xref: /openbmc/linux/kernel/trace/ring_buffer.c (revision b34e08d5)
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 void 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 		cpu_buffer = buffer->buffers[cpu];
561 		work = &cpu_buffer->irq_work;
562 	}
563 
564 
565 	prepare_to_wait(&work->waiters, &wait, TASK_INTERRUPTIBLE);
566 
567 	/*
568 	 * The events can happen in critical sections where
569 	 * checking a work queue can cause deadlocks.
570 	 * After adding a task to the queue, this flag is set
571 	 * only to notify events to try to wake up the queue
572 	 * using irq_work.
573 	 *
574 	 * We don't clear it even if the buffer is no longer
575 	 * empty. The flag only causes the next event to run
576 	 * irq_work to do the work queue wake up. The worse
577 	 * that can happen if we race with !trace_empty() is that
578 	 * an event will cause an irq_work to try to wake up
579 	 * an empty queue.
580 	 *
581 	 * There's no reason to protect this flag either, as
582 	 * the work queue and irq_work logic will do the necessary
583 	 * synchronization for the wake ups. The only thing
584 	 * that is necessary is that the wake up happens after
585 	 * a task has been queued. It's OK for spurious wake ups.
586 	 */
587 	work->waiters_pending = true;
588 
589 	if ((cpu == RING_BUFFER_ALL_CPUS && ring_buffer_empty(buffer)) ||
590 	    (cpu != RING_BUFFER_ALL_CPUS && ring_buffer_empty_cpu(buffer, cpu)))
591 		schedule();
592 
593 	finish_wait(&work->waiters, &wait);
594 }
595 
596 /**
597  * ring_buffer_poll_wait - poll on buffer input
598  * @buffer: buffer to wait on
599  * @cpu: the cpu buffer to wait on
600  * @filp: the file descriptor
601  * @poll_table: The poll descriptor
602  *
603  * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
604  * as data is added to any of the @buffer's cpu buffers. Otherwise
605  * it will wait for data to be added to a specific cpu buffer.
606  *
607  * Returns POLLIN | POLLRDNORM if data exists in the buffers,
608  * zero otherwise.
609  */
610 int ring_buffer_poll_wait(struct ring_buffer *buffer, int cpu,
611 			  struct file *filp, poll_table *poll_table)
612 {
613 	struct ring_buffer_per_cpu *cpu_buffer;
614 	struct rb_irq_work *work;
615 
616 	if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
617 	    (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
618 		return POLLIN | POLLRDNORM;
619 
620 	if (cpu == RING_BUFFER_ALL_CPUS)
621 		work = &buffer->irq_work;
622 	else {
623 		if (!cpumask_test_cpu(cpu, buffer->cpumask))
624 			return -EINVAL;
625 
626 		cpu_buffer = buffer->buffers[cpu];
627 		work = &cpu_buffer->irq_work;
628 	}
629 
630 	work->waiters_pending = true;
631 	poll_wait(filp, &work->waiters, poll_table);
632 
633 	if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
634 	    (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
635 		return POLLIN | POLLRDNORM;
636 	return 0;
637 }
638 
639 /* buffer may be either ring_buffer or ring_buffer_per_cpu */
640 #define RB_WARN_ON(b, cond)						\
641 	({								\
642 		int _____ret = unlikely(cond);				\
643 		if (_____ret) {						\
644 			if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
645 				struct ring_buffer_per_cpu *__b =	\
646 					(void *)b;			\
647 				atomic_inc(&__b->buffer->record_disabled); \
648 			} else						\
649 				atomic_inc(&b->record_disabled);	\
650 			WARN_ON(1);					\
651 		}							\
652 		_____ret;						\
653 	})
654 
655 /* Up this if you want to test the TIME_EXTENTS and normalization */
656 #define DEBUG_SHIFT 0
657 
658 static inline u64 rb_time_stamp(struct ring_buffer *buffer)
659 {
660 	/* shift to debug/test normalization and TIME_EXTENTS */
661 	return buffer->clock() << DEBUG_SHIFT;
662 }
663 
664 u64 ring_buffer_time_stamp(struct ring_buffer *buffer, int cpu)
665 {
666 	u64 time;
667 
668 	preempt_disable_notrace();
669 	time = rb_time_stamp(buffer);
670 	preempt_enable_no_resched_notrace();
671 
672 	return time;
673 }
674 EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
675 
676 void ring_buffer_normalize_time_stamp(struct ring_buffer *buffer,
677 				      int cpu, u64 *ts)
678 {
679 	/* Just stupid testing the normalize function and deltas */
680 	*ts >>= DEBUG_SHIFT;
681 }
682 EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
683 
684 /*
685  * Making the ring buffer lockless makes things tricky.
686  * Although writes only happen on the CPU that they are on,
687  * and they only need to worry about interrupts. Reads can
688  * happen on any CPU.
689  *
690  * The reader page is always off the ring buffer, but when the
691  * reader finishes with a page, it needs to swap its page with
692  * a new one from the buffer. The reader needs to take from
693  * the head (writes go to the tail). But if a writer is in overwrite
694  * mode and wraps, it must push the head page forward.
695  *
696  * Here lies the problem.
697  *
698  * The reader must be careful to replace only the head page, and
699  * not another one. As described at the top of the file in the
700  * ASCII art, the reader sets its old page to point to the next
701  * page after head. It then sets the page after head to point to
702  * the old reader page. But if the writer moves the head page
703  * during this operation, the reader could end up with the tail.
704  *
705  * We use cmpxchg to help prevent this race. We also do something
706  * special with the page before head. We set the LSB to 1.
707  *
708  * When the writer must push the page forward, it will clear the
709  * bit that points to the head page, move the head, and then set
710  * the bit that points to the new head page.
711  *
712  * We also don't want an interrupt coming in and moving the head
713  * page on another writer. Thus we use the second LSB to catch
714  * that too. Thus:
715  *
716  * head->list->prev->next        bit 1          bit 0
717  *                              -------        -------
718  * Normal page                     0              0
719  * Points to head page             0              1
720  * New head page                   1              0
721  *
722  * Note we can not trust the prev pointer of the head page, because:
723  *
724  * +----+       +-----+        +-----+
725  * |    |------>|  T  |---X--->|  N  |
726  * |    |<------|     |        |     |
727  * +----+       +-----+        +-----+
728  *   ^                           ^ |
729  *   |          +-----+          | |
730  *   +----------|  R  |----------+ |
731  *              |     |<-----------+
732  *              +-----+
733  *
734  * Key:  ---X-->  HEAD flag set in pointer
735  *         T      Tail page
736  *         R      Reader page
737  *         N      Next page
738  *
739  * (see __rb_reserve_next() to see where this happens)
740  *
741  *  What the above shows is that the reader just swapped out
742  *  the reader page with a page in the buffer, but before it
743  *  could make the new header point back to the new page added
744  *  it was preempted by a writer. The writer moved forward onto
745  *  the new page added by the reader and is about to move forward
746  *  again.
747  *
748  *  You can see, it is legitimate for the previous pointer of
749  *  the head (or any page) not to point back to itself. But only
750  *  temporarially.
751  */
752 
753 #define RB_PAGE_NORMAL		0UL
754 #define RB_PAGE_HEAD		1UL
755 #define RB_PAGE_UPDATE		2UL
756 
757 
758 #define RB_FLAG_MASK		3UL
759 
760 /* PAGE_MOVED is not part of the mask */
761 #define RB_PAGE_MOVED		4UL
762 
763 /*
764  * rb_list_head - remove any bit
765  */
766 static struct list_head *rb_list_head(struct list_head *list)
767 {
768 	unsigned long val = (unsigned long)list;
769 
770 	return (struct list_head *)(val & ~RB_FLAG_MASK);
771 }
772 
773 /*
774  * rb_is_head_page - test if the given page is the head page
775  *
776  * Because the reader may move the head_page pointer, we can
777  * not trust what the head page is (it may be pointing to
778  * the reader page). But if the next page is a header page,
779  * its flags will be non zero.
780  */
781 static inline int
782 rb_is_head_page(struct ring_buffer_per_cpu *cpu_buffer,
783 		struct buffer_page *page, struct list_head *list)
784 {
785 	unsigned long val;
786 
787 	val = (unsigned long)list->next;
788 
789 	if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
790 		return RB_PAGE_MOVED;
791 
792 	return val & RB_FLAG_MASK;
793 }
794 
795 /*
796  * rb_is_reader_page
797  *
798  * The unique thing about the reader page, is that, if the
799  * writer is ever on it, the previous pointer never points
800  * back to the reader page.
801  */
802 static int rb_is_reader_page(struct buffer_page *page)
803 {
804 	struct list_head *list = page->list.prev;
805 
806 	return rb_list_head(list->next) != &page->list;
807 }
808 
809 /*
810  * rb_set_list_to_head - set a list_head to be pointing to head.
811  */
812 static void rb_set_list_to_head(struct ring_buffer_per_cpu *cpu_buffer,
813 				struct list_head *list)
814 {
815 	unsigned long *ptr;
816 
817 	ptr = (unsigned long *)&list->next;
818 	*ptr |= RB_PAGE_HEAD;
819 	*ptr &= ~RB_PAGE_UPDATE;
820 }
821 
822 /*
823  * rb_head_page_activate - sets up head page
824  */
825 static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
826 {
827 	struct buffer_page *head;
828 
829 	head = cpu_buffer->head_page;
830 	if (!head)
831 		return;
832 
833 	/*
834 	 * Set the previous list pointer to have the HEAD flag.
835 	 */
836 	rb_set_list_to_head(cpu_buffer, head->list.prev);
837 }
838 
839 static void rb_list_head_clear(struct list_head *list)
840 {
841 	unsigned long *ptr = (unsigned long *)&list->next;
842 
843 	*ptr &= ~RB_FLAG_MASK;
844 }
845 
846 /*
847  * rb_head_page_dactivate - clears head page ptr (for free list)
848  */
849 static void
850 rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
851 {
852 	struct list_head *hd;
853 
854 	/* Go through the whole list and clear any pointers found. */
855 	rb_list_head_clear(cpu_buffer->pages);
856 
857 	list_for_each(hd, cpu_buffer->pages)
858 		rb_list_head_clear(hd);
859 }
860 
861 static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
862 			    struct buffer_page *head,
863 			    struct buffer_page *prev,
864 			    int old_flag, int new_flag)
865 {
866 	struct list_head *list;
867 	unsigned long val = (unsigned long)&head->list;
868 	unsigned long ret;
869 
870 	list = &prev->list;
871 
872 	val &= ~RB_FLAG_MASK;
873 
874 	ret = cmpxchg((unsigned long *)&list->next,
875 		      val | old_flag, val | new_flag);
876 
877 	/* check if the reader took the page */
878 	if ((ret & ~RB_FLAG_MASK) != val)
879 		return RB_PAGE_MOVED;
880 
881 	return ret & RB_FLAG_MASK;
882 }
883 
884 static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
885 				   struct buffer_page *head,
886 				   struct buffer_page *prev,
887 				   int old_flag)
888 {
889 	return rb_head_page_set(cpu_buffer, head, prev,
890 				old_flag, RB_PAGE_UPDATE);
891 }
892 
893 static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
894 				 struct buffer_page *head,
895 				 struct buffer_page *prev,
896 				 int old_flag)
897 {
898 	return rb_head_page_set(cpu_buffer, head, prev,
899 				old_flag, RB_PAGE_HEAD);
900 }
901 
902 static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
903 				   struct buffer_page *head,
904 				   struct buffer_page *prev,
905 				   int old_flag)
906 {
907 	return rb_head_page_set(cpu_buffer, head, prev,
908 				old_flag, RB_PAGE_NORMAL);
909 }
910 
911 static inline void rb_inc_page(struct ring_buffer_per_cpu *cpu_buffer,
912 			       struct buffer_page **bpage)
913 {
914 	struct list_head *p = rb_list_head((*bpage)->list.next);
915 
916 	*bpage = list_entry(p, struct buffer_page, list);
917 }
918 
919 static struct buffer_page *
920 rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
921 {
922 	struct buffer_page *head;
923 	struct buffer_page *page;
924 	struct list_head *list;
925 	int i;
926 
927 	if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
928 		return NULL;
929 
930 	/* sanity check */
931 	list = cpu_buffer->pages;
932 	if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
933 		return NULL;
934 
935 	page = head = cpu_buffer->head_page;
936 	/*
937 	 * It is possible that the writer moves the header behind
938 	 * where we started, and we miss in one loop.
939 	 * A second loop should grab the header, but we'll do
940 	 * three loops just because I'm paranoid.
941 	 */
942 	for (i = 0; i < 3; i++) {
943 		do {
944 			if (rb_is_head_page(cpu_buffer, page, page->list.prev)) {
945 				cpu_buffer->head_page = page;
946 				return page;
947 			}
948 			rb_inc_page(cpu_buffer, &page);
949 		} while (page != head);
950 	}
951 
952 	RB_WARN_ON(cpu_buffer, 1);
953 
954 	return NULL;
955 }
956 
957 static int rb_head_page_replace(struct buffer_page *old,
958 				struct buffer_page *new)
959 {
960 	unsigned long *ptr = (unsigned long *)&old->list.prev->next;
961 	unsigned long val;
962 	unsigned long ret;
963 
964 	val = *ptr & ~RB_FLAG_MASK;
965 	val |= RB_PAGE_HEAD;
966 
967 	ret = cmpxchg(ptr, val, (unsigned long)&new->list);
968 
969 	return ret == val;
970 }
971 
972 /*
973  * rb_tail_page_update - move the tail page forward
974  *
975  * Returns 1 if moved tail page, 0 if someone else did.
976  */
977 static int rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
978 			       struct buffer_page *tail_page,
979 			       struct buffer_page *next_page)
980 {
981 	struct buffer_page *old_tail;
982 	unsigned long old_entries;
983 	unsigned long old_write;
984 	int ret = 0;
985 
986 	/*
987 	 * The tail page now needs to be moved forward.
988 	 *
989 	 * We need to reset the tail page, but without messing
990 	 * with possible erasing of data brought in by interrupts
991 	 * that have moved the tail page and are currently on it.
992 	 *
993 	 * We add a counter to the write field to denote this.
994 	 */
995 	old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
996 	old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
997 
998 	/*
999 	 * Just make sure we have seen our old_write and synchronize
1000 	 * with any interrupts that come in.
1001 	 */
1002 	barrier();
1003 
1004 	/*
1005 	 * If the tail page is still the same as what we think
1006 	 * it is, then it is up to us to update the tail
1007 	 * pointer.
1008 	 */
1009 	if (tail_page == cpu_buffer->tail_page) {
1010 		/* Zero the write counter */
1011 		unsigned long val = old_write & ~RB_WRITE_MASK;
1012 		unsigned long eval = old_entries & ~RB_WRITE_MASK;
1013 
1014 		/*
1015 		 * This will only succeed if an interrupt did
1016 		 * not come in and change it. In which case, we
1017 		 * do not want to modify it.
1018 		 *
1019 		 * We add (void) to let the compiler know that we do not care
1020 		 * about the return value of these functions. We use the
1021 		 * cmpxchg to only update if an interrupt did not already
1022 		 * do it for us. If the cmpxchg fails, we don't care.
1023 		 */
1024 		(void)local_cmpxchg(&next_page->write, old_write, val);
1025 		(void)local_cmpxchg(&next_page->entries, old_entries, eval);
1026 
1027 		/*
1028 		 * No need to worry about races with clearing out the commit.
1029 		 * it only can increment when a commit takes place. But that
1030 		 * only happens in the outer most nested commit.
1031 		 */
1032 		local_set(&next_page->page->commit, 0);
1033 
1034 		old_tail = cmpxchg(&cpu_buffer->tail_page,
1035 				   tail_page, next_page);
1036 
1037 		if (old_tail == tail_page)
1038 			ret = 1;
1039 	}
1040 
1041 	return ret;
1042 }
1043 
1044 static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
1045 			  struct buffer_page *bpage)
1046 {
1047 	unsigned long val = (unsigned long)bpage;
1048 
1049 	if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK))
1050 		return 1;
1051 
1052 	return 0;
1053 }
1054 
1055 /**
1056  * rb_check_list - make sure a pointer to a list has the last bits zero
1057  */
1058 static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer,
1059 			 struct list_head *list)
1060 {
1061 	if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev))
1062 		return 1;
1063 	if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next))
1064 		return 1;
1065 	return 0;
1066 }
1067 
1068 /**
1069  * rb_check_pages - integrity check of buffer pages
1070  * @cpu_buffer: CPU buffer with pages to test
1071  *
1072  * As a safety measure we check to make sure the data pages have not
1073  * been corrupted.
1074  */
1075 static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
1076 {
1077 	struct list_head *head = cpu_buffer->pages;
1078 	struct buffer_page *bpage, *tmp;
1079 
1080 	/* Reset the head page if it exists */
1081 	if (cpu_buffer->head_page)
1082 		rb_set_head_page(cpu_buffer);
1083 
1084 	rb_head_page_deactivate(cpu_buffer);
1085 
1086 	if (RB_WARN_ON(cpu_buffer, head->next->prev != head))
1087 		return -1;
1088 	if (RB_WARN_ON(cpu_buffer, head->prev->next != head))
1089 		return -1;
1090 
1091 	if (rb_check_list(cpu_buffer, head))
1092 		return -1;
1093 
1094 	list_for_each_entry_safe(bpage, tmp, head, list) {
1095 		if (RB_WARN_ON(cpu_buffer,
1096 			       bpage->list.next->prev != &bpage->list))
1097 			return -1;
1098 		if (RB_WARN_ON(cpu_buffer,
1099 			       bpage->list.prev->next != &bpage->list))
1100 			return -1;
1101 		if (rb_check_list(cpu_buffer, &bpage->list))
1102 			return -1;
1103 	}
1104 
1105 	rb_head_page_activate(cpu_buffer);
1106 
1107 	return 0;
1108 }
1109 
1110 static int __rb_allocate_pages(int nr_pages, struct list_head *pages, int cpu)
1111 {
1112 	int i;
1113 	struct buffer_page *bpage, *tmp;
1114 
1115 	for (i = 0; i < nr_pages; i++) {
1116 		struct page *page;
1117 		/*
1118 		 * __GFP_NORETRY flag makes sure that the allocation fails
1119 		 * gracefully without invoking oom-killer and the system is
1120 		 * not destabilized.
1121 		 */
1122 		bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1123 				    GFP_KERNEL | __GFP_NORETRY,
1124 				    cpu_to_node(cpu));
1125 		if (!bpage)
1126 			goto free_pages;
1127 
1128 		list_add(&bpage->list, pages);
1129 
1130 		page = alloc_pages_node(cpu_to_node(cpu),
1131 					GFP_KERNEL | __GFP_NORETRY, 0);
1132 		if (!page)
1133 			goto free_pages;
1134 		bpage->page = page_address(page);
1135 		rb_init_page(bpage->page);
1136 	}
1137 
1138 	return 0;
1139 
1140 free_pages:
1141 	list_for_each_entry_safe(bpage, tmp, pages, list) {
1142 		list_del_init(&bpage->list);
1143 		free_buffer_page(bpage);
1144 	}
1145 
1146 	return -ENOMEM;
1147 }
1148 
1149 static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1150 			     unsigned nr_pages)
1151 {
1152 	LIST_HEAD(pages);
1153 
1154 	WARN_ON(!nr_pages);
1155 
1156 	if (__rb_allocate_pages(nr_pages, &pages, cpu_buffer->cpu))
1157 		return -ENOMEM;
1158 
1159 	/*
1160 	 * The ring buffer page list is a circular list that does not
1161 	 * start and end with a list head. All page list items point to
1162 	 * other pages.
1163 	 */
1164 	cpu_buffer->pages = pages.next;
1165 	list_del(&pages);
1166 
1167 	cpu_buffer->nr_pages = nr_pages;
1168 
1169 	rb_check_pages(cpu_buffer);
1170 
1171 	return 0;
1172 }
1173 
1174 static struct ring_buffer_per_cpu *
1175 rb_allocate_cpu_buffer(struct ring_buffer *buffer, int nr_pages, int cpu)
1176 {
1177 	struct ring_buffer_per_cpu *cpu_buffer;
1178 	struct buffer_page *bpage;
1179 	struct page *page;
1180 	int ret;
1181 
1182 	cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1183 				  GFP_KERNEL, cpu_to_node(cpu));
1184 	if (!cpu_buffer)
1185 		return NULL;
1186 
1187 	cpu_buffer->cpu = cpu;
1188 	cpu_buffer->buffer = buffer;
1189 	raw_spin_lock_init(&cpu_buffer->reader_lock);
1190 	lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1191 	cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1192 	INIT_WORK(&cpu_buffer->update_pages_work, update_pages_handler);
1193 	init_completion(&cpu_buffer->update_done);
1194 	init_irq_work(&cpu_buffer->irq_work.work, rb_wake_up_waiters);
1195 	init_waitqueue_head(&cpu_buffer->irq_work.waiters);
1196 
1197 	bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1198 			    GFP_KERNEL, cpu_to_node(cpu));
1199 	if (!bpage)
1200 		goto fail_free_buffer;
1201 
1202 	rb_check_bpage(cpu_buffer, bpage);
1203 
1204 	cpu_buffer->reader_page = bpage;
1205 	page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0);
1206 	if (!page)
1207 		goto fail_free_reader;
1208 	bpage->page = page_address(page);
1209 	rb_init_page(bpage->page);
1210 
1211 	INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1212 	INIT_LIST_HEAD(&cpu_buffer->new_pages);
1213 
1214 	ret = rb_allocate_pages(cpu_buffer, nr_pages);
1215 	if (ret < 0)
1216 		goto fail_free_reader;
1217 
1218 	cpu_buffer->head_page
1219 		= list_entry(cpu_buffer->pages, struct buffer_page, list);
1220 	cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1221 
1222 	rb_head_page_activate(cpu_buffer);
1223 
1224 	return cpu_buffer;
1225 
1226  fail_free_reader:
1227 	free_buffer_page(cpu_buffer->reader_page);
1228 
1229  fail_free_buffer:
1230 	kfree(cpu_buffer);
1231 	return NULL;
1232 }
1233 
1234 static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1235 {
1236 	struct list_head *head = cpu_buffer->pages;
1237 	struct buffer_page *bpage, *tmp;
1238 
1239 	free_buffer_page(cpu_buffer->reader_page);
1240 
1241 	rb_head_page_deactivate(cpu_buffer);
1242 
1243 	if (head) {
1244 		list_for_each_entry_safe(bpage, tmp, head, list) {
1245 			list_del_init(&bpage->list);
1246 			free_buffer_page(bpage);
1247 		}
1248 		bpage = list_entry(head, struct buffer_page, list);
1249 		free_buffer_page(bpage);
1250 	}
1251 
1252 	kfree(cpu_buffer);
1253 }
1254 
1255 #ifdef CONFIG_HOTPLUG_CPU
1256 static int rb_cpu_notify(struct notifier_block *self,
1257 			 unsigned long action, void *hcpu);
1258 #endif
1259 
1260 /**
1261  * __ring_buffer_alloc - allocate a new ring_buffer
1262  * @size: the size in bytes per cpu that is needed.
1263  * @flags: attributes to set for the ring buffer.
1264  *
1265  * Currently the only flag that is available is the RB_FL_OVERWRITE
1266  * flag. This flag means that the buffer will overwrite old data
1267  * when the buffer wraps. If this flag is not set, the buffer will
1268  * drop data when the tail hits the head.
1269  */
1270 struct ring_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1271 					struct lock_class_key *key)
1272 {
1273 	struct ring_buffer *buffer;
1274 	int bsize;
1275 	int cpu, nr_pages;
1276 
1277 	/* keep it in its own cache line */
1278 	buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1279 			 GFP_KERNEL);
1280 	if (!buffer)
1281 		return NULL;
1282 
1283 	if (!alloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1284 		goto fail_free_buffer;
1285 
1286 	nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1287 	buffer->flags = flags;
1288 	buffer->clock = trace_clock_local;
1289 	buffer->reader_lock_key = key;
1290 
1291 	init_irq_work(&buffer->irq_work.work, rb_wake_up_waiters);
1292 	init_waitqueue_head(&buffer->irq_work.waiters);
1293 
1294 	/* need at least two pages */
1295 	if (nr_pages < 2)
1296 		nr_pages = 2;
1297 
1298 	/*
1299 	 * In case of non-hotplug cpu, if the ring-buffer is allocated
1300 	 * in early initcall, it will not be notified of secondary cpus.
1301 	 * In that off case, we need to allocate for all possible cpus.
1302 	 */
1303 #ifdef CONFIG_HOTPLUG_CPU
1304 	cpu_notifier_register_begin();
1305 	cpumask_copy(buffer->cpumask, cpu_online_mask);
1306 #else
1307 	cpumask_copy(buffer->cpumask, cpu_possible_mask);
1308 #endif
1309 	buffer->cpus = nr_cpu_ids;
1310 
1311 	bsize = sizeof(void *) * nr_cpu_ids;
1312 	buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1313 				  GFP_KERNEL);
1314 	if (!buffer->buffers)
1315 		goto fail_free_cpumask;
1316 
1317 	for_each_buffer_cpu(buffer, cpu) {
1318 		buffer->buffers[cpu] =
1319 			rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
1320 		if (!buffer->buffers[cpu])
1321 			goto fail_free_buffers;
1322 	}
1323 
1324 #ifdef CONFIG_HOTPLUG_CPU
1325 	buffer->cpu_notify.notifier_call = rb_cpu_notify;
1326 	buffer->cpu_notify.priority = 0;
1327 	__register_cpu_notifier(&buffer->cpu_notify);
1328 	cpu_notifier_register_done();
1329 #endif
1330 
1331 	mutex_init(&buffer->mutex);
1332 
1333 	return buffer;
1334 
1335  fail_free_buffers:
1336 	for_each_buffer_cpu(buffer, cpu) {
1337 		if (buffer->buffers[cpu])
1338 			rb_free_cpu_buffer(buffer->buffers[cpu]);
1339 	}
1340 	kfree(buffer->buffers);
1341 
1342  fail_free_cpumask:
1343 	free_cpumask_var(buffer->cpumask);
1344 #ifdef CONFIG_HOTPLUG_CPU
1345 	cpu_notifier_register_done();
1346 #endif
1347 
1348  fail_free_buffer:
1349 	kfree(buffer);
1350 	return NULL;
1351 }
1352 EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1353 
1354 /**
1355  * ring_buffer_free - free a ring buffer.
1356  * @buffer: the buffer to free.
1357  */
1358 void
1359 ring_buffer_free(struct ring_buffer *buffer)
1360 {
1361 	int cpu;
1362 
1363 #ifdef CONFIG_HOTPLUG_CPU
1364 	cpu_notifier_register_begin();
1365 	__unregister_cpu_notifier(&buffer->cpu_notify);
1366 #endif
1367 
1368 	for_each_buffer_cpu(buffer, cpu)
1369 		rb_free_cpu_buffer(buffer->buffers[cpu]);
1370 
1371 #ifdef CONFIG_HOTPLUG_CPU
1372 	cpu_notifier_register_done();
1373 #endif
1374 
1375 	kfree(buffer->buffers);
1376 	free_cpumask_var(buffer->cpumask);
1377 
1378 	kfree(buffer);
1379 }
1380 EXPORT_SYMBOL_GPL(ring_buffer_free);
1381 
1382 void ring_buffer_set_clock(struct ring_buffer *buffer,
1383 			   u64 (*clock)(void))
1384 {
1385 	buffer->clock = clock;
1386 }
1387 
1388 static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1389 
1390 static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1391 {
1392 	return local_read(&bpage->entries) & RB_WRITE_MASK;
1393 }
1394 
1395 static inline unsigned long rb_page_write(struct buffer_page *bpage)
1396 {
1397 	return local_read(&bpage->write) & RB_WRITE_MASK;
1398 }
1399 
1400 static int
1401 rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned int nr_pages)
1402 {
1403 	struct list_head *tail_page, *to_remove, *next_page;
1404 	struct buffer_page *to_remove_page, *tmp_iter_page;
1405 	struct buffer_page *last_page, *first_page;
1406 	unsigned int nr_removed;
1407 	unsigned long head_bit;
1408 	int page_entries;
1409 
1410 	head_bit = 0;
1411 
1412 	raw_spin_lock_irq(&cpu_buffer->reader_lock);
1413 	atomic_inc(&cpu_buffer->record_disabled);
1414 	/*
1415 	 * We don't race with the readers since we have acquired the reader
1416 	 * lock. We also don't race with writers after disabling recording.
1417 	 * This makes it easy to figure out the first and the last page to be
1418 	 * removed from the list. We unlink all the pages in between including
1419 	 * the first and last pages. This is done in a busy loop so that we
1420 	 * lose the least number of traces.
1421 	 * The pages are freed after we restart recording and unlock readers.
1422 	 */
1423 	tail_page = &cpu_buffer->tail_page->list;
1424 
1425 	/*
1426 	 * tail page might be on reader page, we remove the next page
1427 	 * from the ring buffer
1428 	 */
1429 	if (cpu_buffer->tail_page == cpu_buffer->reader_page)
1430 		tail_page = rb_list_head(tail_page->next);
1431 	to_remove = tail_page;
1432 
1433 	/* start of pages to remove */
1434 	first_page = list_entry(rb_list_head(to_remove->next),
1435 				struct buffer_page, list);
1436 
1437 	for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
1438 		to_remove = rb_list_head(to_remove)->next;
1439 		head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
1440 	}
1441 
1442 	next_page = rb_list_head(to_remove)->next;
1443 
1444 	/*
1445 	 * Now we remove all pages between tail_page and next_page.
1446 	 * Make sure that we have head_bit value preserved for the
1447 	 * next page
1448 	 */
1449 	tail_page->next = (struct list_head *)((unsigned long)next_page |
1450 						head_bit);
1451 	next_page = rb_list_head(next_page);
1452 	next_page->prev = tail_page;
1453 
1454 	/* make sure pages points to a valid page in the ring buffer */
1455 	cpu_buffer->pages = next_page;
1456 
1457 	/* update head page */
1458 	if (head_bit)
1459 		cpu_buffer->head_page = list_entry(next_page,
1460 						struct buffer_page, list);
1461 
1462 	/*
1463 	 * change read pointer to make sure any read iterators reset
1464 	 * themselves
1465 	 */
1466 	cpu_buffer->read = 0;
1467 
1468 	/* pages are removed, resume tracing and then free the pages */
1469 	atomic_dec(&cpu_buffer->record_disabled);
1470 	raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1471 
1472 	RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
1473 
1474 	/* last buffer page to remove */
1475 	last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
1476 				list);
1477 	tmp_iter_page = first_page;
1478 
1479 	do {
1480 		to_remove_page = tmp_iter_page;
1481 		rb_inc_page(cpu_buffer, &tmp_iter_page);
1482 
1483 		/* update the counters */
1484 		page_entries = rb_page_entries(to_remove_page);
1485 		if (page_entries) {
1486 			/*
1487 			 * If something was added to this page, it was full
1488 			 * since it is not the tail page. So we deduct the
1489 			 * bytes consumed in ring buffer from here.
1490 			 * Increment overrun to account for the lost events.
1491 			 */
1492 			local_add(page_entries, &cpu_buffer->overrun);
1493 			local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1494 		}
1495 
1496 		/*
1497 		 * We have already removed references to this list item, just
1498 		 * free up the buffer_page and its page
1499 		 */
1500 		free_buffer_page(to_remove_page);
1501 		nr_removed--;
1502 
1503 	} while (to_remove_page != last_page);
1504 
1505 	RB_WARN_ON(cpu_buffer, nr_removed);
1506 
1507 	return nr_removed == 0;
1508 }
1509 
1510 static int
1511 rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
1512 {
1513 	struct list_head *pages = &cpu_buffer->new_pages;
1514 	int retries, success;
1515 
1516 	raw_spin_lock_irq(&cpu_buffer->reader_lock);
1517 	/*
1518 	 * We are holding the reader lock, so the reader page won't be swapped
1519 	 * in the ring buffer. Now we are racing with the writer trying to
1520 	 * move head page and the tail page.
1521 	 * We are going to adapt the reader page update process where:
1522 	 * 1. We first splice the start and end of list of new pages between
1523 	 *    the head page and its previous page.
1524 	 * 2. We cmpxchg the prev_page->next to point from head page to the
1525 	 *    start of new pages list.
1526 	 * 3. Finally, we update the head->prev to the end of new list.
1527 	 *
1528 	 * We will try this process 10 times, to make sure that we don't keep
1529 	 * spinning.
1530 	 */
1531 	retries = 10;
1532 	success = 0;
1533 	while (retries--) {
1534 		struct list_head *head_page, *prev_page, *r;
1535 		struct list_head *last_page, *first_page;
1536 		struct list_head *head_page_with_bit;
1537 
1538 		head_page = &rb_set_head_page(cpu_buffer)->list;
1539 		if (!head_page)
1540 			break;
1541 		prev_page = head_page->prev;
1542 
1543 		first_page = pages->next;
1544 		last_page  = pages->prev;
1545 
1546 		head_page_with_bit = (struct list_head *)
1547 				     ((unsigned long)head_page | RB_PAGE_HEAD);
1548 
1549 		last_page->next = head_page_with_bit;
1550 		first_page->prev = prev_page;
1551 
1552 		r = cmpxchg(&prev_page->next, head_page_with_bit, first_page);
1553 
1554 		if (r == head_page_with_bit) {
1555 			/*
1556 			 * yay, we replaced the page pointer to our new list,
1557 			 * now, we just have to update to head page's prev
1558 			 * pointer to point to end of list
1559 			 */
1560 			head_page->prev = last_page;
1561 			success = 1;
1562 			break;
1563 		}
1564 	}
1565 
1566 	if (success)
1567 		INIT_LIST_HEAD(pages);
1568 	/*
1569 	 * If we weren't successful in adding in new pages, warn and stop
1570 	 * tracing
1571 	 */
1572 	RB_WARN_ON(cpu_buffer, !success);
1573 	raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1574 
1575 	/* free pages if they weren't inserted */
1576 	if (!success) {
1577 		struct buffer_page *bpage, *tmp;
1578 		list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1579 					 list) {
1580 			list_del_init(&bpage->list);
1581 			free_buffer_page(bpage);
1582 		}
1583 	}
1584 	return success;
1585 }
1586 
1587 static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
1588 {
1589 	int success;
1590 
1591 	if (cpu_buffer->nr_pages_to_update > 0)
1592 		success = rb_insert_pages(cpu_buffer);
1593 	else
1594 		success = rb_remove_pages(cpu_buffer,
1595 					-cpu_buffer->nr_pages_to_update);
1596 
1597 	if (success)
1598 		cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
1599 }
1600 
1601 static void update_pages_handler(struct work_struct *work)
1602 {
1603 	struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
1604 			struct ring_buffer_per_cpu, update_pages_work);
1605 	rb_update_pages(cpu_buffer);
1606 	complete(&cpu_buffer->update_done);
1607 }
1608 
1609 /**
1610  * ring_buffer_resize - resize the ring buffer
1611  * @buffer: the buffer to resize.
1612  * @size: the new size.
1613  * @cpu_id: the cpu buffer to resize
1614  *
1615  * Minimum size is 2 * BUF_PAGE_SIZE.
1616  *
1617  * Returns 0 on success and < 0 on failure.
1618  */
1619 int ring_buffer_resize(struct ring_buffer *buffer, unsigned long size,
1620 			int cpu_id)
1621 {
1622 	struct ring_buffer_per_cpu *cpu_buffer;
1623 	unsigned nr_pages;
1624 	int cpu, err = 0;
1625 
1626 	/*
1627 	 * Always succeed at resizing a non-existent buffer:
1628 	 */
1629 	if (!buffer)
1630 		return size;
1631 
1632 	/* Make sure the requested buffer exists */
1633 	if (cpu_id != RING_BUFFER_ALL_CPUS &&
1634 	    !cpumask_test_cpu(cpu_id, buffer->cpumask))
1635 		return size;
1636 
1637 	size = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1638 	size *= BUF_PAGE_SIZE;
1639 
1640 	/* we need a minimum of two pages */
1641 	if (size < BUF_PAGE_SIZE * 2)
1642 		size = BUF_PAGE_SIZE * 2;
1643 
1644 	nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1645 
1646 	/*
1647 	 * Don't succeed if resizing is disabled, as a reader might be
1648 	 * manipulating the ring buffer and is expecting a sane state while
1649 	 * this is true.
1650 	 */
1651 	if (atomic_read(&buffer->resize_disabled))
1652 		return -EBUSY;
1653 
1654 	/* prevent another thread from changing buffer sizes */
1655 	mutex_lock(&buffer->mutex);
1656 
1657 	if (cpu_id == RING_BUFFER_ALL_CPUS) {
1658 		/* calculate the pages to update */
1659 		for_each_buffer_cpu(buffer, cpu) {
1660 			cpu_buffer = buffer->buffers[cpu];
1661 
1662 			cpu_buffer->nr_pages_to_update = nr_pages -
1663 							cpu_buffer->nr_pages;
1664 			/*
1665 			 * nothing more to do for removing pages or no update
1666 			 */
1667 			if (cpu_buffer->nr_pages_to_update <= 0)
1668 				continue;
1669 			/*
1670 			 * to add pages, make sure all new pages can be
1671 			 * allocated without receiving ENOMEM
1672 			 */
1673 			INIT_LIST_HEAD(&cpu_buffer->new_pages);
1674 			if (__rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1675 						&cpu_buffer->new_pages, cpu)) {
1676 				/* not enough memory for new pages */
1677 				err = -ENOMEM;
1678 				goto out_err;
1679 			}
1680 		}
1681 
1682 		get_online_cpus();
1683 		/*
1684 		 * Fire off all the required work handlers
1685 		 * We can't schedule on offline CPUs, but it's not necessary
1686 		 * since we can change their buffer sizes without any race.
1687 		 */
1688 		for_each_buffer_cpu(buffer, cpu) {
1689 			cpu_buffer = buffer->buffers[cpu];
1690 			if (!cpu_buffer->nr_pages_to_update)
1691 				continue;
1692 
1693 			/* The update must run on the CPU that is being updated. */
1694 			preempt_disable();
1695 			if (cpu == smp_processor_id() || !cpu_online(cpu)) {
1696 				rb_update_pages(cpu_buffer);
1697 				cpu_buffer->nr_pages_to_update = 0;
1698 			} else {
1699 				/*
1700 				 * Can not disable preemption for schedule_work_on()
1701 				 * on PREEMPT_RT.
1702 				 */
1703 				preempt_enable();
1704 				schedule_work_on(cpu,
1705 						&cpu_buffer->update_pages_work);
1706 				preempt_disable();
1707 			}
1708 			preempt_enable();
1709 		}
1710 
1711 		/* wait for all the updates to complete */
1712 		for_each_buffer_cpu(buffer, cpu) {
1713 			cpu_buffer = buffer->buffers[cpu];
1714 			if (!cpu_buffer->nr_pages_to_update)
1715 				continue;
1716 
1717 			if (cpu_online(cpu))
1718 				wait_for_completion(&cpu_buffer->update_done);
1719 			cpu_buffer->nr_pages_to_update = 0;
1720 		}
1721 
1722 		put_online_cpus();
1723 	} else {
1724 		/* Make sure this CPU has been intitialized */
1725 		if (!cpumask_test_cpu(cpu_id, buffer->cpumask))
1726 			goto out;
1727 
1728 		cpu_buffer = buffer->buffers[cpu_id];
1729 
1730 		if (nr_pages == cpu_buffer->nr_pages)
1731 			goto out;
1732 
1733 		cpu_buffer->nr_pages_to_update = nr_pages -
1734 						cpu_buffer->nr_pages;
1735 
1736 		INIT_LIST_HEAD(&cpu_buffer->new_pages);
1737 		if (cpu_buffer->nr_pages_to_update > 0 &&
1738 			__rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1739 					    &cpu_buffer->new_pages, cpu_id)) {
1740 			err = -ENOMEM;
1741 			goto out_err;
1742 		}
1743 
1744 		get_online_cpus();
1745 
1746 		preempt_disable();
1747 		/* The update must run on the CPU that is being updated. */
1748 		if (cpu_id == smp_processor_id() || !cpu_online(cpu_id))
1749 			rb_update_pages(cpu_buffer);
1750 		else {
1751 			/*
1752 			 * Can not disable preemption for schedule_work_on()
1753 			 * on PREEMPT_RT.
1754 			 */
1755 			preempt_enable();
1756 			schedule_work_on(cpu_id,
1757 					 &cpu_buffer->update_pages_work);
1758 			wait_for_completion(&cpu_buffer->update_done);
1759 			preempt_disable();
1760 		}
1761 		preempt_enable();
1762 
1763 		cpu_buffer->nr_pages_to_update = 0;
1764 		put_online_cpus();
1765 	}
1766 
1767  out:
1768 	/*
1769 	 * The ring buffer resize can happen with the ring buffer
1770 	 * enabled, so that the update disturbs the tracing as little
1771 	 * as possible. But if the buffer is disabled, we do not need
1772 	 * to worry about that, and we can take the time to verify
1773 	 * that the buffer is not corrupt.
1774 	 */
1775 	if (atomic_read(&buffer->record_disabled)) {
1776 		atomic_inc(&buffer->record_disabled);
1777 		/*
1778 		 * Even though the buffer was disabled, we must make sure
1779 		 * that it is truly disabled before calling rb_check_pages.
1780 		 * There could have been a race between checking
1781 		 * record_disable and incrementing it.
1782 		 */
1783 		synchronize_sched();
1784 		for_each_buffer_cpu(buffer, cpu) {
1785 			cpu_buffer = buffer->buffers[cpu];
1786 			rb_check_pages(cpu_buffer);
1787 		}
1788 		atomic_dec(&buffer->record_disabled);
1789 	}
1790 
1791 	mutex_unlock(&buffer->mutex);
1792 	return size;
1793 
1794  out_err:
1795 	for_each_buffer_cpu(buffer, cpu) {
1796 		struct buffer_page *bpage, *tmp;
1797 
1798 		cpu_buffer = buffer->buffers[cpu];
1799 		cpu_buffer->nr_pages_to_update = 0;
1800 
1801 		if (list_empty(&cpu_buffer->new_pages))
1802 			continue;
1803 
1804 		list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1805 					list) {
1806 			list_del_init(&bpage->list);
1807 			free_buffer_page(bpage);
1808 		}
1809 	}
1810 	mutex_unlock(&buffer->mutex);
1811 	return err;
1812 }
1813 EXPORT_SYMBOL_GPL(ring_buffer_resize);
1814 
1815 void ring_buffer_change_overwrite(struct ring_buffer *buffer, int val)
1816 {
1817 	mutex_lock(&buffer->mutex);
1818 	if (val)
1819 		buffer->flags |= RB_FL_OVERWRITE;
1820 	else
1821 		buffer->flags &= ~RB_FL_OVERWRITE;
1822 	mutex_unlock(&buffer->mutex);
1823 }
1824 EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
1825 
1826 static inline void *
1827 __rb_data_page_index(struct buffer_data_page *bpage, unsigned index)
1828 {
1829 	return bpage->data + index;
1830 }
1831 
1832 static inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
1833 {
1834 	return bpage->page->data + index;
1835 }
1836 
1837 static inline struct ring_buffer_event *
1838 rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
1839 {
1840 	return __rb_page_index(cpu_buffer->reader_page,
1841 			       cpu_buffer->reader_page->read);
1842 }
1843 
1844 static inline struct ring_buffer_event *
1845 rb_iter_head_event(struct ring_buffer_iter *iter)
1846 {
1847 	return __rb_page_index(iter->head_page, iter->head);
1848 }
1849 
1850 static inline unsigned rb_page_commit(struct buffer_page *bpage)
1851 {
1852 	return local_read(&bpage->page->commit);
1853 }
1854 
1855 /* Size is determined by what has been committed */
1856 static inline unsigned rb_page_size(struct buffer_page *bpage)
1857 {
1858 	return rb_page_commit(bpage);
1859 }
1860 
1861 static inline unsigned
1862 rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
1863 {
1864 	return rb_page_commit(cpu_buffer->commit_page);
1865 }
1866 
1867 static inline unsigned
1868 rb_event_index(struct ring_buffer_event *event)
1869 {
1870 	unsigned long addr = (unsigned long)event;
1871 
1872 	return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
1873 }
1874 
1875 static inline int
1876 rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
1877 		   struct ring_buffer_event *event)
1878 {
1879 	unsigned long addr = (unsigned long)event;
1880 	unsigned long index;
1881 
1882 	index = rb_event_index(event);
1883 	addr &= PAGE_MASK;
1884 
1885 	return cpu_buffer->commit_page->page == (void *)addr &&
1886 		rb_commit_index(cpu_buffer) == index;
1887 }
1888 
1889 static void
1890 rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
1891 {
1892 	unsigned long max_count;
1893 
1894 	/*
1895 	 * We only race with interrupts and NMIs on this CPU.
1896 	 * If we own the commit event, then we can commit
1897 	 * all others that interrupted us, since the interruptions
1898 	 * are in stack format (they finish before they come
1899 	 * back to us). This allows us to do a simple loop to
1900 	 * assign the commit to the tail.
1901 	 */
1902  again:
1903 	max_count = cpu_buffer->nr_pages * 100;
1904 
1905 	while (cpu_buffer->commit_page != cpu_buffer->tail_page) {
1906 		if (RB_WARN_ON(cpu_buffer, !(--max_count)))
1907 			return;
1908 		if (RB_WARN_ON(cpu_buffer,
1909 			       rb_is_reader_page(cpu_buffer->tail_page)))
1910 			return;
1911 		local_set(&cpu_buffer->commit_page->page->commit,
1912 			  rb_page_write(cpu_buffer->commit_page));
1913 		rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
1914 		cpu_buffer->write_stamp =
1915 			cpu_buffer->commit_page->page->time_stamp;
1916 		/* add barrier to keep gcc from optimizing too much */
1917 		barrier();
1918 	}
1919 	while (rb_commit_index(cpu_buffer) !=
1920 	       rb_page_write(cpu_buffer->commit_page)) {
1921 
1922 		local_set(&cpu_buffer->commit_page->page->commit,
1923 			  rb_page_write(cpu_buffer->commit_page));
1924 		RB_WARN_ON(cpu_buffer,
1925 			   local_read(&cpu_buffer->commit_page->page->commit) &
1926 			   ~RB_WRITE_MASK);
1927 		barrier();
1928 	}
1929 
1930 	/* again, keep gcc from optimizing */
1931 	barrier();
1932 
1933 	/*
1934 	 * If an interrupt came in just after the first while loop
1935 	 * and pushed the tail page forward, we will be left with
1936 	 * a dangling commit that will never go forward.
1937 	 */
1938 	if (unlikely(cpu_buffer->commit_page != cpu_buffer->tail_page))
1939 		goto again;
1940 }
1941 
1942 static void rb_reset_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
1943 {
1944 	cpu_buffer->read_stamp = cpu_buffer->reader_page->page->time_stamp;
1945 	cpu_buffer->reader_page->read = 0;
1946 }
1947 
1948 static void rb_inc_iter(struct ring_buffer_iter *iter)
1949 {
1950 	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
1951 
1952 	/*
1953 	 * The iterator could be on the reader page (it starts there).
1954 	 * But the head could have moved, since the reader was
1955 	 * found. Check for this case and assign the iterator
1956 	 * to the head page instead of next.
1957 	 */
1958 	if (iter->head_page == cpu_buffer->reader_page)
1959 		iter->head_page = rb_set_head_page(cpu_buffer);
1960 	else
1961 		rb_inc_page(cpu_buffer, &iter->head_page);
1962 
1963 	iter->read_stamp = iter->head_page->page->time_stamp;
1964 	iter->head = 0;
1965 }
1966 
1967 /* Slow path, do not inline */
1968 static noinline struct ring_buffer_event *
1969 rb_add_time_stamp(struct ring_buffer_event *event, u64 delta)
1970 {
1971 	event->type_len = RINGBUF_TYPE_TIME_EXTEND;
1972 
1973 	/* Not the first event on the page? */
1974 	if (rb_event_index(event)) {
1975 		event->time_delta = delta & TS_MASK;
1976 		event->array[0] = delta >> TS_SHIFT;
1977 	} else {
1978 		/* nope, just zero it */
1979 		event->time_delta = 0;
1980 		event->array[0] = 0;
1981 	}
1982 
1983 	return skip_time_extend(event);
1984 }
1985 
1986 /**
1987  * rb_update_event - update event type and data
1988  * @event: the even to update
1989  * @type: the type of event
1990  * @length: the size of the event field in the ring buffer
1991  *
1992  * Update the type and data fields of the event. The length
1993  * is the actual size that is written to the ring buffer,
1994  * and with this, we can determine what to place into the
1995  * data field.
1996  */
1997 static void
1998 rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
1999 		struct ring_buffer_event *event, unsigned length,
2000 		int add_timestamp, u64 delta)
2001 {
2002 	/* Only a commit updates the timestamp */
2003 	if (unlikely(!rb_event_is_commit(cpu_buffer, event)))
2004 		delta = 0;
2005 
2006 	/*
2007 	 * If we need to add a timestamp, then we
2008 	 * add it to the start of the resevered space.
2009 	 */
2010 	if (unlikely(add_timestamp)) {
2011 		event = rb_add_time_stamp(event, delta);
2012 		length -= RB_LEN_TIME_EXTEND;
2013 		delta = 0;
2014 	}
2015 
2016 	event->time_delta = delta;
2017 	length -= RB_EVNT_HDR_SIZE;
2018 	if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
2019 		event->type_len = 0;
2020 		event->array[0] = length;
2021 	} else
2022 		event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
2023 }
2024 
2025 /*
2026  * rb_handle_head_page - writer hit the head page
2027  *
2028  * Returns: +1 to retry page
2029  *           0 to continue
2030  *          -1 on error
2031  */
2032 static int
2033 rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
2034 		    struct buffer_page *tail_page,
2035 		    struct buffer_page *next_page)
2036 {
2037 	struct buffer_page *new_head;
2038 	int entries;
2039 	int type;
2040 	int ret;
2041 
2042 	entries = rb_page_entries(next_page);
2043 
2044 	/*
2045 	 * The hard part is here. We need to move the head
2046 	 * forward, and protect against both readers on
2047 	 * other CPUs and writers coming in via interrupts.
2048 	 */
2049 	type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
2050 				       RB_PAGE_HEAD);
2051 
2052 	/*
2053 	 * type can be one of four:
2054 	 *  NORMAL - an interrupt already moved it for us
2055 	 *  HEAD   - we are the first to get here.
2056 	 *  UPDATE - we are the interrupt interrupting
2057 	 *           a current move.
2058 	 *  MOVED  - a reader on another CPU moved the next
2059 	 *           pointer to its reader page. Give up
2060 	 *           and try again.
2061 	 */
2062 
2063 	switch (type) {
2064 	case RB_PAGE_HEAD:
2065 		/*
2066 		 * We changed the head to UPDATE, thus
2067 		 * it is our responsibility to update
2068 		 * the counters.
2069 		 */
2070 		local_add(entries, &cpu_buffer->overrun);
2071 		local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
2072 
2073 		/*
2074 		 * The entries will be zeroed out when we move the
2075 		 * tail page.
2076 		 */
2077 
2078 		/* still more to do */
2079 		break;
2080 
2081 	case RB_PAGE_UPDATE:
2082 		/*
2083 		 * This is an interrupt that interrupt the
2084 		 * previous update. Still more to do.
2085 		 */
2086 		break;
2087 	case RB_PAGE_NORMAL:
2088 		/*
2089 		 * An interrupt came in before the update
2090 		 * and processed this for us.
2091 		 * Nothing left to do.
2092 		 */
2093 		return 1;
2094 	case RB_PAGE_MOVED:
2095 		/*
2096 		 * The reader is on another CPU and just did
2097 		 * a swap with our next_page.
2098 		 * Try again.
2099 		 */
2100 		return 1;
2101 	default:
2102 		RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
2103 		return -1;
2104 	}
2105 
2106 	/*
2107 	 * Now that we are here, the old head pointer is
2108 	 * set to UPDATE. This will keep the reader from
2109 	 * swapping the head page with the reader page.
2110 	 * The reader (on another CPU) will spin till
2111 	 * we are finished.
2112 	 *
2113 	 * We just need to protect against interrupts
2114 	 * doing the job. We will set the next pointer
2115 	 * to HEAD. After that, we set the old pointer
2116 	 * to NORMAL, but only if it was HEAD before.
2117 	 * otherwise we are an interrupt, and only
2118 	 * want the outer most commit to reset it.
2119 	 */
2120 	new_head = next_page;
2121 	rb_inc_page(cpu_buffer, &new_head);
2122 
2123 	ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
2124 				    RB_PAGE_NORMAL);
2125 
2126 	/*
2127 	 * Valid returns are:
2128 	 *  HEAD   - an interrupt came in and already set it.
2129 	 *  NORMAL - One of two things:
2130 	 *            1) We really set it.
2131 	 *            2) A bunch of interrupts came in and moved
2132 	 *               the page forward again.
2133 	 */
2134 	switch (ret) {
2135 	case RB_PAGE_HEAD:
2136 	case RB_PAGE_NORMAL:
2137 		/* OK */
2138 		break;
2139 	default:
2140 		RB_WARN_ON(cpu_buffer, 1);
2141 		return -1;
2142 	}
2143 
2144 	/*
2145 	 * It is possible that an interrupt came in,
2146 	 * set the head up, then more interrupts came in
2147 	 * and moved it again. When we get back here,
2148 	 * the page would have been set to NORMAL but we
2149 	 * just set it back to HEAD.
2150 	 *
2151 	 * How do you detect this? Well, if that happened
2152 	 * the tail page would have moved.
2153 	 */
2154 	if (ret == RB_PAGE_NORMAL) {
2155 		/*
2156 		 * If the tail had moved passed next, then we need
2157 		 * to reset the pointer.
2158 		 */
2159 		if (cpu_buffer->tail_page != tail_page &&
2160 		    cpu_buffer->tail_page != next_page)
2161 			rb_head_page_set_normal(cpu_buffer, new_head,
2162 						next_page,
2163 						RB_PAGE_HEAD);
2164 	}
2165 
2166 	/*
2167 	 * If this was the outer most commit (the one that
2168 	 * changed the original pointer from HEAD to UPDATE),
2169 	 * then it is up to us to reset it to NORMAL.
2170 	 */
2171 	if (type == RB_PAGE_HEAD) {
2172 		ret = rb_head_page_set_normal(cpu_buffer, next_page,
2173 					      tail_page,
2174 					      RB_PAGE_UPDATE);
2175 		if (RB_WARN_ON(cpu_buffer,
2176 			       ret != RB_PAGE_UPDATE))
2177 			return -1;
2178 	}
2179 
2180 	return 0;
2181 }
2182 
2183 static unsigned rb_calculate_event_length(unsigned length)
2184 {
2185 	struct ring_buffer_event event; /* Used only for sizeof array */
2186 
2187 	/* zero length can cause confusions */
2188 	if (!length)
2189 		length = 1;
2190 
2191 	if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
2192 		length += sizeof(event.array[0]);
2193 
2194 	length += RB_EVNT_HDR_SIZE;
2195 	length = ALIGN(length, RB_ARCH_ALIGNMENT);
2196 
2197 	return length;
2198 }
2199 
2200 static inline void
2201 rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2202 	      struct buffer_page *tail_page,
2203 	      unsigned long tail, unsigned long length)
2204 {
2205 	struct ring_buffer_event *event;
2206 
2207 	/*
2208 	 * Only the event that crossed the page boundary
2209 	 * must fill the old tail_page with padding.
2210 	 */
2211 	if (tail >= BUF_PAGE_SIZE) {
2212 		/*
2213 		 * If the page was filled, then we still need
2214 		 * to update the real_end. Reset it to zero
2215 		 * and the reader will ignore it.
2216 		 */
2217 		if (tail == BUF_PAGE_SIZE)
2218 			tail_page->real_end = 0;
2219 
2220 		local_sub(length, &tail_page->write);
2221 		return;
2222 	}
2223 
2224 	event = __rb_page_index(tail_page, tail);
2225 	kmemcheck_annotate_bitfield(event, bitfield);
2226 
2227 	/* account for padding bytes */
2228 	local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes);
2229 
2230 	/*
2231 	 * Save the original length to the meta data.
2232 	 * This will be used by the reader to add lost event
2233 	 * counter.
2234 	 */
2235 	tail_page->real_end = tail;
2236 
2237 	/*
2238 	 * If this event is bigger than the minimum size, then
2239 	 * we need to be careful that we don't subtract the
2240 	 * write counter enough to allow another writer to slip
2241 	 * in on this page.
2242 	 * We put in a discarded commit instead, to make sure
2243 	 * that this space is not used again.
2244 	 *
2245 	 * If we are less than the minimum size, we don't need to
2246 	 * worry about it.
2247 	 */
2248 	if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
2249 		/* No room for any events */
2250 
2251 		/* Mark the rest of the page with padding */
2252 		rb_event_set_padding(event);
2253 
2254 		/* Set the write back to the previous setting */
2255 		local_sub(length, &tail_page->write);
2256 		return;
2257 	}
2258 
2259 	/* Put in a discarded event */
2260 	event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
2261 	event->type_len = RINGBUF_TYPE_PADDING;
2262 	/* time delta must be non zero */
2263 	event->time_delta = 1;
2264 
2265 	/* Set write to end of buffer */
2266 	length = (tail + length) - BUF_PAGE_SIZE;
2267 	local_sub(length, &tail_page->write);
2268 }
2269 
2270 /*
2271  * This is the slow path, force gcc not to inline it.
2272  */
2273 static noinline struct ring_buffer_event *
2274 rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2275 	     unsigned long length, unsigned long tail,
2276 	     struct buffer_page *tail_page, u64 ts)
2277 {
2278 	struct buffer_page *commit_page = cpu_buffer->commit_page;
2279 	struct ring_buffer *buffer = cpu_buffer->buffer;
2280 	struct buffer_page *next_page;
2281 	int ret;
2282 
2283 	next_page = tail_page;
2284 
2285 	rb_inc_page(cpu_buffer, &next_page);
2286 
2287 	/*
2288 	 * If for some reason, we had an interrupt storm that made
2289 	 * it all the way around the buffer, bail, and warn
2290 	 * about it.
2291 	 */
2292 	if (unlikely(next_page == commit_page)) {
2293 		local_inc(&cpu_buffer->commit_overrun);
2294 		goto out_reset;
2295 	}
2296 
2297 	/*
2298 	 * This is where the fun begins!
2299 	 *
2300 	 * We are fighting against races between a reader that
2301 	 * could be on another CPU trying to swap its reader
2302 	 * page with the buffer head.
2303 	 *
2304 	 * We are also fighting against interrupts coming in and
2305 	 * moving the head or tail on us as well.
2306 	 *
2307 	 * If the next page is the head page then we have filled
2308 	 * the buffer, unless the commit page is still on the
2309 	 * reader page.
2310 	 */
2311 	if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) {
2312 
2313 		/*
2314 		 * If the commit is not on the reader page, then
2315 		 * move the header page.
2316 		 */
2317 		if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2318 			/*
2319 			 * If we are not in overwrite mode,
2320 			 * this is easy, just stop here.
2321 			 */
2322 			if (!(buffer->flags & RB_FL_OVERWRITE)) {
2323 				local_inc(&cpu_buffer->dropped_events);
2324 				goto out_reset;
2325 			}
2326 
2327 			ret = rb_handle_head_page(cpu_buffer,
2328 						  tail_page,
2329 						  next_page);
2330 			if (ret < 0)
2331 				goto out_reset;
2332 			if (ret)
2333 				goto out_again;
2334 		} else {
2335 			/*
2336 			 * We need to be careful here too. The
2337 			 * commit page could still be on the reader
2338 			 * page. We could have a small buffer, and
2339 			 * have filled up the buffer with events
2340 			 * from interrupts and such, and wrapped.
2341 			 *
2342 			 * Note, if the tail page is also the on the
2343 			 * reader_page, we let it move out.
2344 			 */
2345 			if (unlikely((cpu_buffer->commit_page !=
2346 				      cpu_buffer->tail_page) &&
2347 				     (cpu_buffer->commit_page ==
2348 				      cpu_buffer->reader_page))) {
2349 				local_inc(&cpu_buffer->commit_overrun);
2350 				goto out_reset;
2351 			}
2352 		}
2353 	}
2354 
2355 	ret = rb_tail_page_update(cpu_buffer, tail_page, next_page);
2356 	if (ret) {
2357 		/*
2358 		 * Nested commits always have zero deltas, so
2359 		 * just reread the time stamp
2360 		 */
2361 		ts = rb_time_stamp(buffer);
2362 		next_page->page->time_stamp = ts;
2363 	}
2364 
2365  out_again:
2366 
2367 	rb_reset_tail(cpu_buffer, tail_page, tail, length);
2368 
2369 	/* fail and let the caller try again */
2370 	return ERR_PTR(-EAGAIN);
2371 
2372  out_reset:
2373 	/* reset write */
2374 	rb_reset_tail(cpu_buffer, tail_page, tail, length);
2375 
2376 	return NULL;
2377 }
2378 
2379 static struct ring_buffer_event *
2380 __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
2381 		  unsigned long length, u64 ts,
2382 		  u64 delta, int add_timestamp)
2383 {
2384 	struct buffer_page *tail_page;
2385 	struct ring_buffer_event *event;
2386 	unsigned long tail, write;
2387 
2388 	/*
2389 	 * If the time delta since the last event is too big to
2390 	 * hold in the time field of the event, then we append a
2391 	 * TIME EXTEND event ahead of the data event.
2392 	 */
2393 	if (unlikely(add_timestamp))
2394 		length += RB_LEN_TIME_EXTEND;
2395 
2396 	tail_page = cpu_buffer->tail_page;
2397 	write = local_add_return(length, &tail_page->write);
2398 
2399 	/* set write to only the index of the write */
2400 	write &= RB_WRITE_MASK;
2401 	tail = write - length;
2402 
2403 	/*
2404 	 * If this is the first commit on the page, then it has the same
2405 	 * timestamp as the page itself.
2406 	 */
2407 	if (!tail)
2408 		delta = 0;
2409 
2410 	/* See if we shot pass the end of this buffer page */
2411 	if (unlikely(write > BUF_PAGE_SIZE))
2412 		return rb_move_tail(cpu_buffer, length, tail,
2413 				    tail_page, ts);
2414 
2415 	/* We reserved something on the buffer */
2416 
2417 	event = __rb_page_index(tail_page, tail);
2418 	kmemcheck_annotate_bitfield(event, bitfield);
2419 	rb_update_event(cpu_buffer, event, length, add_timestamp, delta);
2420 
2421 	local_inc(&tail_page->entries);
2422 
2423 	/*
2424 	 * If this is the first commit on the page, then update
2425 	 * its timestamp.
2426 	 */
2427 	if (!tail)
2428 		tail_page->page->time_stamp = ts;
2429 
2430 	/* account for these added bytes */
2431 	local_add(length, &cpu_buffer->entries_bytes);
2432 
2433 	return event;
2434 }
2435 
2436 static inline int
2437 rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2438 		  struct ring_buffer_event *event)
2439 {
2440 	unsigned long new_index, old_index;
2441 	struct buffer_page *bpage;
2442 	unsigned long index;
2443 	unsigned long addr;
2444 
2445 	new_index = rb_event_index(event);
2446 	old_index = new_index + rb_event_ts_length(event);
2447 	addr = (unsigned long)event;
2448 	addr &= PAGE_MASK;
2449 
2450 	bpage = cpu_buffer->tail_page;
2451 
2452 	if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2453 		unsigned long write_mask =
2454 			local_read(&bpage->write) & ~RB_WRITE_MASK;
2455 		unsigned long event_length = rb_event_length(event);
2456 		/*
2457 		 * This is on the tail page. It is possible that
2458 		 * a write could come in and move the tail page
2459 		 * and write to the next page. That is fine
2460 		 * because we just shorten what is on this page.
2461 		 */
2462 		old_index += write_mask;
2463 		new_index += write_mask;
2464 		index = local_cmpxchg(&bpage->write, old_index, new_index);
2465 		if (index == old_index) {
2466 			/* update counters */
2467 			local_sub(event_length, &cpu_buffer->entries_bytes);
2468 			return 1;
2469 		}
2470 	}
2471 
2472 	/* could not discard */
2473 	return 0;
2474 }
2475 
2476 static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2477 {
2478 	local_inc(&cpu_buffer->committing);
2479 	local_inc(&cpu_buffer->commits);
2480 }
2481 
2482 static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
2483 {
2484 	unsigned long commits;
2485 
2486 	if (RB_WARN_ON(cpu_buffer,
2487 		       !local_read(&cpu_buffer->committing)))
2488 		return;
2489 
2490  again:
2491 	commits = local_read(&cpu_buffer->commits);
2492 	/* synchronize with interrupts */
2493 	barrier();
2494 	if (local_read(&cpu_buffer->committing) == 1)
2495 		rb_set_commit_to_write(cpu_buffer);
2496 
2497 	local_dec(&cpu_buffer->committing);
2498 
2499 	/* synchronize with interrupts */
2500 	barrier();
2501 
2502 	/*
2503 	 * Need to account for interrupts coming in between the
2504 	 * updating of the commit page and the clearing of the
2505 	 * committing counter.
2506 	 */
2507 	if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
2508 	    !local_read(&cpu_buffer->committing)) {
2509 		local_inc(&cpu_buffer->committing);
2510 		goto again;
2511 	}
2512 }
2513 
2514 static struct ring_buffer_event *
2515 rb_reserve_next_event(struct ring_buffer *buffer,
2516 		      struct ring_buffer_per_cpu *cpu_buffer,
2517 		      unsigned long length)
2518 {
2519 	struct ring_buffer_event *event;
2520 	u64 ts, delta;
2521 	int nr_loops = 0;
2522 	int add_timestamp;
2523 	u64 diff;
2524 
2525 	rb_start_commit(cpu_buffer);
2526 
2527 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2528 	/*
2529 	 * Due to the ability to swap a cpu buffer from a buffer
2530 	 * it is possible it was swapped before we committed.
2531 	 * (committing stops a swap). We check for it here and
2532 	 * if it happened, we have to fail the write.
2533 	 */
2534 	barrier();
2535 	if (unlikely(ACCESS_ONCE(cpu_buffer->buffer) != buffer)) {
2536 		local_dec(&cpu_buffer->committing);
2537 		local_dec(&cpu_buffer->commits);
2538 		return NULL;
2539 	}
2540 #endif
2541 
2542 	length = rb_calculate_event_length(length);
2543  again:
2544 	add_timestamp = 0;
2545 	delta = 0;
2546 
2547 	/*
2548 	 * We allow for interrupts to reenter here and do a trace.
2549 	 * If one does, it will cause this original code to loop
2550 	 * back here. Even with heavy interrupts happening, this
2551 	 * should only happen a few times in a row. If this happens
2552 	 * 1000 times in a row, there must be either an interrupt
2553 	 * storm or we have something buggy.
2554 	 * Bail!
2555 	 */
2556 	if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
2557 		goto out_fail;
2558 
2559 	ts = rb_time_stamp(cpu_buffer->buffer);
2560 	diff = ts - cpu_buffer->write_stamp;
2561 
2562 	/* make sure this diff is calculated here */
2563 	barrier();
2564 
2565 	/* Did the write stamp get updated already? */
2566 	if (likely(ts >= cpu_buffer->write_stamp)) {
2567 		delta = diff;
2568 		if (unlikely(test_time_stamp(delta))) {
2569 			int local_clock_stable = 1;
2570 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2571 			local_clock_stable = sched_clock_stable();
2572 #endif
2573 			WARN_ONCE(delta > (1ULL << 59),
2574 				  KERN_WARNING "Delta way too big! %llu ts=%llu write stamp = %llu\n%s",
2575 				  (unsigned long long)delta,
2576 				  (unsigned long long)ts,
2577 				  (unsigned long long)cpu_buffer->write_stamp,
2578 				  local_clock_stable ? "" :
2579 				  "If you just came from a suspend/resume,\n"
2580 				  "please switch to the trace global clock:\n"
2581 				  "  echo global > /sys/kernel/debug/tracing/trace_clock\n");
2582 			add_timestamp = 1;
2583 		}
2584 	}
2585 
2586 	event = __rb_reserve_next(cpu_buffer, length, ts,
2587 				  delta, add_timestamp);
2588 	if (unlikely(PTR_ERR(event) == -EAGAIN))
2589 		goto again;
2590 
2591 	if (!event)
2592 		goto out_fail;
2593 
2594 	return event;
2595 
2596  out_fail:
2597 	rb_end_commit(cpu_buffer);
2598 	return NULL;
2599 }
2600 
2601 #ifdef CONFIG_TRACING
2602 
2603 /*
2604  * The lock and unlock are done within a preempt disable section.
2605  * The current_context per_cpu variable can only be modified
2606  * by the current task between lock and unlock. But it can
2607  * be modified more than once via an interrupt. To pass this
2608  * information from the lock to the unlock without having to
2609  * access the 'in_interrupt()' functions again (which do show
2610  * a bit of overhead in something as critical as function tracing,
2611  * we use a bitmask trick.
2612  *
2613  *  bit 0 =  NMI context
2614  *  bit 1 =  IRQ context
2615  *  bit 2 =  SoftIRQ context
2616  *  bit 3 =  normal context.
2617  *
2618  * This works because this is the order of contexts that can
2619  * preempt other contexts. A SoftIRQ never preempts an IRQ
2620  * context.
2621  *
2622  * When the context is determined, the corresponding bit is
2623  * checked and set (if it was set, then a recursion of that context
2624  * happened).
2625  *
2626  * On unlock, we need to clear this bit. To do so, just subtract
2627  * 1 from the current_context and AND it to itself.
2628  *
2629  * (binary)
2630  *  101 - 1 = 100
2631  *  101 & 100 = 100 (clearing bit zero)
2632  *
2633  *  1010 - 1 = 1001
2634  *  1010 & 1001 = 1000 (clearing bit 1)
2635  *
2636  * The least significant bit can be cleared this way, and it
2637  * just so happens that it is the same bit corresponding to
2638  * the current context.
2639  */
2640 static DEFINE_PER_CPU(unsigned int, current_context);
2641 
2642 static __always_inline int trace_recursive_lock(void)
2643 {
2644 	unsigned int val = this_cpu_read(current_context);
2645 	int bit;
2646 
2647 	if (in_interrupt()) {
2648 		if (in_nmi())
2649 			bit = 0;
2650 		else if (in_irq())
2651 			bit = 1;
2652 		else
2653 			bit = 2;
2654 	} else
2655 		bit = 3;
2656 
2657 	if (unlikely(val & (1 << bit)))
2658 		return 1;
2659 
2660 	val |= (1 << bit);
2661 	this_cpu_write(current_context, val);
2662 
2663 	return 0;
2664 }
2665 
2666 static __always_inline void trace_recursive_unlock(void)
2667 {
2668 	unsigned int val = this_cpu_read(current_context);
2669 
2670 	val--;
2671 	val &= this_cpu_read(current_context);
2672 	this_cpu_write(current_context, val);
2673 }
2674 
2675 #else
2676 
2677 #define trace_recursive_lock()		(0)
2678 #define trace_recursive_unlock()	do { } while (0)
2679 
2680 #endif
2681 
2682 /**
2683  * ring_buffer_lock_reserve - reserve a part of the buffer
2684  * @buffer: the ring buffer to reserve from
2685  * @length: the length of the data to reserve (excluding event header)
2686  *
2687  * Returns a reseverd event on the ring buffer to copy directly to.
2688  * The user of this interface will need to get the body to write into
2689  * and can use the ring_buffer_event_data() interface.
2690  *
2691  * The length is the length of the data needed, not the event length
2692  * which also includes the event header.
2693  *
2694  * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2695  * If NULL is returned, then nothing has been allocated or locked.
2696  */
2697 struct ring_buffer_event *
2698 ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length)
2699 {
2700 	struct ring_buffer_per_cpu *cpu_buffer;
2701 	struct ring_buffer_event *event;
2702 	int cpu;
2703 
2704 	if (ring_buffer_flags != RB_BUFFERS_ON)
2705 		return NULL;
2706 
2707 	/* If we are tracing schedule, we don't want to recurse */
2708 	preempt_disable_notrace();
2709 
2710 	if (atomic_read(&buffer->record_disabled))
2711 		goto out_nocheck;
2712 
2713 	if (trace_recursive_lock())
2714 		goto out_nocheck;
2715 
2716 	cpu = raw_smp_processor_id();
2717 
2718 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
2719 		goto out;
2720 
2721 	cpu_buffer = buffer->buffers[cpu];
2722 
2723 	if (atomic_read(&cpu_buffer->record_disabled))
2724 		goto out;
2725 
2726 	if (length > BUF_MAX_DATA_SIZE)
2727 		goto out;
2728 
2729 	event = rb_reserve_next_event(buffer, cpu_buffer, length);
2730 	if (!event)
2731 		goto out;
2732 
2733 	return event;
2734 
2735  out:
2736 	trace_recursive_unlock();
2737 
2738  out_nocheck:
2739 	preempt_enable_notrace();
2740 	return NULL;
2741 }
2742 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
2743 
2744 static void
2745 rb_update_write_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2746 		      struct ring_buffer_event *event)
2747 {
2748 	u64 delta;
2749 
2750 	/*
2751 	 * The event first in the commit queue updates the
2752 	 * time stamp.
2753 	 */
2754 	if (rb_event_is_commit(cpu_buffer, event)) {
2755 		/*
2756 		 * A commit event that is first on a page
2757 		 * updates the write timestamp with the page stamp
2758 		 */
2759 		if (!rb_event_index(event))
2760 			cpu_buffer->write_stamp =
2761 				cpu_buffer->commit_page->page->time_stamp;
2762 		else if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
2763 			delta = event->array[0];
2764 			delta <<= TS_SHIFT;
2765 			delta += event->time_delta;
2766 			cpu_buffer->write_stamp += delta;
2767 		} else
2768 			cpu_buffer->write_stamp += event->time_delta;
2769 	}
2770 }
2771 
2772 static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
2773 		      struct ring_buffer_event *event)
2774 {
2775 	local_inc(&cpu_buffer->entries);
2776 	rb_update_write_stamp(cpu_buffer, event);
2777 	rb_end_commit(cpu_buffer);
2778 }
2779 
2780 static __always_inline void
2781 rb_wakeups(struct ring_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer)
2782 {
2783 	if (buffer->irq_work.waiters_pending) {
2784 		buffer->irq_work.waiters_pending = false;
2785 		/* irq_work_queue() supplies it's own memory barriers */
2786 		irq_work_queue(&buffer->irq_work.work);
2787 	}
2788 
2789 	if (cpu_buffer->irq_work.waiters_pending) {
2790 		cpu_buffer->irq_work.waiters_pending = false;
2791 		/* irq_work_queue() supplies it's own memory barriers */
2792 		irq_work_queue(&cpu_buffer->irq_work.work);
2793 	}
2794 }
2795 
2796 /**
2797  * ring_buffer_unlock_commit - commit a reserved
2798  * @buffer: The buffer to commit to
2799  * @event: The event pointer to commit.
2800  *
2801  * This commits the data to the ring buffer, and releases any locks held.
2802  *
2803  * Must be paired with ring_buffer_lock_reserve.
2804  */
2805 int ring_buffer_unlock_commit(struct ring_buffer *buffer,
2806 			      struct ring_buffer_event *event)
2807 {
2808 	struct ring_buffer_per_cpu *cpu_buffer;
2809 	int cpu = raw_smp_processor_id();
2810 
2811 	cpu_buffer = buffer->buffers[cpu];
2812 
2813 	rb_commit(cpu_buffer, event);
2814 
2815 	rb_wakeups(buffer, cpu_buffer);
2816 
2817 	trace_recursive_unlock();
2818 
2819 	preempt_enable_notrace();
2820 
2821 	return 0;
2822 }
2823 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
2824 
2825 static inline void rb_event_discard(struct ring_buffer_event *event)
2826 {
2827 	if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
2828 		event = skip_time_extend(event);
2829 
2830 	/* array[0] holds the actual length for the discarded event */
2831 	event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
2832 	event->type_len = RINGBUF_TYPE_PADDING;
2833 	/* time delta must be non zero */
2834 	if (!event->time_delta)
2835 		event->time_delta = 1;
2836 }
2837 
2838 /*
2839  * Decrement the entries to the page that an event is on.
2840  * The event does not even need to exist, only the pointer
2841  * to the page it is on. This may only be called before the commit
2842  * takes place.
2843  */
2844 static inline void
2845 rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
2846 		   struct ring_buffer_event *event)
2847 {
2848 	unsigned long addr = (unsigned long)event;
2849 	struct buffer_page *bpage = cpu_buffer->commit_page;
2850 	struct buffer_page *start;
2851 
2852 	addr &= PAGE_MASK;
2853 
2854 	/* Do the likely case first */
2855 	if (likely(bpage->page == (void *)addr)) {
2856 		local_dec(&bpage->entries);
2857 		return;
2858 	}
2859 
2860 	/*
2861 	 * Because the commit page may be on the reader page we
2862 	 * start with the next page and check the end loop there.
2863 	 */
2864 	rb_inc_page(cpu_buffer, &bpage);
2865 	start = bpage;
2866 	do {
2867 		if (bpage->page == (void *)addr) {
2868 			local_dec(&bpage->entries);
2869 			return;
2870 		}
2871 		rb_inc_page(cpu_buffer, &bpage);
2872 	} while (bpage != start);
2873 
2874 	/* commit not part of this buffer?? */
2875 	RB_WARN_ON(cpu_buffer, 1);
2876 }
2877 
2878 /**
2879  * ring_buffer_commit_discard - discard an event that has not been committed
2880  * @buffer: the ring buffer
2881  * @event: non committed event to discard
2882  *
2883  * Sometimes an event that is in the ring buffer needs to be ignored.
2884  * This function lets the user discard an event in the ring buffer
2885  * and then that event will not be read later.
2886  *
2887  * This function only works if it is called before the the item has been
2888  * committed. It will try to free the event from the ring buffer
2889  * if another event has not been added behind it.
2890  *
2891  * If another event has been added behind it, it will set the event
2892  * up as discarded, and perform the commit.
2893  *
2894  * If this function is called, do not call ring_buffer_unlock_commit on
2895  * the event.
2896  */
2897 void ring_buffer_discard_commit(struct ring_buffer *buffer,
2898 				struct ring_buffer_event *event)
2899 {
2900 	struct ring_buffer_per_cpu *cpu_buffer;
2901 	int cpu;
2902 
2903 	/* The event is discarded regardless */
2904 	rb_event_discard(event);
2905 
2906 	cpu = smp_processor_id();
2907 	cpu_buffer = buffer->buffers[cpu];
2908 
2909 	/*
2910 	 * This must only be called if the event has not been
2911 	 * committed yet. Thus we can assume that preemption
2912 	 * is still disabled.
2913 	 */
2914 	RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
2915 
2916 	rb_decrement_entry(cpu_buffer, event);
2917 	if (rb_try_to_discard(cpu_buffer, event))
2918 		goto out;
2919 
2920 	/*
2921 	 * The commit is still visible by the reader, so we
2922 	 * must still update the timestamp.
2923 	 */
2924 	rb_update_write_stamp(cpu_buffer, event);
2925  out:
2926 	rb_end_commit(cpu_buffer);
2927 
2928 	trace_recursive_unlock();
2929 
2930 	preempt_enable_notrace();
2931 
2932 }
2933 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
2934 
2935 /**
2936  * ring_buffer_write - write data to the buffer without reserving
2937  * @buffer: The ring buffer to write to.
2938  * @length: The length of the data being written (excluding the event header)
2939  * @data: The data to write to the buffer.
2940  *
2941  * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
2942  * one function. If you already have the data to write to the buffer, it
2943  * may be easier to simply call this function.
2944  *
2945  * Note, like ring_buffer_lock_reserve, the length is the length of the data
2946  * and not the length of the event which would hold the header.
2947  */
2948 int ring_buffer_write(struct ring_buffer *buffer,
2949 		      unsigned long length,
2950 		      void *data)
2951 {
2952 	struct ring_buffer_per_cpu *cpu_buffer;
2953 	struct ring_buffer_event *event;
2954 	void *body;
2955 	int ret = -EBUSY;
2956 	int cpu;
2957 
2958 	if (ring_buffer_flags != RB_BUFFERS_ON)
2959 		return -EBUSY;
2960 
2961 	preempt_disable_notrace();
2962 
2963 	if (atomic_read(&buffer->record_disabled))
2964 		goto out;
2965 
2966 	cpu = raw_smp_processor_id();
2967 
2968 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
2969 		goto out;
2970 
2971 	cpu_buffer = buffer->buffers[cpu];
2972 
2973 	if (atomic_read(&cpu_buffer->record_disabled))
2974 		goto out;
2975 
2976 	if (length > BUF_MAX_DATA_SIZE)
2977 		goto out;
2978 
2979 	event = rb_reserve_next_event(buffer, cpu_buffer, length);
2980 	if (!event)
2981 		goto out;
2982 
2983 	body = rb_event_data(event);
2984 
2985 	memcpy(body, data, length);
2986 
2987 	rb_commit(cpu_buffer, event);
2988 
2989 	rb_wakeups(buffer, cpu_buffer);
2990 
2991 	ret = 0;
2992  out:
2993 	preempt_enable_notrace();
2994 
2995 	return ret;
2996 }
2997 EXPORT_SYMBOL_GPL(ring_buffer_write);
2998 
2999 static int rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
3000 {
3001 	struct buffer_page *reader = cpu_buffer->reader_page;
3002 	struct buffer_page *head = rb_set_head_page(cpu_buffer);
3003 	struct buffer_page *commit = cpu_buffer->commit_page;
3004 
3005 	/* In case of error, head will be NULL */
3006 	if (unlikely(!head))
3007 		return 1;
3008 
3009 	return reader->read == rb_page_commit(reader) &&
3010 		(commit == reader ||
3011 		 (commit == head &&
3012 		  head->read == rb_page_commit(commit)));
3013 }
3014 
3015 /**
3016  * ring_buffer_record_disable - stop all writes into the buffer
3017  * @buffer: The ring buffer to stop writes to.
3018  *
3019  * This prevents all writes to the buffer. Any attempt to write
3020  * to the buffer after this will fail and return NULL.
3021  *
3022  * The caller should call synchronize_sched() after this.
3023  */
3024 void ring_buffer_record_disable(struct ring_buffer *buffer)
3025 {
3026 	atomic_inc(&buffer->record_disabled);
3027 }
3028 EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
3029 
3030 /**
3031  * ring_buffer_record_enable - enable writes to the buffer
3032  * @buffer: The ring buffer to enable writes
3033  *
3034  * Note, multiple disables will need the same number of enables
3035  * to truly enable the writing (much like preempt_disable).
3036  */
3037 void ring_buffer_record_enable(struct ring_buffer *buffer)
3038 {
3039 	atomic_dec(&buffer->record_disabled);
3040 }
3041 EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
3042 
3043 /**
3044  * ring_buffer_record_off - stop all writes into the buffer
3045  * @buffer: The ring buffer to stop writes to.
3046  *
3047  * This prevents all writes to the buffer. Any attempt to write
3048  * to the buffer after this will fail and return NULL.
3049  *
3050  * This is different than ring_buffer_record_disable() as
3051  * it works like an on/off switch, where as the disable() version
3052  * must be paired with a enable().
3053  */
3054 void ring_buffer_record_off(struct ring_buffer *buffer)
3055 {
3056 	unsigned int rd;
3057 	unsigned int new_rd;
3058 
3059 	do {
3060 		rd = atomic_read(&buffer->record_disabled);
3061 		new_rd = rd | RB_BUFFER_OFF;
3062 	} while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3063 }
3064 EXPORT_SYMBOL_GPL(ring_buffer_record_off);
3065 
3066 /**
3067  * ring_buffer_record_on - restart writes into the buffer
3068  * @buffer: The ring buffer to start writes to.
3069  *
3070  * This enables all writes to the buffer that was disabled by
3071  * ring_buffer_record_off().
3072  *
3073  * This is different than ring_buffer_record_enable() as
3074  * it works like an on/off switch, where as the enable() version
3075  * must be paired with a disable().
3076  */
3077 void ring_buffer_record_on(struct ring_buffer *buffer)
3078 {
3079 	unsigned int rd;
3080 	unsigned int new_rd;
3081 
3082 	do {
3083 		rd = atomic_read(&buffer->record_disabled);
3084 		new_rd = rd & ~RB_BUFFER_OFF;
3085 	} while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3086 }
3087 EXPORT_SYMBOL_GPL(ring_buffer_record_on);
3088 
3089 /**
3090  * ring_buffer_record_is_on - return true if the ring buffer can write
3091  * @buffer: The ring buffer to see if write is enabled
3092  *
3093  * Returns true if the ring buffer is in a state that it accepts writes.
3094  */
3095 int ring_buffer_record_is_on(struct ring_buffer *buffer)
3096 {
3097 	return !atomic_read(&buffer->record_disabled);
3098 }
3099 
3100 /**
3101  * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
3102  * @buffer: The ring buffer to stop writes to.
3103  * @cpu: The CPU buffer to stop
3104  *
3105  * This prevents all writes to the buffer. Any attempt to write
3106  * to the buffer after this will fail and return NULL.
3107  *
3108  * The caller should call synchronize_sched() after this.
3109  */
3110 void ring_buffer_record_disable_cpu(struct ring_buffer *buffer, int cpu)
3111 {
3112 	struct ring_buffer_per_cpu *cpu_buffer;
3113 
3114 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3115 		return;
3116 
3117 	cpu_buffer = buffer->buffers[cpu];
3118 	atomic_inc(&cpu_buffer->record_disabled);
3119 }
3120 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
3121 
3122 /**
3123  * ring_buffer_record_enable_cpu - enable writes to the buffer
3124  * @buffer: The ring buffer to enable writes
3125  * @cpu: The CPU to enable.
3126  *
3127  * Note, multiple disables will need the same number of enables
3128  * to truly enable the writing (much like preempt_disable).
3129  */
3130 void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu)
3131 {
3132 	struct ring_buffer_per_cpu *cpu_buffer;
3133 
3134 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3135 		return;
3136 
3137 	cpu_buffer = buffer->buffers[cpu];
3138 	atomic_dec(&cpu_buffer->record_disabled);
3139 }
3140 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
3141 
3142 /*
3143  * The total entries in the ring buffer is the running counter
3144  * of entries entered into the ring buffer, minus the sum of
3145  * the entries read from the ring buffer and the number of
3146  * entries that were overwritten.
3147  */
3148 static inline unsigned long
3149 rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
3150 {
3151 	return local_read(&cpu_buffer->entries) -
3152 		(local_read(&cpu_buffer->overrun) + cpu_buffer->read);
3153 }
3154 
3155 /**
3156  * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
3157  * @buffer: The ring buffer
3158  * @cpu: The per CPU buffer to read from.
3159  */
3160 u64 ring_buffer_oldest_event_ts(struct ring_buffer *buffer, int cpu)
3161 {
3162 	unsigned long flags;
3163 	struct ring_buffer_per_cpu *cpu_buffer;
3164 	struct buffer_page *bpage;
3165 	u64 ret = 0;
3166 
3167 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3168 		return 0;
3169 
3170 	cpu_buffer = buffer->buffers[cpu];
3171 	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3172 	/*
3173 	 * if the tail is on reader_page, oldest time stamp is on the reader
3174 	 * page
3175 	 */
3176 	if (cpu_buffer->tail_page == cpu_buffer->reader_page)
3177 		bpage = cpu_buffer->reader_page;
3178 	else
3179 		bpage = rb_set_head_page(cpu_buffer);
3180 	if (bpage)
3181 		ret = bpage->page->time_stamp;
3182 	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3183 
3184 	return ret;
3185 }
3186 EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
3187 
3188 /**
3189  * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
3190  * @buffer: The ring buffer
3191  * @cpu: The per CPU buffer to read from.
3192  */
3193 unsigned long ring_buffer_bytes_cpu(struct ring_buffer *buffer, int cpu)
3194 {
3195 	struct ring_buffer_per_cpu *cpu_buffer;
3196 	unsigned long ret;
3197 
3198 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3199 		return 0;
3200 
3201 	cpu_buffer = buffer->buffers[cpu];
3202 	ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
3203 
3204 	return ret;
3205 }
3206 EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
3207 
3208 /**
3209  * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
3210  * @buffer: The ring buffer
3211  * @cpu: The per CPU buffer to get the entries from.
3212  */
3213 unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu)
3214 {
3215 	struct ring_buffer_per_cpu *cpu_buffer;
3216 
3217 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3218 		return 0;
3219 
3220 	cpu_buffer = buffer->buffers[cpu];
3221 
3222 	return rb_num_of_entries(cpu_buffer);
3223 }
3224 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
3225 
3226 /**
3227  * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
3228  * buffer wrapping around (only if RB_FL_OVERWRITE is on).
3229  * @buffer: The ring buffer
3230  * @cpu: The per CPU buffer to get the number of overruns from
3231  */
3232 unsigned long ring_buffer_overrun_cpu(struct ring_buffer *buffer, int cpu)
3233 {
3234 	struct ring_buffer_per_cpu *cpu_buffer;
3235 	unsigned long ret;
3236 
3237 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3238 		return 0;
3239 
3240 	cpu_buffer = buffer->buffers[cpu];
3241 	ret = local_read(&cpu_buffer->overrun);
3242 
3243 	return ret;
3244 }
3245 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
3246 
3247 /**
3248  * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
3249  * commits failing due to the buffer wrapping around while there are uncommitted
3250  * events, such as during an interrupt storm.
3251  * @buffer: The ring buffer
3252  * @cpu: The per CPU buffer to get the number of overruns from
3253  */
3254 unsigned long
3255 ring_buffer_commit_overrun_cpu(struct ring_buffer *buffer, int cpu)
3256 {
3257 	struct ring_buffer_per_cpu *cpu_buffer;
3258 	unsigned long ret;
3259 
3260 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3261 		return 0;
3262 
3263 	cpu_buffer = buffer->buffers[cpu];
3264 	ret = local_read(&cpu_buffer->commit_overrun);
3265 
3266 	return ret;
3267 }
3268 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
3269 
3270 /**
3271  * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
3272  * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
3273  * @buffer: The ring buffer
3274  * @cpu: The per CPU buffer to get the number of overruns from
3275  */
3276 unsigned long
3277 ring_buffer_dropped_events_cpu(struct ring_buffer *buffer, int cpu)
3278 {
3279 	struct ring_buffer_per_cpu *cpu_buffer;
3280 	unsigned long ret;
3281 
3282 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3283 		return 0;
3284 
3285 	cpu_buffer = buffer->buffers[cpu];
3286 	ret = local_read(&cpu_buffer->dropped_events);
3287 
3288 	return ret;
3289 }
3290 EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
3291 
3292 /**
3293  * ring_buffer_read_events_cpu - get the number of events successfully read
3294  * @buffer: The ring buffer
3295  * @cpu: The per CPU buffer to get the number of events read
3296  */
3297 unsigned long
3298 ring_buffer_read_events_cpu(struct ring_buffer *buffer, int cpu)
3299 {
3300 	struct ring_buffer_per_cpu *cpu_buffer;
3301 
3302 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3303 		return 0;
3304 
3305 	cpu_buffer = buffer->buffers[cpu];
3306 	return cpu_buffer->read;
3307 }
3308 EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu);
3309 
3310 /**
3311  * ring_buffer_entries - get the number of entries in a buffer
3312  * @buffer: The ring buffer
3313  *
3314  * Returns the total number of entries in the ring buffer
3315  * (all CPU entries)
3316  */
3317 unsigned long ring_buffer_entries(struct ring_buffer *buffer)
3318 {
3319 	struct ring_buffer_per_cpu *cpu_buffer;
3320 	unsigned long entries = 0;
3321 	int cpu;
3322 
3323 	/* if you care about this being correct, lock the buffer */
3324 	for_each_buffer_cpu(buffer, cpu) {
3325 		cpu_buffer = buffer->buffers[cpu];
3326 		entries += rb_num_of_entries(cpu_buffer);
3327 	}
3328 
3329 	return entries;
3330 }
3331 EXPORT_SYMBOL_GPL(ring_buffer_entries);
3332 
3333 /**
3334  * ring_buffer_overruns - get the number of overruns in buffer
3335  * @buffer: The ring buffer
3336  *
3337  * Returns the total number of overruns in the ring buffer
3338  * (all CPU entries)
3339  */
3340 unsigned long ring_buffer_overruns(struct ring_buffer *buffer)
3341 {
3342 	struct ring_buffer_per_cpu *cpu_buffer;
3343 	unsigned long overruns = 0;
3344 	int cpu;
3345 
3346 	/* if you care about this being correct, lock the buffer */
3347 	for_each_buffer_cpu(buffer, cpu) {
3348 		cpu_buffer = buffer->buffers[cpu];
3349 		overruns += local_read(&cpu_buffer->overrun);
3350 	}
3351 
3352 	return overruns;
3353 }
3354 EXPORT_SYMBOL_GPL(ring_buffer_overruns);
3355 
3356 static void rb_iter_reset(struct ring_buffer_iter *iter)
3357 {
3358 	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3359 
3360 	/* Iterator usage is expected to have record disabled */
3361 	if (list_empty(&cpu_buffer->reader_page->list)) {
3362 		iter->head_page = rb_set_head_page(cpu_buffer);
3363 		if (unlikely(!iter->head_page))
3364 			return;
3365 		iter->head = iter->head_page->read;
3366 	} else {
3367 		iter->head_page = cpu_buffer->reader_page;
3368 		iter->head = cpu_buffer->reader_page->read;
3369 	}
3370 	if (iter->head)
3371 		iter->read_stamp = cpu_buffer->read_stamp;
3372 	else
3373 		iter->read_stamp = iter->head_page->page->time_stamp;
3374 	iter->cache_reader_page = cpu_buffer->reader_page;
3375 	iter->cache_read = cpu_buffer->read;
3376 }
3377 
3378 /**
3379  * ring_buffer_iter_reset - reset an iterator
3380  * @iter: The iterator to reset
3381  *
3382  * Resets the iterator, so that it will start from the beginning
3383  * again.
3384  */
3385 void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
3386 {
3387 	struct ring_buffer_per_cpu *cpu_buffer;
3388 	unsigned long flags;
3389 
3390 	if (!iter)
3391 		return;
3392 
3393 	cpu_buffer = iter->cpu_buffer;
3394 
3395 	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3396 	rb_iter_reset(iter);
3397 	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3398 }
3399 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
3400 
3401 /**
3402  * ring_buffer_iter_empty - check if an iterator has no more to read
3403  * @iter: The iterator to check
3404  */
3405 int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
3406 {
3407 	struct ring_buffer_per_cpu *cpu_buffer;
3408 
3409 	cpu_buffer = iter->cpu_buffer;
3410 
3411 	return iter->head_page == cpu_buffer->commit_page &&
3412 		iter->head == rb_commit_index(cpu_buffer);
3413 }
3414 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
3415 
3416 static void
3417 rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
3418 		     struct ring_buffer_event *event)
3419 {
3420 	u64 delta;
3421 
3422 	switch (event->type_len) {
3423 	case RINGBUF_TYPE_PADDING:
3424 		return;
3425 
3426 	case RINGBUF_TYPE_TIME_EXTEND:
3427 		delta = event->array[0];
3428 		delta <<= TS_SHIFT;
3429 		delta += event->time_delta;
3430 		cpu_buffer->read_stamp += delta;
3431 		return;
3432 
3433 	case RINGBUF_TYPE_TIME_STAMP:
3434 		/* FIXME: not implemented */
3435 		return;
3436 
3437 	case RINGBUF_TYPE_DATA:
3438 		cpu_buffer->read_stamp += event->time_delta;
3439 		return;
3440 
3441 	default:
3442 		BUG();
3443 	}
3444 	return;
3445 }
3446 
3447 static void
3448 rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
3449 			  struct ring_buffer_event *event)
3450 {
3451 	u64 delta;
3452 
3453 	switch (event->type_len) {
3454 	case RINGBUF_TYPE_PADDING:
3455 		return;
3456 
3457 	case RINGBUF_TYPE_TIME_EXTEND:
3458 		delta = event->array[0];
3459 		delta <<= TS_SHIFT;
3460 		delta += event->time_delta;
3461 		iter->read_stamp += delta;
3462 		return;
3463 
3464 	case RINGBUF_TYPE_TIME_STAMP:
3465 		/* FIXME: not implemented */
3466 		return;
3467 
3468 	case RINGBUF_TYPE_DATA:
3469 		iter->read_stamp += event->time_delta;
3470 		return;
3471 
3472 	default:
3473 		BUG();
3474 	}
3475 	return;
3476 }
3477 
3478 static struct buffer_page *
3479 rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
3480 {
3481 	struct buffer_page *reader = NULL;
3482 	unsigned long overwrite;
3483 	unsigned long flags;
3484 	int nr_loops = 0;
3485 	int ret;
3486 
3487 	local_irq_save(flags);
3488 	arch_spin_lock(&cpu_buffer->lock);
3489 
3490  again:
3491 	/*
3492 	 * This should normally only loop twice. But because the
3493 	 * start of the reader inserts an empty page, it causes
3494 	 * a case where we will loop three times. There should be no
3495 	 * reason to loop four times (that I know of).
3496 	 */
3497 	if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
3498 		reader = NULL;
3499 		goto out;
3500 	}
3501 
3502 	reader = cpu_buffer->reader_page;
3503 
3504 	/* If there's more to read, return this page */
3505 	if (cpu_buffer->reader_page->read < rb_page_size(reader))
3506 		goto out;
3507 
3508 	/* Never should we have an index greater than the size */
3509 	if (RB_WARN_ON(cpu_buffer,
3510 		       cpu_buffer->reader_page->read > rb_page_size(reader)))
3511 		goto out;
3512 
3513 	/* check if we caught up to the tail */
3514 	reader = NULL;
3515 	if (cpu_buffer->commit_page == cpu_buffer->reader_page)
3516 		goto out;
3517 
3518 	/* Don't bother swapping if the ring buffer is empty */
3519 	if (rb_num_of_entries(cpu_buffer) == 0)
3520 		goto out;
3521 
3522 	/*
3523 	 * Reset the reader page to size zero.
3524 	 */
3525 	local_set(&cpu_buffer->reader_page->write, 0);
3526 	local_set(&cpu_buffer->reader_page->entries, 0);
3527 	local_set(&cpu_buffer->reader_page->page->commit, 0);
3528 	cpu_buffer->reader_page->real_end = 0;
3529 
3530  spin:
3531 	/*
3532 	 * Splice the empty reader page into the list around the head.
3533 	 */
3534 	reader = rb_set_head_page(cpu_buffer);
3535 	if (!reader)
3536 		goto out;
3537 	cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
3538 	cpu_buffer->reader_page->list.prev = reader->list.prev;
3539 
3540 	/*
3541 	 * cpu_buffer->pages just needs to point to the buffer, it
3542 	 *  has no specific buffer page to point to. Lets move it out
3543 	 *  of our way so we don't accidentally swap it.
3544 	 */
3545 	cpu_buffer->pages = reader->list.prev;
3546 
3547 	/* The reader page will be pointing to the new head */
3548 	rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list);
3549 
3550 	/*
3551 	 * We want to make sure we read the overruns after we set up our
3552 	 * pointers to the next object. The writer side does a
3553 	 * cmpxchg to cross pages which acts as the mb on the writer
3554 	 * side. Note, the reader will constantly fail the swap
3555 	 * while the writer is updating the pointers, so this
3556 	 * guarantees that the overwrite recorded here is the one we
3557 	 * want to compare with the last_overrun.
3558 	 */
3559 	smp_mb();
3560 	overwrite = local_read(&(cpu_buffer->overrun));
3561 
3562 	/*
3563 	 * Here's the tricky part.
3564 	 *
3565 	 * We need to move the pointer past the header page.
3566 	 * But we can only do that if a writer is not currently
3567 	 * moving it. The page before the header page has the
3568 	 * flag bit '1' set if it is pointing to the page we want.
3569 	 * but if the writer is in the process of moving it
3570 	 * than it will be '2' or already moved '0'.
3571 	 */
3572 
3573 	ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
3574 
3575 	/*
3576 	 * If we did not convert it, then we must try again.
3577 	 */
3578 	if (!ret)
3579 		goto spin;
3580 
3581 	/*
3582 	 * Yeah! We succeeded in replacing the page.
3583 	 *
3584 	 * Now make the new head point back to the reader page.
3585 	 */
3586 	rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
3587 	rb_inc_page(cpu_buffer, &cpu_buffer->head_page);
3588 
3589 	/* Finally update the reader page to the new head */
3590 	cpu_buffer->reader_page = reader;
3591 	rb_reset_reader_page(cpu_buffer);
3592 
3593 	if (overwrite != cpu_buffer->last_overrun) {
3594 		cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
3595 		cpu_buffer->last_overrun = overwrite;
3596 	}
3597 
3598 	goto again;
3599 
3600  out:
3601 	arch_spin_unlock(&cpu_buffer->lock);
3602 	local_irq_restore(flags);
3603 
3604 	return reader;
3605 }
3606 
3607 static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
3608 {
3609 	struct ring_buffer_event *event;
3610 	struct buffer_page *reader;
3611 	unsigned length;
3612 
3613 	reader = rb_get_reader_page(cpu_buffer);
3614 
3615 	/* This function should not be called when buffer is empty */
3616 	if (RB_WARN_ON(cpu_buffer, !reader))
3617 		return;
3618 
3619 	event = rb_reader_event(cpu_buffer);
3620 
3621 	if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
3622 		cpu_buffer->read++;
3623 
3624 	rb_update_read_stamp(cpu_buffer, event);
3625 
3626 	length = rb_event_length(event);
3627 	cpu_buffer->reader_page->read += length;
3628 }
3629 
3630 static void rb_advance_iter(struct ring_buffer_iter *iter)
3631 {
3632 	struct ring_buffer_per_cpu *cpu_buffer;
3633 	struct ring_buffer_event *event;
3634 	unsigned length;
3635 
3636 	cpu_buffer = iter->cpu_buffer;
3637 
3638 	/*
3639 	 * Check if we are at the end of the buffer.
3640 	 */
3641 	if (iter->head >= rb_page_size(iter->head_page)) {
3642 		/* discarded commits can make the page empty */
3643 		if (iter->head_page == cpu_buffer->commit_page)
3644 			return;
3645 		rb_inc_iter(iter);
3646 		return;
3647 	}
3648 
3649 	event = rb_iter_head_event(iter);
3650 
3651 	length = rb_event_length(event);
3652 
3653 	/*
3654 	 * This should not be called to advance the header if we are
3655 	 * at the tail of the buffer.
3656 	 */
3657 	if (RB_WARN_ON(cpu_buffer,
3658 		       (iter->head_page == cpu_buffer->commit_page) &&
3659 		       (iter->head + length > rb_commit_index(cpu_buffer))))
3660 		return;
3661 
3662 	rb_update_iter_read_stamp(iter, event);
3663 
3664 	iter->head += length;
3665 
3666 	/* check for end of page padding */
3667 	if ((iter->head >= rb_page_size(iter->head_page)) &&
3668 	    (iter->head_page != cpu_buffer->commit_page))
3669 		rb_inc_iter(iter);
3670 }
3671 
3672 static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
3673 {
3674 	return cpu_buffer->lost_events;
3675 }
3676 
3677 static struct ring_buffer_event *
3678 rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
3679 	       unsigned long *lost_events)
3680 {
3681 	struct ring_buffer_event *event;
3682 	struct buffer_page *reader;
3683 	int nr_loops = 0;
3684 
3685  again:
3686 	/*
3687 	 * We repeat when a time extend is encountered.
3688 	 * Since the time extend is always attached to a data event,
3689 	 * we should never loop more than once.
3690 	 * (We never hit the following condition more than twice).
3691 	 */
3692 	if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3693 		return NULL;
3694 
3695 	reader = rb_get_reader_page(cpu_buffer);
3696 	if (!reader)
3697 		return NULL;
3698 
3699 	event = rb_reader_event(cpu_buffer);
3700 
3701 	switch (event->type_len) {
3702 	case RINGBUF_TYPE_PADDING:
3703 		if (rb_null_event(event))
3704 			RB_WARN_ON(cpu_buffer, 1);
3705 		/*
3706 		 * Because the writer could be discarding every
3707 		 * event it creates (which would probably be bad)
3708 		 * if we were to go back to "again" then we may never
3709 		 * catch up, and will trigger the warn on, or lock
3710 		 * the box. Return the padding, and we will release
3711 		 * the current locks, and try again.
3712 		 */
3713 		return event;
3714 
3715 	case RINGBUF_TYPE_TIME_EXTEND:
3716 		/* Internal data, OK to advance */
3717 		rb_advance_reader(cpu_buffer);
3718 		goto again;
3719 
3720 	case RINGBUF_TYPE_TIME_STAMP:
3721 		/* FIXME: not implemented */
3722 		rb_advance_reader(cpu_buffer);
3723 		goto again;
3724 
3725 	case RINGBUF_TYPE_DATA:
3726 		if (ts) {
3727 			*ts = cpu_buffer->read_stamp + event->time_delta;
3728 			ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3729 							 cpu_buffer->cpu, ts);
3730 		}
3731 		if (lost_events)
3732 			*lost_events = rb_lost_events(cpu_buffer);
3733 		return event;
3734 
3735 	default:
3736 		BUG();
3737 	}
3738 
3739 	return NULL;
3740 }
3741 EXPORT_SYMBOL_GPL(ring_buffer_peek);
3742 
3743 static struct ring_buffer_event *
3744 rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3745 {
3746 	struct ring_buffer *buffer;
3747 	struct ring_buffer_per_cpu *cpu_buffer;
3748 	struct ring_buffer_event *event;
3749 	int nr_loops = 0;
3750 
3751 	cpu_buffer = iter->cpu_buffer;
3752 	buffer = cpu_buffer->buffer;
3753 
3754 	/*
3755 	 * Check if someone performed a consuming read to
3756 	 * the buffer. A consuming read invalidates the iterator
3757 	 * and we need to reset the iterator in this case.
3758 	 */
3759 	if (unlikely(iter->cache_read != cpu_buffer->read ||
3760 		     iter->cache_reader_page != cpu_buffer->reader_page))
3761 		rb_iter_reset(iter);
3762 
3763  again:
3764 	if (ring_buffer_iter_empty(iter))
3765 		return NULL;
3766 
3767 	/*
3768 	 * We repeat when a time extend is encountered.
3769 	 * Since the time extend is always attached to a data event,
3770 	 * we should never loop more than once.
3771 	 * (We never hit the following condition more than twice).
3772 	 */
3773 	if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3774 		return NULL;
3775 
3776 	if (rb_per_cpu_empty(cpu_buffer))
3777 		return NULL;
3778 
3779 	if (iter->head >= local_read(&iter->head_page->page->commit)) {
3780 		rb_inc_iter(iter);
3781 		goto again;
3782 	}
3783 
3784 	event = rb_iter_head_event(iter);
3785 
3786 	switch (event->type_len) {
3787 	case RINGBUF_TYPE_PADDING:
3788 		if (rb_null_event(event)) {
3789 			rb_inc_iter(iter);
3790 			goto again;
3791 		}
3792 		rb_advance_iter(iter);
3793 		return event;
3794 
3795 	case RINGBUF_TYPE_TIME_EXTEND:
3796 		/* Internal data, OK to advance */
3797 		rb_advance_iter(iter);
3798 		goto again;
3799 
3800 	case RINGBUF_TYPE_TIME_STAMP:
3801 		/* FIXME: not implemented */
3802 		rb_advance_iter(iter);
3803 		goto again;
3804 
3805 	case RINGBUF_TYPE_DATA:
3806 		if (ts) {
3807 			*ts = iter->read_stamp + event->time_delta;
3808 			ring_buffer_normalize_time_stamp(buffer,
3809 							 cpu_buffer->cpu, ts);
3810 		}
3811 		return event;
3812 
3813 	default:
3814 		BUG();
3815 	}
3816 
3817 	return NULL;
3818 }
3819 EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
3820 
3821 static inline int rb_ok_to_lock(void)
3822 {
3823 	/*
3824 	 * If an NMI die dumps out the content of the ring buffer
3825 	 * do not grab locks. We also permanently disable the ring
3826 	 * buffer too. A one time deal is all you get from reading
3827 	 * the ring buffer from an NMI.
3828 	 */
3829 	if (likely(!in_nmi()))
3830 		return 1;
3831 
3832 	tracing_off_permanent();
3833 	return 0;
3834 }
3835 
3836 /**
3837  * ring_buffer_peek - peek at the next event to be read
3838  * @buffer: The ring buffer to read
3839  * @cpu: The cpu to peak at
3840  * @ts: The timestamp counter of this event.
3841  * @lost_events: a variable to store if events were lost (may be NULL)
3842  *
3843  * This will return the event that will be read next, but does
3844  * not consume the data.
3845  */
3846 struct ring_buffer_event *
3847 ring_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts,
3848 		 unsigned long *lost_events)
3849 {
3850 	struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3851 	struct ring_buffer_event *event;
3852 	unsigned long flags;
3853 	int dolock;
3854 
3855 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3856 		return NULL;
3857 
3858 	dolock = rb_ok_to_lock();
3859  again:
3860 	local_irq_save(flags);
3861 	if (dolock)
3862 		raw_spin_lock(&cpu_buffer->reader_lock);
3863 	event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3864 	if (event && event->type_len == RINGBUF_TYPE_PADDING)
3865 		rb_advance_reader(cpu_buffer);
3866 	if (dolock)
3867 		raw_spin_unlock(&cpu_buffer->reader_lock);
3868 	local_irq_restore(flags);
3869 
3870 	if (event && event->type_len == RINGBUF_TYPE_PADDING)
3871 		goto again;
3872 
3873 	return event;
3874 }
3875 
3876 /**
3877  * ring_buffer_iter_peek - peek at the next event to be read
3878  * @iter: The ring buffer iterator
3879  * @ts: The timestamp counter of this event.
3880  *
3881  * This will return the event that will be read next, but does
3882  * not increment the iterator.
3883  */
3884 struct ring_buffer_event *
3885 ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3886 {
3887 	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3888 	struct ring_buffer_event *event;
3889 	unsigned long flags;
3890 
3891  again:
3892 	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3893 	event = rb_iter_peek(iter, ts);
3894 	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3895 
3896 	if (event && event->type_len == RINGBUF_TYPE_PADDING)
3897 		goto again;
3898 
3899 	return event;
3900 }
3901 
3902 /**
3903  * ring_buffer_consume - return an event and consume it
3904  * @buffer: The ring buffer to get the next event from
3905  * @cpu: the cpu to read the buffer from
3906  * @ts: a variable to store the timestamp (may be NULL)
3907  * @lost_events: a variable to store if events were lost (may be NULL)
3908  *
3909  * Returns the next event in the ring buffer, and that event is consumed.
3910  * Meaning, that sequential reads will keep returning a different event,
3911  * and eventually empty the ring buffer if the producer is slower.
3912  */
3913 struct ring_buffer_event *
3914 ring_buffer_consume(struct ring_buffer *buffer, int cpu, u64 *ts,
3915 		    unsigned long *lost_events)
3916 {
3917 	struct ring_buffer_per_cpu *cpu_buffer;
3918 	struct ring_buffer_event *event = NULL;
3919 	unsigned long flags;
3920 	int dolock;
3921 
3922 	dolock = rb_ok_to_lock();
3923 
3924  again:
3925 	/* might be called in atomic */
3926 	preempt_disable();
3927 
3928 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3929 		goto out;
3930 
3931 	cpu_buffer = buffer->buffers[cpu];
3932 	local_irq_save(flags);
3933 	if (dolock)
3934 		raw_spin_lock(&cpu_buffer->reader_lock);
3935 
3936 	event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3937 	if (event) {
3938 		cpu_buffer->lost_events = 0;
3939 		rb_advance_reader(cpu_buffer);
3940 	}
3941 
3942 	if (dolock)
3943 		raw_spin_unlock(&cpu_buffer->reader_lock);
3944 	local_irq_restore(flags);
3945 
3946  out:
3947 	preempt_enable();
3948 
3949 	if (event && event->type_len == RINGBUF_TYPE_PADDING)
3950 		goto again;
3951 
3952 	return event;
3953 }
3954 EXPORT_SYMBOL_GPL(ring_buffer_consume);
3955 
3956 /**
3957  * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
3958  * @buffer: The ring buffer to read from
3959  * @cpu: The cpu buffer to iterate over
3960  *
3961  * This performs the initial preparations necessary to iterate
3962  * through the buffer.  Memory is allocated, buffer recording
3963  * is disabled, and the iterator pointer is returned to the caller.
3964  *
3965  * Disabling buffer recordng prevents the reading from being
3966  * corrupted. This is not a consuming read, so a producer is not
3967  * expected.
3968  *
3969  * After a sequence of ring_buffer_read_prepare calls, the user is
3970  * expected to make at least one call to ring_buffer_read_prepare_sync.
3971  * Afterwards, ring_buffer_read_start is invoked to get things going
3972  * for real.
3973  *
3974  * This overall must be paired with ring_buffer_read_finish.
3975  */
3976 struct ring_buffer_iter *
3977 ring_buffer_read_prepare(struct ring_buffer *buffer, int cpu)
3978 {
3979 	struct ring_buffer_per_cpu *cpu_buffer;
3980 	struct ring_buffer_iter *iter;
3981 
3982 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3983 		return NULL;
3984 
3985 	iter = kmalloc(sizeof(*iter), GFP_KERNEL);
3986 	if (!iter)
3987 		return NULL;
3988 
3989 	cpu_buffer = buffer->buffers[cpu];
3990 
3991 	iter->cpu_buffer = cpu_buffer;
3992 
3993 	atomic_inc(&buffer->resize_disabled);
3994 	atomic_inc(&cpu_buffer->record_disabled);
3995 
3996 	return iter;
3997 }
3998 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
3999 
4000 /**
4001  * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
4002  *
4003  * All previously invoked ring_buffer_read_prepare calls to prepare
4004  * iterators will be synchronized.  Afterwards, read_buffer_read_start
4005  * calls on those iterators are allowed.
4006  */
4007 void
4008 ring_buffer_read_prepare_sync(void)
4009 {
4010 	synchronize_sched();
4011 }
4012 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
4013 
4014 /**
4015  * ring_buffer_read_start - start a non consuming read of the buffer
4016  * @iter: The iterator returned by ring_buffer_read_prepare
4017  *
4018  * This finalizes the startup of an iteration through the buffer.
4019  * The iterator comes from a call to ring_buffer_read_prepare and
4020  * an intervening ring_buffer_read_prepare_sync must have been
4021  * performed.
4022  *
4023  * Must be paired with ring_buffer_read_finish.
4024  */
4025 void
4026 ring_buffer_read_start(struct ring_buffer_iter *iter)
4027 {
4028 	struct ring_buffer_per_cpu *cpu_buffer;
4029 	unsigned long flags;
4030 
4031 	if (!iter)
4032 		return;
4033 
4034 	cpu_buffer = iter->cpu_buffer;
4035 
4036 	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4037 	arch_spin_lock(&cpu_buffer->lock);
4038 	rb_iter_reset(iter);
4039 	arch_spin_unlock(&cpu_buffer->lock);
4040 	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4041 }
4042 EXPORT_SYMBOL_GPL(ring_buffer_read_start);
4043 
4044 /**
4045  * ring_buffer_read_finish - finish reading the iterator of the buffer
4046  * @iter: The iterator retrieved by ring_buffer_start
4047  *
4048  * This re-enables the recording to the buffer, and frees the
4049  * iterator.
4050  */
4051 void
4052 ring_buffer_read_finish(struct ring_buffer_iter *iter)
4053 {
4054 	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4055 	unsigned long flags;
4056 
4057 	/*
4058 	 * Ring buffer is disabled from recording, here's a good place
4059 	 * to check the integrity of the ring buffer.
4060 	 * Must prevent readers from trying to read, as the check
4061 	 * clears the HEAD page and readers require it.
4062 	 */
4063 	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4064 	rb_check_pages(cpu_buffer);
4065 	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4066 
4067 	atomic_dec(&cpu_buffer->record_disabled);
4068 	atomic_dec(&cpu_buffer->buffer->resize_disabled);
4069 	kfree(iter);
4070 }
4071 EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
4072 
4073 /**
4074  * ring_buffer_read - read the next item in the ring buffer by the iterator
4075  * @iter: The ring buffer iterator
4076  * @ts: The time stamp of the event read.
4077  *
4078  * This reads the next event in the ring buffer and increments the iterator.
4079  */
4080 struct ring_buffer_event *
4081 ring_buffer_read(struct ring_buffer_iter *iter, u64 *ts)
4082 {
4083 	struct ring_buffer_event *event;
4084 	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4085 	unsigned long flags;
4086 
4087 	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4088  again:
4089 	event = rb_iter_peek(iter, ts);
4090 	if (!event)
4091 		goto out;
4092 
4093 	if (event->type_len == RINGBUF_TYPE_PADDING)
4094 		goto again;
4095 
4096 	rb_advance_iter(iter);
4097  out:
4098 	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4099 
4100 	return event;
4101 }
4102 EXPORT_SYMBOL_GPL(ring_buffer_read);
4103 
4104 /**
4105  * ring_buffer_size - return the size of the ring buffer (in bytes)
4106  * @buffer: The ring buffer.
4107  */
4108 unsigned long ring_buffer_size(struct ring_buffer *buffer, int cpu)
4109 {
4110 	/*
4111 	 * Earlier, this method returned
4112 	 *	BUF_PAGE_SIZE * buffer->nr_pages
4113 	 * Since the nr_pages field is now removed, we have converted this to
4114 	 * return the per cpu buffer value.
4115 	 */
4116 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4117 		return 0;
4118 
4119 	return BUF_PAGE_SIZE * buffer->buffers[cpu]->nr_pages;
4120 }
4121 EXPORT_SYMBOL_GPL(ring_buffer_size);
4122 
4123 static void
4124 rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
4125 {
4126 	rb_head_page_deactivate(cpu_buffer);
4127 
4128 	cpu_buffer->head_page
4129 		= list_entry(cpu_buffer->pages, struct buffer_page, list);
4130 	local_set(&cpu_buffer->head_page->write, 0);
4131 	local_set(&cpu_buffer->head_page->entries, 0);
4132 	local_set(&cpu_buffer->head_page->page->commit, 0);
4133 
4134 	cpu_buffer->head_page->read = 0;
4135 
4136 	cpu_buffer->tail_page = cpu_buffer->head_page;
4137 	cpu_buffer->commit_page = cpu_buffer->head_page;
4138 
4139 	INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
4140 	INIT_LIST_HEAD(&cpu_buffer->new_pages);
4141 	local_set(&cpu_buffer->reader_page->write, 0);
4142 	local_set(&cpu_buffer->reader_page->entries, 0);
4143 	local_set(&cpu_buffer->reader_page->page->commit, 0);
4144 	cpu_buffer->reader_page->read = 0;
4145 
4146 	local_set(&cpu_buffer->entries_bytes, 0);
4147 	local_set(&cpu_buffer->overrun, 0);
4148 	local_set(&cpu_buffer->commit_overrun, 0);
4149 	local_set(&cpu_buffer->dropped_events, 0);
4150 	local_set(&cpu_buffer->entries, 0);
4151 	local_set(&cpu_buffer->committing, 0);
4152 	local_set(&cpu_buffer->commits, 0);
4153 	cpu_buffer->read = 0;
4154 	cpu_buffer->read_bytes = 0;
4155 
4156 	cpu_buffer->write_stamp = 0;
4157 	cpu_buffer->read_stamp = 0;
4158 
4159 	cpu_buffer->lost_events = 0;
4160 	cpu_buffer->last_overrun = 0;
4161 
4162 	rb_head_page_activate(cpu_buffer);
4163 }
4164 
4165 /**
4166  * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
4167  * @buffer: The ring buffer to reset a per cpu buffer of
4168  * @cpu: The CPU buffer to be reset
4169  */
4170 void ring_buffer_reset_cpu(struct ring_buffer *buffer, int cpu)
4171 {
4172 	struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4173 	unsigned long flags;
4174 
4175 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4176 		return;
4177 
4178 	atomic_inc(&buffer->resize_disabled);
4179 	atomic_inc(&cpu_buffer->record_disabled);
4180 
4181 	/* Make sure all commits have finished */
4182 	synchronize_sched();
4183 
4184 	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4185 
4186 	if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
4187 		goto out;
4188 
4189 	arch_spin_lock(&cpu_buffer->lock);
4190 
4191 	rb_reset_cpu(cpu_buffer);
4192 
4193 	arch_spin_unlock(&cpu_buffer->lock);
4194 
4195  out:
4196 	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4197 
4198 	atomic_dec(&cpu_buffer->record_disabled);
4199 	atomic_dec(&buffer->resize_disabled);
4200 }
4201 EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
4202 
4203 /**
4204  * ring_buffer_reset - reset a ring buffer
4205  * @buffer: The ring buffer to reset all cpu buffers
4206  */
4207 void ring_buffer_reset(struct ring_buffer *buffer)
4208 {
4209 	int cpu;
4210 
4211 	for_each_buffer_cpu(buffer, cpu)
4212 		ring_buffer_reset_cpu(buffer, cpu);
4213 }
4214 EXPORT_SYMBOL_GPL(ring_buffer_reset);
4215 
4216 /**
4217  * rind_buffer_empty - is the ring buffer empty?
4218  * @buffer: The ring buffer to test
4219  */
4220 int ring_buffer_empty(struct ring_buffer *buffer)
4221 {
4222 	struct ring_buffer_per_cpu *cpu_buffer;
4223 	unsigned long flags;
4224 	int dolock;
4225 	int cpu;
4226 	int ret;
4227 
4228 	dolock = rb_ok_to_lock();
4229 
4230 	/* yes this is racy, but if you don't like the race, lock the buffer */
4231 	for_each_buffer_cpu(buffer, cpu) {
4232 		cpu_buffer = buffer->buffers[cpu];
4233 		local_irq_save(flags);
4234 		if (dolock)
4235 			raw_spin_lock(&cpu_buffer->reader_lock);
4236 		ret = rb_per_cpu_empty(cpu_buffer);
4237 		if (dolock)
4238 			raw_spin_unlock(&cpu_buffer->reader_lock);
4239 		local_irq_restore(flags);
4240 
4241 		if (!ret)
4242 			return 0;
4243 	}
4244 
4245 	return 1;
4246 }
4247 EXPORT_SYMBOL_GPL(ring_buffer_empty);
4248 
4249 /**
4250  * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
4251  * @buffer: The ring buffer
4252  * @cpu: The CPU buffer to test
4253  */
4254 int ring_buffer_empty_cpu(struct ring_buffer *buffer, int cpu)
4255 {
4256 	struct ring_buffer_per_cpu *cpu_buffer;
4257 	unsigned long flags;
4258 	int dolock;
4259 	int ret;
4260 
4261 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4262 		return 1;
4263 
4264 	dolock = rb_ok_to_lock();
4265 
4266 	cpu_buffer = buffer->buffers[cpu];
4267 	local_irq_save(flags);
4268 	if (dolock)
4269 		raw_spin_lock(&cpu_buffer->reader_lock);
4270 	ret = rb_per_cpu_empty(cpu_buffer);
4271 	if (dolock)
4272 		raw_spin_unlock(&cpu_buffer->reader_lock);
4273 	local_irq_restore(flags);
4274 
4275 	return ret;
4276 }
4277 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
4278 
4279 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
4280 /**
4281  * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
4282  * @buffer_a: One buffer to swap with
4283  * @buffer_b: The other buffer to swap with
4284  *
4285  * This function is useful for tracers that want to take a "snapshot"
4286  * of a CPU buffer and has another back up buffer lying around.
4287  * it is expected that the tracer handles the cpu buffer not being
4288  * used at the moment.
4289  */
4290 int ring_buffer_swap_cpu(struct ring_buffer *buffer_a,
4291 			 struct ring_buffer *buffer_b, int cpu)
4292 {
4293 	struct ring_buffer_per_cpu *cpu_buffer_a;
4294 	struct ring_buffer_per_cpu *cpu_buffer_b;
4295 	int ret = -EINVAL;
4296 
4297 	if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
4298 	    !cpumask_test_cpu(cpu, buffer_b->cpumask))
4299 		goto out;
4300 
4301 	cpu_buffer_a = buffer_a->buffers[cpu];
4302 	cpu_buffer_b = buffer_b->buffers[cpu];
4303 
4304 	/* At least make sure the two buffers are somewhat the same */
4305 	if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
4306 		goto out;
4307 
4308 	ret = -EAGAIN;
4309 
4310 	if (ring_buffer_flags != RB_BUFFERS_ON)
4311 		goto out;
4312 
4313 	if (atomic_read(&buffer_a->record_disabled))
4314 		goto out;
4315 
4316 	if (atomic_read(&buffer_b->record_disabled))
4317 		goto out;
4318 
4319 	if (atomic_read(&cpu_buffer_a->record_disabled))
4320 		goto out;
4321 
4322 	if (atomic_read(&cpu_buffer_b->record_disabled))
4323 		goto out;
4324 
4325 	/*
4326 	 * We can't do a synchronize_sched here because this
4327 	 * function can be called in atomic context.
4328 	 * Normally this will be called from the same CPU as cpu.
4329 	 * If not it's up to the caller to protect this.
4330 	 */
4331 	atomic_inc(&cpu_buffer_a->record_disabled);
4332 	atomic_inc(&cpu_buffer_b->record_disabled);
4333 
4334 	ret = -EBUSY;
4335 	if (local_read(&cpu_buffer_a->committing))
4336 		goto out_dec;
4337 	if (local_read(&cpu_buffer_b->committing))
4338 		goto out_dec;
4339 
4340 	buffer_a->buffers[cpu] = cpu_buffer_b;
4341 	buffer_b->buffers[cpu] = cpu_buffer_a;
4342 
4343 	cpu_buffer_b->buffer = buffer_a;
4344 	cpu_buffer_a->buffer = buffer_b;
4345 
4346 	ret = 0;
4347 
4348 out_dec:
4349 	atomic_dec(&cpu_buffer_a->record_disabled);
4350 	atomic_dec(&cpu_buffer_b->record_disabled);
4351 out:
4352 	return ret;
4353 }
4354 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
4355 #endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
4356 
4357 /**
4358  * ring_buffer_alloc_read_page - allocate a page to read from buffer
4359  * @buffer: the buffer to allocate for.
4360  * @cpu: the cpu buffer to allocate.
4361  *
4362  * This function is used in conjunction with ring_buffer_read_page.
4363  * When reading a full page from the ring buffer, these functions
4364  * can be used to speed up the process. The calling function should
4365  * allocate a few pages first with this function. Then when it
4366  * needs to get pages from the ring buffer, it passes the result
4367  * of this function into ring_buffer_read_page, which will swap
4368  * the page that was allocated, with the read page of the buffer.
4369  *
4370  * Returns:
4371  *  The page allocated, or NULL on error.
4372  */
4373 void *ring_buffer_alloc_read_page(struct ring_buffer *buffer, int cpu)
4374 {
4375 	struct buffer_data_page *bpage;
4376 	struct page *page;
4377 
4378 	page = alloc_pages_node(cpu_to_node(cpu),
4379 				GFP_KERNEL | __GFP_NORETRY, 0);
4380 	if (!page)
4381 		return NULL;
4382 
4383 	bpage = page_address(page);
4384 
4385 	rb_init_page(bpage);
4386 
4387 	return bpage;
4388 }
4389 EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
4390 
4391 /**
4392  * ring_buffer_free_read_page - free an allocated read page
4393  * @buffer: the buffer the page was allocate for
4394  * @data: the page to free
4395  *
4396  * Free a page allocated from ring_buffer_alloc_read_page.
4397  */
4398 void ring_buffer_free_read_page(struct ring_buffer *buffer, void *data)
4399 {
4400 	free_page((unsigned long)data);
4401 }
4402 EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
4403 
4404 /**
4405  * ring_buffer_read_page - extract a page from the ring buffer
4406  * @buffer: buffer to extract from
4407  * @data_page: the page to use allocated from ring_buffer_alloc_read_page
4408  * @len: amount to extract
4409  * @cpu: the cpu of the buffer to extract
4410  * @full: should the extraction only happen when the page is full.
4411  *
4412  * This function will pull out a page from the ring buffer and consume it.
4413  * @data_page must be the address of the variable that was returned
4414  * from ring_buffer_alloc_read_page. This is because the page might be used
4415  * to swap with a page in the ring buffer.
4416  *
4417  * for example:
4418  *	rpage = ring_buffer_alloc_read_page(buffer, cpu);
4419  *	if (!rpage)
4420  *		return error;
4421  *	ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
4422  *	if (ret >= 0)
4423  *		process_page(rpage, ret);
4424  *
4425  * When @full is set, the function will not return true unless
4426  * the writer is off the reader page.
4427  *
4428  * Note: it is up to the calling functions to handle sleeps and wakeups.
4429  *  The ring buffer can be used anywhere in the kernel and can not
4430  *  blindly call wake_up. The layer that uses the ring buffer must be
4431  *  responsible for that.
4432  *
4433  * Returns:
4434  *  >=0 if data has been transferred, returns the offset of consumed data.
4435  *  <0 if no data has been transferred.
4436  */
4437 int ring_buffer_read_page(struct ring_buffer *buffer,
4438 			  void **data_page, size_t len, int cpu, int full)
4439 {
4440 	struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4441 	struct ring_buffer_event *event;
4442 	struct buffer_data_page *bpage;
4443 	struct buffer_page *reader;
4444 	unsigned long missed_events;
4445 	unsigned long flags;
4446 	unsigned int commit;
4447 	unsigned int read;
4448 	u64 save_timestamp;
4449 	int ret = -1;
4450 
4451 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4452 		goto out;
4453 
4454 	/*
4455 	 * If len is not big enough to hold the page header, then
4456 	 * we can not copy anything.
4457 	 */
4458 	if (len <= BUF_PAGE_HDR_SIZE)
4459 		goto out;
4460 
4461 	len -= BUF_PAGE_HDR_SIZE;
4462 
4463 	if (!data_page)
4464 		goto out;
4465 
4466 	bpage = *data_page;
4467 	if (!bpage)
4468 		goto out;
4469 
4470 	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4471 
4472 	reader = rb_get_reader_page(cpu_buffer);
4473 	if (!reader)
4474 		goto out_unlock;
4475 
4476 	event = rb_reader_event(cpu_buffer);
4477 
4478 	read = reader->read;
4479 	commit = rb_page_commit(reader);
4480 
4481 	/* Check if any events were dropped */
4482 	missed_events = cpu_buffer->lost_events;
4483 
4484 	/*
4485 	 * If this page has been partially read or
4486 	 * if len is not big enough to read the rest of the page or
4487 	 * a writer is still on the page, then
4488 	 * we must copy the data from the page to the buffer.
4489 	 * Otherwise, we can simply swap the page with the one passed in.
4490 	 */
4491 	if (read || (len < (commit - read)) ||
4492 	    cpu_buffer->reader_page == cpu_buffer->commit_page) {
4493 		struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
4494 		unsigned int rpos = read;
4495 		unsigned int pos = 0;
4496 		unsigned int size;
4497 
4498 		if (full)
4499 			goto out_unlock;
4500 
4501 		if (len > (commit - read))
4502 			len = (commit - read);
4503 
4504 		/* Always keep the time extend and data together */
4505 		size = rb_event_ts_length(event);
4506 
4507 		if (len < size)
4508 			goto out_unlock;
4509 
4510 		/* save the current timestamp, since the user will need it */
4511 		save_timestamp = cpu_buffer->read_stamp;
4512 
4513 		/* Need to copy one event at a time */
4514 		do {
4515 			/* We need the size of one event, because
4516 			 * rb_advance_reader only advances by one event,
4517 			 * whereas rb_event_ts_length may include the size of
4518 			 * one or two events.
4519 			 * We have already ensured there's enough space if this
4520 			 * is a time extend. */
4521 			size = rb_event_length(event);
4522 			memcpy(bpage->data + pos, rpage->data + rpos, size);
4523 
4524 			len -= size;
4525 
4526 			rb_advance_reader(cpu_buffer);
4527 			rpos = reader->read;
4528 			pos += size;
4529 
4530 			if (rpos >= commit)
4531 				break;
4532 
4533 			event = rb_reader_event(cpu_buffer);
4534 			/* Always keep the time extend and data together */
4535 			size = rb_event_ts_length(event);
4536 		} while (len >= size);
4537 
4538 		/* update bpage */
4539 		local_set(&bpage->commit, pos);
4540 		bpage->time_stamp = save_timestamp;
4541 
4542 		/* we copied everything to the beginning */
4543 		read = 0;
4544 	} else {
4545 		/* update the entry counter */
4546 		cpu_buffer->read += rb_page_entries(reader);
4547 		cpu_buffer->read_bytes += BUF_PAGE_SIZE;
4548 
4549 		/* swap the pages */
4550 		rb_init_page(bpage);
4551 		bpage = reader->page;
4552 		reader->page = *data_page;
4553 		local_set(&reader->write, 0);
4554 		local_set(&reader->entries, 0);
4555 		reader->read = 0;
4556 		*data_page = bpage;
4557 
4558 		/*
4559 		 * Use the real_end for the data size,
4560 		 * This gives us a chance to store the lost events
4561 		 * on the page.
4562 		 */
4563 		if (reader->real_end)
4564 			local_set(&bpage->commit, reader->real_end);
4565 	}
4566 	ret = read;
4567 
4568 	cpu_buffer->lost_events = 0;
4569 
4570 	commit = local_read(&bpage->commit);
4571 	/*
4572 	 * Set a flag in the commit field if we lost events
4573 	 */
4574 	if (missed_events) {
4575 		/* If there is room at the end of the page to save the
4576 		 * missed events, then record it there.
4577 		 */
4578 		if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
4579 			memcpy(&bpage->data[commit], &missed_events,
4580 			       sizeof(missed_events));
4581 			local_add(RB_MISSED_STORED, &bpage->commit);
4582 			commit += sizeof(missed_events);
4583 		}
4584 		local_add(RB_MISSED_EVENTS, &bpage->commit);
4585 	}
4586 
4587 	/*
4588 	 * This page may be off to user land. Zero it out here.
4589 	 */
4590 	if (commit < BUF_PAGE_SIZE)
4591 		memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
4592 
4593  out_unlock:
4594 	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4595 
4596  out:
4597 	return ret;
4598 }
4599 EXPORT_SYMBOL_GPL(ring_buffer_read_page);
4600 
4601 #ifdef CONFIG_HOTPLUG_CPU
4602 static int rb_cpu_notify(struct notifier_block *self,
4603 			 unsigned long action, void *hcpu)
4604 {
4605 	struct ring_buffer *buffer =
4606 		container_of(self, struct ring_buffer, cpu_notify);
4607 	long cpu = (long)hcpu;
4608 	int cpu_i, nr_pages_same;
4609 	unsigned int nr_pages;
4610 
4611 	switch (action) {
4612 	case CPU_UP_PREPARE:
4613 	case CPU_UP_PREPARE_FROZEN:
4614 		if (cpumask_test_cpu(cpu, buffer->cpumask))
4615 			return NOTIFY_OK;
4616 
4617 		nr_pages = 0;
4618 		nr_pages_same = 1;
4619 		/* check if all cpu sizes are same */
4620 		for_each_buffer_cpu(buffer, cpu_i) {
4621 			/* fill in the size from first enabled cpu */
4622 			if (nr_pages == 0)
4623 				nr_pages = buffer->buffers[cpu_i]->nr_pages;
4624 			if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
4625 				nr_pages_same = 0;
4626 				break;
4627 			}
4628 		}
4629 		/* allocate minimum pages, user can later expand it */
4630 		if (!nr_pages_same)
4631 			nr_pages = 2;
4632 		buffer->buffers[cpu] =
4633 			rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
4634 		if (!buffer->buffers[cpu]) {
4635 			WARN(1, "failed to allocate ring buffer on CPU %ld\n",
4636 			     cpu);
4637 			return NOTIFY_OK;
4638 		}
4639 		smp_wmb();
4640 		cpumask_set_cpu(cpu, buffer->cpumask);
4641 		break;
4642 	case CPU_DOWN_PREPARE:
4643 	case CPU_DOWN_PREPARE_FROZEN:
4644 		/*
4645 		 * Do nothing.
4646 		 *  If we were to free the buffer, then the user would
4647 		 *  lose any trace that was in the buffer.
4648 		 */
4649 		break;
4650 	default:
4651 		break;
4652 	}
4653 	return NOTIFY_OK;
4654 }
4655 #endif
4656 
4657 #ifdef CONFIG_RING_BUFFER_STARTUP_TEST
4658 /*
4659  * This is a basic integrity check of the ring buffer.
4660  * Late in the boot cycle this test will run when configured in.
4661  * It will kick off a thread per CPU that will go into a loop
4662  * writing to the per cpu ring buffer various sizes of data.
4663  * Some of the data will be large items, some small.
4664  *
4665  * Another thread is created that goes into a spin, sending out
4666  * IPIs to the other CPUs to also write into the ring buffer.
4667  * this is to test the nesting ability of the buffer.
4668  *
4669  * Basic stats are recorded and reported. If something in the
4670  * ring buffer should happen that's not expected, a big warning
4671  * is displayed and all ring buffers are disabled.
4672  */
4673 static struct task_struct *rb_threads[NR_CPUS] __initdata;
4674 
4675 struct rb_test_data {
4676 	struct ring_buffer	*buffer;
4677 	unsigned long		events;
4678 	unsigned long		bytes_written;
4679 	unsigned long		bytes_alloc;
4680 	unsigned long		bytes_dropped;
4681 	unsigned long		events_nested;
4682 	unsigned long		bytes_written_nested;
4683 	unsigned long		bytes_alloc_nested;
4684 	unsigned long		bytes_dropped_nested;
4685 	int			min_size_nested;
4686 	int			max_size_nested;
4687 	int			max_size;
4688 	int			min_size;
4689 	int			cpu;
4690 	int			cnt;
4691 };
4692 
4693 static struct rb_test_data rb_data[NR_CPUS] __initdata;
4694 
4695 /* 1 meg per cpu */
4696 #define RB_TEST_BUFFER_SIZE	1048576
4697 
4698 static char rb_string[] __initdata =
4699 	"abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
4700 	"?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
4701 	"!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
4702 
4703 static bool rb_test_started __initdata;
4704 
4705 struct rb_item {
4706 	int size;
4707 	char str[];
4708 };
4709 
4710 static __init int rb_write_something(struct rb_test_data *data, bool nested)
4711 {
4712 	struct ring_buffer_event *event;
4713 	struct rb_item *item;
4714 	bool started;
4715 	int event_len;
4716 	int size;
4717 	int len;
4718 	int cnt;
4719 
4720 	/* Have nested writes different that what is written */
4721 	cnt = data->cnt + (nested ? 27 : 0);
4722 
4723 	/* Multiply cnt by ~e, to make some unique increment */
4724 	size = (data->cnt * 68 / 25) % (sizeof(rb_string) - 1);
4725 
4726 	len = size + sizeof(struct rb_item);
4727 
4728 	started = rb_test_started;
4729 	/* read rb_test_started before checking buffer enabled */
4730 	smp_rmb();
4731 
4732 	event = ring_buffer_lock_reserve(data->buffer, len);
4733 	if (!event) {
4734 		/* Ignore dropped events before test starts. */
4735 		if (started) {
4736 			if (nested)
4737 				data->bytes_dropped += len;
4738 			else
4739 				data->bytes_dropped_nested += len;
4740 		}
4741 		return len;
4742 	}
4743 
4744 	event_len = ring_buffer_event_length(event);
4745 
4746 	if (RB_WARN_ON(data->buffer, event_len < len))
4747 		goto out;
4748 
4749 	item = ring_buffer_event_data(event);
4750 	item->size = size;
4751 	memcpy(item->str, rb_string, size);
4752 
4753 	if (nested) {
4754 		data->bytes_alloc_nested += event_len;
4755 		data->bytes_written_nested += len;
4756 		data->events_nested++;
4757 		if (!data->min_size_nested || len < data->min_size_nested)
4758 			data->min_size_nested = len;
4759 		if (len > data->max_size_nested)
4760 			data->max_size_nested = len;
4761 	} else {
4762 		data->bytes_alloc += event_len;
4763 		data->bytes_written += len;
4764 		data->events++;
4765 		if (!data->min_size || len < data->min_size)
4766 			data->max_size = len;
4767 		if (len > data->max_size)
4768 			data->max_size = len;
4769 	}
4770 
4771  out:
4772 	ring_buffer_unlock_commit(data->buffer, event);
4773 
4774 	return 0;
4775 }
4776 
4777 static __init int rb_test(void *arg)
4778 {
4779 	struct rb_test_data *data = arg;
4780 
4781 	while (!kthread_should_stop()) {
4782 		rb_write_something(data, false);
4783 		data->cnt++;
4784 
4785 		set_current_state(TASK_INTERRUPTIBLE);
4786 		/* Now sleep between a min of 100-300us and a max of 1ms */
4787 		usleep_range(((data->cnt % 3) + 1) * 100, 1000);
4788 	}
4789 
4790 	return 0;
4791 }
4792 
4793 static __init void rb_ipi(void *ignore)
4794 {
4795 	struct rb_test_data *data;
4796 	int cpu = smp_processor_id();
4797 
4798 	data = &rb_data[cpu];
4799 	rb_write_something(data, true);
4800 }
4801 
4802 static __init int rb_hammer_test(void *arg)
4803 {
4804 	while (!kthread_should_stop()) {
4805 
4806 		/* Send an IPI to all cpus to write data! */
4807 		smp_call_function(rb_ipi, NULL, 1);
4808 		/* No sleep, but for non preempt, let others run */
4809 		schedule();
4810 	}
4811 
4812 	return 0;
4813 }
4814 
4815 static __init int test_ringbuffer(void)
4816 {
4817 	struct task_struct *rb_hammer;
4818 	struct ring_buffer *buffer;
4819 	int cpu;
4820 	int ret = 0;
4821 
4822 	pr_info("Running ring buffer tests...\n");
4823 
4824 	buffer = ring_buffer_alloc(RB_TEST_BUFFER_SIZE, RB_FL_OVERWRITE);
4825 	if (WARN_ON(!buffer))
4826 		return 0;
4827 
4828 	/* Disable buffer so that threads can't write to it yet */
4829 	ring_buffer_record_off(buffer);
4830 
4831 	for_each_online_cpu(cpu) {
4832 		rb_data[cpu].buffer = buffer;
4833 		rb_data[cpu].cpu = cpu;
4834 		rb_data[cpu].cnt = cpu;
4835 		rb_threads[cpu] = kthread_create(rb_test, &rb_data[cpu],
4836 						 "rbtester/%d", cpu);
4837 		if (WARN_ON(!rb_threads[cpu])) {
4838 			pr_cont("FAILED\n");
4839 			ret = -1;
4840 			goto out_free;
4841 		}
4842 
4843 		kthread_bind(rb_threads[cpu], cpu);
4844  		wake_up_process(rb_threads[cpu]);
4845 	}
4846 
4847 	/* Now create the rb hammer! */
4848 	rb_hammer = kthread_run(rb_hammer_test, NULL, "rbhammer");
4849 	if (WARN_ON(!rb_hammer)) {
4850 		pr_cont("FAILED\n");
4851 		ret = -1;
4852 		goto out_free;
4853 	}
4854 
4855 	ring_buffer_record_on(buffer);
4856 	/*
4857 	 * Show buffer is enabled before setting rb_test_started.
4858 	 * Yes there's a small race window where events could be
4859 	 * dropped and the thread wont catch it. But when a ring
4860 	 * buffer gets enabled, there will always be some kind of
4861 	 * delay before other CPUs see it. Thus, we don't care about
4862 	 * those dropped events. We care about events dropped after
4863 	 * the threads see that the buffer is active.
4864 	 */
4865 	smp_wmb();
4866 	rb_test_started = true;
4867 
4868 	set_current_state(TASK_INTERRUPTIBLE);
4869 	/* Just run for 10 seconds */;
4870 	schedule_timeout(10 * HZ);
4871 
4872 	kthread_stop(rb_hammer);
4873 
4874  out_free:
4875 	for_each_online_cpu(cpu) {
4876 		if (!rb_threads[cpu])
4877 			break;
4878 		kthread_stop(rb_threads[cpu]);
4879 	}
4880 	if (ret) {
4881 		ring_buffer_free(buffer);
4882 		return ret;
4883 	}
4884 
4885 	/* Report! */
4886 	pr_info("finished\n");
4887 	for_each_online_cpu(cpu) {
4888 		struct ring_buffer_event *event;
4889 		struct rb_test_data *data = &rb_data[cpu];
4890 		struct rb_item *item;
4891 		unsigned long total_events;
4892 		unsigned long total_dropped;
4893 		unsigned long total_written;
4894 		unsigned long total_alloc;
4895 		unsigned long total_read = 0;
4896 		unsigned long total_size = 0;
4897 		unsigned long total_len = 0;
4898 		unsigned long total_lost = 0;
4899 		unsigned long lost;
4900 		int big_event_size;
4901 		int small_event_size;
4902 
4903 		ret = -1;
4904 
4905 		total_events = data->events + data->events_nested;
4906 		total_written = data->bytes_written + data->bytes_written_nested;
4907 		total_alloc = data->bytes_alloc + data->bytes_alloc_nested;
4908 		total_dropped = data->bytes_dropped + data->bytes_dropped_nested;
4909 
4910 		big_event_size = data->max_size + data->max_size_nested;
4911 		small_event_size = data->min_size + data->min_size_nested;
4912 
4913 		pr_info("CPU %d:\n", cpu);
4914 		pr_info("              events:    %ld\n", total_events);
4915 		pr_info("       dropped bytes:    %ld\n", total_dropped);
4916 		pr_info("       alloced bytes:    %ld\n", total_alloc);
4917 		pr_info("       written bytes:    %ld\n", total_written);
4918 		pr_info("       biggest event:    %d\n", big_event_size);
4919 		pr_info("      smallest event:    %d\n", small_event_size);
4920 
4921 		if (RB_WARN_ON(buffer, total_dropped))
4922 			break;
4923 
4924 		ret = 0;
4925 
4926 		while ((event = ring_buffer_consume(buffer, cpu, NULL, &lost))) {
4927 			total_lost += lost;
4928 			item = ring_buffer_event_data(event);
4929 			total_len += ring_buffer_event_length(event);
4930 			total_size += item->size + sizeof(struct rb_item);
4931 			if (memcmp(&item->str[0], rb_string, item->size) != 0) {
4932 				pr_info("FAILED!\n");
4933 				pr_info("buffer had: %.*s\n", item->size, item->str);
4934 				pr_info("expected:   %.*s\n", item->size, rb_string);
4935 				RB_WARN_ON(buffer, 1);
4936 				ret = -1;
4937 				break;
4938 			}
4939 			total_read++;
4940 		}
4941 		if (ret)
4942 			break;
4943 
4944 		ret = -1;
4945 
4946 		pr_info("         read events:   %ld\n", total_read);
4947 		pr_info("         lost events:   %ld\n", total_lost);
4948 		pr_info("        total events:   %ld\n", total_lost + total_read);
4949 		pr_info("  recorded len bytes:   %ld\n", total_len);
4950 		pr_info(" recorded size bytes:   %ld\n", total_size);
4951 		if (total_lost)
4952 			pr_info(" With dropped events, record len and size may not match\n"
4953 				" alloced and written from above\n");
4954 		if (!total_lost) {
4955 			if (RB_WARN_ON(buffer, total_len != total_alloc ||
4956 				       total_size != total_written))
4957 				break;
4958 		}
4959 		if (RB_WARN_ON(buffer, total_lost + total_read != total_events))
4960 			break;
4961 
4962 		ret = 0;
4963 	}
4964 	if (!ret)
4965 		pr_info("Ring buffer PASSED!\n");
4966 
4967 	ring_buffer_free(buffer);
4968 	return 0;
4969 }
4970 
4971 late_initcall(test_ringbuffer);
4972 #endif /* CONFIG_RING_BUFFER_STARTUP_TEST */
4973