +---+   +---+
87  *                     ^               |
88  *                     |               |
89  *                     +---------------+
90  *
91  *
92  *   +------+
93  *   |reader|          RING BUFFER
94  *   |page  |------------------v
95  *   +------+        +---+   +---+   +---+
96  *      ^            |   |-->|   |-->|   |
97  *      |            +---+   +---+   +---+
98  *      |                              |
99  *      |                              |
100  *      +------------------------------+
101  *
102  *
103  *   +------+
104  *   |buffer|          RING BUFFER
105  *   |page  |------------------v
106  *   +------+        +---+   +---+   +---+
107  *      ^            |   |   |   |-->|   |
108  *      |   New      +---+   +---+   +---+
109  *      |  Reader------^               |
110  *      |   page                       |
111  *      +------------------------------+
112  *
113  *
114  * After we make this swap, the reader can hand this page off to the splice
115  * code and be done with it. It can even allocate a new page if it needs to
116  * and swap that into the ring buffer.
117  *
118  * We will be using cmpxchg soon to make all this lockless.
119  *
120  */
121 
122 /*
123  * A fast way to enable or disable all ring buffers is to
124  * call tracing_on or tracing_off. Turning off the ring buffers
125  * prevents all ring buffers from being recorded to.
126  * Turning this switch on, makes it OK to write to the
127  * ring buffer, if the ring buffer is enabled itself.
128  *
129  * There's three layers that must be on in order to write
130  * to the ring buffer.
131  *
132  * 1) This global flag must be set.
133  * 2) The ring buffer must be enabled for recording.
134  * 3) The per cpu buffer must be enabled for recording.
135  *
136  * In case of an anomaly, this global flag has a bit set that
137  * will permantly disable all ring buffers.
138  */
139 
140 /*
141  * Global flag to disable all recording to ring buffers
142  *  This has two bits: ON, DISABLED
143  *
144  *  ON   DISABLED
145  * ---- ----------
146  *   0      0        : ring buffers are off
147  *   1      0        : ring buffers are on
148  *   X      1        : ring buffers are permanently disabled
149  */
150 
151 enum {
152 	RB_BUFFERS_ON_BIT	= 0,
153 	RB_BUFFERS_DISABLED_BIT	= 1,
154 };
155 
156 enum {
157 	RB_BUFFERS_ON		= 1 << RB_BUFFERS_ON_BIT,
158 	RB_BUFFERS_DISABLED	= 1 << RB_BUFFERS_DISABLED_BIT,
159 };
160 
161 static unsigned long ring_buffer_flags __read_mostly = RB_BUFFERS_ON;
162 
163 /* Used for individual buffers (after the counter) */
164 #define RB_BUFFER_OFF		(1 << 20)
165 
166 #define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
167 
168 /**
169  * tracing_off_permanent - permanently disable ring buffers
170  *
171  * This function, once called, will disable all ring buffers
172  * permanently.
173  */
174 void tracing_off_permanent(void)
175 {
176 	set_bit(RB_BUFFERS_DISABLED_BIT, &ring_buffer_flags);
177 }
178 
179 #define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
180 #define RB_ALIGNMENT		4U
181 #define RB_MAX_SMALL_DATA	(RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
182 #define RB_EVNT_MIN_SIZE	8U	/* two 32bit words */
183 
184 #ifndef CONFIG_HAVE_64BIT_ALIGNED_ACCESS
185 # define RB_FORCE_8BYTE_ALIGNMENT	0
186 # define RB_ARCH_ALIGNMENT		RB_ALIGNMENT
187 #else
188 # define RB_FORCE_8BYTE_ALIGNMENT	1
189 # define RB_ARCH_ALIGNMENT		8U
190 #endif
191 
192 #define RB_ALIGN_DATA		__aligned(RB_ARCH_ALIGNMENT)
193 
194 /* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
195 #define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
196 
197 enum {
198 	RB_LEN_TIME_EXTEND = 8,
199 	RB_LEN_TIME_STAMP = 16,
200 };
201 
202 #define skip_time_extend(event) \
203 	((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
204 
205 static inline int rb_null_event(struct ring_buffer_event *event)
206 {
207 	return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
208 }
209 
210 static void rb_event_set_padding(struct ring_buffer_event *event)
211 {
212 	/* padding has a NULL time_delta */
213 	event->type_len = RINGBUF_TYPE_PADDING;
214 	event->time_delta = 0;
215 }
216 
217 static unsigned
218 rb_event_data_length(struct ring_buffer_event *event)
219 {
220 	unsigned length;
221 
222 	if (event->type_len)
223 		length = event->type_len * RB_ALIGNMENT;
224 	else
225 		length = event->array[0];
226 	return length + RB_EVNT_HDR_SIZE;
227 }
228 
229 /*
230  * Return the length of the given event. Will return
231  * the length of the time extend if the event is a
232  * time extend.
233  */
234 static inline unsigned
235 rb_event_length(struct ring_buffer_event *event)
236 {
237 	switch (event->type_len) {
238 	case RINGBUF_TYPE_PADDING:
239 		if (rb_null_event(event))
240 			/* undefined */
241 			return -1;
242 		return  event->array[0] + RB_EVNT_HDR_SIZE;
243 
244 	case RINGBUF_TYPE_TIME_EXTEND:
245 		return RB_LEN_TIME_EXTEND;
246 
247 	case RINGBUF_TYPE_TIME_STAMP:
248 		return RB_LEN_TIME_STAMP;
249 
250 	case RINGBUF_TYPE_DATA:
251 		return rb_event_data_length(event);
252 	default:
253 		BUG();
254 	}
255 	/* not hit */
256 	return 0;
257 }
258 
259 /*
260  * Return total length of time extend and data,
261  *   or just the event length for all other events.
262  */
263 static inline unsigned
264 rb_event_ts_length(struct ring_buffer_event *event)
265 {
266 	unsigned len = 0;
267 
268 	if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
269 		/* time extends include the data event after it */
270 		len = RB_LEN_TIME_EXTEND;
271 		event = skip_time_extend(event);
272 	}
273 	return len + rb_event_length(event);
274 }
275 
276 /**
277  * ring_buffer_event_length - return the length of the event
278  * @event: the event to get the length of
279  *
280  * Returns the size of the data load of a data event.
281  * If the event is something other than a data event, it
282  * returns the size of the event itself. With the exception
283  * of a TIME EXTEND, where it still returns the size of the
284  * data load of the data event after it.
285  */
286 unsigned ring_buffer_event_length(struct ring_buffer_event *event)
287 {
288 	unsigned length;
289 
290 	if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
291 		event = skip_time_extend(event);
292 
293 	length = rb_event_length(event);
294 	if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
295 		return length;
296 	length -= RB_EVNT_HDR_SIZE;
297 	if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
298                 length -= sizeof(event->array[0]);
299 	return length;
300 }
301 EXPORT_SYMBOL_GPL(ring_buffer_event_length);
302 
303 /* inline for ring buffer fast paths */
304 static void *
305 rb_event_data(struct ring_buffer_event *event)
306 {
307 	if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
308 		event = skip_time_extend(event);
309 	BUG_ON(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
310 	/* If length is in len field, then array[0] has the data */
311 	if (event->type_len)
312 		return (void *)&event->array[0];
313 	/* Otherwise length is in array[0] and array[1] has the data */
314 	return (void *)&event->array[1];
315 }
316 
317 /**
318  * ring_buffer_event_data - return the data of the event
319  * @event: the event to get the data from
320  */
321 void *ring_buffer_event_data(struct ring_buffer_event *event)
322 {
323 	return rb_event_data(event);
324 }
325 EXPORT_SYMBOL_GPL(ring_buffer_event_data);
326 
327 #define for_each_buffer_cpu(buffer, cpu)		\
328 	for_each_cpu(cpu, buffer->cpumask)
329 
330 #define TS_SHIFT	27
331 #define TS_MASK		((1ULL << TS_SHIFT) - 1)
332 #define TS_DELTA_TEST	(~TS_MASK)
333 
334 /* Flag when events were overwritten */
335 #define RB_MISSED_EVENTS	(1 << 31)
336 /* Missed count stored at end */
337 #define RB_MISSED_STORED	(1 << 30)
338 
339 struct buffer_data_page {
340 	u64		 time_stamp;	/* page time stamp */
341 	local_t		 commit;	/* write committed index */
342 	unsigned char	 data[] RB_ALIGN_DATA;	/* data of buffer page */
343 };
344 
345 /*
346  * Note, the buffer_page list must be first. The buffer pages
347  * are allocated in cache lines, which means that each buffer
348  * page will be at the beginning of a cache line, and thus
349  * the least significant bits will be zero. We use this to
350  * add flags in the list struct pointers, to make the ring buffer
351  * lockless.
352  */
353 struct buffer_page {
354 	struct list_head list;		/* list of buffer pages */
355 	local_t		 write;		/* index for next write */
356 	unsigned	 read;		/* index for next read */
357 	local_t		 entries;	/* entries on this page */
358 	unsigned long	 real_end;	/* real end of data */
359 	struct buffer_data_page *page;	/* Actual data page */
360 };
361 
362 /*
363  * The buffer page counters, write and entries, must be reset
364  * atomically when crossing page boundaries. To synchronize this
365  * update, two counters are inserted into the number. One is
366  * the actual counter for the write position or count on the page.
367  *
368  * The other is a counter of updaters. Before an update happens
369  * the update partition of the counter is incremented. This will
370  * allow the updater to update the counter atomically.
371  *
372  * The counter is 20 bits, and the state data is 12.
373  */
374 #define RB_WRITE_MASK		0xfffff
375 #define RB_WRITE_INTCNT		(1 << 20)
376 
377 static void rb_init_page(struct buffer_data_page *bpage)
378 {
379 	local_set(&bpage->commit, 0);
380 }
381 
382 /**
383  * ring_buffer_page_len - the size of data on the page.
384  * @page: The page to read
385  *
386  * Returns the amount of data on the page, including buffer page header.
387  */
388 size_t ring_buffer_page_len(void *page)
389 {
390 	return local_read(&((struct buffer_data_page *)page)->commit)
391 		+ BUF_PAGE_HDR_SIZE;
392 }
393 
394 /*
395  * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
396  * this issue out.
397  */
398 static void free_buffer_page(struct buffer_page *bpage)
399 {
400 	free_page((unsigned long)bpage->page);
401 	kfree(bpage);
402 }
403 
404 /*
405  * We need to fit the time_stamp delta into 27 bits.
406  */
407 static inline int test_time_stamp(u64 delta)
408 {
409 	if (delta & TS_DELTA_TEST)
410 		return 1;
411 	return 0;
412 }
413 
414 #define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
415 
416 /* Max payload is BUF_PAGE_SIZE - header (8bytes) */
417 #define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
418 
419 int ring_buffer_print_page_header(struct trace_seq *s)
420 {
421 	struct buffer_data_page field;
422 	int ret;
423 
424 	ret = trace_seq_printf(s, "\tfield: u64 timestamp;\t"
425 			       "offset:0;\tsize:%u;\tsigned:%u;\n",
426 			       (unsigned int)sizeof(field.time_stamp),
427 			       (unsigned int)is_signed_type(u64));
428 
429 	ret = trace_seq_printf(s, "\tfield: local_t commit;\t"
430 			       "offset:%u;\tsize:%u;\tsigned:%u;\n",
431 			       (unsigned int)offsetof(typeof(field), commit),
432 			       (unsigned int)sizeof(field.commit),
433 			       (unsigned int)is_signed_type(long));
434 
435 	ret = trace_seq_printf(s, "\tfield: int overwrite;\t"
436 			       "offset:%u;\tsize:%u;\tsigned:%u;\n",
437 			       (unsigned int)offsetof(typeof(field), commit),
438 			       1,
439 			       (unsigned int)is_signed_type(long));
440 
441 	ret = trace_seq_printf(s, "\tfield: char data;\t"
442 			       "offset:%u;\tsize:%u;\tsigned:%u;\n",
443 			       (unsigned int)offsetof(typeof(field), data),
444 			       (unsigned int)BUF_PAGE_SIZE,
445 			       (unsigned int)is_signed_type(char));
446 
447 	return ret;
448 }
449 
450 struct rb_irq_work {
451 	struct irq_work			work;
452 	wait_queue_head_t		waiters;
453 	bool				waiters_pending;
454 };
455 
456 /*
457  * head_page == tail_page && head == tail then buffer is empty.
458  */
459 struct ring_buffer_per_cpu {
460 	int				cpu;
461 	atomic_t			record_disabled;
462 	struct ring_buffer		*buffer;
463 	raw_spinlock_t			reader_lock;	/* serialize readers */
464 	arch_spinlock_t			lock;
465 	struct lock_class_key		lock_key;
466 	unsigned int			nr_pages;
467 	struct list_head		*pages;
468 	struct buffer_page		*head_page;	/* read from head */
469 	struct buffer_page		*tail_page;	/* write to tail */
470 	struct buffer_page		*commit_page;	/* committed pages */
471 	struct buffer_page		*reader_page;
472 	unsigned long			lost_events;
473 	unsigned long			last_overrun;
474 	local_t				entries_bytes;
475 	local_t				entries;
476 	local_t				overrun;
477 	local_t				commit_overrun;
478 	local_t				dropped_events;
479 	local_t				committing;
480 	local_t				commits;
481 	unsigned long			read;
482 	unsigned long			read_bytes;
483 	u64				write_stamp;
484 	u64				read_stamp;
485 	/* ring buffer pages to update, > 0 to add, < 0 to remove */
486 	int				nr_pages_to_update;
487 	struct list_head		new_pages; /* new pages to add */
488 	struct work_struct		update_pages_work;
489 	struct completion		update_done;
490 
491 	struct rb_irq_work		irq_work;
492 };
493 
494 struct ring_buffer {
495 	unsigned			flags;
496 	int				cpus;
497 	atomic_t			record_disabled;
498 	atomic_t			resize_disabled;
499 	cpumask_var_t			cpumask;
500 
501 	struct lock_class_key		*reader_lock_key;
502 
503 	struct mutex			mutex;
504 
505 	struct ring_buffer_per_cpu	**buffers;
506 
507 #ifdef CONFIG_HOTPLUG_CPU
508 	struct notifier_block		cpu_notify;
509 #endif
510 	u64				(*clock)(void);
511 
512 	struct rb_irq_work		irq_work;
513 };
514 
515 struct ring_buffer_iter {
516 	struct ring_buffer_per_cpu	*cpu_buffer;
517 	unsigned long			head;
518 	struct buffer_page		*head_page;
519 	struct buffer_page		*cache_reader_page;
520 	unsigned long			cache_read;
521 	u64				read_stamp;
522 };
523 
524 /*
525  * rb_wake_up_waiters - wake up tasks waiting for ring buffer input
526  *
527  * Schedules a delayed work to wake up any task that is blocked on the
528  * ring buffer waiters queue.
529  */
530 static void rb_wake_up_waiters(struct irq_work *work)
531 {
532 	struct rb_irq_work *rbwork = container_of(work, struct rb_irq_work, work);
533 
534 	wake_up_all(&rbwork->waiters);
535 }
536 
537 /**
538  * ring_buffer_wait - wait for input to the ring buffer
539  * @buffer: buffer to wait on
540  * @cpu: the cpu buffer to wait on
541  *
542  * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
543  * as data is added to any of the @buffer's cpu buffers. Otherwise
544  * it will wait for data to be added to a specific cpu buffer.
545  */
546 int ring_buffer_wait(struct ring_buffer *buffer, int cpu)
547 {
548 	struct ring_buffer_per_cpu *cpu_buffer;
549 	DEFINE_WAIT(wait);
550 	struct rb_irq_work *work;
551 
552 	/*
553 	 * Depending on what the caller is waiting for, either any
554 	 * data in any cpu buffer, or a specific buffer, put the
555 	 * caller on the appropriate wait queue.
556 	 */
557 	if (cpu == RING_BUFFER_ALL_CPUS)
558 		work = &buffer->irq_work;
559 	else {
560 		if (!cpumask_test_cpu(cpu, buffer->cpumask))
561 			return -ENODEV;
562 		cpu_buffer = buffer->buffers[cpu];
563 		work = &cpu_buffer->irq_work;
564 	}
565 
566 
567 	prepare_to_wait(&work->waiters, &wait, TASK_INTERRUPTIBLE);
568 
569 	/*
570 	 * The events can happen in critical sections where
571 	 * checking a work queue can cause deadlocks.
572 	 * After adding a task to the queue, this flag is set
573 	 * only to notify events to try to wake up the queue
574 	 * using irq_work.
575 	 *
576 	 * We don't clear it even if the buffer is no longer
577 	 * empty. The flag only causes the next event to run
578 	 * irq_work to do the work queue wake up. The worse
579 	 * that can happen if we race with !trace_empty() is that
580 	 * an event will cause an irq_work to try to wake up
581 	 * an empty queue.
582 	 *
583 	 * There's no reason to protect this flag either, as
584 	 * the work queue and irq_work logic will do the necessary
585 	 * synchronization for the wake ups. The only thing
586 	 * that is necessary is that the wake up happens after
587 	 * a task has been queued. It's OK for spurious wake ups.
588 	 */
589 	work->waiters_pending = true;
590 
591 	if ((cpu == RING_BUFFER_ALL_CPUS && ring_buffer_empty(buffer)) ||
592 	    (cpu != RING_BUFFER_ALL_CPUS && ring_buffer_empty_cpu(buffer, cpu)))
593 		schedule();
594 
595 	finish_wait(&work->waiters, &wait);
596 	return 0;
597 }
598 
599 /**
600  * ring_buffer_poll_wait - poll on buffer input
601  * @buffer: buffer to wait on
602  * @cpu: the cpu buffer to wait on
603  * @filp: the file descriptor
604  * @poll_table: The poll descriptor
605  *
606  * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
607  * as data is added to any of the @buffer's cpu buffers. Otherwise
608  * it will wait for data to be added to a specific cpu buffer.
609  *
610  * Returns POLLIN | POLLRDNORM if data exists in the buffers,
611  * zero otherwise.
612  */
613 int ring_buffer_poll_wait(struct ring_buffer *buffer, int cpu,
614 			  struct file *filp, poll_table *poll_table)
615 {
616 	struct ring_buffer_per_cpu *cpu_buffer;
617 	struct rb_irq_work *work;
618 
619 	if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
620 	    (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
621 		return POLLIN | POLLRDNORM;
622 
623 	if (cpu == RING_BUFFER_ALL_CPUS)
624 		work = &buffer->irq_work;
625 	else {
626 		if (!cpumask_test_cpu(cpu, buffer->cpumask))
627 			return -EINVAL;
628 
629 		cpu_buffer = buffer->buffers[cpu];
630 		work = &cpu_buffer->irq_work;
631 	}
632 
633 	work->waiters_pending = true;
634 	poll_wait(filp, &work->waiters, poll_table);
635 
636 	if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
637 	    (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
638 		return POLLIN | POLLRDNORM;
639 	return 0;
640 }
641 
642 /* buffer may be either ring_buffer or ring_buffer_per_cpu */
643 #define RB_WARN_ON(b, cond)						\
644 	({								\
645 		int _____ret = unlikely(cond);				\
646 		if (_____ret) {						\
647 			if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
648 				struct ring_buffer_per_cpu *__b =	\
649 					(void *)b;			\
650 				atomic_inc(&__b->buffer->record_disabled); \
651 			} else						\
652 				atomic_inc(&b->record_disabled);	\
653 			WARN_ON(1);					\
654 		}							\
655 		_____ret;						\
656 	})
657 
658 /* Up this if you want to test the TIME_EXTENTS and normalization */
659 #define DEBUG_SHIFT 0
660 
661 static inline u64 rb_time_stamp(struct ring_buffer *buffer)
662 {
663 	/* shift to debug/test normalization and TIME_EXTENTS */
664 	return buffer->clock() << DEBUG_SHIFT;
665 }
666 
667 u64 ring_buffer_time_stamp(struct ring_buffer *buffer, int cpu)
668 {
669 	u64 time;
670 
671 	preempt_disable_notrace();
672 	time = rb_time_stamp(buffer);
673 	preempt_enable_no_resched_notrace();
674 
675 	return time;
676 }
677 EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
678 
679 void ring_buffer_normalize_time_stamp(struct ring_buffer *buffer,
680 				      int cpu, u64 *ts)
681 {
682 	/* Just stupid testing the normalize function and deltas */
683 	*ts >>= DEBUG_SHIFT;
684 }
685 EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
686 
687 /*
688  * Making the ring buffer lockless makes things tricky.
689  * Although writes only happen on the CPU that they are on,
690  * and they only need to worry about interrupts. Reads can
691  * happen on any CPU.
692  *
693  * The reader page is always off the ring buffer, but when the
694  * reader finishes with a page, it needs to swap its page with
695  * a new one from the buffer. The reader needs to take from
696  * the head (writes go to the tail). But if a writer is in overwrite
697  * mode and wraps, it must push the head page forward.
698  *
699  * Here lies the problem.
700  *
701  * The reader must be careful to replace only the head page, and
702  * not another one. As described at the top of the file in the
703  * ASCII art, the reader sets its old page to point to the next
704  * page after head. It then sets the page after head to point to
705  * the old reader page. But if the writer moves the head page
706  * during this operation, the reader could end up with the tail.
707  *
708  * We use cmpxchg to help prevent this race. We also do something
709  * special with the page before head. We set the LSB to 1.
710  *
711  * When the writer must push the page forward, it will clear the
712  * bit that points to the head page, move the head, and then set
713  * the bit that points to the new head page.
714  *
715  * We also don't want an interrupt coming in and moving the head
716  * page on another writer. Thus we use the second LSB to catch
717  * that too. Thus:
718  *
719  * head->list->prev->next        bit 1          bit 0
720  *                              -------        -------
721  * Normal page                     0              0
722  * Points to head page             0              1
723  * New head page                   1              0
724  *
725  * Note we can not trust the prev pointer of the head page, because:
726  *
727  * +----+       +-----+        +-----+
728  * |    |------>|  T  |---X--->|  N  |
729  * |    |<------|     |        |     |
730  * +----+       +-----+        +-----+
731  *   ^                           ^ |
732  *   |          +-----+          | |
733  *   +----------|  R  |----------+ |
734  *              |     |<-----------+
735  *              +-----+
736  *
737  * Key:  ---X-->  HEAD flag set in pointer
738  *         T      Tail page
739  *         R      Reader page
740  *         N      Next page
741  *
742  * (see __rb_reserve_next() to see where this happens)
743  *
744  *  What the above shows is that the reader just swapped out
745  *  the reader page with a page in the buffer, but before it
746  *  could make the new header point back to the new page added
747  *  it was preempted by a writer. The writer moved forward onto
748  *  the new page added by the reader and is about to move forward
749  *  again.
750  *
751  *  You can see, it is legitimate for the previous pointer of
752  *  the head (or any page) not to point back to itself. But only
753  *  temporarially.
754  */
755 
756 #define RB_PAGE_NORMAL		0UL
757 #define RB_PAGE_HEAD		1UL
758 #define RB_PAGE_UPDATE		2UL
759 
760 
761 #define RB_FLAG_MASK		3UL
762 
763 /* PAGE_MOVED is not part of the mask */
764 #define RB_PAGE_MOVED		4UL
765 
766 /*
767  * rb_list_head - remove any bit
768  */
769 static struct list_head *rb_list_head(struct list_head *list)
770 {
771 	unsigned long val = (unsigned long)list;
772 
773 	return (struct list_head *)(val & ~RB_FLAG_MASK);
774 }
775 
776 /*
777  * rb_is_head_page - test if the given page is the head page
778  *
779  * Because the reader may move the head_page pointer, we can
780  * not trust what the head page is (it may be pointing to
781  * the reader page). But if the next page is a header page,
782  * its flags will be non zero.
783  */
784 static inline int
785 rb_is_head_page(struct ring_buffer_per_cpu *cpu_buffer,
786 		struct buffer_page *page, struct list_head *list)
787 {
788 	unsigned long val;
789 
790 	val = (unsigned long)list->next;
791 
792 	if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
793 		return RB_PAGE_MOVED;
794 
795 	return val & RB_FLAG_MASK;
796 }
797 
798 /*
799  * rb_is_reader_page
800  *
801  * The unique thing about the reader page, is that, if the
802  * writer is ever on it, the previous pointer never points
803  * back to the reader page.
804  */
805 static int rb_is_reader_page(struct buffer_page *page)
806 {
807 	struct list_head *list = page->list.prev;
808 
809 	return rb_list_head(list->next) != &page->list;
810 }
811 
812 /*
813  * rb_set_list_to_head - set a list_head to be pointing to head.
814  */
815 static void rb_set_list_to_head(struct ring_buffer_per_cpu *cpu_buffer,
816 				struct list_head *list)
817 {
818 	unsigned long *ptr;
819 
820 	ptr = (unsigned long *)&list->next;
821 	*ptr |= RB_PAGE_HEAD;
822 	*ptr &= ~RB_PAGE_UPDATE;
823 }
824 
825 /*
826  * rb_head_page_activate - sets up head page
827  */
828 static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
829 {
830 	struct buffer_page *head;
831 
832 	head = cpu_buffer->head_page;
833 	if (!head)
834 		return;
835 
836 	/*
837 	 * Set the previous list pointer to have the HEAD flag.
838 	 */
839 	rb_set_list_to_head(cpu_buffer, head->list.prev);
840 }
841 
842 static void rb_list_head_clear(struct list_head *list)
843 {
844 	unsigned long *ptr = (unsigned long *)&list->next;
845 
846 	*ptr &= ~RB_FLAG_MASK;
847 }
848 
849 /*
850  * rb_head_page_dactivate - clears head page ptr (for free list)
851  */
852 static void
853 rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
854 {
855 	struct list_head *hd;
856 
857 	/* Go through the whole list and clear any pointers found. */
858 	rb_list_head_clear(cpu_buffer->pages);
859 
860 	list_for_each(hd, cpu_buffer->pages)
861 		rb_list_head_clear(hd);
862 }
863 
864 static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
865 			    struct buffer_page *head,
866 			    struct buffer_page *prev,
867 			    int old_flag, int new_flag)
868 {
869 	struct list_head *list;
870 	unsigned long val = (unsigned long)&head->list;
871 	unsigned long ret;
872 
873 	list = &prev->list;
874 
875 	val &= ~RB_FLAG_MASK;
876 
877 	ret = cmpxchg((unsigned long *)&list->next,
878 		      val | old_flag, val | new_flag);
879 
880 	/* check if the reader took the page */
881 	if ((ret & ~RB_FLAG_MASK) != val)
882 		return RB_PAGE_MOVED;
883 
884 	return ret & RB_FLAG_MASK;
885 }
886 
887 static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
888 				   struct buffer_page *head,
889 				   struct buffer_page *prev,
890 				   int old_flag)
891 {
892 	return rb_head_page_set(cpu_buffer, head, prev,
893 				old_flag, RB_PAGE_UPDATE);
894 }
895 
896 static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
897 				 struct buffer_page *head,
898 				 struct buffer_page *prev,
899 				 int old_flag)
900 {
901 	return rb_head_page_set(cpu_buffer, head, prev,
902 				old_flag, RB_PAGE_HEAD);
903 }
904 
905 static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
906 				   struct buffer_page *head,
907 				   struct buffer_page *prev,
908 				   int old_flag)
909 {
910 	return rb_head_page_set(cpu_buffer, head, prev,
911 				old_flag, RB_PAGE_NORMAL);
912 }
913 
914 static inline void rb_inc_page(struct ring_buffer_per_cpu *cpu_buffer,
915 			       struct buffer_page **bpage)
916 {
917 	struct list_head *p = rb_list_head((*bpage)->list.next);
918 
919 	*bpage = list_entry(p, struct buffer_page, list);
920 }
921 
922 static struct buffer_page *
923 rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
924 {
925 	struct buffer_page *head;
926 	struct buffer_page *page;
927 	struct list_head *list;
928 	int i;
929 
930 	if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
931 		return NULL;
932 
933 	/* sanity check */
934 	list = cpu_buffer->pages;
935 	if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
936 		return NULL;
937 
938 	page = head = cpu_buffer->head_page;
939 	/*
940 	 * It is possible that the writer moves the header behind
941 	 * where we started, and we miss in one loop.
942 	 * A second loop should grab the header, but we'll do
943 	 * three loops just because I'm paranoid.
944 	 */
945 	for (i = 0; i < 3; i++) {
946 		do {
947 			if (rb_is_head_page(cpu_buffer, page, page->list.prev)) {
948 				cpu_buffer->head_page = page;
949 				return page;
950 			}
951 			rb_inc_page(cpu_buffer, &page);
952 		} while (page != head);
953 	}
954 
955 	RB_WARN_ON(cpu_buffer, 1);
956 
957 	return NULL;
958 }
959 
960 static int rb_head_page_replace(struct buffer_page *old,
961 				struct buffer_page *new)
962 {
963 	unsigned long *ptr = (unsigned long *)&old->list.prev->next;
964 	unsigned long val;
965 	unsigned long ret;
966 
967 	val = *ptr & ~RB_FLAG_MASK;
968 	val |= RB_PAGE_HEAD;
969 
970 	ret = cmpxchg(ptr, val, (unsigned long)&new->list);
971 
972 	return ret == val;
973 }
974 
975 /*
976  * rb_tail_page_update - move the tail page forward
977  *
978  * Returns 1 if moved tail page, 0 if someone else did.
979  */
980 static int rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
981 			       struct buffer_page *tail_page,
982 			       struct buffer_page *next_page)
983 {
984 	struct buffer_page *old_tail;
985 	unsigned long old_entries;
986 	unsigned long old_write;
987 	int ret = 0;
988 
989 	/*
990 	 * The tail page now needs to be moved forward.
991 	 *
992 	 * We need to reset the tail page, but without messing
993 	 * with possible erasing of data brought in by interrupts
994 	 * that have moved the tail page and are currently on it.
995 	 *
996 	 * We add a counter to the write field to denote this.
997 	 */
998 	old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
999 	old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
1000 
1001 	/*
1002 	 * Just make sure we have seen our old_write and synchronize
1003 	 * with any interrupts that come in.
1004 	 */
1005 	barrier();
1006 
1007 	/*
1008 	 * If the tail page is still the same as what we think
1009 	 * it is, then it is up to us to update the tail
1010 	 * pointer.
1011 	 */
1012 	if (tail_page == cpu_buffer->tail_page) {
1013 		/* Zero the write counter */
1014 		unsigned long val = old_write & ~RB_WRITE_MASK;
1015 		unsigned long eval = old_entries & ~RB_WRITE_MASK;
1016 
1017 		/*
1018 		 * This will only succeed if an interrupt did
1019 		 * not come in and change it. In which case, we
1020 		 * do not want to modify it.
1021 		 *
1022 		 * We add (void) to let the compiler know that we do not care
1023 		 * about the return value of these functions. We use the
1024 		 * cmpxchg to only update if an interrupt did not already
1025 		 * do it for us. If the cmpxchg fails, we don't care.
1026 		 */
1027 		(void)local_cmpxchg(&next_page->write, old_write, val);
1028 		(void)local_cmpxchg(&next_page->entries, old_entries, eval);
1029 
1030 		/*
1031 		 * No need to worry about races with clearing out the commit.
1032 		 * it only can increment when a commit takes place. But that
1033 		 * only happens in the outer most nested commit.
1034 		 */
1035 		local_set(&next_page->page->commit, 0);
1036 
1037 		old_tail = cmpxchg(&cpu_buffer->tail_page,
1038 				   tail_page, next_page);
1039 
1040 		if (old_tail == tail_page)
1041 			ret = 1;
1042 	}
1043 
1044 	return ret;
1045 }
1046 
1047 static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
1048 			  struct buffer_page *bpage)
1049 {
1050 	unsigned long val = (unsigned long)bpage;
1051 
1052 	if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK))
1053 		return 1;
1054 
1055 	return 0;
1056 }
1057 
1058 /**
1059  * rb_check_list - make sure a pointer to a list has the last bits zero
1060  */
1061 static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer,
1062 			 struct list_head *list)
1063 {
1064 	if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev))
1065 		return 1;
1066 	if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next))
1067 		return 1;
1068 	return 0;
1069 }
1070 
1071 /**
1072  * rb_check_pages - integrity check of buffer pages
1073  * @cpu_buffer: CPU buffer with pages to test
1074  *
1075  * As a safety measure we check to make sure the data pages have not
1076  * been corrupted.
1077  */
1078 static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
1079 {
1080 	struct list_head *head = cpu_buffer->pages;
1081 	struct buffer_page *bpage, *tmp;
1082 
1083 	/* Reset the head page if it exists */
1084 	if (cpu_buffer->head_page)
1085 		rb_set_head_page(cpu_buffer);
1086 
1087 	rb_head_page_deactivate(cpu_buffer);
1088 
1089 	if (RB_WARN_ON(cpu_buffer, head->next->prev != head))
1090 		return -1;
1091 	if (RB_WARN_ON(cpu_buffer, head->prev->next != head))
1092 		return -1;
1093 
1094 	if (rb_check_list(cpu_buffer, head))
1095 		return -1;
1096 
1097 	list_for_each_entry_safe(bpage, tmp, head, list) {
1098 		if (RB_WARN_ON(cpu_buffer,
1099 			       bpage->list.next->prev != &bpage->list))
1100 			return -1;
1101 		if (RB_WARN_ON(cpu_buffer,
1102 			       bpage->list.prev->next != &bpage->list))
1103 			return -1;
1104 		if (rb_check_list(cpu_buffer, &bpage->list))
1105 			return -1;
1106 	}
1107 
1108 	rb_head_page_activate(cpu_buffer);
1109 
1110 	return 0;
1111 }
1112 
1113 static int __rb_allocate_pages(int nr_pages, struct list_head *pages, int cpu)
1114 {
1115 	int i;
1116 	struct buffer_page *bpage, *tmp;
1117 
1118 	for (i = 0; i < nr_pages; i++) {
1119 		struct page *page;
1120 		/*
1121 		 * __GFP_NORETRY flag makes sure that the allocation fails
1122 		 * gracefully without invoking oom-killer and the system is
1123 		 * not destabilized.
1124 		 */
1125 		bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1126 				    GFP_KERNEL | __GFP_NORETRY,
1127 				    cpu_to_node(cpu));
1128 		if (!bpage)
1129 			goto free_pages;
1130 
1131 		list_add(&bpage->list, pages);
1132 
1133 		page = alloc_pages_node(cpu_to_node(cpu),
1134 					GFP_KERNEL | __GFP_NORETRY, 0);
1135 		if (!page)
1136 			goto free_pages;
1137 		bpage->page = page_address(page);
1138 		rb_init_page(bpage->page);
1139 	}
1140 
1141 	return 0;
1142 
1143 free_pages:
1144 	list_for_each_entry_safe(bpage, tmp, pages, list) {
1145 		list_del_init(&bpage->list);
1146 		free_buffer_page(bpage);
1147 	}
1148 
1149 	return -ENOMEM;
1150 }
1151 
1152 static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1153 			     unsigned nr_pages)
1154 {
1155 	LIST_HEAD(pages);
1156 
1157 	WARN_ON(!nr_pages);
1158 
1159 	if (__rb_allocate_pages(nr_pages, &pages, cpu_buffer->cpu))
1160 		return -ENOMEM;
1161 
1162 	/*
1163 	 * The ring buffer page list is a circular list that does not
1164 	 * start and end with a list head. All page list items point to
1165 	 * other pages.
1166 	 */
1167 	cpu_buffer->pages = pages.next;
1168 	list_del(&pages);
1169 
1170 	cpu_buffer->nr_pages = nr_pages;
1171 
1172 	rb_check_pages(cpu_buffer);
1173 
1174 	return 0;
1175 }
1176 
1177 static struct ring_buffer_per_cpu *
1178 rb_allocate_cpu_buffer(struct ring_buffer *buffer, int nr_pages, int cpu)
1179 {
1180 	struct ring_buffer_per_cpu *cpu_buffer;
1181 	struct buffer_page *bpage;
1182 	struct page *page;
1183 	int ret;
1184 
1185 	cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1186 				  GFP_KERNEL, cpu_to_node(cpu));
1187 	if (!cpu_buffer)
1188 		return NULL;
1189 
1190 	cpu_buffer->cpu = cpu;
1191 	cpu_buffer->buffer = buffer;
1192 	raw_spin_lock_init(&cpu_buffer->reader_lock);
1193 	lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1194 	cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1195 	INIT_WORK(&cpu_buffer->update_pages_work, update_pages_handler);
1196 	init_completion(&cpu_buffer->update_done);
1197 	init_irq_work(&cpu_buffer->irq_work.work, rb_wake_up_waiters);
1198 	init_waitqueue_head(&cpu_buffer->irq_work.waiters);
1199 
1200 	bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1201 			    GFP_KERNEL, cpu_to_node(cpu));
1202 	if (!bpage)
1203 		goto fail_free_buffer;
1204 
1205 	rb_check_bpage(cpu_buffer, bpage);
1206 
1207 	cpu_buffer->reader_page = bpage;
1208 	page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0);
1209 	if (!page)
1210 		goto fail_free_reader;
1211 	bpage->page = page_address(page);
1212 	rb_init_page(bpage->page);
1213 
1214 	INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1215 	INIT_LIST_HEAD(&cpu_buffer->new_pages);
1216 
1217 	ret = rb_allocate_pages(cpu_buffer, nr_pages);
1218 	if (ret < 0)
1219 		goto fail_free_reader;
1220 
1221 	cpu_buffer->head_page
1222 		= list_entry(cpu_buffer->pages, struct buffer_page, list);
1223 	cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1224 
1225 	rb_head_page_activate(cpu_buffer);
1226 
1227 	return cpu_buffer;
1228 
1229  fail_free_reader:
1230 	free_buffer_page(cpu_buffer->reader_page);
1231 
1232  fail_free_buffer:
1233 	kfree(cpu_buffer);
1234 	return NULL;
1235 }
1236 
1237 static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1238 {
1239 	struct list_head *head = cpu_buffer->pages;
1240 	struct buffer_page *bpage, *tmp;
1241 
1242 	free_buffer_page(cpu_buffer->reader_page);
1243 
1244 	rb_head_page_deactivate(cpu_buffer);
1245 
1246 	if (head) {
1247 		list_for_each_entry_safe(bpage, tmp, head, list) {
1248 			list_del_init(&bpage->list);
1249 			free_buffer_page(bpage);
1250 		}
1251 		bpage = list_entry(head, struct buffer_page, list);
1252 		free_buffer_page(bpage);
1253 	}
1254 
1255 	kfree(cpu_buffer);
1256 }
1257 
1258 #ifdef CONFIG_HOTPLUG_CPU
1259 static int rb_cpu_notify(struct notifier_block *self,
1260 			 unsigned long action, void *hcpu);
1261 #endif
1262 
1263 /**
1264  * __ring_buffer_alloc - allocate a new ring_buffer
1265  * @size: the size in bytes per cpu that is needed.
1266  * @flags: attributes to set for the ring buffer.
1267  *
1268  * Currently the only flag that is available is the RB_FL_OVERWRITE
1269  * flag. This flag means that the buffer will overwrite old data
1270  * when the buffer wraps. If this flag is not set, the buffer will
1271  * drop data when the tail hits the head.
1272  */
1273 struct ring_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1274 					struct lock_class_key *key)
1275 {
1276 	struct ring_buffer *buffer;
1277 	int bsize;
1278 	int cpu, nr_pages;
1279 
1280 	/* keep it in its own cache line */
1281 	buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1282 			 GFP_KERNEL);
1283 	if (!buffer)
1284 		return NULL;
1285 
1286 	if (!alloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1287 		goto fail_free_buffer;
1288 
1289 	nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1290 	buffer->flags = flags;
1291 	buffer->clock = trace_clock_local;
1292 	buffer->reader_lock_key = key;
1293 
1294 	init_irq_work(&buffer->irq_work.work, rb_wake_up_waiters);
1295 	init_waitqueue_head(&buffer->irq_work.waiters);
1296 
1297 	/* need at least two pages */
1298 	if (nr_pages < 2)
1299 		nr_pages = 2;
1300 
1301 	/*
1302 	 * In case of non-hotplug cpu, if the ring-buffer is allocated
1303 	 * in early initcall, it will not be notified of secondary cpus.
1304 	 * In that off case, we need to allocate for all possible cpus.
1305 	 */
1306 #ifdef CONFIG_HOTPLUG_CPU
1307 	cpu_notifier_register_begin();
1308 	cpumask_copy(buffer->cpumask, cpu_online_mask);
1309 #else
1310 	cpumask_copy(buffer->cpumask, cpu_possible_mask);
1311 #endif
1312 	buffer->cpus = nr_cpu_ids;
1313 
1314 	bsize = sizeof(void *) * nr_cpu_ids;
1315 	buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1316 				  GFP_KERNEL);
1317 	if (!buffer->buffers)
1318 		goto fail_free_cpumask;
1319 
1320 	for_each_buffer_cpu(buffer, cpu) {
1321 		buffer->buffers[cpu] =
1322 			rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
1323 		if (!buffer->buffers[cpu])
1324 			goto fail_free_buffers;
1325 	}
1326 
1327 #ifdef CONFIG_HOTPLUG_CPU
1328 	buffer->cpu_notify.notifier_call = rb_cpu_notify;
1329 	buffer->cpu_notify.priority = 0;
1330 	__register_cpu_notifier(&buffer->cpu_notify);
1331 	cpu_notifier_register_done();
1332 #endif
1333 
1334 	mutex_init(&buffer->mutex);
1335 
1336 	return buffer;
1337 
1338  fail_free_buffers:
1339 	for_each_buffer_cpu(buffer, cpu) {
1340 		if (buffer->buffers[cpu])
1341 			rb_free_cpu_buffer(buffer->buffers[cpu]);
1342 	}
1343 	kfree(buffer->buffers);
1344 
1345  fail_free_cpumask:
1346 	free_cpumask_var(buffer->cpumask);
1347 #ifdef CONFIG_HOTPLUG_CPU
1348 	cpu_notifier_register_done();
1349 #endif
1350 
1351  fail_free_buffer:
1352 	kfree(buffer);
1353 	return NULL;
1354 }
1355 EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1356 
1357 /**
1358  * ring_buffer_free - free a ring buffer.
1359  * @buffer: the buffer to free.
1360  */
1361 void
1362 ring_buffer_free(struct ring_buffer *buffer)
1363 {
1364 	int cpu;
1365 
1366 #ifdef CONFIG_HOTPLUG_CPU
1367 	cpu_notifier_register_begin();
1368 	__unregister_cpu_notifier(&buffer->cpu_notify);
1369 #endif
1370 
1371 	for_each_buffer_cpu(buffer, cpu)
1372 		rb_free_cpu_buffer(buffer->buffers[cpu]);
1373 
1374 #ifdef CONFIG_HOTPLUG_CPU
1375 	cpu_notifier_register_done();
1376 #endif
1377 
1378 	kfree(buffer->buffers);
1379 	free_cpumask_var(buffer->cpumask);
1380 
1381 	kfree(buffer);
1382 }
1383 EXPORT_SYMBOL_GPL(ring_buffer_free);
1384 
1385 void ring_buffer_set_clock(struct ring_buffer *buffer,
1386 			   u64 (*clock)(void))
1387 {
1388 	buffer->clock = clock;
1389 }
1390 
1391 static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1392 
1393 static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1394 {
1395 	return local_read(&bpage->entries) & RB_WRITE_MASK;
1396 }
1397 
1398 static inline unsigned long rb_page_write(struct buffer_page *bpage)
1399 {
1400 	return local_read(&bpage->write) & RB_WRITE_MASK;
1401 }
1402 
1403 static int
1404 rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned int nr_pages)
1405 {
1406 	struct list_head *tail_page, *to_remove, *next_page;
1407 	struct buffer_page *to_remove_page, *tmp_iter_page;
1408 	struct buffer_page *last_page, *first_page;
1409 	unsigned int nr_removed;
1410 	unsigned long head_bit;
1411 	int page_entries;
1412 
1413 	head_bit = 0;
1414 
1415 	raw_spin_lock_irq(&cpu_buffer->reader_lock);
1416 	atomic_inc(&cpu_buffer->record_disabled);
1417 	/*
1418 	 * We don't race with the readers since we have acquired the reader
1419 	 * lock. We also don't race with writers after disabling recording.
1420 	 * This makes it easy to figure out the first and the last page to be
1421 	 * removed from the list. We unlink all the pages in between including
1422 	 * the first and last pages. This is done in a busy loop so that we
1423 	 * lose the least number of traces.
1424 	 * The pages are freed after we restart recording and unlock readers.
1425 	 */
1426 	tail_page = &cpu_buffer->tail_page->list;
1427 
1428 	/*
1429 	 * tail page might be on reader page, we remove the next page
1430 	 * from the ring buffer
1431 	 */
1432 	if (cpu_buffer->tail_page == cpu_buffer->reader_page)
1433 		tail_page = rb_list_head(tail_page->next);
1434 	to_remove = tail_page;
1435 
1436 	/* start of pages to remove */
1437 	first_page = list_entry(rb_list_head(to_remove->next),
1438 				struct buffer_page, list);
1439 
1440 	for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
1441 		to_remove = rb_list_head(to_remove)->next;
1442 		head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
1443 	}
1444 
1445 	next_page = rb_list_head(to_remove)->next;
1446 
1447 	/*
1448 	 * Now we remove all pages between tail_page and next_page.
1449 	 * Make sure that we have head_bit value preserved for the
1450 	 * next page
1451 	 */
1452 	tail_page->next = (struct list_head *)((unsigned long)next_page |
1453 						head_bit);
1454 	next_page = rb_list_head(next_page);
1455 	next_page->prev = tail_page;
1456 
1457 	/* make sure pages points to a valid page in the ring buffer */
1458 	cpu_buffer->pages = next_page;
1459 
1460 	/* update head page */
1461 	if (head_bit)
1462 		cpu_buffer->head_page = list_entry(next_page,
1463 						struct buffer_page, list);
1464 
1465 	/*
1466 	 * change read pointer to make sure any read iterators reset
1467 	 * themselves
1468 	 */
1469 	cpu_buffer->read = 0;
1470 
1471 	/* pages are removed, resume tracing and then free the pages */
1472 	atomic_dec(&cpu_buffer->record_disabled);
1473 	raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1474 
1475 	RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
1476 
1477 	/* last buffer page to remove */
1478 	last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
1479 				list);
1480 	tmp_iter_page = first_page;
1481 
1482 	do {
1483 		to_remove_page = tmp_iter_page;
1484 		rb_inc_page(cpu_buffer, &tmp_iter_page);
1485 
1486 		/* update the counters */
1487 		page_entries = rb_page_entries(to_remove_page);
1488 		if (page_entries) {
1489 			/*
1490 			 * If something was added to this page, it was full
1491 			 * since it is not the tail page. So we deduct the
1492 			 * bytes consumed in ring buffer from here.
1493 			 * Increment overrun to account for the lost events.
1494 			 */
1495 			local_add(page_entries, &cpu_buffer->overrun);
1496 			local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1497 		}
1498 
1499 		/*
1500 		 * We have already removed references to this list item, just
1501 		 * free up the buffer_page and its page
1502 		 */
1503 		free_buffer_page(to_remove_page);
1504 		nr_removed--;
1505 
1506 	} while (to_remove_page != last_page);
1507 
1508 	RB_WARN_ON(cpu_buffer, nr_removed);
1509 
1510 	return nr_removed == 0;
1511 }
1512 
1513 static int
1514 rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
1515 {
1516 	struct list_head *pages = &cpu_buffer->new_pages;
1517 	int retries, success;
1518 
1519 	raw_spin_lock_irq(&cpu_buffer->reader_lock);
1520 	/*
1521 	 * We are holding the reader lock, so the reader page won't be swapped
1522 	 * in the ring buffer. Now we are racing with the writer trying to
1523 	 * move head page and the tail page.
1524 	 * We are going to adapt the reader page update process where:
1525 	 * 1. We first splice the start and end of list of new pages between
1526 	 *    the head page and its previous page.
1527 	 * 2. We cmpxchg the prev_page->next to point from head page to the
1528 	 *    start of new pages list.
1529 	 * 3. Finally, we update the head->prev to the end of new list.
1530 	 *
1531 	 * We will try this process 10 times, to make sure that we don't keep
1532 	 * spinning.
1533 	 */
1534 	retries = 10;
1535 	success = 0;
1536 	while (retries--) {
1537 		struct list_head *head_page, *prev_page, *r;
1538 		struct list_head *last_page, *first_page;
1539 		struct list_head *head_page_with_bit;
1540 
1541 		head_page = &rb_set_head_page(cpu_buffer)->list;
1542 		if (!head_page)
1543 			break;
1544 		prev_page = head_page->prev;
1545 
1546 		first_page = pages->next;
1547 		last_page  = pages->prev;
1548 
1549 		head_page_with_bit = (struct list_head *)
1550 				     ((unsigned long)head_page | RB_PAGE_HEAD);
1551 
1552 		last_page->next = head_page_with_bit;
1553 		first_page->prev = prev_page;
1554 
1555 		r = cmpxchg(&prev_page->next, head_page_with_bit, first_page);
1556 
1557 		if (r == head_page_with_bit) {
1558 			/*
1559 			 * yay, we replaced the page pointer to our new list,
1560 			 * now, we just have to update to head page's prev
1561 			 * pointer to point to end of list
1562 			 */
1563 			head_page->prev = last_page;
1564 			success = 1;
1565 			break;
1566 		}
1567 	}
1568 
1569 	if (success)
1570 		INIT_LIST_HEAD(pages);
1571 	/*
1572 	 * If we weren't successful in adding in new pages, warn and stop
1573 	 * tracing
1574 	 */
1575 	RB_WARN_ON(cpu_buffer, !success);
1576 	raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1577 
1578 	/* free pages if they weren't inserted */
1579 	if (!success) {
1580 		struct buffer_page *bpage, *tmp;
1581 		list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1582 					 list) {
1583 			list_del_init(&bpage->list);
1584 			free_buffer_page(bpage);
1585 		}
1586 	}
1587 	return success;
1588 }
1589 
1590 static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
1591 {
1592 	int success;
1593 
1594 	if (cpu_buffer->nr_pages_to_update > 0)
1595 		success = rb_insert_pages(cpu_buffer);
1596 	else
1597 		success = rb_remove_pages(cpu_buffer,
1598 					-cpu_buffer->nr_pages_to_update);
1599 
1600 	if (success)
1601 		cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
1602 }
1603 
1604 static void update_pages_handler(struct work_struct *work)
1605 {
1606 	struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
1607 			struct ring_buffer_per_cpu, update_pages_work);
1608 	rb_update_pages(cpu_buffer);
1609 	complete(&cpu_buffer->update_done);
1610 }
1611 
1612 /**
1613  * ring_buffer_resize - resize the ring buffer
1614  * @buffer: the buffer to resize.
1615  * @size: the new size.
1616  * @cpu_id: the cpu buffer to resize
1617  *
1618  * Minimum size is 2 * BUF_PAGE_SIZE.
1619  *
1620  * Returns 0 on success and < 0 on failure.
1621  */
1622 int ring_buffer_resize(struct ring_buffer *buffer, unsigned long size,
1623 			int cpu_id)
1624 {
1625 	struct ring_buffer_per_cpu *cpu_buffer;
1626 	unsigned nr_pages;
1627 	int cpu, err = 0;
1628 
1629 	/*
1630 	 * Always succeed at resizing a non-existent buffer:
1631 	 */
1632 	if (!buffer)
1633 		return size;
1634 
1635 	/* Make sure the requested buffer exists */
1636 	if (cpu_id != RING_BUFFER_ALL_CPUS &&
1637 	    !cpumask_test_cpu(cpu_id, buffer->cpumask))
1638 		return size;
1639 
1640 	size = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1641 	size *= BUF_PAGE_SIZE;
1642 
1643 	/* we need a minimum of two pages */
1644 	if (size < BUF_PAGE_SIZE * 2)
1645 		size = BUF_PAGE_SIZE * 2;
1646 
1647 	nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1648 
1649 	/*
1650 	 * Don't succeed if resizing is disabled, as a reader might be
1651 	 * manipulating the ring buffer and is expecting a sane state while
1652 	 * this is true.
1653 	 */
1654 	if (atomic_read(&buffer->resize_disabled))
1655 		return -EBUSY;
1656 
1657 	/* prevent another thread from changing buffer sizes */
1658 	mutex_lock(&buffer->mutex);
1659 
1660 	if (cpu_id == RING_BUFFER_ALL_CPUS) {
1661 		/* calculate the pages to update */
1662 		for_each_buffer_cpu(buffer, cpu) {
1663 			cpu_buffer = buffer->buffers[cpu];
1664 
1665 			cpu_buffer->nr_pages_to_update = nr_pages -
1666 							cpu_buffer->nr_pages;
1667 			/*
1668 			 * nothing more to do for removing pages or no update
1669 			 */
1670 			if (cpu_buffer->nr_pages_to_update <= 0)
1671 				continue;
1672 			/*
1673 			 * to add pages, make sure all new pages can be
1674 			 * allocated without receiving ENOMEM
1675 			 */
1676 			INIT_LIST_HEAD(&cpu_buffer->new_pages);
1677 			if (__rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1678 						&cpu_buffer->new_pages, cpu)) {
1679 				/* not enough memory for new pages */
1680 				err = -ENOMEM;
1681 				goto out_err;
1682 			}
1683 		}
1684 
1685 		get_online_cpus();
1686 		/*
1687 		 * Fire off all the required work handlers
1688 		 * We can't schedule on offline CPUs, but it's not necessary
1689 		 * since we can change their buffer sizes without any race.
1690 		 */
1691 		for_each_buffer_cpu(buffer, cpu) {
1692 			cpu_buffer = buffer->buffers[cpu];
1693 			if (!cpu_buffer->nr_pages_to_update)
1694 				continue;
1695 
1696 			/* The update must run on the CPU that is being updated. */
1697 			preempt_disable();
1698 			if (cpu == smp_processor_id() || !cpu_online(cpu)) {
1699 				rb_update_pages(cpu_buffer);
1700 				cpu_buffer->nr_pages_to_update = 0;
1701 			} else {
1702 				/*
1703 				 * Can not disable preemption for schedule_work_on()
1704 				 * on PREEMPT_RT.
1705 				 */
1706 				preempt_enable();
1707 				schedule_work_on(cpu,
1708 						&cpu_buffer->update_pages_work);
1709 				preempt_disable();
1710 			}
1711 			preempt_enable();
1712 		}
1713 
1714 		/* wait for all the updates to complete */
1715 		for_each_buffer_cpu(buffer, cpu) {
1716 			cpu_buffer = buffer->buffers[cpu];
1717 			if (!cpu_buffer->nr_pages_to_update)
1718 				continue;
1719 
1720 			if (cpu_online(cpu))
1721 				wait_for_completion(&cpu_buffer->update_done);
1722 			cpu_buffer->nr_pages_to_update = 0;
1723 		}
1724 
1725 		put_online_cpus();
1726 	} else {
1727 		/* Make sure this CPU has been intitialized */
1728 		if (!cpumask_test_cpu(cpu_id, buffer->cpumask))
1729 			goto out;
1730 
1731 		cpu_buffer = buffer->buffers[cpu_id];
1732 
1733 		if (nr_pages == cpu_buffer->nr_pages)
1734 			goto out;
1735 
1736 		cpu_buffer->nr_pages_to_update = nr_pages -
1737 						cpu_buffer->nr_pages;
1738 
1739 		INIT_LIST_HEAD(&cpu_buffer->new_pages);
1740 		if (cpu_buffer->nr_pages_to_update > 0 &&
1741 			__rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1742 					    &cpu_buffer->new_pages, cpu_id)) {
1743 			err = -ENOMEM;
1744 			goto out_err;
1745 		}
1746 
1747 		get_online_cpus();
1748 
1749 		preempt_disable();
1750 		/* The update must run on the CPU that is being updated. */
1751 		if (cpu_id == smp_processor_id() || !cpu_online(cpu_id))
1752 			rb_update_pages(cpu_buffer);
1753 		else {
1754 			/*
1755 			 * Can not disable preemption for schedule_work_on()
1756 			 * on PREEMPT_RT.
1757 			 */
1758 			preempt_enable();
1759 			schedule_work_on(cpu_id,
1760 					 &cpu_buffer->update_pages_work);
1761 			wait_for_completion(&cpu_buffer->update_done);
1762 			preempt_disable();
1763 		}
1764 		preempt_enable();
1765 
1766 		cpu_buffer->nr_pages_to_update = 0;
1767 		put_online_cpus();
1768 	}
1769 
1770  out:
1771 	/*
1772 	 * The ring buffer resize can happen with the ring buffer
1773 	 * enabled, so that the update disturbs the tracing as little
1774 	 * as possible. But if the buffer is disabled, we do not need
1775 	 * to worry about that, and we can take the time to verify
1776 	 * that the buffer is not corrupt.
1777 	 */
1778 	if (atomic_read(&buffer->record_disabled)) {
1779 		atomic_inc(&buffer->record_disabled);
1780 		/*
1781 		 * Even though the buffer was disabled, we must make sure
1782 		 * that it is truly disabled before calling rb_check_pages.
1783 		 * There could have been a race between checking
1784 		 * record_disable and incrementing it.
1785 		 */
1786 		synchronize_sched();
1787 		for_each_buffer_cpu(buffer, cpu) {
1788 			cpu_buffer = buffer->buffers[cpu];
1789 			rb_check_pages(cpu_buffer);
1790 		}
1791 		atomic_dec(&buffer->record_disabled);
1792 	}
1793 
1794 	mutex_unlock(&buffer->mutex);
1795 	return size;
1796 
1797  out_err:
1798 	for_each_buffer_cpu(buffer, cpu) {
1799 		struct buffer_page *bpage, *tmp;
1800 
1801 		cpu_buffer = buffer->buffers[cpu];
1802 		cpu_buffer->nr_pages_to_update = 0;
1803 
1804 		if (list_empty(&cpu_buffer->new_pages))
1805 			continue;
1806 
1807 		list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1808 					list) {
1809 			list_del_init(&bpage->list);
1810 			free_buffer_page(bpage);
1811 		}
1812 	}
1813 	mutex_unlock(&buffer->mutex);
1814 	return err;
1815 }
1816 EXPORT_SYMBOL_GPL(ring_buffer_resize);
1817 
1818 void ring_buffer_change_overwrite(struct ring_buffer *buffer, int val)
1819 {
1820 	mutex_lock(&buffer->mutex);
1821 	if (val)
1822 		buffer->flags |= RB_FL_OVERWRITE;
1823 	else
1824 		buffer->flags &= ~RB_FL_OVERWRITE;
1825 	mutex_unlock(&buffer->mutex);
1826 }
1827 EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
1828 
1829 static inline void *
1830 __rb_data_page_index(struct buffer_data_page *bpage, unsigned index)
1831 {
1832 	return bpage->data + index;
1833 }
1834 
1835 static inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
1836 {
1837 	return bpage->page->data + index;
1838 }
1839 
1840 static inline struct ring_buffer_event *
1841 rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
1842 {
1843 	return __rb_page_index(cpu_buffer->reader_page,
1844 			       cpu_buffer->reader_page->read);
1845 }
1846 
1847 static inline struct ring_buffer_event *
1848 rb_iter_head_event(struct ring_buffer_iter *iter)
1849 {
1850 	return __rb_page_index(iter->head_page, iter->head);
1851 }
1852 
1853 static inline unsigned rb_page_commit(struct buffer_page *bpage)
1854 {
1855 	return local_read(&bpage->page->commit);
1856 }
1857 
1858 /* Size is determined by what has been committed */
1859 static inline unsigned rb_page_size(struct buffer_page *bpage)
1860 {
1861 	return rb_page_commit(bpage);
1862 }
1863 
1864 static inline unsigned
1865 rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
1866 {
1867 	return rb_page_commit(cpu_buffer->commit_page);
1868 }
1869 
1870 static inline unsigned
1871 rb_event_index(struct ring_buffer_event *event)
1872 {
1873 	unsigned long addr = (unsigned long)event;
1874 
1875 	return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
1876 }
1877 
1878 static inline int
1879 rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
1880 		   struct ring_buffer_event *event)
1881 {
1882 	unsigned long addr = (unsigned long)event;
1883 	unsigned long index;
1884 
1885 	index = rb_event_index(event);
1886 	addr &= PAGE_MASK;
1887 
1888 	return cpu_buffer->commit_page->page == (void *)addr &&
1889 		rb_commit_index(cpu_buffer) == index;
1890 }
1891 
1892 static void
1893 rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
1894 {
1895 	unsigned long max_count;
1896 
1897 	/*
1898 	 * We only race with interrupts and NMIs on this CPU.
1899 	 * If we own the commit event, then we can commit
1900 	 * all others that interrupted us, since the interruptions
1901 	 * are in stack format (they finish before they come
1902 	 * back to us). This allows us to do a simple loop to
1903 	 * assign the commit to the tail.
1904 	 */
1905  again:
1906 	max_count = cpu_buffer->nr_pages * 100;
1907 
1908 	while (cpu_buffer->commit_page != cpu_buffer->tail_page) {
1909 		if (RB_WARN_ON(cpu_buffer, !(--max_count)))
1910 			return;
1911 		if (RB_WARN_ON(cpu_buffer,
1912 			       rb_is_reader_page(cpu_buffer->tail_page)))
1913 			return;
1914 		local_set(&cpu_buffer->commit_page->page->commit,
1915 			  rb_page_write(cpu_buffer->commit_page));
1916 		rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
1917 		cpu_buffer->write_stamp =
1918 			cpu_buffer->commit_page->page->time_stamp;
1919 		/* add barrier to keep gcc from optimizing too much */
1920 		barrier();
1921 	}
1922 	while (rb_commit_index(cpu_buffer) !=
1923 	       rb_page_write(cpu_buffer->commit_page)) {
1924 
1925 		local_set(&cpu_buffer->commit_page->page->commit,
1926 			  rb_page_write(cpu_buffer->commit_page));
1927 		RB_WARN_ON(cpu_buffer,
1928 			   local_read(&cpu_buffer->commit_page->page->commit) &
1929 			   ~RB_WRITE_MASK);
1930 		barrier();
1931 	}
1932 
1933 	/* again, keep gcc from optimizing */
1934 	barrier();
1935 
1936 	/*
1937 	 * If an interrupt came in just after the first while loop
1938 	 * and pushed the tail page forward, we will be left with
1939 	 * a dangling commit that will never go forward.
1940 	 */
1941 	if (unlikely(cpu_buffer->commit_page != cpu_buffer->tail_page))
1942 		goto again;
1943 }
1944 
1945 static void rb_reset_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
1946 {
1947 	cpu_buffer->read_stamp = cpu_buffer->reader_page->page->time_stamp;
1948 	cpu_buffer->reader_page->read = 0;
1949 }
1950 
1951 static void rb_inc_iter(struct ring_buffer_iter *iter)
1952 {
1953 	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
1954 
1955 	/*
1956 	 * The iterator could be on the reader page (it starts there).
1957 	 * But the head could have moved, since the reader was
1958 	 * found. Check for this case and assign the iterator
1959 	 * to the head page instead of next.
1960 	 */
1961 	if (iter->head_page == cpu_buffer->reader_page)
1962 		iter->head_page = rb_set_head_page(cpu_buffer);
1963 	else
1964 		rb_inc_page(cpu_buffer, &iter->head_page);
1965 
1966 	iter->read_stamp = iter->head_page->page->time_stamp;
1967 	iter->head = 0;
1968 }
1969 
1970 /* Slow path, do not inline */
1971 static noinline struct ring_buffer_event *
1972 rb_add_time_stamp(struct ring_buffer_event *event, u64 delta)
1973 {
1974 	event->type_len = RINGBUF_TYPE_TIME_EXTEND;
1975 
1976 	/* Not the first event on the page? */
1977 	if (rb_event_index(event)) {
1978 		event->time_delta = delta & TS_MASK;
1979 		event->array[0] = delta >> TS_SHIFT;
1980 	} else {
1981 		/* nope, just zero it */
1982 		event->time_delta = 0;
1983 		event->array[0] = 0;
1984 	}
1985 
1986 	return skip_time_extend(event);
1987 }
1988 
1989 /**
1990  * rb_update_event - update event type and data
1991  * @event: the even to update
1992  * @type: the type of event
1993  * @length: the size of the event field in the ring buffer
1994  *
1995  * Update the type and data fields of the event. The length
1996  * is the actual size that is written to the ring buffer,
1997  * and with this, we can determine what to place into the
1998  * data field.
1999  */
2000 static void
2001 rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
2002 		struct ring_buffer_event *event, unsigned length,
2003 		int add_timestamp, u64 delta)
2004 {
2005 	/* Only a commit updates the timestamp */
2006 	if (unlikely(!rb_event_is_commit(cpu_buffer, event)))
2007 		delta = 0;
2008 
2009 	/*
2010 	 * If we need to add a timestamp, then we
2011 	 * add it to the start of the resevered space.
2012 	 */
2013 	if (unlikely(add_timestamp)) {
2014 		event = rb_add_time_stamp(event, delta);
2015 		length -= RB_LEN_TIME_EXTEND;
2016 		delta = 0;
2017 	}
2018 
2019 	event->time_delta = delta;
2020 	length -= RB_EVNT_HDR_SIZE;
2021 	if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
2022 		event->type_len = 0;
2023 		event->array[0] = length;
2024 	} else
2025 		event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
2026 }
2027 
2028 /*
2029  * rb_handle_head_page - writer hit the head page
2030  *
2031  * Returns: +1 to retry page
2032  *           0 to continue
2033  *          -1 on error
2034  */
2035 static int
2036 rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
2037 		    struct buffer_page *tail_page,
2038 		    struct buffer_page *next_page)
2039 {
2040 	struct buffer_page *new_head;
2041 	int entries;
2042 	int type;
2043 	int ret;
2044 
2045 	entries = rb_page_entries(next_page);
2046 
2047 	/*
2048 	 * The hard part is here. We need to move the head
2049 	 * forward, and protect against both readers on
2050 	 * other CPUs and writers coming in via interrupts.
2051 	 */
2052 	type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
2053 				       RB_PAGE_HEAD);
2054 
2055 	/*
2056 	 * type can be one of four:
2057 	 *  NORMAL - an interrupt already moved it for us
2058 	 *  HEAD   - we are the first to get here.
2059 	 *  UPDATE - we are the interrupt interrupting
2060 	 *           a current move.
2061 	 *  MOVED  - a reader on another CPU moved the next
2062 	 *           pointer to its reader page. Give up
2063 	 *           and try again.
2064 	 */
2065 
2066 	switch (type) {
2067 	case RB_PAGE_HEAD:
2068 		/*
2069 		 * We changed the head to UPDATE, thus
2070 		 * it is our responsibility to update
2071 		 * the counters.
2072 		 */
2073 		local_add(entries, &cpu_buffer->overrun);
2074 		local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
2075 
2076 		/*
2077 		 * The entries will be zeroed out when we move the
2078 		 * tail page.
2079 		 */
2080 
2081 		/* still more to do */
2082 		break;
2083 
2084 	case RB_PAGE_UPDATE:
2085 		/*
2086 		 * This is an interrupt that interrupt the
2087 		 * previous update. Still more to do.
2088 		 */
2089 		break;
2090 	case RB_PAGE_NORMAL:
2091 		/*
2092 		 * An interrupt came in before the update
2093 		 * and processed this for us.
2094 		 * Nothing left to do.
2095 		 */
2096 		return 1;
2097 	case RB_PAGE_MOVED:
2098 		/*
2099 		 * The reader is on another CPU and just did
2100 		 * a swap with our next_page.
2101 		 * Try again.
2102 		 */
2103 		return 1;
2104 	default:
2105 		RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
2106 		return -1;
2107 	}
2108 
2109 	/*
2110 	 * Now that we are here, the old head pointer is
2111 	 * set to UPDATE. This will keep the reader from
2112 	 * swapping the head page with the reader page.
2113 	 * The reader (on another CPU) will spin till
2114 	 * we are finished.
2115 	 *
2116 	 * We just need to protect against interrupts
2117 	 * doing the job. We will set the next pointer
2118 	 * to HEAD. After that, we set the old pointer
2119 	 * to NORMAL, but only if it was HEAD before.
2120 	 * otherwise we are an interrupt, and only
2121 	 * want the outer most commit to reset it.
2122 	 */
2123 	new_head = next_page;
2124 	rb_inc_page(cpu_buffer, &new_head);
2125 
2126 	ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
2127 				    RB_PAGE_NORMAL);
2128 
2129 	/*
2130 	 * Valid returns are:
2131 	 *  HEAD   - an interrupt came in and already set it.
2132 	 *  NORMAL - One of two things:
2133 	 *            1) We really set it.
2134 	 *            2) A bunch of interrupts came in and moved
2135 	 *               the page forward again.
2136 	 */
2137 	switch (ret) {
2138 	case RB_PAGE_HEAD:
2139 	case RB_PAGE_NORMAL:
2140 		/* OK */
2141 		break;
2142 	default:
2143 		RB_WARN_ON(cpu_buffer, 1);
2144 		return -1;
2145 	}
2146 
2147 	/*
2148 	 * It is possible that an interrupt came in,
2149 	 * set the head up, then more interrupts came in
2150 	 * and moved it again. When we get back here,
2151 	 * the page would have been set to NORMAL but we
2152 	 * just set it back to HEAD.
2153 	 *
2154 	 * How do you detect this? Well, if that happened
2155 	 * the tail page would have moved.
2156 	 */
2157 	if (ret == RB_PAGE_NORMAL) {
2158 		/*
2159 		 * If the tail had moved passed next, then we need
2160 		 * to reset the pointer.
2161 		 */
2162 		if (cpu_buffer->tail_page != tail_page &&
2163 		    cpu_buffer->tail_page != next_page)
2164 			rb_head_page_set_normal(cpu_buffer, new_head,
2165 						next_page,
2166 						RB_PAGE_HEAD);
2167 	}
2168 
2169 	/*
2170 	 * If this was the outer most commit (the one that
2171 	 * changed the original pointer from HEAD to UPDATE),
2172 	 * then it is up to us to reset it to NORMAL.
2173 	 */
2174 	if (type == RB_PAGE_HEAD) {
2175 		ret = rb_head_page_set_normal(cpu_buffer, next_page,
2176 					      tail_page,
2177 					      RB_PAGE_UPDATE);
2178 		if (RB_WARN_ON(cpu_buffer,
2179 			       ret != RB_PAGE_UPDATE))
2180 			return -1;
2181 	}
2182 
2183 	return 0;
2184 }
2185 
2186 static unsigned rb_calculate_event_length(unsigned length)
2187 {
2188 	struct ring_buffer_event event; /* Used only for sizeof array */
2189 
2190 	/* zero length can cause confusions */
2191 	if (!length)
2192 		length = 1;
2193 
2194 	if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
2195 		length += sizeof(event.array[0]);
2196 
2197 	length += RB_EVNT_HDR_SIZE;
2198 	length = ALIGN(length, RB_ARCH_ALIGNMENT);
2199 
2200 	return length;
2201 }
2202 
2203 static inline void
2204 rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2205 	      struct buffer_page *tail_page,
2206 	      unsigned long tail, unsigned long length)
2207 {
2208 	struct ring_buffer_event *event;
2209 
2210 	/*
2211 	 * Only the event that crossed the page boundary
2212 	 * must fill the old tail_page with padding.
2213 	 */
2214 	if (tail >= BUF_PAGE_SIZE) {
2215 		/*
2216 		 * If the page was filled, then we still need
2217 		 * to update the real_end. Reset it to zero
2218 		 * and the reader will ignore it.
2219 		 */
2220 		if (tail == BUF_PAGE_SIZE)
2221 			tail_page->real_end = 0;
2222 
2223 		local_sub(length, &tail_page->write);
2224 		return;
2225 	}
2226 
2227 	event = __rb_page_index(tail_page, tail);
2228 	kmemcheck_annotate_bitfield(event, bitfield);
2229 
2230 	/* account for padding bytes */
2231 	local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes);
2232 
2233 	/*
2234 	 * Save the original length to the meta data.
2235 	 * This will be used by the reader to add lost event
2236 	 * counter.
2237 	 */
2238 	tail_page->real_end = tail;
2239 
2240 	/*
2241 	 * If this event is bigger than the minimum size, then
2242 	 * we need to be careful that we don't subtract the
2243 	 * write counter enough to allow another writer to slip
2244 	 * in on this page.
2245 	 * We put in a discarded commit instead, to make sure
2246 	 * that this space is not used again.
2247 	 *
2248 	 * If we are less than the minimum size, we don't need to
2249 	 * worry about it.
2250 	 */
2251 	if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
2252 		/* No room for any events */
2253 
2254 		/* Mark the rest of the page with padding */
2255 		rb_event_set_padding(event);
2256 
2257 		/* Set the write back to the previous setting */
2258 		local_sub(length, &tail_page->write);
2259 		return;
2260 	}
2261 
2262 	/* Put in a discarded event */
2263 	event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
2264 	event->type_len = RINGBUF_TYPE_PADDING;
2265 	/* time delta must be non zero */
2266 	event->time_delta = 1;
2267 
2268 	/* Set write to end of buffer */
2269 	length = (tail + length) - BUF_PAGE_SIZE;
2270 	local_sub(length, &tail_page->write);
2271 }
2272 
2273 /*
2274  * This is the slow path, force gcc not to inline it.
2275  */
2276 static noinline struct ring_buffer_event *
2277 rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2278 	     unsigned long length, unsigned long tail,
2279 	     struct buffer_page *tail_page, u64 ts)
2280 {
2281 	struct buffer_page *commit_page = cpu_buffer->commit_page;
2282 	struct ring_buffer *buffer = cpu_buffer->buffer;
2283 	struct buffer_page *next_page;
2284 	int ret;
2285 
2286 	next_page = tail_page;
2287 
2288 	rb_inc_page(cpu_buffer, &next_page);
2289 
2290 	/*
2291 	 * If for some reason, we had an interrupt storm that made
2292 	 * it all the way around the buffer, bail, and warn
2293 	 * about it.
2294 	 */
2295 	if (unlikely(next_page == commit_page)) {
2296 		local_inc(&cpu_buffer->commit_overrun);
2297 		goto out_reset;
2298 	}
2299 
2300 	/*
2301 	 * This is where the fun begins!
2302 	 *
2303 	 * We are fighting against races between a reader that
2304 	 * could be on another CPU trying to swap its reader
2305 	 * page with the buffer head.
2306 	 *
2307 	 * We are also fighting against interrupts coming in and
2308 	 * moving the head or tail on us as well.
2309 	 *
2310 	 * If the next page is the head page then we have filled
2311 	 * the buffer, unless the commit page is still on the
2312 	 * reader page.
2313 	 */
2314 	if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) {
2315 
2316 		/*
2317 		 * If the commit is not on the reader page, then
2318 		 * move the header page.
2319 		 */
2320 		if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2321 			/*
2322 			 * If we are not in overwrite mode,
2323 			 * this is easy, just stop here.
2324 			 */
2325 			if (!(buffer->flags & RB_FL_OVERWRITE)) {
2326 				local_inc(&cpu_buffer->dropped_events);
2327 				goto out_reset;
2328 			}
2329 
2330 			ret = rb_handle_head_page(cpu_buffer,
2331 						  tail_page,
2332 						  next_page);
2333 			if (ret < 0)
2334 				goto out_reset;
2335 			if (ret)
2336 				goto out_again;
2337 		} else {
2338 			/*
2339 			 * We need to be careful here too. The
2340 			 * commit page could still be on the reader
2341 			 * page. We could have a small buffer, and
2342 			 * have filled up the buffer with events
2343 			 * from interrupts and such, and wrapped.
2344 			 *
2345 			 * Note, if the tail page is also the on the
2346 			 * reader_page, we let it move out.
2347 			 */
2348 			if (unlikely((cpu_buffer->commit_page !=
2349 				      cpu_buffer->tail_page) &&
2350 				     (cpu_buffer->commit_page ==
2351 				      cpu_buffer->reader_page))) {
2352 				local_inc(&cpu_buffer->commit_overrun);
2353 				goto out_reset;
2354 			}
2355 		}
2356 	}
2357 
2358 	ret = rb_tail_page_update(cpu_buffer, tail_page, next_page);
2359 	if (ret) {
2360 		/*
2361 		 * Nested commits always have zero deltas, so
2362 		 * just reread the time stamp
2363 		 */
2364 		ts = rb_time_stamp(buffer);
2365 		next_page->page->time_stamp = ts;
2366 	}
2367 
2368  out_again:
2369 
2370 	rb_reset_tail(cpu_buffer, tail_page, tail, length);
2371 
2372 	/* fail and let the caller try again */
2373 	return ERR_PTR(-EAGAIN);
2374 
2375  out_reset:
2376 	/* reset write */
2377 	rb_reset_tail(cpu_buffer, tail_page, tail, length);
2378 
2379 	return NULL;
2380 }
2381 
2382 static struct ring_buffer_event *
2383 __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
2384 		  unsigned long length, u64 ts,
2385 		  u64 delta, int add_timestamp)
2386 {
2387 	struct buffer_page *tail_page;
2388 	struct ring_buffer_event *event;
2389 	unsigned long tail, write;
2390 
2391 	/*
2392 	 * If the time delta since the last event is too big to
2393 	 * hold in the time field of the event, then we append a
2394 	 * TIME EXTEND event ahead of the data event.
2395 	 */
2396 	if (unlikely(add_timestamp))
2397 		length += RB_LEN_TIME_EXTEND;
2398 
2399 	tail_page = cpu_buffer->tail_page;
2400 	write = local_add_return(length, &tail_page->write);
2401 
2402 	/* set write to only the index of the write */
2403 	write &= RB_WRITE_MASK;
2404 	tail = write - length;
2405 
2406 	/*
2407 	 * If this is the first commit on the page, then it has the same
2408 	 * timestamp as the page itself.
2409 	 */
2410 	if (!tail)
2411 		delta = 0;
2412 
2413 	/* See if we shot pass the end of this buffer page */
2414 	if (unlikely(write > BUF_PAGE_SIZE))
2415 		return rb_move_tail(cpu_buffer, length, tail,
2416 				    tail_page, ts);
2417 
2418 	/* We reserved something on the buffer */
2419 
2420 	event = __rb_page_index(tail_page, tail);
2421 	kmemcheck_annotate_bitfield(event, bitfield);
2422 	rb_update_event(cpu_buffer, event, length, add_timestamp, delta);
2423 
2424 	local_inc(&tail_page->entries);
2425 
2426 	/*
2427 	 * If this is the first commit on the page, then update
2428 	 * its timestamp.
2429 	 */
2430 	if (!tail)
2431 		tail_page->page->time_stamp = ts;
2432 
2433 	/* account for these added bytes */
2434 	local_add(length, &cpu_buffer->entries_bytes);
2435 
2436 	return event;
2437 }
2438 
2439 static inline int
2440 rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2441 		  struct ring_buffer_event *event)
2442 {
2443 	unsigned long new_index, old_index;
2444 	struct buffer_page *bpage;
2445 	unsigned long index;
2446 	unsigned long addr;
2447 
2448 	new_index = rb_event_index(event);
2449 	old_index = new_index + rb_event_ts_length(event);
2450 	addr = (unsigned long)event;
2451 	addr &= PAGE_MASK;
2452 
2453 	bpage = cpu_buffer->tail_page;
2454 
2455 	if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2456 		unsigned long write_mask =
2457 			local_read(&bpage->write) & ~RB_WRITE_MASK;
2458 		unsigned long event_length = rb_event_length(event);
2459 		/*
2460 		 * This is on the tail page. It is possible that
2461 		 * a write could come in and move the tail page
2462 		 * and write to the next page. That is fine
2463 		 * because we just shorten what is on this page.
2464 		 */
2465 		old_index += write_mask;
2466 		new_index += write_mask;
2467 		index = local_cmpxchg(&bpage->write, old_index, new_index);
2468 		if (index == old_index) {
2469 			/* update counters */
2470 			local_sub(event_length, &cpu_buffer->entries_bytes);
2471 			return 1;
2472 		}
2473 	}
2474 
2475 	/* could not discard */
2476 	return 0;
2477 }
2478 
2479 static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2480 {
2481 	local_inc(&cpu_buffer->committing);
2482 	local_inc(&cpu_buffer->commits);
2483 }
2484 
2485 static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
2486 {
2487 	unsigned long commits;
2488 
2489 	if (RB_WARN_ON(cpu_buffer,
2490 		       !local_read(&cpu_buffer->committing)))
2491 		return;
2492 
2493  again:
2494 	commits = local_read(&cpu_buffer->commits);
2495 	/* synchronize with interrupts */
2496 	barrier();
2497 	if (local_read(&cpu_buffer->committing) == 1)
2498 		rb_set_commit_to_write(cpu_buffer);
2499 
2500 	local_dec(&cpu_buffer->committing);
2501 
2502 	/* synchronize with interrupts */
2503 	barrier();
2504 
2505 	/*
2506 	 * Need to account for interrupts coming in between the
2507 	 * updating of the commit page and the clearing of the
2508 	 * committing counter.
2509 	 */
2510 	if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
2511 	    !local_read(&cpu_buffer->committing)) {
2512 		local_inc(&cpu_buffer->committing);
2513 		goto again;
2514 	}
2515 }
2516 
2517 static struct ring_buffer_event *
2518 rb_reserve_next_event(struct ring_buffer *buffer,
2519 		      struct ring_buffer_per_cpu *cpu_buffer,
2520 		      unsigned long length)
2521 {
2522 	struct ring_buffer_event *event;
2523 	u64 ts, delta;
2524 	int nr_loops = 0;
2525 	int add_timestamp;
2526 	u64 diff;
2527 
2528 	rb_start_commit(cpu_buffer);
2529 
2530 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2531 	/*
2532 	 * Due to the ability to swap a cpu buffer from a buffer
2533 	 * it is possible it was swapped before we committed.
2534 	 * (committing stops a swap). We check for it here and
2535 	 * if it happened, we have to fail the write.
2536 	 */
2537 	barrier();
2538 	if (unlikely(ACCESS_ONCE(cpu_buffer->buffer) != buffer)) {
2539 		local_dec(&cpu_buffer->committing);
2540 		local_dec(&cpu_buffer->commits);
2541 		return NULL;
2542 	}
2543 #endif
2544 
2545 	length = rb_calculate_event_length(length);
2546  again:
2547 	add_timestamp = 0;
2548 	delta = 0;
2549 
2550 	/*
2551 	 * We allow for interrupts to reenter here and do a trace.
2552 	 * If one does, it will cause this original code to loop
2553 	 * back here. Even with heavy interrupts happening, this
2554 	 * should only happen a few times in a row. If this happens
2555 	 * 1000 times in a row, there must be either an interrupt
2556 	 * storm or we have something buggy.
2557 	 * Bail!
2558 	 */
2559 	if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
2560 		goto out_fail;
2561 
2562 	ts = rb_time_stamp(cpu_buffer->buffer);
2563 	diff = ts - cpu_buffer->write_stamp;
2564 
2565 	/* make sure this diff is calculated here */
2566 	barrier();
2567 
2568 	/* Did the write stamp get updated already? */
2569 	if (likely(ts >= cpu_buffer->write_stamp)) {
2570 		delta = diff;
2571 		if (unlikely(test_time_stamp(delta))) {
2572 			int local_clock_stable = 1;
2573 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2574 			local_clock_stable = sched_clock_stable();
2575 #endif
2576 			WARN_ONCE(delta > (1ULL << 59),
2577 				  KERN_WARNING "Delta way too big! %llu ts=%llu write stamp = %llu\n%s",
2578 				  (unsigned long long)delta,
2579 				  (unsigned long long)ts,
2580 				  (unsigned long long)cpu_buffer->write_stamp,
2581 				  local_clock_stable ? "" :
2582 				  "If you just came from a suspend/resume,\n"
2583 				  "please switch to the trace global clock:\n"
2584 				  "  echo global > /sys/kernel/debug/tracing/trace_clock\n");
2585 			add_timestamp = 1;
2586 		}
2587 	}
2588 
2589 	event = __rb_reserve_next(cpu_buffer, length, ts,
2590 				  delta, add_timestamp);
2591 	if (unlikely(PTR_ERR(event) == -EAGAIN))
2592 		goto again;
2593 
2594 	if (!event)
2595 		goto out_fail;
2596 
2597 	return event;
2598 
2599  out_fail:
2600 	rb_end_commit(cpu_buffer);
2601 	return NULL;
2602 }
2603 
2604 #ifdef CONFIG_TRACING
2605 
2606 /*
2607  * The lock and unlock are done within a preempt disable section.
2608  * The current_context per_cpu variable can only be modified
2609  * by the current task between lock and unlock. But it can
2610  * be modified more than once via an interrupt. To pass this
2611  * information from the lock to the unlock without having to
2612  * access the 'in_interrupt()' functions again (which do show
2613  * a bit of overhead in something as critical as function tracing,
2614  * we use a bitmask trick.
2615  *
2616  *  bit 0 =  NMI context
2617  *  bit 1 =  IRQ context
2618  *  bit 2 =  SoftIRQ context
2619  *  bit 3 =  normal context.
2620  *
2621  * This works because this is the order of contexts that can
2622  * preempt other contexts. A SoftIRQ never preempts an IRQ
2623  * context.
2624  *
2625  * When the context is determined, the corresponding bit is
2626  * checked and set (if it was set, then a recursion of that context
2627  * happened).
2628  *
2629  * On unlock, we need to clear this bit. To do so, just subtract
2630  * 1 from the current_context and AND it to itself.
2631  *
2632  * (binary)
2633  *  101 - 1 = 100
2634  *  101 & 100 = 100 (clearing bit zero)
2635  *
2636  *  1010 - 1 = 1001
2637  *  1010 & 1001 = 1000 (clearing bit 1)
2638  *
2639  * The least significant bit can be cleared this way, and it
2640  * just so happens that it is the same bit corresponding to
2641  * the current context.
2642  */
2643 static DEFINE_PER_CPU(unsigned int, current_context);
2644 
2645 static __always_inline int trace_recursive_lock(void)
2646 {
2647 	unsigned int val = this_cpu_read(current_context);
2648 	int bit;
2649 
2650 	if (in_interrupt()) {
2651 		if (in_nmi())
2652 			bit = 0;
2653 		else if (in_irq())
2654 			bit = 1;
2655 		else
2656 			bit = 2;
2657 	} else
2658 		bit = 3;
2659 
2660 	if (unlikely(val & (1 << bit)))
2661 		return 1;
2662 
2663 	val |= (1 << bit);
2664 	this_cpu_write(current_context, val);
2665 
2666 	return 0;
2667 }
2668 
2669 static __always_inline void trace_recursive_unlock(void)
2670 {
2671 	unsigned int val = this_cpu_read(current_context);
2672 
2673 	val--;
2674 	val &= this_cpu_read(current_context);
2675 	this_cpu_write(current_context, val);
2676 }
2677 
2678 #else
2679 
2680 #define trace_recursive_lock()		(0)
2681 #define trace_recursive_unlock()	do { } while (0)
2682 
2683 #endif
2684 
2685 /**
2686  * ring_buffer_lock_reserve - reserve a part of the buffer
2687  * @buffer: the ring buffer to reserve from
2688  * @length: the length of the data to reserve (excluding event header)
2689  *
2690  * Returns a reseverd event on the ring buffer to copy directly to.
2691  * The user of this interface will need to get the body to write into
2692  * and can use the ring_buffer_event_data() interface.
2693  *
2694  * The length is the length of the data needed, not the event length
2695  * which also includes the event header.
2696  *
2697  * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2698  * If NULL is returned, then nothing has been allocated or locked.
2699  */
2700 struct ring_buffer_event *
2701 ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length)
2702 {
2703 	struct ring_buffer_per_cpu *cpu_buffer;
2704 	struct ring_buffer_event *event;
2705 	int cpu;
2706 
2707 	if (ring_buffer_flags != RB_BUFFERS_ON)
2708 		return NULL;
2709 
2710 	/* If we are tracing schedule, we don't want to recurse */
2711 	preempt_disable_notrace();
2712 
2713 	if (atomic_read(&buffer->record_disabled))
2714 		goto out_nocheck;
2715 
2716 	if (trace_recursive_lock())
2717 		goto out_nocheck;
2718 
2719 	cpu = raw_smp_processor_id();
2720 
2721 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
2722 		goto out;
2723 
2724 	cpu_buffer = buffer->buffers[cpu];
2725 
2726 	if (atomic_read(&cpu_buffer->record_disabled))
2727 		goto out;
2728 
2729 	if (length > BUF_MAX_DATA_SIZE)
2730 		goto out;
2731 
2732 	event = rb_reserve_next_event(buffer, cpu_buffer, length);
2733 	if (!event)
2734 		goto out;
2735 
2736 	return event;
2737 
2738  out:
2739 	trace_recursive_unlock();
2740 
2741  out_nocheck:
2742 	preempt_enable_notrace();
2743 	return NULL;
2744 }
2745 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
2746 
2747 static void
2748 rb_update_write_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2749 		      struct ring_buffer_event *event)
2750 {
2751 	u64 delta;
2752 
2753 	/*
2754 	 * The event first in the commit queue updates the
2755 	 * time stamp.
2756 	 */
2757 	if (rb_event_is_commit(cpu_buffer, event)) {
2758 		/*
2759 		 * A commit event that is first on a page
2760 		 * updates the write timestamp with the page stamp
2761 		 */
2762 		if (!rb_event_index(event))
2763 			cpu_buffer->write_stamp =
2764 				cpu_buffer->commit_page->page->time_stamp;
2765 		else if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
2766 			delta = event->array[0];
2767 			delta <<= TS_SHIFT;
2768 			delta += event->time_delta;
2769 			cpu_buffer->write_stamp += delta;
2770 		} else
2771 			cpu_buffer->write_stamp += event->time_delta;
2772 	}
2773 }
2774 
2775 static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
2776 		      struct ring_buffer_event *event)
2777 {
2778 	local_inc(&cpu_buffer->entries);
2779 	rb_update_write_stamp(cpu_buffer, event);
2780 	rb_end_commit(cpu_buffer);
2781 }
2782 
2783 static __always_inline void
2784 rb_wakeups(struct ring_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer)
2785 {
2786 	if (buffer->irq_work.waiters_pending) {
2787 		buffer->irq_work.waiters_pending = false;
2788 		/* irq_work_queue() supplies it's own memory barriers */
2789 		irq_work_queue(&buffer->irq_work.work);
2790 	}
2791 
2792 	if (cpu_buffer->irq_work.waiters_pending) {
2793 		cpu_buffer->irq_work.waiters_pending = false;
2794 		/* irq_work_queue() supplies it's own memory barriers */
2795 		irq_work_queue(&cpu_buffer->irq_work.work);
2796 	}
2797 }
2798 
2799 /**
2800  * ring_buffer_unlock_commit - commit a reserved
2801  * @buffer: The buffer to commit to
2802  * @event: The event pointer to commit.
2803  *
2804  * This commits the data to the ring buffer, and releases any locks held.
2805  *
2806  * Must be paired with ring_buffer_lock_reserve.
2807  */
2808 int ring_buffer_unlock_commit(struct ring_buffer *buffer,
2809 			      struct ring_buffer_event *event)
2810 {
2811 	struct ring_buffer_per_cpu *cpu_buffer;
2812 	int cpu = raw_smp_processor_id();
2813 
2814 	cpu_buffer = buffer->buffers[cpu];
2815 
2816 	rb_commit(cpu_buffer, event);
2817 
2818 	rb_wakeups(buffer, cpu_buffer);
2819 
2820 	trace_recursive_unlock();
2821 
2822 	preempt_enable_notrace();
2823 
2824 	return 0;
2825 }
2826 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
2827 
2828 static inline void rb_event_discard(struct ring_buffer_event *event)
2829 {
2830 	if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
2831 		event = skip_time_extend(event);
2832 
2833 	/* array[0] holds the actual length for the discarded event */
2834 	event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
2835 	event->type_len = RINGBUF_TYPE_PADDING;
2836 	/* time delta must be non zero */
2837 	if (!event->time_delta)
2838 		event->time_delta = 1;
2839 }
2840 
2841 /*
2842  * Decrement the entries to the page that an event is on.
2843  * The event does not even need to exist, only the pointer
2844  * to the page it is on. This may only be called before the commit
2845  * takes place.
2846  */
2847 static inline void
2848 rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
2849 		   struct ring_buffer_event *event)
2850 {
2851 	unsigned long addr = (unsigned long)event;
2852 	struct buffer_page *bpage = cpu_buffer->commit_page;
2853 	struct buffer_page *start;
2854 
2855 	addr &= PAGE_MASK;
2856 
2857 	/* Do the likely case first */
2858 	if (likely(bpage->page == (void *)addr)) {
2859 		local_dec(&bpage->entries);
2860 		return;
2861 	}
2862 
2863 	/*
2864 	 * Because the commit page may be on the reader page we
2865 	 * start with the next page and check the end loop there.
2866 	 */
2867 	rb_inc_page(cpu_buffer, &bpage);
2868 	start = bpage;
2869 	do {
2870 		if (bpage->page == (void *)addr) {
2871 			local_dec(&bpage->entries);
2872 			return;
2873 		}
2874 		rb_inc_page(cpu_buffer, &bpage);
2875 	} while (bpage != start);
2876 
2877 	/* commit not part of this buffer?? */
2878 	RB_WARN_ON(cpu_buffer, 1);
2879 }
2880 
2881 /**
2882  * ring_buffer_commit_discard - discard an event that has not been committed
2883  * @buffer: the ring buffer
2884  * @event: non committed event to discard
2885  *
2886  * Sometimes an event that is in the ring buffer needs to be ignored.
2887  * This function lets the user discard an event in the ring buffer
2888  * and then that event will not be read later.
2889  *
2890  * This function only works if it is called before the the item has been
2891  * committed. It will try to free the event from the ring buffer
2892  * if another event has not been added behind it.
2893  *
2894  * If another event has been added behind it, it will set the event
2895  * up as discarded, and perform the commit.
2896  *
2897  * If this function is called, do not call ring_buffer_unlock_commit on
2898  * the event.
2899  */
2900 void ring_buffer_discard_commit(struct ring_buffer *buffer,
2901 				struct ring_buffer_event *event)
2902 {
2903 	struct ring_buffer_per_cpu *cpu_buffer;
2904 	int cpu;
2905 
2906 	/* The event is discarded regardless */
2907 	rb_event_discard(event);
2908 
2909 	cpu = smp_processor_id();
2910 	cpu_buffer = buffer->buffers[cpu];
2911 
2912 	/*
2913 	 * This must only be called if the event has not been
2914 	 * committed yet. Thus we can assume that preemption
2915 	 * is still disabled.
2916 	 */
2917 	RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
2918 
2919 	rb_decrement_entry(cpu_buffer, event);
2920 	if (rb_try_to_discard(cpu_buffer, event))
2921 		goto out;
2922 
2923 	/*
2924 	 * The commit is still visible by the reader, so we
2925 	 * must still update the timestamp.
2926 	 */
2927 	rb_update_write_stamp(cpu_buffer, event);
2928  out:
2929 	rb_end_commit(cpu_buffer);
2930 
2931 	trace_recursive_unlock();
2932 
2933 	preempt_enable_notrace();
2934 
2935 }
2936 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
2937 
2938 /**
2939  * ring_buffer_write - write data to the buffer without reserving
2940  * @buffer: The ring buffer to write to.
2941  * @length: The length of the data being written (excluding the event header)
2942  * @data: The data to write to the buffer.
2943  *
2944  * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
2945  * one function. If you already have the data to write to the buffer, it
2946  * may be easier to simply call this function.
2947  *
2948  * Note, like ring_buffer_lock_reserve, the length is the length of the data
2949  * and not the length of the event which would hold the header.
2950  */
2951 int ring_buffer_write(struct ring_buffer *buffer,
2952 		      unsigned long length,
2953 		      void *data)
2954 {
2955 	struct ring_buffer_per_cpu *cpu_buffer;
2956 	struct ring_buffer_event *event;
2957 	void *body;
2958 	int ret = -EBUSY;
2959 	int cpu;
2960 
2961 	if (ring_buffer_flags != RB_BUFFERS_ON)
2962 		return -EBUSY;
2963 
2964 	preempt_disable_notrace();
2965 
2966 	if (atomic_read(&buffer->record_disabled))
2967 		goto out;
2968 
2969 	cpu = raw_smp_processor_id();
2970 
2971 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
2972 		goto out;
2973 
2974 	cpu_buffer = buffer->buffers[cpu];
2975 
2976 	if (atomic_read(&cpu_buffer->record_disabled))
2977 		goto out;
2978 
2979 	if (length > BUF_MAX_DATA_SIZE)
2980 		goto out;
2981 
2982 	event = rb_reserve_next_event(buffer, cpu_buffer, length);
2983 	if (!event)
2984 		goto out;
2985 
2986 	body = rb_event_data(event);
2987 
2988 	memcpy(body, data, length);
2989 
2990 	rb_commit(cpu_buffer, event);
2991 
2992 	rb_wakeups(buffer, cpu_buffer);
2993 
2994 	ret = 0;
2995  out:
2996 	preempt_enable_notrace();
2997 
2998 	return ret;
2999 }
3000 EXPORT_SYMBOL_GPL(ring_buffer_write);
3001 
3002 static int rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
3003 {
3004 	struct buffer_page *reader = cpu_buffer->reader_page;
3005 	struct buffer_page *head = rb_set_head_page(cpu_buffer);
3006 	struct buffer_page *commit = cpu_buffer->commit_page;
3007 
3008 	/* In case of error, head will be NULL */
3009 	if (unlikely(!head))
3010 		return 1;
3011 
3012 	return reader->read == rb_page_commit(reader) &&
3013 		(commit == reader ||
3014 		 (commit == head &&
3015 		  head->read == rb_page_commit(commit)));
3016 }
3017 
3018 /**
3019  * ring_buffer_record_disable - stop all writes into the buffer
3020  * @buffer: The ring buffer to stop writes to.
3021  *
3022  * This prevents all writes to the buffer. Any attempt to write
3023  * to the buffer after this will fail and return NULL.
3024  *
3025  * The caller should call synchronize_sched() after this.
3026  */
3027 void ring_buffer_record_disable(struct ring_buffer *buffer)
3028 {
3029 	atomic_inc(&buffer->record_disabled);
3030 }
3031 EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
3032 
3033 /**
3034  * ring_buffer_record_enable - enable writes to the buffer
3035  * @buffer: The ring buffer to enable writes
3036  *
3037  * Note, multiple disables will need the same number of enables
3038  * to truly enable the writing (much like preempt_disable).
3039  */
3040 void ring_buffer_record_enable(struct ring_buffer *buffer)
3041 {
3042 	atomic_dec(&buffer->record_disabled);
3043 }
3044 EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
3045 
3046 /**
3047  * ring_buffer_record_off - stop all writes into the buffer
3048  * @buffer: The ring buffer to stop writes to.
3049  *
3050  * This prevents all writes to the buffer. Any attempt to write
3051  * to the buffer after this will fail and return NULL.
3052  *
3053  * This is different than ring_buffer_record_disable() as
3054  * it works like an on/off switch, where as the disable() version
3055  * must be paired with a enable().
3056  */
3057 void ring_buffer_record_off(struct ring_buffer *buffer)
3058 {
3059 	unsigned int rd;
3060 	unsigned int new_rd;
3061 
3062 	do {
3063 		rd = atomic_read(&buffer->record_disabled);
3064 		new_rd = rd | RB_BUFFER_OFF;
3065 	} while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3066 }
3067 EXPORT_SYMBOL_GPL(ring_buffer_record_off);
3068 
3069 /**
3070  * ring_buffer_record_on - restart writes into the buffer
3071  * @buffer: The ring buffer to start writes to.
3072  *
3073  * This enables all writes to the buffer that was disabled by
3074  * ring_buffer_record_off().
3075  *
3076  * This is different than ring_buffer_record_enable() as
3077  * it works like an on/off switch, where as the enable() version
3078  * must be paired with a disable().
3079  */
3080 void ring_buffer_record_on(struct ring_buffer *buffer)
3081 {
3082 	unsigned int rd;
3083 	unsigned int new_rd;
3084 
3085 	do {
3086 		rd = atomic_read(&buffer->record_disabled);
3087 		new_rd = rd & ~RB_BUFFER_OFF;
3088 	} while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3089 }
3090 EXPORT_SYMBOL_GPL(ring_buffer_record_on);
3091 
3092 /**
3093  * ring_buffer_record_is_on - return true if the ring buffer can write
3094  * @buffer: The ring buffer to see if write is enabled
3095  *
3096  * Returns true if the ring buffer is in a state that it accepts writes.
3097  */
3098 int ring_buffer_record_is_on(struct ring_buffer *buffer)
3099 {
3100 	return !atomic_read(&buffer->record_disabled);
3101 }
3102 
3103 /**
3104  * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
3105  * @buffer: The ring buffer to stop writes to.
3106  * @cpu: The CPU buffer to stop
3107  *
3108  * This prevents all writes to the buffer. Any attempt to write
3109  * to the buffer after this will fail and return NULL.
3110  *
3111  * The caller should call synchronize_sched() after this.
3112  */
3113 void ring_buffer_record_disable_cpu(struct ring_buffer *buffer, int cpu)
3114 {
3115 	struct ring_buffer_per_cpu *cpu_buffer;
3116 
3117 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3118 		return;
3119 
3120 	cpu_buffer = buffer->buffers[cpu];
3121 	atomic_inc(&cpu_buffer->record_disabled);
3122 }
3123 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
3124 
3125 /**
3126  * ring_buffer_record_enable_cpu - enable writes to the buffer
3127  * @buffer: The ring buffer to enable writes
3128  * @cpu: The CPU to enable.
3129  *
3130  * Note, multiple disables will need the same number of enables
3131  * to truly enable the writing (much like preempt_disable).
3132  */
3133 void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu)
3134 {
3135 	struct ring_buffer_per_cpu *cpu_buffer;
3136 
3137 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3138 		return;
3139 
3140 	cpu_buffer = buffer->buffers[cpu];
3141 	atomic_dec(&cpu_buffer->record_disabled);
3142 }
3143 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
3144 
3145 /*
3146  * The total entries in the ring buffer is the running counter
3147  * of entries entered into the ring buffer, minus the sum of
3148  * the entries read from the ring buffer and the number of
3149  * entries that were overwritten.
3150  */
3151 static inline unsigned long
3152 rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
3153 {
3154 	return local_read(&cpu_buffer->entries) -
3155 		(local_read(&cpu_buffer->overrun) + cpu_buffer->read);
3156 }
3157 
3158 /**
3159  * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
3160  * @buffer: The ring buffer
3161  * @cpu: The per CPU buffer to read from.
3162  */
3163 u64 ring_buffer_oldest_event_ts(struct ring_buffer *buffer, int cpu)
3164 {
3165 	unsigned long flags;
3166 	struct ring_buffer_per_cpu *cpu_buffer;
3167 	struct buffer_page *bpage;
3168 	u64 ret = 0;
3169 
3170 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3171 		return 0;
3172 
3173 	cpu_buffer = buffer->buffers[cpu];
3174 	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3175 	/*
3176 	 * if the tail is on reader_page, oldest time stamp is on the reader
3177 	 * page
3178 	 */
3179 	if (cpu_buffer->tail_page == cpu_buffer->reader_page)
3180 		bpage = cpu_buffer->reader_page;
3181 	else
3182 		bpage = rb_set_head_page(cpu_buffer);
3183 	if (bpage)
3184 		ret = bpage->page->time_stamp;
3185 	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3186 
3187 	return ret;
3188 }
3189 EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
3190 
3191 /**
3192  * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
3193  * @buffer: The ring buffer
3194  * @cpu: The per CPU buffer to read from.
3195  */
3196 unsigned long ring_buffer_bytes_cpu(struct ring_buffer *buffer, int cpu)
3197 {
3198 	struct ring_buffer_per_cpu *cpu_buffer;
3199 	unsigned long ret;
3200 
3201 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3202 		return 0;
3203 
3204 	cpu_buffer = buffer->buffers[cpu];
3205 	ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
3206 
3207 	return ret;
3208 }
3209 EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
3210 
3211 /**
3212  * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
3213  * @buffer: The ring buffer
3214  * @cpu: The per CPU buffer to get the entries from.
3215  */
3216 unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu)
3217 {
3218 	struct ring_buffer_per_cpu *cpu_buffer;
3219 
3220 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3221 		return 0;
3222 
3223 	cpu_buffer = buffer->buffers[cpu];
3224 
3225 	return rb_num_of_entries(cpu_buffer);
3226 }
3227 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
3228 
3229 /**
3230  * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
3231  * buffer wrapping around (only if RB_FL_OVERWRITE is on).
3232  * @buffer: The ring buffer
3233  * @cpu: The per CPU buffer to get the number of overruns from
3234  */
3235 unsigned long ring_buffer_overrun_cpu(struct ring_buffer *buffer, int cpu)
3236 {
3237 	struct ring_buffer_per_cpu *cpu_buffer;
3238 	unsigned long ret;
3239 
3240 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3241 		return 0;
3242 
3243 	cpu_buffer = buffer->buffers[cpu];
3244 	ret = local_read(&cpu_buffer->overrun);
3245 
3246 	return ret;
3247 }
3248 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
3249 
3250 /**
3251  * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
3252  * commits failing due to the buffer wrapping around while there are uncommitted
3253  * events, such as during an interrupt storm.
3254  * @buffer: The ring buffer
3255  * @cpu: The per CPU buffer to get the number of overruns from
3256  */
3257 unsigned long
3258 ring_buffer_commit_overrun_cpu(struct ring_buffer *buffer, int cpu)
3259 {
3260 	struct ring_buffer_per_cpu *cpu_buffer;
3261 	unsigned long ret;
3262 
3263 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3264 		return 0;
3265 
3266 	cpu_buffer = buffer->buffers[cpu];
3267 	ret = local_read(&cpu_buffer->commit_overrun);
3268 
3269 	return ret;
3270 }
3271 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
3272 
3273 /**
3274  * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
3275  * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
3276  * @buffer: The ring buffer
3277  * @cpu: The per CPU buffer to get the number of overruns from
3278  */
3279 unsigned long
3280 ring_buffer_dropped_events_cpu(struct ring_buffer *buffer, int cpu)
3281 {
3282 	struct ring_buffer_per_cpu *cpu_buffer;
3283 	unsigned long ret;
3284 
3285 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3286 		return 0;
3287 
3288 	cpu_buffer = buffer->buffers[cpu];
3289 	ret = local_read(&cpu_buffer->dropped_events);
3290 
3291 	return ret;
3292 }
3293 EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
3294 
3295 /**
3296  * ring_buffer_read_events_cpu - get the number of events successfully read
3297  * @buffer: The ring buffer
3298  * @cpu: The per CPU buffer to get the number of events read
3299  */
3300 unsigned long
3301 ring_buffer_read_events_cpu(struct ring_buffer *buffer, int cpu)
3302 {
3303 	struct ring_buffer_per_cpu *cpu_buffer;
3304 
3305 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3306 		return 0;
3307 
3308 	cpu_buffer = buffer->buffers[cpu];
3309 	return cpu_buffer->read;
3310 }
3311 EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu);
3312 
3313 /**
3314  * ring_buffer_entries - get the number of entries in a buffer
3315  * @buffer: The ring buffer
3316  *
3317  * Returns the total number of entries in the ring buffer
3318  * (all CPU entries)
3319  */
3320 unsigned long ring_buffer_entries(struct ring_buffer *buffer)
3321 {
3322 	struct ring_buffer_per_cpu *cpu_buffer;
3323 	unsigned long entries = 0;
3324 	int cpu;
3325 
3326 	/* if you care about this being correct, lock the buffer */
3327 	for_each_buffer_cpu(buffer, cpu) {
3328 		cpu_buffer = buffer->buffers[cpu];
3329 		entries += rb_num_of_entries(cpu_buffer);
3330 	}
3331 
3332 	return entries;
3333 }
3334 EXPORT_SYMBOL_GPL(ring_buffer_entries);
3335 
3336 /**
3337  * ring_buffer_overruns - get the number of overruns in buffer
3338  * @buffer: The ring buffer
3339  *
3340  * Returns the total number of overruns in the ring buffer
3341  * (all CPU entries)
3342  */
3343 unsigned long ring_buffer_overruns(struct ring_buffer *buffer)
3344 {
3345 	struct ring_buffer_per_cpu *cpu_buffer;
3346 	unsigned long overruns = 0;
3347 	int cpu;
3348 
3349 	/* if you care about this being correct, lock the buffer */
3350 	for_each_buffer_cpu(buffer, cpu) {
3351 		cpu_buffer = buffer->buffers[cpu];
3352 		overruns += local_read(&cpu_buffer->overrun);
3353 	}
3354 
3355 	return overruns;
3356 }
3357 EXPORT_SYMBOL_GPL(ring_buffer_overruns);
3358 
3359 static void rb_iter_reset(struct ring_buffer_iter *iter)
3360 {
3361 	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3362 
3363 	/* Iterator usage is expected to have record disabled */
3364 	if (list_empty(&cpu_buffer->reader_page->list)) {
3365 		iter->head_page = rb_set_head_page(cpu_buffer);
3366 		if (unlikely(!iter->head_page))
3367 			return;
3368 		iter->head = iter->head_page->read;
3369 	} else {
3370 		iter->head_page = cpu_buffer->reader_page;
3371 		iter->head = cpu_buffer->reader_page->read;
3372 	}
3373 	if (iter->head)
3374 		iter->read_stamp = cpu_buffer->read_stamp;
3375 	else
3376 		iter->read_stamp = iter->head_page->page->time_stamp;
3377 	iter->cache_reader_page = cpu_buffer->reader_page;
3378 	iter->cache_read = cpu_buffer->read;
3379 }
3380 
3381 /**
3382  * ring_buffer_iter_reset - reset an iterator
3383  * @iter: The iterator to reset
3384  *
3385  * Resets the iterator, so that it will start from the beginning
3386  * again.
3387  */
3388 void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
3389 {
3390 	struct ring_buffer_per_cpu *cpu_buffer;
3391 	unsigned long flags;
3392 
3393 	if (!iter)
3394 		return;
3395 
3396 	cpu_buffer = iter->cpu_buffer;
3397 
3398 	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3399 	rb_iter_reset(iter);
3400 	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3401 }
3402 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
3403 
3404 /**
3405  * ring_buffer_iter_empty - check if an iterator has no more to read
3406  * @iter: The iterator to check
3407  */
3408 int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
3409 {
3410 	struct ring_buffer_per_cpu *cpu_buffer;
3411 
3412 	cpu_buffer = iter->cpu_buffer;
3413 
3414 	return iter->head_page == cpu_buffer->commit_page &&
3415 		iter->head == rb_commit_index(cpu_buffer);
3416 }
3417 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
3418 
3419 static void
3420 rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
3421 		     struct ring_buffer_event *event)
3422 {
3423 	u64 delta;
3424 
3425 	switch (event->type_len) {
3426 	case RINGBUF_TYPE_PADDING:
3427 		return;
3428 
3429 	case RINGBUF_TYPE_TIME_EXTEND:
3430 		delta = event->array[0];
3431 		delta <<= TS_SHIFT;
3432 		delta += event->time_delta;
3433 		cpu_buffer->read_stamp += delta;
3434 		return;
3435 
3436 	case RINGBUF_TYPE_TIME_STAMP:
3437 		/* FIXME: not implemented */
3438 		return;
3439 
3440 	case RINGBUF_TYPE_DATA:
3441 		cpu_buffer->read_stamp += event->time_delta;
3442 		return;
3443 
3444 	default:
3445 		BUG();
3446 	}
3447 	return;
3448 }
3449 
3450 static void
3451 rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
3452 			  struct ring_buffer_event *event)
3453 {
3454 	u64 delta;
3455 
3456 	switch (event->type_len) {
3457 	case RINGBUF_TYPE_PADDING:
3458 		return;
3459 
3460 	case RINGBUF_TYPE_TIME_EXTEND:
3461 		delta = event->array[0];
3462 		delta <<= TS_SHIFT;
3463 		delta += event->time_delta;
3464 		iter->read_stamp += delta;
3465 		return;
3466 
3467 	case RINGBUF_TYPE_TIME_STAMP:
3468 		/* FIXME: not implemented */
3469 		return;
3470 
3471 	case RINGBUF_TYPE_DATA:
3472 		iter->read_stamp += event->time_delta;
3473 		return;
3474 
3475 	default:
3476 		BUG();
3477 	}
3478 	return;
3479 }
3480 
3481 static struct buffer_page *
3482 rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
3483 {
3484 	struct buffer_page *reader = NULL;
3485 	unsigned long overwrite;
3486 	unsigned long flags;
3487 	int nr_loops = 0;
3488 	int ret;
3489 
3490 	local_irq_save(flags);
3491 	arch_spin_lock(&cpu_buffer->lock);
3492 
3493  again:
3494 	/*
3495 	 * This should normally only loop twice. But because the
3496 	 * start of the reader inserts an empty page, it causes
3497 	 * a case where we will loop three times. There should be no
3498 	 * reason to loop four times (that I know of).
3499 	 */
3500 	if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
3501 		reader = NULL;
3502 		goto out;
3503 	}
3504 
3505 	reader = cpu_buffer->reader_page;
3506 
3507 	/* If there's more to read, return this page */
3508 	if (cpu_buffer->reader_page->read < rb_page_size(reader))
3509 		goto out;
3510 
3511 	/* Never should we have an index greater than the size */
3512 	if (RB_WARN_ON(cpu_buffer,
3513 		       cpu_buffer->reader_page->read > rb_page_size(reader)))
3514 		goto out;
3515 
3516 	/* check if we caught up to the tail */
3517 	reader = NULL;
3518 	if (cpu_buffer->commit_page == cpu_buffer->reader_page)
3519 		goto out;
3520 
3521 	/* Don't bother swapping if the ring buffer is empty */
3522 	if (rb_num_of_entries(cpu_buffer) == 0)
3523 		goto out;
3524 
3525 	/*
3526 	 * Reset the reader page to size zero.
3527 	 */
3528 	local_set(&cpu_buffer->reader_page->write, 0);
3529 	local_set(&cpu_buffer->reader_page->entries, 0);
3530 	local_set(&cpu_buffer->reader_page->page->commit, 0);
3531 	cpu_buffer->reader_page->real_end = 0;
3532 
3533  spin:
3534 	/*
3535 	 * Splice the empty reader page into the list around the head.
3536 	 */
3537 	reader = rb_set_head_page(cpu_buffer);
3538 	if (!reader)
3539 		goto out;
3540 	cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
3541 	cpu_buffer->reader_page->list.prev = reader->list.prev;
3542 
3543 	/*
3544 	 * cpu_buffer->pages just needs to point to the buffer, it
3545 	 *  has no specific buffer page to point to. Lets move it out
3546 	 *  of our way so we don't accidentally swap it.
3547 	 */
3548 	cpu_buffer->pages = reader->list.prev;
3549 
3550 	/* The reader page will be pointing to the new head */
3551 	rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list);
3552 
3553 	/*
3554 	 * We want to make sure we read the overruns after we set up our
3555 	 * pointers to the next object. The writer side does a
3556 	 * cmpxchg to cross pages which acts as the mb on the writer
3557 	 * side. Note, the reader will constantly fail the swap
3558 	 * while the writer is updating the pointers, so this
3559 	 * guarantees that the overwrite recorded here is the one we
3560 	 * want to compare with the last_overrun.
3561 	 */
3562 	smp_mb();
3563 	overwrite = local_read(&(cpu_buffer->overrun));
3564 
3565 	/*
3566 	 * Here's the tricky part.
3567 	 *
3568 	 * We need to move the pointer past the header page.
3569 	 * But we can only do that if a writer is not currently
3570 	 * moving it. The page before the header page has the
3571 	 * flag bit '1' set if it is pointing to the page we want.
3572 	 * but if the writer is in the process of moving it
3573 	 * than it will be '2' or already moved '0'.
3574 	 */
3575 
3576 	ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
3577 
3578 	/*
3579 	 * If we did not convert it, then we must try again.
3580 	 */
3581 	if (!ret)
3582 		goto spin;
3583 
3584 	/*
3585 	 * Yeah! We succeeded in replacing the page.
3586 	 *
3587 	 * Now make the new head point back to the reader page.
3588 	 */
3589 	rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
3590 	rb_inc_page(cpu_buffer, &cpu_buffer->head_page);
3591 
3592 	/* Finally update the reader page to the new head */
3593 	cpu_buffer->reader_page = reader;
3594 	rb_reset_reader_page(cpu_buffer);
3595 
3596 	if (overwrite != cpu_buffer->last_overrun) {
3597 		cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
3598 		cpu_buffer->last_overrun = overwrite;
3599 	}
3600 
3601 	goto again;
3602 
3603  out:
3604 	arch_spin_unlock(&cpu_buffer->lock);
3605 	local_irq_restore(flags);
3606 
3607 	return reader;
3608 }
3609 
3610 static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
3611 {
3612 	struct ring_buffer_event *event;
3613 	struct buffer_page *reader;
3614 	unsigned length;
3615 
3616 	reader = rb_get_reader_page(cpu_buffer);
3617 
3618 	/* This function should not be called when buffer is empty */
3619 	if (RB_WARN_ON(cpu_buffer, !reader))
3620 		return;
3621 
3622 	event = rb_reader_event(cpu_buffer);
3623 
3624 	if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
3625 		cpu_buffer->read++;
3626 
3627 	rb_update_read_stamp(cpu_buffer, event);
3628 
3629 	length = rb_event_length(event);
3630 	cpu_buffer->reader_page->read += length;
3631 }
3632 
3633 static void rb_advance_iter(struct ring_buffer_iter *iter)
3634 {
3635 	struct ring_buffer_per_cpu *cpu_buffer;
3636 	struct ring_buffer_event *event;
3637 	unsigned length;
3638 
3639 	cpu_buffer = iter->cpu_buffer;
3640 
3641 	/*
3642 	 * Check if we are at the end of the buffer.
3643 	 */
3644 	if (iter->head >= rb_page_size(iter->head_page)) {
3645 		/* discarded commits can make the page empty */
3646 		if (iter->head_page == cpu_buffer->commit_page)
3647 			return;
3648 		rb_inc_iter(iter);
3649 		return;
3650 	}
3651 
3652 	event = rb_iter_head_event(iter);
3653 
3654 	length = rb_event_length(event);
3655 
3656 	/*
3657 	 * This should not be called to advance the header if we are
3658 	 * at the tail of the buffer.
3659 	 */
3660 	if (RB_WARN_ON(cpu_buffer,
3661 		       (iter->head_page == cpu_buffer->commit_page) &&
3662 		       (iter->head + length > rb_commit_index(cpu_buffer))))
3663 		return;
3664 
3665 	rb_update_iter_read_stamp(iter, event);
3666 
3667 	iter->head += length;
3668 
3669 	/* check for end of page padding */
3670 	if ((iter->head >= rb_page_size(iter->head_page)) &&
3671 	    (iter->head_page != cpu_buffer->commit_page))
3672 		rb_inc_iter(iter);
3673 }
3674 
3675 static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
3676 {
3677 	return cpu_buffer->lost_events;
3678 }
3679 
3680 static struct ring_buffer_event *
3681 rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
3682 	       unsigned long *lost_events)
3683 {
3684 	struct ring_buffer_event *event;
3685 	struct buffer_page *reader;
3686 	int nr_loops = 0;
3687 
3688  again:
3689 	/*
3690 	 * We repeat when a time extend is encountered.
3691 	 * Since the time extend is always attached to a data event,
3692 	 * we should never loop more than once.
3693 	 * (We never hit the following condition more than twice).
3694 	 */
3695 	if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3696 		return NULL;
3697 
3698 	reader = rb_get_reader_page(cpu_buffer);
3699 	if (!reader)
3700 		return NULL;
3701 
3702 	event = rb_reader_event(cpu_buffer);
3703 
3704 	switch (event->type_len) {
3705 	case RINGBUF_TYPE_PADDING:
3706 		if (rb_null_event(event))
3707 			RB_WARN_ON(cpu_buffer, 1);
3708 		/*
3709 		 * Because the writer could be discarding every
3710 		 * event it creates (which would probably be bad)
3711 		 * if we were to go back to "again" then we may never
3712 		 * catch up, and will trigger the warn on, or lock
3713 		 * the box. Return the padding, and we will release
3714 		 * the current locks, and try again.
3715 		 */
3716 		return event;
3717 
3718 	case RINGBUF_TYPE_TIME_EXTEND:
3719 		/* Internal data, OK to advance */
3720 		rb_advance_reader(cpu_buffer);
3721 		goto again;
3722 
3723 	case RINGBUF_TYPE_TIME_STAMP:
3724 		/* FIXME: not implemented */
3725 		rb_advance_reader(cpu_buffer);
3726 		goto again;
3727 
3728 	case RINGBUF_TYPE_DATA:
3729 		if (ts) {
3730 			*ts = cpu_buffer->read_stamp + event->time_delta;
3731 			ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3732 							 cpu_buffer->cpu, ts);
3733 		}
3734 		if (lost_events)
3735 			*lost_events = rb_lost_events(cpu_buffer);
3736 		return event;
3737 
3738 	default:
3739 		BUG();
3740 	}
3741 
3742 	return NULL;
3743 }
3744 EXPORT_SYMBOL_GPL(ring_buffer_peek);
3745 
3746 static struct ring_buffer_event *
3747 rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3748 {
3749 	struct ring_buffer *buffer;
3750 	struct ring_buffer_per_cpu *cpu_buffer;
3751 	struct ring_buffer_event *event;
3752 	int nr_loops = 0;
3753 
3754 	cpu_buffer = iter->cpu_buffer;
3755 	buffer = cpu_buffer->buffer;
3756 
3757 	/*
3758 	 * Check if someone performed a consuming read to
3759 	 * the buffer. A consuming read invalidates the iterator
3760 	 * and we need to reset the iterator in this case.
3761 	 */
3762 	if (unlikely(iter->cache_read != cpu_buffer->read ||
3763 		     iter->cache_reader_page != cpu_buffer->reader_page))
3764 		rb_iter_reset(iter);
3765 
3766  again:
3767 	if (ring_buffer_iter_empty(iter))
3768 		return NULL;
3769 
3770 	/*
3771 	 * We repeat when a time extend is encountered.
3772 	 * Since the time extend is always attached to a data event,
3773 	 * we should never loop more than once.
3774 	 * (We never hit the following condition more than twice).
3775 	 */
3776 	if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3777 		return NULL;
3778 
3779 	if (rb_per_cpu_empty(cpu_buffer))
3780 		return NULL;
3781 
3782 	if (iter->head >= local_read(&iter->head_page->page->commit)) {
3783 		rb_inc_iter(iter);
3784 		goto again;
3785 	}
3786 
3787 	event = rb_iter_head_event(iter);
3788 
3789 	switch (event->type_len) {
3790 	case RINGBUF_TYPE_PADDING:
3791 		if (rb_null_event(event)) {
3792 			rb_inc_iter(iter);
3793 			goto again;
3794 		}
3795 		rb_advance_iter(iter);
3796 		return event;
3797 
3798 	case RINGBUF_TYPE_TIME_EXTEND:
3799 		/* Internal data, OK to advance */
3800 		rb_advance_iter(iter);
3801 		goto again;
3802 
3803 	case RINGBUF_TYPE_TIME_STAMP:
3804 		/* FIXME: not implemented */
3805 		rb_advance_iter(iter);
3806 		goto again;
3807 
3808 	case RINGBUF_TYPE_DATA:
3809 		if (ts) {
3810 			*ts = iter->read_stamp + event->time_delta;
3811 			ring_buffer_normalize_time_stamp(buffer,
3812 							 cpu_buffer->cpu, ts);
3813 		}
3814 		return event;
3815 
3816 	default:
3817 		BUG();
3818 	}
3819 
3820 	return NULL;
3821 }
3822 EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
3823 
3824 static inline int rb_ok_to_lock(void)
3825 {
3826 	/*
3827 	 * If an NMI die dumps out the content of the ring buffer
3828 	 * do not grab locks. We also permanently disable the ring
3829 	 * buffer too. A one time deal is all you get from reading
3830 	 * the ring buffer from an NMI.
3831 	 */
3832 	if (likely(!in_nmi()))
3833 		return 1;
3834 
3835 	tracing_off_permanent();
3836 	return 0;
3837 }
3838 
3839 /**
3840  * ring_buffer_peek - peek at the next event to be read
3841  * @buffer: The ring buffer to read
3842  * @cpu: The cpu to peak at
3843  * @ts: The timestamp counter of this event.
3844  * @lost_events: a variable to store if events were lost (may be NULL)
3845  *
3846  * This will return the event that will be read next, but does
3847  * not consume the data.
3848  */
3849 struct ring_buffer_event *
3850 ring_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts,
3851 		 unsigned long *lost_events)
3852 {
3853 	struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3854 	struct ring_buffer_event *event;
3855 	unsigned long flags;
3856 	int dolock;
3857 
3858 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3859 		return NULL;
3860 
3861 	dolock = rb_ok_to_lock();
3862  again:
3863 	local_irq_save(flags);
3864 	if (dolock)
3865 		raw_spin_lock(&cpu_buffer->reader_lock);
3866 	event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3867 	if (event && event->type_len == RINGBUF_TYPE_PADDING)
3868 		rb_advance_reader(cpu_buffer);
3869 	if (dolock)
3870 		raw_spin_unlock(&cpu_buffer->reader_lock);
3871 	local_irq_restore(flags);
3872 
3873 	if (event && event->type_len == RINGBUF_TYPE_PADDING)
3874 		goto again;
3875 
3876 	return event;
3877 }
3878 
3879 /**
3880  * ring_buffer_iter_peek - peek at the next event to be read
3881  * @iter: The ring buffer iterator
3882  * @ts: The timestamp counter of this event.
3883  *
3884  * This will return the event that will be read next, but does
3885  * not increment the iterator.
3886  */
3887 struct ring_buffer_event *
3888 ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3889 {
3890 	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3891 	struct ring_buffer_event *event;
3892 	unsigned long flags;
3893 
3894  again:
3895 	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3896 	event = rb_iter_peek(iter, ts);
3897 	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3898 
3899 	if (event && event->type_len == RINGBUF_TYPE_PADDING)
3900 		goto again;
3901 
3902 	return event;
3903 }
3904 
3905 /**
3906  * ring_buffer_consume - return an event and consume it
3907  * @buffer: The ring buffer to get the next event from
3908  * @cpu: the cpu to read the buffer from
3909  * @ts: a variable to store the timestamp (may be NULL)
3910  * @lost_events: a variable to store if events were lost (may be NULL)
3911  *
3912  * Returns the next event in the ring buffer, and that event is consumed.
3913  * Meaning, that sequential reads will keep returning a different event,
3914  * and eventually empty the ring buffer if the producer is slower.
3915  */
3916 struct ring_buffer_event *
3917 ring_buffer_consume(struct ring_buffer *buffer, int cpu, u64 *ts,
3918 		    unsigned long *lost_events)
3919 {
3920 	struct ring_buffer_per_cpu *cpu_buffer;
3921 	struct ring_buffer_event *event = NULL;
3922 	unsigned long flags;
3923 	int dolock;
3924 
3925 	dolock = rb_ok_to_lock();
3926 
3927  again:
3928 	/* might be called in atomic */
3929 	preempt_disable();
3930 
3931 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3932 		goto out;
3933 
3934 	cpu_buffer = buffer->buffers[cpu];
3935 	local_irq_save(flags);
3936 	if (dolock)
3937 		raw_spin_lock(&cpu_buffer->reader_lock);
3938 
3939 	event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3940 	if (event) {
3941 		cpu_buffer->lost_events = 0;
3942 		rb_advance_reader(cpu_buffer);
3943 	}
3944 
3945 	if (dolock)
3946 		raw_spin_unlock(&cpu_buffer->reader_lock);
3947 	local_irq_restore(flags);
3948 
3949  out:
3950 	preempt_enable();
3951 
3952 	if (event && event->type_len == RINGBUF_TYPE_PADDING)
3953 		goto again;
3954 
3955 	return event;
3956 }
3957 EXPORT_SYMBOL_GPL(ring_buffer_consume);
3958 
3959 /**
3960  * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
3961  * @buffer: The ring buffer to read from
3962  * @cpu: The cpu buffer to iterate over
3963  *
3964  * This performs the initial preparations necessary to iterate
3965  * through the buffer.  Memory is allocated, buffer recording
3966  * is disabled, and the iterator pointer is returned to the caller.
3967  *
3968  * Disabling buffer recordng prevents the reading from being
3969  * corrupted. This is not a consuming read, so a producer is not
3970  * expected.
3971  *
3972  * After a sequence of ring_buffer_read_prepare calls, the user is
3973  * expected to make at least one call to ring_buffer_read_prepare_sync.
3974  * Afterwards, ring_buffer_read_start is invoked to get things going
3975  * for real.
3976  *
3977  * This overall must be paired with ring_buffer_read_finish.
3978  */
3979 struct ring_buffer_iter *
3980 ring_buffer_read_prepare(struct ring_buffer *buffer, int cpu)
3981 {
3982 	struct ring_buffer_per_cpu *cpu_buffer;
3983 	struct ring_buffer_iter *iter;
3984 
3985 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3986 		return NULL;
3987 
3988 	iter = kmalloc(sizeof(*iter), GFP_KERNEL);
3989 	if (!iter)
3990 		return NULL;
3991 
3992 	cpu_buffer = buffer->buffers[cpu];
3993 
3994 	iter->cpu_buffer = cpu_buffer;
3995 
3996 	atomic_inc(&buffer->resize_disabled);
3997 	atomic_inc(&cpu_buffer->record_disabled);
3998 
3999 	return iter;
4000 }
4001 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
4002 
4003 /**
4004  * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
4005  *
4006  * All previously invoked ring_buffer_read_prepare calls to prepare
4007  * iterators will be synchronized.  Afterwards, read_buffer_read_start
4008  * calls on those iterators are allowed.
4009  */
4010 void
4011 ring_buffer_read_prepare_sync(void)
4012 {
4013 	synchronize_sched();
4014 }
4015 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
4016 
4017 /**
4018  * ring_buffer_read_start - start a non consuming read of the buffer
4019  * @iter: The iterator returned by ring_buffer_read_prepare
4020  *
4021  * This finalizes the startup of an iteration through the buffer.
4022  * The iterator comes from a call to ring_buffer_read_prepare and
4023  * an intervening ring_buffer_read_prepare_sync must have been
4024  * performed.
4025  *
4026  * Must be paired with ring_buffer_read_finish.
4027  */
4028 void
4029 ring_buffer_read_start(struct ring_buffer_iter *iter)
4030 {
4031 	struct ring_buffer_per_cpu *cpu_buffer;
4032 	unsigned long flags;
4033 
4034 	if (!iter)
4035 		return;
4036 
4037 	cpu_buffer = iter->cpu_buffer;
4038 
4039 	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4040 	arch_spin_lock(&cpu_buffer->lock);
4041 	rb_iter_reset(iter);
4042 	arch_spin_unlock(&cpu_buffer->lock);
4043 	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4044 }
4045 EXPORT_SYMBOL_GPL(ring_buffer_read_start);
4046 
4047 /**
4048  * ring_buffer_read_finish - finish reading the iterator of the buffer
4049  * @iter: The iterator retrieved by ring_buffer_start
4050  *
4051  * This re-enables the recording to the buffer, and frees the
4052  * iterator.
4053  */
4054 void
4055 ring_buffer_read_finish(struct ring_buffer_iter *iter)
4056 {
4057 	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4058 	unsigned long flags;
4059 
4060 	/*
4061 	 * Ring buffer is disabled from recording, here's a good place
4062 	 * to check the integrity of the ring buffer.
4063 	 * Must prevent readers from trying to read, as the check
4064 	 * clears the HEAD page and readers require it.
4065 	 */
4066 	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4067 	rb_check_pages(cpu_buffer);
4068 	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4069 
4070 	atomic_dec(&cpu_buffer->record_disabled);
4071 	atomic_dec(&cpu_buffer->buffer->resize_disabled);
4072 	kfree(iter);
4073 }
4074 EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
4075 
4076 /**
4077  * ring_buffer_read - read the next item in the ring buffer by the iterator
4078  * @iter: The ring buffer iterator
4079  * @ts: The time stamp of the event read.
4080  *
4081  * This reads the next event in the ring buffer and increments the iterator.
4082  */
4083 struct ring_buffer_event *
4084 ring_buffer_read(struct ring_buffer_iter *iter, u64 *ts)
4085 {
4086 	struct ring_buffer_event *event;
4087 	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4088 	unsigned long flags;
4089 
4090 	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4091  again:
4092 	event = rb_iter_peek(iter, ts);
4093 	if (!event)
4094 		goto out;
4095 
4096 	if (event->type_len == RINGBUF_TYPE_PADDING)
4097 		goto again;
4098 
4099 	rb_advance_iter(iter);
4100  out:
4101 	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4102 
4103 	return event;
4104 }
4105 EXPORT_SYMBOL_GPL(ring_buffer_read);
4106 
4107 /**
4108  * ring_buffer_size - return the size of the ring buffer (in bytes)
4109  * @buffer: The ring buffer.
4110  */
4111 unsigned long ring_buffer_size(struct ring_buffer *buffer, int cpu)
4112 {
4113 	/*
4114 	 * Earlier, this method returned
4115 	 *	BUF_PAGE_SIZE * buffer->nr_pages
4116 	 * Since the nr_pages field is now removed, we have converted this to
4117 	 * return the per cpu buffer value.
4118 	 */
4119 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4120 		return 0;
4121 
4122 	return BUF_PAGE_SIZE * buffer->buffers[cpu]->nr_pages;
4123 }
4124 EXPORT_SYMBOL_GPL(ring_buffer_size);
4125 
4126 static void
4127 rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
4128 {
4129 	rb_head_page_deactivate(cpu_buffer);
4130 
4131 	cpu_buffer->head_page
4132 		= list_entry(cpu_buffer->pages, struct buffer_page, list);
4133 	local_set(&cpu_buffer->head_page->write, 0);
4134 	local_set(&cpu_buffer->head_page->entries, 0);
4135 	local_set(&cpu_buffer->head_page->page->commit, 0);
4136 
4137 	cpu_buffer->head_page->read = 0;
4138 
4139 	cpu_buffer->tail_page = cpu_buffer->head_page;
4140 	cpu_buffer->commit_page = cpu_buffer->head_page;
4141 
4142 	INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
4143 	INIT_LIST_HEAD(&cpu_buffer->new_pages);
4144 	local_set(&cpu_buffer->reader_page->write, 0);
4145 	local_set(&cpu_buffer->reader_page->entries, 0);
4146 	local_set(&cpu_buffer->reader_page->page->commit, 0);
4147 	cpu_buffer->reader_page->read = 0;
4148 
4149 	local_set(&cpu_buffer->entries_bytes, 0);
4150 	local_set(&cpu_buffer->overrun, 0);
4151 	local_set(&cpu_buffer->commit_overrun, 0);
4152 	local_set(&cpu_buffer->dropped_events, 0);
4153 	local_set(&cpu_buffer->entries, 0);
4154 	local_set(&cpu_buffer->committing, 0);
4155 	local_set(&cpu_buffer->commits, 0);
4156 	cpu_buffer->read = 0;
4157 	cpu_buffer->read_bytes = 0;
4158 
4159 	cpu_buffer->write_stamp = 0;
4160 	cpu_buffer->read_stamp = 0;
4161 
4162 	cpu_buffer->lost_events = 0;
4163 	cpu_buffer->last_overrun = 0;
4164 
4165 	rb_head_page_activate(cpu_buffer);
4166 }
4167 
4168 /**
4169  * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
4170  * @buffer: The ring buffer to reset a per cpu buffer of
4171  * @cpu: The CPU buffer to be reset
4172  */
4173 void ring_buffer_reset_cpu(struct ring_buffer *buffer, int cpu)
4174 {
4175 	struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4176 	unsigned long flags;
4177 
4178 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4179 		return;
4180 
4181 	atomic_inc(&buffer->resize_disabled);
4182 	atomic_inc(&cpu_buffer->record_disabled);
4183 
4184 	/* Make sure all commits have finished */
4185 	synchronize_sched();
4186 
4187 	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4188 
4189 	if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
4190 		goto out;
4191 
4192 	arch_spin_lock(&cpu_buffer->lock);
4193 
4194 	rb_reset_cpu(cpu_buffer);
4195 
4196 	arch_spin_unlock(&cpu_buffer->lock);
4197 
4198  out:
4199 	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4200 
4201 	atomic_dec(&cpu_buffer->record_disabled);
4202 	atomic_dec(&buffer->resize_disabled);
4203 }
4204 EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
4205 
4206 /**
4207  * ring_buffer_reset - reset a ring buffer
4208  * @buffer: The ring buffer to reset all cpu buffers
4209  */
4210 void ring_buffer_reset(struct ring_buffer *buffer)
4211 {
4212 	int cpu;
4213 
4214 	for_each_buffer_cpu(buffer, cpu)
4215 		ring_buffer_reset_cpu(buffer, cpu);
4216 }
4217 EXPORT_SYMBOL_GPL(ring_buffer_reset);
4218 
4219 /**
4220  * rind_buffer_empty - is the ring buffer empty?
4221  * @buffer: The ring buffer to test
4222  */
4223 int ring_buffer_empty(struct ring_buffer *buffer)
4224 {
4225 	struct ring_buffer_per_cpu *cpu_buffer;
4226 	unsigned long flags;
4227 	int dolock;
4228 	int cpu;
4229 	int ret;
4230 
4231 	dolock = rb_ok_to_lock();
4232 
4233 	/* yes this is racy, but if you don't like the race, lock the buffer */
4234 	for_each_buffer_cpu(buffer, cpu) {
4235 		cpu_buffer = buffer->buffers[cpu];
4236 		local_irq_save(flags);
4237 		if (dolock)
4238 			raw_spin_lock(&cpu_buffer->reader_lock);
4239 		ret = rb_per_cpu_empty(cpu_buffer);
4240 		if (dolock)
4241 			raw_spin_unlock(&cpu_buffer->reader_lock);
4242 		local_irq_restore(flags);
4243 
4244 		if (!ret)
4245 			return 0;
4246 	}
4247 
4248 	return 1;
4249 }
4250 EXPORT_SYMBOL_GPL(ring_buffer_empty);
4251 
4252 /**
4253  * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
4254  * @buffer: The ring buffer
4255  * @cpu: The CPU buffer to test
4256  */
4257 int ring_buffer_empty_cpu(struct ring_buffer *buffer, int cpu)
4258 {
4259 	struct ring_buffer_per_cpu *cpu_buffer;
4260 	unsigned long flags;
4261 	int dolock;
4262 	int ret;
4263 
4264 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4265 		return 1;
4266 
4267 	dolock = rb_ok_to_lock();
4268 
4269 	cpu_buffer = buffer->buffers[cpu];
4270 	local_irq_save(flags);
4271 	if (dolock)
4272 		raw_spin_lock(&cpu_buffer->reader_lock);
4273 	ret = rb_per_cpu_empty(cpu_buffer);
4274 	if (dolock)
4275 		raw_spin_unlock(&cpu_buffer->reader_lock);
4276 	local_irq_restore(flags);
4277 
4278 	return ret;
4279 }
4280 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
4281 
4282 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
4283 /**
4284  * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
4285  * @buffer_a: One buffer to swap with
4286  * @buffer_b: The other buffer to swap with
4287  *
4288  * This function is useful for tracers that want to take a "snapshot"
4289  * of a CPU buffer and has another back up buffer lying around.
4290  * it is expected that the tracer handles the cpu buffer not being
4291  * used at the moment.
4292  */
4293 int ring_buffer_swap_cpu(struct ring_buffer *buffer_a,
4294 			 struct ring_buffer *buffer_b, int cpu)
4295 {
4296 	struct ring_buffer_per_cpu *cpu_buffer_a;
4297 	struct ring_buffer_per_cpu *cpu_buffer_b;
4298 	int ret = -EINVAL;
4299 
4300 	if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
4301 	    !cpumask_test_cpu(cpu, buffer_b->cpumask))
4302 		goto out;
4303 
4304 	cpu_buffer_a = buffer_a->buffers[cpu];
4305 	cpu_buffer_b = buffer_b->buffers[cpu];
4306 
4307 	/* At least make sure the two buffers are somewhat the same */
4308 	if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
4309 		goto out;
4310 
4311 	ret = -EAGAIN;
4312 
4313 	if (ring_buffer_flags != RB_BUFFERS_ON)
4314 		goto out;
4315 
4316 	if (atomic_read(&buffer_a->record_disabled))
4317 		goto out;
4318 
4319 	if (atomic_read(&buffer_b->record_disabled))
4320 		goto out;
4321 
4322 	if (atomic_read(&cpu_buffer_a->record_disabled))
4323 		goto out;
4324 
4325 	if (atomic_read(&cpu_buffer_b->record_disabled))
4326 		goto out;
4327 
4328 	/*
4329 	 * We can't do a synchronize_sched here because this
4330 	 * function can be called in atomic context.
4331 	 * Normally this will be called from the same CPU as cpu.
4332 	 * If not it's up to the caller to protect this.
4333 	 */
4334 	atomic_inc(&cpu_buffer_a->record_disabled);
4335 	atomic_inc(&cpu_buffer_b->record_disabled);
4336 
4337 	ret = -EBUSY;
4338 	if (local_read(&cpu_buffer_a->committing))
4339 		goto out_dec;
4340 	if (local_read(&cpu_buffer_b->committing))
4341 		goto out_dec;
4342 
4343 	buffer_a->buffers[cpu] = cpu_buffer_b;
4344 	buffer_b->buffers[cpu] = cpu_buffer_a;
4345 
4346 	cpu_buffer_b->buffer = buffer_a;
4347 	cpu_buffer_a->buffer = buffer_b;
4348 
4349 	ret = 0;
4350 
4351 out_dec:
4352 	atomic_dec(&cpu_buffer_a->record_disabled);
4353 	atomic_dec(&cpu_buffer_b->record_disabled);
4354 out:
4355 	return ret;
4356 }
4357 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
4358 #endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
4359 
4360 /**
4361  * ring_buffer_alloc_read_page - allocate a page to read from buffer
4362  * @buffer: the buffer to allocate for.
4363  * @cpu: the cpu buffer to allocate.
4364  *
4365  * This function is used in conjunction with ring_buffer_read_page.
4366  * When reading a full page from the ring buffer, these functions
4367  * can be used to speed up the process. The calling function should
4368  * allocate a few pages first with this function. Then when it
4369  * needs to get pages from the ring buffer, it passes the result
4370  * of this function into ring_buffer_read_page, which will swap
4371  * the page that was allocated, with the read page of the buffer.
4372  *
4373  * Returns:
4374  *  The page allocated, or NULL on error.
4375  */
4376 void *ring_buffer_alloc_read_page(struct ring_buffer *buffer, int cpu)
4377 {
4378 	struct buffer_data_page *bpage;
4379 	struct page *page;
4380 
4381 	page = alloc_pages_node(cpu_to_node(cpu),
4382 				GFP_KERNEL | __GFP_NORETRY, 0);
4383 	if (!page)
4384 		return NULL;
4385 
4386 	bpage = page_address(page);
4387 
4388 	rb_init_page(bpage);
4389 
4390 	return bpage;
4391 }
4392 EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
4393 
4394 /**
4395  * ring_buffer_free_read_page - free an allocated read page
4396  * @buffer: the buffer the page was allocate for
4397  * @data: the page to free
4398  *
4399  * Free a page allocated from ring_buffer_alloc_read_page.
4400  */
4401 void ring_buffer_free_read_page(struct ring_buffer *buffer, void *data)
4402 {
4403 	free_page((unsigned long)data);
4404 }
4405 EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
4406 
4407 /**
4408  * ring_buffer_read_page - extract a page from the ring buffer
4409  * @buffer: buffer to extract from
4410  * @data_page: the page to use allocated from ring_buffer_alloc_read_page
4411  * @len: amount to extract
4412  * @cpu: the cpu of the buffer to extract
4413  * @full: should the extraction only happen when the page is full.
4414  *
4415  * This function will pull out a page from the ring buffer and consume it.
4416  * @data_page must be the address of the variable that was returned
4417  * from ring_buffer_alloc_read_page. This is because the page might be used
4418  * to swap with a page in the ring buffer.
4419  *
4420  * for example:
4421  *	rpage = ring_buffer_alloc_read_page(buffer, cpu);
4422  *	if (!rpage)
4423  *		return error;
4424  *	ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
4425  *	if (ret >= 0)
4426  *		process_page(rpage, ret);
4427  *
4428  * When @full is set, the function will not return true unless
4429  * the writer is off the reader page.
4430  *
4431  * Note: it is up to the calling functions to handle sleeps and wakeups.
4432  *  The ring buffer can be used anywhere in the kernel and can not
4433  *  blindly call wake_up. The layer that uses the ring buffer must be
4434  *  responsible for that.
4435  *
4436  * Returns:
4437  *  >=0 if data has been transferred, returns the offset of consumed data.
4438  *  <0 if no data has been transferred.
4439  */
4440 int ring_buffer_read_page(struct ring_buffer *buffer,
4441 			  void **data_page, size_t len, int cpu, int full)
4442 {
4443 	struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4444 	struct ring_buffer_event *event;
4445 	struct buffer_data_page *bpage;
4446 	struct buffer_page *reader;
4447 	unsigned long missed_events;
4448 	unsigned long flags;
4449 	unsigned int commit;
4450 	unsigned int read;
4451 	u64 save_timestamp;
4452 	int ret = -1;
4453 
4454 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4455 		goto out;
4456 
4457 	/*
4458 	 * If len is not big enough to hold the page header, then
4459 	 * we can not copy anything.
4460 	 */
4461 	if (len <= BUF_PAGE_HDR_SIZE)
4462 		goto out;
4463 
4464 	len -= BUF_PAGE_HDR_SIZE;
4465 
4466 	if (!data_page)
4467 		goto out;
4468 
4469 	bpage = *data_page;
4470 	if (!bpage)
4471 		goto out;
4472 
4473 	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4474 
4475 	reader = rb_get_reader_page(cpu_buffer);
4476 	if (!reader)
4477 		goto out_unlock;
4478 
4479 	event = rb_reader_event(cpu_buffer);
4480 
4481 	read = reader->read;
4482 	commit = rb_page_commit(reader);
4483 
4484 	/* Check if any events were dropped */
4485 	missed_events = cpu_buffer->lost_events;
4486 
4487 	/*
4488 	 * If this page has been partially read or
4489 	 * if len is not big enough to read the rest of the page or
4490 	 * a writer is still on the page, then
4491 	 * we must copy the data from the page to the buffer.
4492 	 * Otherwise, we can simply swap the page with the one passed in.
4493 	 */
4494 	if (read || (len < (commit - read)) ||
4495 	    cpu_buffer->reader_page == cpu_buffer->commit_page) {
4496 		struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
4497 		unsigned int rpos = read;
4498 		unsigned int pos = 0;
4499 		unsigned int size;
4500 
4501 		if (full)
4502 			goto out_unlock;
4503 
4504 		if (len > (commit - read))
4505 			len = (commit - read);
4506 
4507 		/* Always keep the time extend and data together */
4508 		size = rb_event_ts_length(event);
4509 
4510 		if (len < size)
4511 			goto out_unlock;
4512 
4513 		/* save the current timestamp, since the user will need it */
4514 		save_timestamp = cpu_buffer->read_stamp;
4515 
4516 		/* Need to copy one event at a time */
4517 		do {
4518 			/* We need the size of one event, because
4519 			 * rb_advance_reader only advances by one event,
4520 			 * whereas rb_event_ts_length may include the size of
4521 			 * one or two events.
4522 			 * We have already ensured there's enough space if this
4523 			 * is a time extend. */
4524 			size = rb_event_length(event);
4525 			memcpy(bpage->data + pos, rpage->data + rpos, size);
4526 
4527 			len -= size;
4528 
4529 			rb_advance_reader(cpu_buffer);
4530 			rpos = reader->read;
4531 			pos += size;
4532 
4533 			if (rpos >= commit)
4534 				break;
4535 
4536 			event = rb_reader_event(cpu_buffer);
4537 			/* Always keep the time extend and data together */
4538 			size = rb_event_ts_length(event);
4539 		} while (len >= size);
4540 
4541 		/* update bpage */
4542 		local_set(&bpage->commit, pos);
4543 		bpage->time_stamp = save_timestamp;
4544 
4545 		/* we copied everything to the beginning */
4546 		read = 0;
4547 	} else {
4548 		/* update the entry counter */
4549 		cpu_buffer->read += rb_page_entries(reader);
4550 		cpu_buffer->read_bytes += BUF_PAGE_SIZE;
4551 
4552 		/* swap the pages */
4553 		rb_init_page(bpage);
4554 		bpage = reader->page;
4555 		reader->page = *data_page;
4556 		local_set(&reader->write, 0);
4557 		local_set(&reader->entries, 0);
4558 		reader->read = 0;
4559 		*data_page = bpage;
4560 
4561 		/*
4562 		 * Use the real_end for the data size,
4563 		 * This gives us a chance to store the lost events
4564 		 * on the page.
4565 		 */
4566 		if (reader->real_end)
4567 			local_set(&bpage->commit, reader->real_end);
4568 	}
4569 	ret = read;
4570 
4571 	cpu_buffer->lost_events = 0;
4572 
4573 	commit = local_read(&bpage->commit);
4574 	/*
4575 	 * Set a flag in the commit field if we lost events
4576 	 */
4577 	if (missed_events) {
4578 		/* If there is room at the end of the page to save the
4579 		 * missed events, then record it there.
4580 		 */
4581 		if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
4582 			memcpy(&bpage->data[commit], &missed_events,
4583 			       sizeof(missed_events));
4584 			local_add(RB_MISSED_STORED, &bpage->commit);
4585 			commit += sizeof(missed_events);
4586 		}
4587 		local_add(RB_MISSED_EVENTS, &bpage->commit);
4588 	}
4589 
4590 	/*
4591 	 * This page may be off to user land. Zero it out here.
4592 	 */
4593 	if (commit < BUF_PAGE_SIZE)
4594 		memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
4595 
4596  out_unlock:
4597 	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4598 
4599  out:
4600 	return ret;
4601 }
4602 EXPORT_SYMBOL_GPL(ring_buffer_read_page);
4603 
4604 #ifdef CONFIG_HOTPLUG_CPU
4605 static int rb_cpu_notify(struct notifier_block *self,
4606 			 unsigned long action, void *hcpu)
4607 {
4608 	struct ring_buffer *buffer =
4609 		container_of(self, struct ring_buffer, cpu_notify);
4610 	long cpu = (long)hcpu;
4611 	int cpu_i, nr_pages_same;
4612 	unsigned int nr_pages;
4613 
4614 	switch (action) {
4615 	case CPU_UP_PREPARE:
4616 	case CPU_UP_PREPARE_FROZEN:
4617 		if (cpumask_test_cpu(cpu, buffer->cpumask))
4618 			return NOTIFY_OK;
4619 
4620 		nr_pages = 0;
4621 		nr_pages_same = 1;
4622 		/* check if all cpu sizes are same */
4623 		for_each_buffer_cpu(buffer, cpu_i) {
4624 			/* fill in the size from first enabled cpu */
4625 			if (nr_pages == 0)
4626 				nr_pages = buffer->buffers[cpu_i]->nr_pages;
4627 			if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
4628 				nr_pages_same = 0;
4629 				break;
4630 			}
4631 		}
4632 		/* allocate minimum pages, user can later expand it */
4633 		if (!nr_pages_same)
4634 			nr_pages = 2;
4635 		buffer->buffers[cpu] =
4636 			rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
4637 		if (!buffer->buffers[cpu]) {
4638 			WARN(1, "failed to allocate ring buffer on CPU %ld\n",
4639 			     cpu);
4640 			return NOTIFY_OK;
4641 		}
4642 		smp_wmb();
4643 		cpumask_set_cpu(cpu, buffer->cpumask);
4644 		break;
4645 	case CPU_DOWN_PREPARE:
4646 	case CPU_DOWN_PREPARE_FROZEN:
4647 		/*
4648 		 * Do nothing.
4649 		 *  If we were to free the buffer, then the user would
4650 		 *  lose any trace that was in the buffer.
4651 		 */
4652 		break;
4653 	default:
4654 		break;
4655 	}
4656 	return NOTIFY_OK;
4657 }
4658 #endif
4659 
4660 #ifdef CONFIG_RING_BUFFER_STARTUP_TEST
4661 /*
4662  * This is a basic integrity check of the ring buffer.
4663  * Late in the boot cycle this test will run when configured in.
4664  * It will kick off a thread per CPU that will go into a loop
4665  * writing to the per cpu ring buffer various sizes of data.
4666  * Some of the data will be large items, some small.
4667  *
4668  * Another thread is created that goes into a spin, sending out
4669  * IPIs to the other CPUs to also write into the ring buffer.
4670  * this is to test the nesting ability of the buffer.
4671  *
4672  * Basic stats are recorded and reported. If something in the
4673  * ring buffer should happen that's not expected, a big warning
4674  * is displayed and all ring buffers are disabled.
4675  */
4676 static struct task_struct *rb_threads[NR_CPUS] __initdata;
4677 
4678 struct rb_test_data {
4679 	struct ring_buffer	*buffer;
4680 	unsigned long		events;
4681 	unsigned long		bytes_written;
4682 	unsigned long		bytes_alloc;
4683 	unsigned long		bytes_dropped;
4684 	unsigned long		events_nested;
4685 	unsigned long		bytes_written_nested;
4686 	unsigned long		bytes_alloc_nested;
4687 	unsigned long		bytes_dropped_nested;
4688 	int			min_size_nested;
4689 	int			max_size_nested;
4690 	int			max_size;
4691 	int			min_size;
4692 	int			cpu;
4693 	int			cnt;
4694 };
4695 
4696 static struct rb_test_data rb_data[NR_CPUS] __initdata;
4697 
4698 /* 1 meg per cpu */
4699 #define RB_TEST_BUFFER_SIZE	1048576
4700 
4701 static char rb_string[] __initdata =
4702 	"abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
4703 	"?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
4704 	"!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
4705 
4706 static bool rb_test_started __initdata;
4707 
4708 struct rb_item {
4709 	int size;
4710 	char str[];
4711 };
4712 
4713 static __init int rb_write_something(struct rb_test_data *data, bool nested)
4714 {
4715 	struct ring_buffer_event *event;
4716 	struct rb_item *item;
4717 	bool started;
4718 	int event_len;
4719 	int size;
4720 	int len;
4721 	int cnt;
4722 
4723 	/* Have nested writes different that what is written */
4724 	cnt = data->cnt + (nested ? 27 : 0);
4725 
4726 	/* Multiply cnt by ~e, to make some unique increment */
4727 	size = (data->cnt * 68 / 25) % (sizeof(rb_string) - 1);
4728 
4729 	len = size + sizeof(struct rb_item);
4730 
4731 	started = rb_test_started;
4732 	/* read rb_test_started before checking buffer enabled */
4733 	smp_rmb();
4734 
4735 	event = ring_buffer_lock_reserve(data->buffer, len);
4736 	if (!event) {
4737 		/* Ignore dropped events before test starts. */
4738 		if (started) {
4739 			if (nested)
4740 				data->bytes_dropped += len;
4741 			else
4742 				data->bytes_dropped_nested += len;
4743 		}
4744 		return len;
4745 	}
4746 
4747 	event_len = ring_buffer_event_length(event);
4748 
4749 	if (RB_WARN_ON(data->buffer, event_len < len))
4750 		goto out;
4751 
4752 	item = ring_buffer_event_data(event);
4753 	item->size = size;
4754 	memcpy(item->str, rb_string, size);
4755 
4756 	if (nested) {
4757 		data->bytes_alloc_nested += event_len;
4758 		data->bytes_written_nested += len;
4759 		data->events_nested++;
4760 		if (!data->min_size_nested || len < data->min_size_nested)
4761 			data->min_size_nested = len;
4762 		if (len > data->max_size_nested)
4763 			data->max_size_nested = len;
4764 	} else {
4765 		data->bytes_alloc += event_len;
4766 		data->bytes_written += len;
4767 		data->events++;
4768 		if (!data->min_size || len < data->min_size)
4769 			data->max_size = len;
4770 		if (len > data->max_size)
4771 			data->max_size = len;
4772 	}
4773 
4774  out:
4775 	ring_buffer_unlock_commit(data->buffer, event);
4776 
4777 	return 0;
4778 }
4779 
4780 static __init int rb_test(void *arg)
4781 {
4782 	struct rb_test_data *data = arg;
4783 
4784 	while (!kthread_should_stop()) {
4785 		rb_write_something(data, false);
4786 		data->cnt++;
4787 
4788 		set_current_state(TASK_INTERRUPTIBLE);
4789 		/* Now sleep between a min of 100-300us and a max of 1ms */
4790 		usleep_range(((data->cnt % 3) + 1) * 100, 1000);
4791 	}
4792 
4793 	return 0;
4794 }
4795 
4796 static __init void rb_ipi(void *ignore)
4797 {
4798 	struct rb_test_data *data;
4799 	int cpu = smp_processor_id();
4800 
4801 	data = &rb_data[cpu];
4802 	rb_write_something(data, true);
4803 }
4804 
4805 static __init int rb_hammer_test(void *arg)
4806 {
4807 	while (!kthread_should_stop()) {
4808 
4809 		/* Send an IPI to all cpus to write data! */
4810 		smp_call_function(rb_ipi, NULL, 1);
4811 		/* No sleep, but for non preempt, let others run */
4812 		schedule();
4813 	}
4814 
4815 	return 0;
4816 }
4817 
4818 static __init int test_ringbuffer(void)
4819 {
4820 	struct task_struct *rb_hammer;
4821 	struct ring_buffer *buffer;
4822 	int cpu;
4823 	int ret = 0;
4824 
4825 	pr_info("Running ring buffer tests...\n");
4826 
4827 	buffer = ring_buffer_alloc(RB_TEST_BUFFER_SIZE, RB_FL_OVERWRITE);
4828 	if (WARN_ON(!buffer))
4829 		return 0;
4830 
4831 	/* Disable buffer so that threads can't write to it yet */
4832 	ring_buffer_record_off(buffer);
4833 
4834 	for_each_online_cpu(cpu) {
4835 		rb_data[cpu].buffer = buffer;
4836 		rb_data[cpu].cpu = cpu;
4837 		rb_data[cpu].cnt = cpu;
4838 		rb_threads[cpu] = kthread_create(rb_test, &rb_data[cpu],
4839 						 "rbtester/%d", cpu);
4840 		if (WARN_ON(!rb_threads[cpu])) {
4841 			pr_cont("FAILED\n");
4842 			ret = -1;
4843 			goto out_free;
4844 		}
4845 
4846 		kthread_bind(rb_threads[cpu], cpu);
4847  		wake_up_process(rb_threads[cpu]);
4848 	}
4849 
4850 	/* Now create the rb hammer! */
4851 	rb_hammer = kthread_run(rb_hammer_test, NULL, "rbhammer");
4852 	if (WARN_ON(!rb_hammer)) {
4853 		pr_cont("FAILED\n");
4854 		ret = -1;
4855 		goto out_free;
4856 	}
4857 
4858 	ring_buffer_record_on(buffer);
4859 	/*
4860 	 * Show buffer is enabled before setting rb_test_started.
4861 	 * Yes there's a small race window where events could be
4862 	 * dropped and the thread wont catch it. But when a ring
4863 	 * buffer gets enabled, there will always be some kind of
4864 	 * delay before other CPUs see it. Thus, we don't care about
4865 	 * those dropped events. We care about events dropped after
4866 	 * the threads see that the buffer is active.
4867 	 */
4868 	smp_wmb();
4869 	rb_test_started = true;
4870 
4871 	set_current_state(TASK_INTERRUPTIBLE);
4872 	/* Just run for 10 seconds */;
4873 	schedule_timeout(10 * HZ);
4874 
4875 	kthread_stop(rb_hammer);
4876 
4877  out_free:
4878 	for_each_online_cpu(cpu) {
4879 		if (!rb_threads[cpu])
4880 			break;
4881 		kthread_stop(rb_threads[cpu]);
4882 	}
4883 	if (ret) {
4884 		ring_buffer_free(buffer);
4885 		return ret;
4886 	}
4887 
4888 	/* Report! */
4889 	pr_info("finished\n");
4890 	for_each_online_cpu(cpu) {
4891 		struct ring_buffer_event *event;
4892 		struct rb_test_data *data = &rb_data[cpu];
4893 		struct rb_item *item;
4894 		unsigned long total_events;
4895 		unsigned long total_dropped;
4896 		unsigned long total_written;
4897 		unsigned long total_alloc;
4898 		unsigned long total_read = 0;
4899 		unsigned long total_size = 0;
4900 		unsigned long total_len = 0;
4901 		unsigned long total_lost = 0;
4902 		unsigned long lost;
4903 		int big_event_size;
4904 		int small_event_size;
4905 
4906 		ret = -1;
4907 
4908 		total_events = data->events + data->events_nested;
4909 		total_written = data->bytes_written + data->bytes_written_nested;
4910 		total_alloc = data->bytes_alloc + data->bytes_alloc_nested;
4911 		total_dropped = data->bytes_dropped + data->bytes_dropped_nested;
4912 
4913 		big_event_size = data->max_size + data->max_size_nested;
4914 		small_event_size = data->min_size + data->min_size_nested;
4915 
4916 		pr_info("CPU %d:\n", cpu);
4917 		pr_info("              events:    %ld\n", total_events);
4918 		pr_info("       dropped bytes:    %ld\n", total_dropped);
4919 		pr_info("       alloced bytes:    %ld\n", total_alloc);
4920 		pr_info("       written bytes:    %ld\n", total_written);
4921 		pr_info("       biggest event:    %d\n", big_event_size);
4922 		pr_info("      smallest event:    %d\n", small_event_size);
4923 
4924 		if (RB_WARN_ON(buffer, total_dropped))
4925 			break;
4926 
4927 		ret = 0;
4928 
4929 		while ((event = ring_buffer_consume(buffer, cpu, NULL, &lost))) {
4930 			total_lost += lost;
4931 			item = ring_buffer_event_data(event);
4932 			total_len += ring_buffer_event_length(event);
4933 			total_size += item->size + sizeof(struct rb_item);
4934 			if (memcmp(&item->str[0], rb_string, item->size) != 0) {
4935 				pr_info("FAILED!\n");
4936 				pr_info("buffer had: %.*s\n", item->size, item->str);
4937 				pr_info("expected:   %.*s\n", item->size, rb_string);
4938 				RB_WARN_ON(buffer, 1);
4939 				ret = -1;
4940 				break;
4941 			}
4942 			total_read++;
4943 		}
4944 		if (ret)
4945 			break;
4946 
4947 		ret = -1;
4948 
4949 		pr_info("         read events:   %ld\n", total_read);
4950 		pr_info("         lost events:   %ld\n", total_lost);
4951 		pr_info("        total events:   %ld\n", total_lost + total_read);
4952 		pr_info("  recorded len bytes:   %ld\n", total_len);
4953 		pr_info(" recorded size bytes:   %ld\n", total_size);
4954 		if (total_lost)
4955 			pr_info(" With dropped events, record len and size may not match\n"
4956 				" alloced and written from above\n");
4957 		if (!total_lost) {
4958 			if (RB_WARN_ON(buffer, total_len != total_alloc ||
4959 				       total_size != total_written))
4960 				break;
4961 		}
4962 		if (RB_WARN_ON(buffer, total_lost + total_read != total_events))
4963 			break;
4964 
4965 		ret = 0;
4966 	}
4967 	if (!ret)
4968 		pr_info("Ring buffer PASSED!\n");
4969 
4970 	ring_buffer_free(buffer);
4971 	return 0;
4972 }
4973 
4974 late_initcall(test_ringbuffer);
4975 #endif /* CONFIG_RING_BUFFER_STARTUP_TEST */
4976