xref: /linux/kernel/trace/ring_buffer.c (revision fdfc374af5dc345fbb9686921fa60176c1c41da0)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Generic ring buffer
4  *
5  * Copyright (C) 2008 Steven Rostedt <srostedt@redhat.com>
6  */
7 #include <linux/trace_recursion.h>
8 #include <linux/trace_events.h>
9 #include <linux/ring_buffer.h>
10 #include <linux/trace_clock.h>
11 #include <linux/sched/clock.h>
12 #include <linux/trace_seq.h>
13 #include <linux/spinlock.h>
14 #include <linux/irq_work.h>
15 #include <linux/security.h>
16 #include <linux/uaccess.h>
17 #include <linux/hardirq.h>
18 #include <linux/kthread.h>	/* for self test */
19 #include <linux/module.h>
20 #include <linux/percpu.h>
21 #include <linux/mutex.h>
22 #include <linux/delay.h>
23 #include <linux/slab.h>
24 #include <linux/init.h>
25 #include <linux/hash.h>
26 #include <linux/list.h>
27 #include <linux/cpu.h>
28 #include <linux/oom.h>
29 
30 #include <asm/local.h>
31 
32 /*
33  * The "absolute" timestamp in the buffer is only 59 bits.
34  * If a clock has the 5 MSBs set, it needs to be saved and
35  * reinserted.
36  */
37 #define TS_MSB		(0xf8ULL << 56)
38 #define ABS_TS_MASK	(~TS_MSB)
39 
40 static void update_pages_handler(struct work_struct *work);
41 
42 /*
43  * The ring buffer header is special. We must manually up keep it.
44  */
45 int ring_buffer_print_entry_header(struct trace_seq *s)
46 {
47 	trace_seq_puts(s, "# compressed entry header\n");
48 	trace_seq_puts(s, "\ttype_len    :    5 bits\n");
49 	trace_seq_puts(s, "\ttime_delta  :   27 bits\n");
50 	trace_seq_puts(s, "\tarray       :   32 bits\n");
51 	trace_seq_putc(s, '\n');
52 	trace_seq_printf(s, "\tpadding     : type == %d\n",
53 			 RINGBUF_TYPE_PADDING);
54 	trace_seq_printf(s, "\ttime_extend : type == %d\n",
55 			 RINGBUF_TYPE_TIME_EXTEND);
56 	trace_seq_printf(s, "\ttime_stamp : type == %d\n",
57 			 RINGBUF_TYPE_TIME_STAMP);
58 	trace_seq_printf(s, "\tdata max type_len  == %d\n",
59 			 RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
60 
61 	return !trace_seq_has_overflowed(s);
62 }
63 
64 /*
65  * The ring buffer is made up of a list of pages. A separate list of pages is
66  * allocated for each CPU. A writer may only write to a buffer that is
67  * associated with the CPU it is currently executing on.  A reader may read
68  * from any per cpu buffer.
69  *
70  * The reader is special. For each per cpu buffer, the reader has its own
71  * reader page. When a reader has read the entire reader page, this reader
72  * page is swapped with another page in the ring buffer.
73  *
74  * Now, as long as the writer is off the reader page, the reader can do what
75  * ever it wants with that page. The writer will never write to that page
76  * again (as long as it is out of the ring buffer).
77  *
78  * Here's some silly ASCII art.
79  *
80  *   +------+
81  *   |reader|          RING BUFFER
82  *   |page  |
83  *   +------+        +---+   +---+   +---+
84  *                   |   |-->|   |-->|   |
85  *                   +---+   +---+   +---+
86  *                     ^               |
87  *                     |               |
88  *                     +---------------+
89  *
90  *
91  *   +------+
92  *   |reader|          RING BUFFER
93  *   |page  |------------------v
94  *   +------+        +---+   +---+   +---+
95  *                   |   |-->|   |-->|   |
96  *                   +---+   +---+   +---+
97  *                     ^               |
98  *                     |               |
99  *                     +---------------+
100  *
101  *
102  *   +------+
103  *   |reader|          RING BUFFER
104  *   |page  |------------------v
105  *   +------+        +---+   +---+   +---+
106  *      ^            |   |-->|   |-->|   |
107  *      |            +---+   +---+   +---+
108  *      |                              |
109  *      |                              |
110  *      +------------------------------+
111  *
112  *
113  *   +------+
114  *   |buffer|          RING BUFFER
115  *   |page  |------------------v
116  *   +------+        +---+   +---+   +---+
117  *      ^            |   |   |   |-->|   |
118  *      |   New      +---+   +---+   +---+
119  *      |  Reader------^               |
120  *      |   page                       |
121  *      +------------------------------+
122  *
123  *
124  * After we make this swap, the reader can hand this page off to the splice
125  * code and be done with it. It can even allocate a new page if it needs to
126  * and swap that into the ring buffer.
127  *
128  * We will be using cmpxchg soon to make all this lockless.
129  *
130  */
131 
132 /* Used for individual buffers (after the counter) */
133 #define RB_BUFFER_OFF		(1 << 20)
134 
135 #define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
136 
137 #define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
138 #define RB_ALIGNMENT		4U
139 #define RB_MAX_SMALL_DATA	(RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
140 #define RB_EVNT_MIN_SIZE	8U	/* two 32bit words */
141 
142 #ifndef CONFIG_HAVE_64BIT_ALIGNED_ACCESS
143 # define RB_FORCE_8BYTE_ALIGNMENT	0
144 # define RB_ARCH_ALIGNMENT		RB_ALIGNMENT
145 #else
146 # define RB_FORCE_8BYTE_ALIGNMENT	1
147 # define RB_ARCH_ALIGNMENT		8U
148 #endif
149 
150 #define RB_ALIGN_DATA		__aligned(RB_ARCH_ALIGNMENT)
151 
152 /* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
153 #define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
154 
155 enum {
156 	RB_LEN_TIME_EXTEND = 8,
157 	RB_LEN_TIME_STAMP =  8,
158 };
159 
160 #define skip_time_extend(event) \
161 	((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
162 
163 #define extended_time(event) \
164 	(event->type_len >= RINGBUF_TYPE_TIME_EXTEND)
165 
166 static inline bool rb_null_event(struct ring_buffer_event *event)
167 {
168 	return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
169 }
170 
171 static void rb_event_set_padding(struct ring_buffer_event *event)
172 {
173 	/* padding has a NULL time_delta */
174 	event->type_len = RINGBUF_TYPE_PADDING;
175 	event->time_delta = 0;
176 }
177 
178 static unsigned
179 rb_event_data_length(struct ring_buffer_event *event)
180 {
181 	unsigned length;
182 
183 	if (event->type_len)
184 		length = event->type_len * RB_ALIGNMENT;
185 	else
186 		length = event->array[0];
187 	return length + RB_EVNT_HDR_SIZE;
188 }
189 
190 /*
191  * Return the length of the given event. Will return
192  * the length of the time extend if the event is a
193  * time extend.
194  */
195 static inline unsigned
196 rb_event_length(struct ring_buffer_event *event)
197 {
198 	switch (event->type_len) {
199 	case RINGBUF_TYPE_PADDING:
200 		if (rb_null_event(event))
201 			/* undefined */
202 			return -1;
203 		return  event->array[0] + RB_EVNT_HDR_SIZE;
204 
205 	case RINGBUF_TYPE_TIME_EXTEND:
206 		return RB_LEN_TIME_EXTEND;
207 
208 	case RINGBUF_TYPE_TIME_STAMP:
209 		return RB_LEN_TIME_STAMP;
210 
211 	case RINGBUF_TYPE_DATA:
212 		return rb_event_data_length(event);
213 	default:
214 		WARN_ON_ONCE(1);
215 	}
216 	/* not hit */
217 	return 0;
218 }
219 
220 /*
221  * Return total length of time extend and data,
222  *   or just the event length for all other events.
223  */
224 static inline unsigned
225 rb_event_ts_length(struct ring_buffer_event *event)
226 {
227 	unsigned len = 0;
228 
229 	if (extended_time(event)) {
230 		/* time extends include the data event after it */
231 		len = RB_LEN_TIME_EXTEND;
232 		event = skip_time_extend(event);
233 	}
234 	return len + rb_event_length(event);
235 }
236 
237 /**
238  * ring_buffer_event_length - return the length of the event
239  * @event: the event to get the length of
240  *
241  * Returns the size of the data load of a data event.
242  * If the event is something other than a data event, it
243  * returns the size of the event itself. With the exception
244  * of a TIME EXTEND, where it still returns the size of the
245  * data load of the data event after it.
246  */
247 unsigned ring_buffer_event_length(struct ring_buffer_event *event)
248 {
249 	unsigned length;
250 
251 	if (extended_time(event))
252 		event = skip_time_extend(event);
253 
254 	length = rb_event_length(event);
255 	if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
256 		return length;
257 	length -= RB_EVNT_HDR_SIZE;
258 	if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
259                 length -= sizeof(event->array[0]);
260 	return length;
261 }
262 EXPORT_SYMBOL_GPL(ring_buffer_event_length);
263 
264 /* inline for ring buffer fast paths */
265 static __always_inline void *
266 rb_event_data(struct ring_buffer_event *event)
267 {
268 	if (extended_time(event))
269 		event = skip_time_extend(event);
270 	WARN_ON_ONCE(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
271 	/* If length is in len field, then array[0] has the data */
272 	if (event->type_len)
273 		return (void *)&event->array[0];
274 	/* Otherwise length is in array[0] and array[1] has the data */
275 	return (void *)&event->array[1];
276 }
277 
278 /**
279  * ring_buffer_event_data - return the data of the event
280  * @event: the event to get the data from
281  */
282 void *ring_buffer_event_data(struct ring_buffer_event *event)
283 {
284 	return rb_event_data(event);
285 }
286 EXPORT_SYMBOL_GPL(ring_buffer_event_data);
287 
288 #define for_each_buffer_cpu(buffer, cpu)		\
289 	for_each_cpu(cpu, buffer->cpumask)
290 
291 #define for_each_online_buffer_cpu(buffer, cpu)		\
292 	for_each_cpu_and(cpu, buffer->cpumask, cpu_online_mask)
293 
294 #define TS_SHIFT	27
295 #define TS_MASK		((1ULL << TS_SHIFT) - 1)
296 #define TS_DELTA_TEST	(~TS_MASK)
297 
298 static u64 rb_event_time_stamp(struct ring_buffer_event *event)
299 {
300 	u64 ts;
301 
302 	ts = event->array[0];
303 	ts <<= TS_SHIFT;
304 	ts += event->time_delta;
305 
306 	return ts;
307 }
308 
309 /* Flag when events were overwritten */
310 #define RB_MISSED_EVENTS	(1 << 31)
311 /* Missed count stored at end */
312 #define RB_MISSED_STORED	(1 << 30)
313 
314 struct buffer_data_page {
315 	u64		 time_stamp;	/* page time stamp */
316 	local_t		 commit;	/* write committed index */
317 	unsigned char	 data[] RB_ALIGN_DATA;	/* data of buffer page */
318 };
319 
320 /*
321  * Note, the buffer_page list must be first. The buffer pages
322  * are allocated in cache lines, which means that each buffer
323  * page will be at the beginning of a cache line, and thus
324  * the least significant bits will be zero. We use this to
325  * add flags in the list struct pointers, to make the ring buffer
326  * lockless.
327  */
328 struct buffer_page {
329 	struct list_head list;		/* list of buffer pages */
330 	local_t		 write;		/* index for next write */
331 	unsigned	 read;		/* index for next read */
332 	local_t		 entries;	/* entries on this page */
333 	unsigned long	 real_end;	/* real end of data */
334 	struct buffer_data_page *page;	/* Actual data page */
335 };
336 
337 /*
338  * The buffer page counters, write and entries, must be reset
339  * atomically when crossing page boundaries. To synchronize this
340  * update, two counters are inserted into the number. One is
341  * the actual counter for the write position or count on the page.
342  *
343  * The other is a counter of updaters. Before an update happens
344  * the update partition of the counter is incremented. This will
345  * allow the updater to update the counter atomically.
346  *
347  * The counter is 20 bits, and the state data is 12.
348  */
349 #define RB_WRITE_MASK		0xfffff
350 #define RB_WRITE_INTCNT		(1 << 20)
351 
352 static void rb_init_page(struct buffer_data_page *bpage)
353 {
354 	local_set(&bpage->commit, 0);
355 }
356 
357 static void free_buffer_page(struct buffer_page *bpage)
358 {
359 	free_page((unsigned long)bpage->page);
360 	kfree(bpage);
361 }
362 
363 /*
364  * We need to fit the time_stamp delta into 27 bits.
365  */
366 static inline bool test_time_stamp(u64 delta)
367 {
368 	return !!(delta & TS_DELTA_TEST);
369 }
370 
371 #define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
372 
373 /* Max payload is BUF_PAGE_SIZE - header (8bytes) */
374 #define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
375 
376 int ring_buffer_print_page_header(struct trace_seq *s)
377 {
378 	struct buffer_data_page field;
379 
380 	trace_seq_printf(s, "\tfield: u64 timestamp;\t"
381 			 "offset:0;\tsize:%u;\tsigned:%u;\n",
382 			 (unsigned int)sizeof(field.time_stamp),
383 			 (unsigned int)is_signed_type(u64));
384 
385 	trace_seq_printf(s, "\tfield: local_t commit;\t"
386 			 "offset:%u;\tsize:%u;\tsigned:%u;\n",
387 			 (unsigned int)offsetof(typeof(field), commit),
388 			 (unsigned int)sizeof(field.commit),
389 			 (unsigned int)is_signed_type(long));
390 
391 	trace_seq_printf(s, "\tfield: int overwrite;\t"
392 			 "offset:%u;\tsize:%u;\tsigned:%u;\n",
393 			 (unsigned int)offsetof(typeof(field), commit),
394 			 1,
395 			 (unsigned int)is_signed_type(long));
396 
397 	trace_seq_printf(s, "\tfield: char data;\t"
398 			 "offset:%u;\tsize:%u;\tsigned:%u;\n",
399 			 (unsigned int)offsetof(typeof(field), data),
400 			 (unsigned int)BUF_PAGE_SIZE,
401 			 (unsigned int)is_signed_type(char));
402 
403 	return !trace_seq_has_overflowed(s);
404 }
405 
406 struct rb_irq_work {
407 	struct irq_work			work;
408 	wait_queue_head_t		waiters;
409 	wait_queue_head_t		full_waiters;
410 	long				wait_index;
411 	bool				waiters_pending;
412 	bool				full_waiters_pending;
413 	bool				wakeup_full;
414 };
415 
416 /*
417  * Structure to hold event state and handle nested events.
418  */
419 struct rb_event_info {
420 	u64			ts;
421 	u64			delta;
422 	u64			before;
423 	u64			after;
424 	unsigned long		length;
425 	struct buffer_page	*tail_page;
426 	int			add_timestamp;
427 };
428 
429 /*
430  * Used for the add_timestamp
431  *  NONE
432  *  EXTEND - wants a time extend
433  *  ABSOLUTE - the buffer requests all events to have absolute time stamps
434  *  FORCE - force a full time stamp.
435  */
436 enum {
437 	RB_ADD_STAMP_NONE		= 0,
438 	RB_ADD_STAMP_EXTEND		= BIT(1),
439 	RB_ADD_STAMP_ABSOLUTE		= BIT(2),
440 	RB_ADD_STAMP_FORCE		= BIT(3)
441 };
442 /*
443  * Used for which event context the event is in.
444  *  TRANSITION = 0
445  *  NMI     = 1
446  *  IRQ     = 2
447  *  SOFTIRQ = 3
448  *  NORMAL  = 4
449  *
450  * See trace_recursive_lock() comment below for more details.
451  */
452 enum {
453 	RB_CTX_TRANSITION,
454 	RB_CTX_NMI,
455 	RB_CTX_IRQ,
456 	RB_CTX_SOFTIRQ,
457 	RB_CTX_NORMAL,
458 	RB_CTX_MAX
459 };
460 
461 #if BITS_PER_LONG == 32
462 #define RB_TIME_32
463 #endif
464 
465 /* To test on 64 bit machines */
466 //#define RB_TIME_32
467 
468 #ifdef RB_TIME_32
469 
470 struct rb_time_struct {
471 	local_t		cnt;
472 	local_t		top;
473 	local_t		bottom;
474 	local_t		msb;
475 };
476 #else
477 #include <asm/local64.h>
478 struct rb_time_struct {
479 	local64_t	time;
480 };
481 #endif
482 typedef struct rb_time_struct rb_time_t;
483 
484 #define MAX_NEST	5
485 
486 /*
487  * head_page == tail_page && head == tail then buffer is empty.
488  */
489 struct ring_buffer_per_cpu {
490 	int				cpu;
491 	atomic_t			record_disabled;
492 	atomic_t			resize_disabled;
493 	struct trace_buffer	*buffer;
494 	raw_spinlock_t			reader_lock;	/* serialize readers */
495 	arch_spinlock_t			lock;
496 	struct lock_class_key		lock_key;
497 	struct buffer_data_page		*free_page;
498 	unsigned long			nr_pages;
499 	unsigned int			current_context;
500 	struct list_head		*pages;
501 	struct buffer_page		*head_page;	/* read from head */
502 	struct buffer_page		*tail_page;	/* write to tail */
503 	struct buffer_page		*commit_page;	/* committed pages */
504 	struct buffer_page		*reader_page;
505 	unsigned long			lost_events;
506 	unsigned long			last_overrun;
507 	unsigned long			nest;
508 	local_t				entries_bytes;
509 	local_t				entries;
510 	local_t				overrun;
511 	local_t				commit_overrun;
512 	local_t				dropped_events;
513 	local_t				committing;
514 	local_t				commits;
515 	local_t				pages_touched;
516 	local_t				pages_lost;
517 	local_t				pages_read;
518 	long				last_pages_touch;
519 	size_t				shortest_full;
520 	unsigned long			read;
521 	unsigned long			read_bytes;
522 	rb_time_t			write_stamp;
523 	rb_time_t			before_stamp;
524 	u64				event_stamp[MAX_NEST];
525 	u64				read_stamp;
526 	/* pages removed since last reset */
527 	unsigned long			pages_removed;
528 	/* ring buffer pages to update, > 0 to add, < 0 to remove */
529 	long				nr_pages_to_update;
530 	struct list_head		new_pages; /* new pages to add */
531 	struct work_struct		update_pages_work;
532 	struct completion		update_done;
533 
534 	struct rb_irq_work		irq_work;
535 };
536 
537 struct trace_buffer {
538 	unsigned			flags;
539 	int				cpus;
540 	atomic_t			record_disabled;
541 	atomic_t			resizing;
542 	cpumask_var_t			cpumask;
543 
544 	struct lock_class_key		*reader_lock_key;
545 
546 	struct mutex			mutex;
547 
548 	struct ring_buffer_per_cpu	**buffers;
549 
550 	struct hlist_node		node;
551 	u64				(*clock)(void);
552 
553 	struct rb_irq_work		irq_work;
554 	bool				time_stamp_abs;
555 };
556 
557 struct ring_buffer_iter {
558 	struct ring_buffer_per_cpu	*cpu_buffer;
559 	unsigned long			head;
560 	unsigned long			next_event;
561 	struct buffer_page		*head_page;
562 	struct buffer_page		*cache_reader_page;
563 	unsigned long			cache_read;
564 	unsigned long			cache_pages_removed;
565 	u64				read_stamp;
566 	u64				page_stamp;
567 	struct ring_buffer_event	*event;
568 	int				missed_events;
569 };
570 
571 #ifdef RB_TIME_32
572 
573 /*
574  * On 32 bit machines, local64_t is very expensive. As the ring
575  * buffer doesn't need all the features of a true 64 bit atomic,
576  * on 32 bit, it uses these functions (64 still uses local64_t).
577  *
578  * For the ring buffer, 64 bit required operations for the time is
579  * the following:
580  *
581  *  - Reads may fail if it interrupted a modification of the time stamp.
582  *      It will succeed if it did not interrupt another write even if
583  *      the read itself is interrupted by a write.
584  *      It returns whether it was successful or not.
585  *
586  *  - Writes always succeed and will overwrite other writes and writes
587  *      that were done by events interrupting the current write.
588  *
589  *  - A write followed by a read of the same time stamp will always succeed,
590  *      but may not contain the same value.
591  *
592  *  - A cmpxchg will fail if it interrupted another write or cmpxchg.
593  *      Other than that, it acts like a normal cmpxchg.
594  *
595  * The 60 bit time stamp is broken up by 30 bits in a top and bottom half
596  *  (bottom being the least significant 30 bits of the 60 bit time stamp).
597  *
598  * The two most significant bits of each half holds a 2 bit counter (0-3).
599  * Each update will increment this counter by one.
600  * When reading the top and bottom, if the two counter bits match then the
601  *  top and bottom together make a valid 60 bit number.
602  */
603 #define RB_TIME_SHIFT	30
604 #define RB_TIME_VAL_MASK ((1 << RB_TIME_SHIFT) - 1)
605 #define RB_TIME_MSB_SHIFT	 60
606 
607 static inline int rb_time_cnt(unsigned long val)
608 {
609 	return (val >> RB_TIME_SHIFT) & 3;
610 }
611 
612 static inline u64 rb_time_val(unsigned long top, unsigned long bottom)
613 {
614 	u64 val;
615 
616 	val = top & RB_TIME_VAL_MASK;
617 	val <<= RB_TIME_SHIFT;
618 	val |= bottom & RB_TIME_VAL_MASK;
619 
620 	return val;
621 }
622 
623 static inline bool __rb_time_read(rb_time_t *t, u64 *ret, unsigned long *cnt)
624 {
625 	unsigned long top, bottom, msb;
626 	unsigned long c;
627 
628 	/*
629 	 * If the read is interrupted by a write, then the cnt will
630 	 * be different. Loop until both top and bottom have been read
631 	 * without interruption.
632 	 */
633 	do {
634 		c = local_read(&t->cnt);
635 		top = local_read(&t->top);
636 		bottom = local_read(&t->bottom);
637 		msb = local_read(&t->msb);
638 	} while (c != local_read(&t->cnt));
639 
640 	*cnt = rb_time_cnt(top);
641 
642 	/* If top and bottom counts don't match, this interrupted a write */
643 	if (*cnt != rb_time_cnt(bottom))
644 		return false;
645 
646 	/* The shift to msb will lose its cnt bits */
647 	*ret = rb_time_val(top, bottom) | ((u64)msb << RB_TIME_MSB_SHIFT);
648 	return true;
649 }
650 
651 static bool rb_time_read(rb_time_t *t, u64 *ret)
652 {
653 	unsigned long cnt;
654 
655 	return __rb_time_read(t, ret, &cnt);
656 }
657 
658 static inline unsigned long rb_time_val_cnt(unsigned long val, unsigned long cnt)
659 {
660 	return (val & RB_TIME_VAL_MASK) | ((cnt & 3) << RB_TIME_SHIFT);
661 }
662 
663 static inline void rb_time_split(u64 val, unsigned long *top, unsigned long *bottom,
664 				 unsigned long *msb)
665 {
666 	*top = (unsigned long)((val >> RB_TIME_SHIFT) & RB_TIME_VAL_MASK);
667 	*bottom = (unsigned long)(val & RB_TIME_VAL_MASK);
668 	*msb = (unsigned long)(val >> RB_TIME_MSB_SHIFT);
669 }
670 
671 static inline void rb_time_val_set(local_t *t, unsigned long val, unsigned long cnt)
672 {
673 	val = rb_time_val_cnt(val, cnt);
674 	local_set(t, val);
675 }
676 
677 static void rb_time_set(rb_time_t *t, u64 val)
678 {
679 	unsigned long cnt, top, bottom, msb;
680 
681 	rb_time_split(val, &top, &bottom, &msb);
682 
683 	/* Writes always succeed with a valid number even if it gets interrupted. */
684 	do {
685 		cnt = local_inc_return(&t->cnt);
686 		rb_time_val_set(&t->top, top, cnt);
687 		rb_time_val_set(&t->bottom, bottom, cnt);
688 		rb_time_val_set(&t->msb, val >> RB_TIME_MSB_SHIFT, cnt);
689 	} while (cnt != local_read(&t->cnt));
690 }
691 
692 static inline bool
693 rb_time_read_cmpxchg(local_t *l, unsigned long expect, unsigned long set)
694 {
695 	return local_try_cmpxchg(l, &expect, set);
696 }
697 
698 static bool rb_time_cmpxchg(rb_time_t *t, u64 expect, u64 set)
699 {
700 	unsigned long cnt, top, bottom, msb;
701 	unsigned long cnt2, top2, bottom2, msb2;
702 	u64 val;
703 
704 	/* The cmpxchg always fails if it interrupted an update */
705 	 if (!__rb_time_read(t, &val, &cnt2))
706 		 return false;
707 
708 	 if (val != expect)
709 		 return false;
710 
711 	 cnt = local_read(&t->cnt);
712 	 if ((cnt & 3) != cnt2)
713 		 return false;
714 
715 	 cnt2 = cnt + 1;
716 
717 	 rb_time_split(val, &top, &bottom, &msb);
718 	 top = rb_time_val_cnt(top, cnt);
719 	 bottom = rb_time_val_cnt(bottom, cnt);
720 
721 	 rb_time_split(set, &top2, &bottom2, &msb2);
722 	 top2 = rb_time_val_cnt(top2, cnt2);
723 	 bottom2 = rb_time_val_cnt(bottom2, cnt2);
724 
725 	if (!rb_time_read_cmpxchg(&t->cnt, cnt, cnt2))
726 		return false;
727 	if (!rb_time_read_cmpxchg(&t->msb, msb, msb2))
728 		return false;
729 	if (!rb_time_read_cmpxchg(&t->top, top, top2))
730 		return false;
731 	if (!rb_time_read_cmpxchg(&t->bottom, bottom, bottom2))
732 		return false;
733 	return true;
734 }
735 
736 #else /* 64 bits */
737 
738 /* local64_t always succeeds */
739 
740 static inline bool rb_time_read(rb_time_t *t, u64 *ret)
741 {
742 	*ret = local64_read(&t->time);
743 	return true;
744 }
745 static void rb_time_set(rb_time_t *t, u64 val)
746 {
747 	local64_set(&t->time, val);
748 }
749 
750 static bool rb_time_cmpxchg(rb_time_t *t, u64 expect, u64 set)
751 {
752 	return local64_try_cmpxchg(&t->time, &expect, set);
753 }
754 #endif
755 
756 /*
757  * Enable this to make sure that the event passed to
758  * ring_buffer_event_time_stamp() is not committed and also
759  * is on the buffer that it passed in.
760  */
761 //#define RB_VERIFY_EVENT
762 #ifdef RB_VERIFY_EVENT
763 static struct list_head *rb_list_head(struct list_head *list);
764 static void verify_event(struct ring_buffer_per_cpu *cpu_buffer,
765 			 void *event)
766 {
767 	struct buffer_page *page = cpu_buffer->commit_page;
768 	struct buffer_page *tail_page = READ_ONCE(cpu_buffer->tail_page);
769 	struct list_head *next;
770 	long commit, write;
771 	unsigned long addr = (unsigned long)event;
772 	bool done = false;
773 	int stop = 0;
774 
775 	/* Make sure the event exists and is not committed yet */
776 	do {
777 		if (page == tail_page || WARN_ON_ONCE(stop++ > 100))
778 			done = true;
779 		commit = local_read(&page->page->commit);
780 		write = local_read(&page->write);
781 		if (addr >= (unsigned long)&page->page->data[commit] &&
782 		    addr < (unsigned long)&page->page->data[write])
783 			return;
784 
785 		next = rb_list_head(page->list.next);
786 		page = list_entry(next, struct buffer_page, list);
787 	} while (!done);
788 	WARN_ON_ONCE(1);
789 }
790 #else
791 static inline void verify_event(struct ring_buffer_per_cpu *cpu_buffer,
792 			 void *event)
793 {
794 }
795 #endif
796 
797 /*
798  * The absolute time stamp drops the 5 MSBs and some clocks may
799  * require them. The rb_fix_abs_ts() will take a previous full
800  * time stamp, and add the 5 MSB of that time stamp on to the
801  * saved absolute time stamp. Then they are compared in case of
802  * the unlikely event that the latest time stamp incremented
803  * the 5 MSB.
804  */
805 static inline u64 rb_fix_abs_ts(u64 abs, u64 save_ts)
806 {
807 	if (save_ts & TS_MSB) {
808 		abs |= save_ts & TS_MSB;
809 		/* Check for overflow */
810 		if (unlikely(abs < save_ts))
811 			abs += 1ULL << 59;
812 	}
813 	return abs;
814 }
815 
816 static inline u64 rb_time_stamp(struct trace_buffer *buffer);
817 
818 /**
819  * ring_buffer_event_time_stamp - return the event's current time stamp
820  * @buffer: The buffer that the event is on
821  * @event: the event to get the time stamp of
822  *
823  * Note, this must be called after @event is reserved, and before it is
824  * committed to the ring buffer. And must be called from the same
825  * context where the event was reserved (normal, softirq, irq, etc).
826  *
827  * Returns the time stamp associated with the current event.
828  * If the event has an extended time stamp, then that is used as
829  * the time stamp to return.
830  * In the highly unlikely case that the event was nested more than
831  * the max nesting, then the write_stamp of the buffer is returned,
832  * otherwise  current time is returned, but that really neither of
833  * the last two cases should ever happen.
834  */
835 u64 ring_buffer_event_time_stamp(struct trace_buffer *buffer,
836 				 struct ring_buffer_event *event)
837 {
838 	struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[smp_processor_id()];
839 	unsigned int nest;
840 	u64 ts;
841 
842 	/* If the event includes an absolute time, then just use that */
843 	if (event->type_len == RINGBUF_TYPE_TIME_STAMP) {
844 		ts = rb_event_time_stamp(event);
845 		return rb_fix_abs_ts(ts, cpu_buffer->tail_page->page->time_stamp);
846 	}
847 
848 	nest = local_read(&cpu_buffer->committing);
849 	verify_event(cpu_buffer, event);
850 	if (WARN_ON_ONCE(!nest))
851 		goto fail;
852 
853 	/* Read the current saved nesting level time stamp */
854 	if (likely(--nest < MAX_NEST))
855 		return cpu_buffer->event_stamp[nest];
856 
857 	/* Shouldn't happen, warn if it does */
858 	WARN_ONCE(1, "nest (%d) greater than max", nest);
859 
860  fail:
861 	/* Can only fail on 32 bit */
862 	if (!rb_time_read(&cpu_buffer->write_stamp, &ts))
863 		/* Screw it, just read the current time */
864 		ts = rb_time_stamp(cpu_buffer->buffer);
865 
866 	return ts;
867 }
868 
869 /**
870  * ring_buffer_nr_pages - get the number of buffer pages in the ring buffer
871  * @buffer: The ring_buffer to get the number of pages from
872  * @cpu: The cpu of the ring_buffer to get the number of pages from
873  *
874  * Returns the number of pages used by a per_cpu buffer of the ring buffer.
875  */
876 size_t ring_buffer_nr_pages(struct trace_buffer *buffer, int cpu)
877 {
878 	return buffer->buffers[cpu]->nr_pages;
879 }
880 
881 /**
882  * ring_buffer_nr_dirty_pages - get the number of used pages in the ring buffer
883  * @buffer: The ring_buffer to get the number of pages from
884  * @cpu: The cpu of the ring_buffer to get the number of pages from
885  *
886  * Returns the number of pages that have content in the ring buffer.
887  */
888 size_t ring_buffer_nr_dirty_pages(struct trace_buffer *buffer, int cpu)
889 {
890 	size_t read;
891 	size_t lost;
892 	size_t cnt;
893 
894 	read = local_read(&buffer->buffers[cpu]->pages_read);
895 	lost = local_read(&buffer->buffers[cpu]->pages_lost);
896 	cnt = local_read(&buffer->buffers[cpu]->pages_touched);
897 
898 	if (WARN_ON_ONCE(cnt < lost))
899 		return 0;
900 
901 	cnt -= lost;
902 
903 	/* The reader can read an empty page, but not more than that */
904 	if (cnt < read) {
905 		WARN_ON_ONCE(read > cnt + 1);
906 		return 0;
907 	}
908 
909 	return cnt - read;
910 }
911 
912 static __always_inline bool full_hit(struct trace_buffer *buffer, int cpu, int full)
913 {
914 	struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
915 	size_t nr_pages;
916 	size_t dirty;
917 
918 	nr_pages = cpu_buffer->nr_pages;
919 	if (!nr_pages || !full)
920 		return true;
921 
922 	dirty = ring_buffer_nr_dirty_pages(buffer, cpu);
923 
924 	return (dirty * 100) > (full * nr_pages);
925 }
926 
927 /*
928  * rb_wake_up_waiters - wake up tasks waiting for ring buffer input
929  *
930  * Schedules a delayed work to wake up any task that is blocked on the
931  * ring buffer waiters queue.
932  */
933 static void rb_wake_up_waiters(struct irq_work *work)
934 {
935 	struct rb_irq_work *rbwork = container_of(work, struct rb_irq_work, work);
936 
937 	wake_up_all(&rbwork->waiters);
938 	if (rbwork->full_waiters_pending || rbwork->wakeup_full) {
939 		rbwork->wakeup_full = false;
940 		rbwork->full_waiters_pending = false;
941 		wake_up_all(&rbwork->full_waiters);
942 	}
943 }
944 
945 /**
946  * ring_buffer_wake_waiters - wake up any waiters on this ring buffer
947  * @buffer: The ring buffer to wake waiters on
948  * @cpu: The CPU buffer to wake waiters on
949  *
950  * In the case of a file that represents a ring buffer is closing,
951  * it is prudent to wake up any waiters that are on this.
952  */
953 void ring_buffer_wake_waiters(struct trace_buffer *buffer, int cpu)
954 {
955 	struct ring_buffer_per_cpu *cpu_buffer;
956 	struct rb_irq_work *rbwork;
957 
958 	if (!buffer)
959 		return;
960 
961 	if (cpu == RING_BUFFER_ALL_CPUS) {
962 
963 		/* Wake up individual ones too. One level recursion */
964 		for_each_buffer_cpu(buffer, cpu)
965 			ring_buffer_wake_waiters(buffer, cpu);
966 
967 		rbwork = &buffer->irq_work;
968 	} else {
969 		if (WARN_ON_ONCE(!buffer->buffers))
970 			return;
971 		if (WARN_ON_ONCE(cpu >= nr_cpu_ids))
972 			return;
973 
974 		cpu_buffer = buffer->buffers[cpu];
975 		/* The CPU buffer may not have been initialized yet */
976 		if (!cpu_buffer)
977 			return;
978 		rbwork = &cpu_buffer->irq_work;
979 	}
980 
981 	rbwork->wait_index++;
982 	/* make sure the waiters see the new index */
983 	smp_wmb();
984 
985 	rb_wake_up_waiters(&rbwork->work);
986 }
987 
988 /**
989  * ring_buffer_wait - wait for input to the ring buffer
990  * @buffer: buffer to wait on
991  * @cpu: the cpu buffer to wait on
992  * @full: wait until the percentage of pages are available, if @cpu != RING_BUFFER_ALL_CPUS
993  *
994  * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
995  * as data is added to any of the @buffer's cpu buffers. Otherwise
996  * it will wait for data to be added to a specific cpu buffer.
997  */
998 int ring_buffer_wait(struct trace_buffer *buffer, int cpu, int full)
999 {
1000 	struct ring_buffer_per_cpu *cpu_buffer;
1001 	DEFINE_WAIT(wait);
1002 	struct rb_irq_work *work;
1003 	long wait_index;
1004 	int ret = 0;
1005 
1006 	/*
1007 	 * Depending on what the caller is waiting for, either any
1008 	 * data in any cpu buffer, or a specific buffer, put the
1009 	 * caller on the appropriate wait queue.
1010 	 */
1011 	if (cpu == RING_BUFFER_ALL_CPUS) {
1012 		work = &buffer->irq_work;
1013 		/* Full only makes sense on per cpu reads */
1014 		full = 0;
1015 	} else {
1016 		if (!cpumask_test_cpu(cpu, buffer->cpumask))
1017 			return -ENODEV;
1018 		cpu_buffer = buffer->buffers[cpu];
1019 		work = &cpu_buffer->irq_work;
1020 	}
1021 
1022 	wait_index = READ_ONCE(work->wait_index);
1023 
1024 	while (true) {
1025 		if (full)
1026 			prepare_to_wait(&work->full_waiters, &wait, TASK_INTERRUPTIBLE);
1027 		else
1028 			prepare_to_wait(&work->waiters, &wait, TASK_INTERRUPTIBLE);
1029 
1030 		/*
1031 		 * The events can happen in critical sections where
1032 		 * checking a work queue can cause deadlocks.
1033 		 * After adding a task to the queue, this flag is set
1034 		 * only to notify events to try to wake up the queue
1035 		 * using irq_work.
1036 		 *
1037 		 * We don't clear it even if the buffer is no longer
1038 		 * empty. The flag only causes the next event to run
1039 		 * irq_work to do the work queue wake up. The worse
1040 		 * that can happen if we race with !trace_empty() is that
1041 		 * an event will cause an irq_work to try to wake up
1042 		 * an empty queue.
1043 		 *
1044 		 * There's no reason to protect this flag either, as
1045 		 * the work queue and irq_work logic will do the necessary
1046 		 * synchronization for the wake ups. The only thing
1047 		 * that is necessary is that the wake up happens after
1048 		 * a task has been queued. It's OK for spurious wake ups.
1049 		 */
1050 		if (full)
1051 			work->full_waiters_pending = true;
1052 		else
1053 			work->waiters_pending = true;
1054 
1055 		if (signal_pending(current)) {
1056 			ret = -EINTR;
1057 			break;
1058 		}
1059 
1060 		if (cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer))
1061 			break;
1062 
1063 		if (cpu != RING_BUFFER_ALL_CPUS &&
1064 		    !ring_buffer_empty_cpu(buffer, cpu)) {
1065 			unsigned long flags;
1066 			bool pagebusy;
1067 			bool done;
1068 
1069 			if (!full)
1070 				break;
1071 
1072 			raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
1073 			pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
1074 			done = !pagebusy && full_hit(buffer, cpu, full);
1075 
1076 			if (!cpu_buffer->shortest_full ||
1077 			    cpu_buffer->shortest_full > full)
1078 				cpu_buffer->shortest_full = full;
1079 			raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
1080 			if (done)
1081 				break;
1082 		}
1083 
1084 		schedule();
1085 
1086 		/* Make sure to see the new wait index */
1087 		smp_rmb();
1088 		if (wait_index != work->wait_index)
1089 			break;
1090 	}
1091 
1092 	if (full)
1093 		finish_wait(&work->full_waiters, &wait);
1094 	else
1095 		finish_wait(&work->waiters, &wait);
1096 
1097 	return ret;
1098 }
1099 
1100 /**
1101  * ring_buffer_poll_wait - poll on buffer input
1102  * @buffer: buffer to wait on
1103  * @cpu: the cpu buffer to wait on
1104  * @filp: the file descriptor
1105  * @poll_table: The poll descriptor
1106  * @full: wait until the percentage of pages are available, if @cpu != RING_BUFFER_ALL_CPUS
1107  *
1108  * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
1109  * as data is added to any of the @buffer's cpu buffers. Otherwise
1110  * it will wait for data to be added to a specific cpu buffer.
1111  *
1112  * Returns EPOLLIN | EPOLLRDNORM if data exists in the buffers,
1113  * zero otherwise.
1114  */
1115 __poll_t ring_buffer_poll_wait(struct trace_buffer *buffer, int cpu,
1116 			  struct file *filp, poll_table *poll_table, int full)
1117 {
1118 	struct ring_buffer_per_cpu *cpu_buffer;
1119 	struct rb_irq_work *work;
1120 
1121 	if (cpu == RING_BUFFER_ALL_CPUS) {
1122 		work = &buffer->irq_work;
1123 		full = 0;
1124 	} else {
1125 		if (!cpumask_test_cpu(cpu, buffer->cpumask))
1126 			return -EINVAL;
1127 
1128 		cpu_buffer = buffer->buffers[cpu];
1129 		work = &cpu_buffer->irq_work;
1130 	}
1131 
1132 	if (full) {
1133 		poll_wait(filp, &work->full_waiters, poll_table);
1134 		work->full_waiters_pending = true;
1135 	} else {
1136 		poll_wait(filp, &work->waiters, poll_table);
1137 		work->waiters_pending = true;
1138 	}
1139 
1140 	/*
1141 	 * There's a tight race between setting the waiters_pending and
1142 	 * checking if the ring buffer is empty.  Once the waiters_pending bit
1143 	 * is set, the next event will wake the task up, but we can get stuck
1144 	 * if there's only a single event in.
1145 	 *
1146 	 * FIXME: Ideally, we need a memory barrier on the writer side as well,
1147 	 * but adding a memory barrier to all events will cause too much of a
1148 	 * performance hit in the fast path.  We only need a memory barrier when
1149 	 * the buffer goes from empty to having content.  But as this race is
1150 	 * extremely small, and it's not a problem if another event comes in, we
1151 	 * will fix it later.
1152 	 */
1153 	smp_mb();
1154 
1155 	if (full)
1156 		return full_hit(buffer, cpu, full) ? EPOLLIN | EPOLLRDNORM : 0;
1157 
1158 	if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
1159 	    (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
1160 		return EPOLLIN | EPOLLRDNORM;
1161 	return 0;
1162 }
1163 
1164 /* buffer may be either ring_buffer or ring_buffer_per_cpu */
1165 #define RB_WARN_ON(b, cond)						\
1166 	({								\
1167 		int _____ret = unlikely(cond);				\
1168 		if (_____ret) {						\
1169 			if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
1170 				struct ring_buffer_per_cpu *__b =	\
1171 					(void *)b;			\
1172 				atomic_inc(&__b->buffer->record_disabled); \
1173 			} else						\
1174 				atomic_inc(&b->record_disabled);	\
1175 			WARN_ON(1);					\
1176 		}							\
1177 		_____ret;						\
1178 	})
1179 
1180 /* Up this if you want to test the TIME_EXTENTS and normalization */
1181 #define DEBUG_SHIFT 0
1182 
1183 static inline u64 rb_time_stamp(struct trace_buffer *buffer)
1184 {
1185 	u64 ts;
1186 
1187 	/* Skip retpolines :-( */
1188 	if (IS_ENABLED(CONFIG_RETPOLINE) && likely(buffer->clock == trace_clock_local))
1189 		ts = trace_clock_local();
1190 	else
1191 		ts = buffer->clock();
1192 
1193 	/* shift to debug/test normalization and TIME_EXTENTS */
1194 	return ts << DEBUG_SHIFT;
1195 }
1196 
1197 u64 ring_buffer_time_stamp(struct trace_buffer *buffer)
1198 {
1199 	u64 time;
1200 
1201 	preempt_disable_notrace();
1202 	time = rb_time_stamp(buffer);
1203 	preempt_enable_notrace();
1204 
1205 	return time;
1206 }
1207 EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
1208 
1209 void ring_buffer_normalize_time_stamp(struct trace_buffer *buffer,
1210 				      int cpu, u64 *ts)
1211 {
1212 	/* Just stupid testing the normalize function and deltas */
1213 	*ts >>= DEBUG_SHIFT;
1214 }
1215 EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
1216 
1217 /*
1218  * Making the ring buffer lockless makes things tricky.
1219  * Although writes only happen on the CPU that they are on,
1220  * and they only need to worry about interrupts. Reads can
1221  * happen on any CPU.
1222  *
1223  * The reader page is always off the ring buffer, but when the
1224  * reader finishes with a page, it needs to swap its page with
1225  * a new one from the buffer. The reader needs to take from
1226  * the head (writes go to the tail). But if a writer is in overwrite
1227  * mode and wraps, it must push the head page forward.
1228  *
1229  * Here lies the problem.
1230  *
1231  * The reader must be careful to replace only the head page, and
1232  * not another one. As described at the top of the file in the
1233  * ASCII art, the reader sets its old page to point to the next
1234  * page after head. It then sets the page after head to point to
1235  * the old reader page. But if the writer moves the head page
1236  * during this operation, the reader could end up with the tail.
1237  *
1238  * We use cmpxchg to help prevent this race. We also do something
1239  * special with the page before head. We set the LSB to 1.
1240  *
1241  * When the writer must push the page forward, it will clear the
1242  * bit that points to the head page, move the head, and then set
1243  * the bit that points to the new head page.
1244  *
1245  * We also don't want an interrupt coming in and moving the head
1246  * page on another writer. Thus we use the second LSB to catch
1247  * that too. Thus:
1248  *
1249  * head->list->prev->next        bit 1          bit 0
1250  *                              -------        -------
1251  * Normal page                     0              0
1252  * Points to head page             0              1
1253  * New head page                   1              0
1254  *
1255  * Note we can not trust the prev pointer of the head page, because:
1256  *
1257  * +----+       +-----+        +-----+
1258  * |    |------>|  T  |---X--->|  N  |
1259  * |    |<------|     |        |     |
1260  * +----+       +-----+        +-----+
1261  *   ^                           ^ |
1262  *   |          +-----+          | |
1263  *   +----------|  R  |----------+ |
1264  *              |     |<-----------+
1265  *              +-----+
1266  *
1267  * Key:  ---X-->  HEAD flag set in pointer
1268  *         T      Tail page
1269  *         R      Reader page
1270  *         N      Next page
1271  *
1272  * (see __rb_reserve_next() to see where this happens)
1273  *
1274  *  What the above shows is that the reader just swapped out
1275  *  the reader page with a page in the buffer, but before it
1276  *  could make the new header point back to the new page added
1277  *  it was preempted by a writer. The writer moved forward onto
1278  *  the new page added by the reader and is about to move forward
1279  *  again.
1280  *
1281  *  You can see, it is legitimate for the previous pointer of
1282  *  the head (or any page) not to point back to itself. But only
1283  *  temporarily.
1284  */
1285 
1286 #define RB_PAGE_NORMAL		0UL
1287 #define RB_PAGE_HEAD		1UL
1288 #define RB_PAGE_UPDATE		2UL
1289 
1290 
1291 #define RB_FLAG_MASK		3UL
1292 
1293 /* PAGE_MOVED is not part of the mask */
1294 #define RB_PAGE_MOVED		4UL
1295 
1296 /*
1297  * rb_list_head - remove any bit
1298  */
1299 static struct list_head *rb_list_head(struct list_head *list)
1300 {
1301 	unsigned long val = (unsigned long)list;
1302 
1303 	return (struct list_head *)(val & ~RB_FLAG_MASK);
1304 }
1305 
1306 /*
1307  * rb_is_head_page - test if the given page is the head page
1308  *
1309  * Because the reader may move the head_page pointer, we can
1310  * not trust what the head page is (it may be pointing to
1311  * the reader page). But if the next page is a header page,
1312  * its flags will be non zero.
1313  */
1314 static inline int
1315 rb_is_head_page(struct buffer_page *page, struct list_head *list)
1316 {
1317 	unsigned long val;
1318 
1319 	val = (unsigned long)list->next;
1320 
1321 	if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
1322 		return RB_PAGE_MOVED;
1323 
1324 	return val & RB_FLAG_MASK;
1325 }
1326 
1327 /*
1328  * rb_is_reader_page
1329  *
1330  * The unique thing about the reader page, is that, if the
1331  * writer is ever on it, the previous pointer never points
1332  * back to the reader page.
1333  */
1334 static bool rb_is_reader_page(struct buffer_page *page)
1335 {
1336 	struct list_head *list = page->list.prev;
1337 
1338 	return rb_list_head(list->next) != &page->list;
1339 }
1340 
1341 /*
1342  * rb_set_list_to_head - set a list_head to be pointing to head.
1343  */
1344 static void rb_set_list_to_head(struct list_head *list)
1345 {
1346 	unsigned long *ptr;
1347 
1348 	ptr = (unsigned long *)&list->next;
1349 	*ptr |= RB_PAGE_HEAD;
1350 	*ptr &= ~RB_PAGE_UPDATE;
1351 }
1352 
1353 /*
1354  * rb_head_page_activate - sets up head page
1355  */
1356 static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
1357 {
1358 	struct buffer_page *head;
1359 
1360 	head = cpu_buffer->head_page;
1361 	if (!head)
1362 		return;
1363 
1364 	/*
1365 	 * Set the previous list pointer to have the HEAD flag.
1366 	 */
1367 	rb_set_list_to_head(head->list.prev);
1368 }
1369 
1370 static void rb_list_head_clear(struct list_head *list)
1371 {
1372 	unsigned long *ptr = (unsigned long *)&list->next;
1373 
1374 	*ptr &= ~RB_FLAG_MASK;
1375 }
1376 
1377 /*
1378  * rb_head_page_deactivate - clears head page ptr (for free list)
1379  */
1380 static void
1381 rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
1382 {
1383 	struct list_head *hd;
1384 
1385 	/* Go through the whole list and clear any pointers found. */
1386 	rb_list_head_clear(cpu_buffer->pages);
1387 
1388 	list_for_each(hd, cpu_buffer->pages)
1389 		rb_list_head_clear(hd);
1390 }
1391 
1392 static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
1393 			    struct buffer_page *head,
1394 			    struct buffer_page *prev,
1395 			    int old_flag, int new_flag)
1396 {
1397 	struct list_head *list;
1398 	unsigned long val = (unsigned long)&head->list;
1399 	unsigned long ret;
1400 
1401 	list = &prev->list;
1402 
1403 	val &= ~RB_FLAG_MASK;
1404 
1405 	ret = cmpxchg((unsigned long *)&list->next,
1406 		      val | old_flag, val | new_flag);
1407 
1408 	/* check if the reader took the page */
1409 	if ((ret & ~RB_FLAG_MASK) != val)
1410 		return RB_PAGE_MOVED;
1411 
1412 	return ret & RB_FLAG_MASK;
1413 }
1414 
1415 static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
1416 				   struct buffer_page *head,
1417 				   struct buffer_page *prev,
1418 				   int old_flag)
1419 {
1420 	return rb_head_page_set(cpu_buffer, head, prev,
1421 				old_flag, RB_PAGE_UPDATE);
1422 }
1423 
1424 static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
1425 				 struct buffer_page *head,
1426 				 struct buffer_page *prev,
1427 				 int old_flag)
1428 {
1429 	return rb_head_page_set(cpu_buffer, head, prev,
1430 				old_flag, RB_PAGE_HEAD);
1431 }
1432 
1433 static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
1434 				   struct buffer_page *head,
1435 				   struct buffer_page *prev,
1436 				   int old_flag)
1437 {
1438 	return rb_head_page_set(cpu_buffer, head, prev,
1439 				old_flag, RB_PAGE_NORMAL);
1440 }
1441 
1442 static inline void rb_inc_page(struct buffer_page **bpage)
1443 {
1444 	struct list_head *p = rb_list_head((*bpage)->list.next);
1445 
1446 	*bpage = list_entry(p, struct buffer_page, list);
1447 }
1448 
1449 static struct buffer_page *
1450 rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
1451 {
1452 	struct buffer_page *head;
1453 	struct buffer_page *page;
1454 	struct list_head *list;
1455 	int i;
1456 
1457 	if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
1458 		return NULL;
1459 
1460 	/* sanity check */
1461 	list = cpu_buffer->pages;
1462 	if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
1463 		return NULL;
1464 
1465 	page = head = cpu_buffer->head_page;
1466 	/*
1467 	 * It is possible that the writer moves the header behind
1468 	 * where we started, and we miss in one loop.
1469 	 * A second loop should grab the header, but we'll do
1470 	 * three loops just because I'm paranoid.
1471 	 */
1472 	for (i = 0; i < 3; i++) {
1473 		do {
1474 			if (rb_is_head_page(page, page->list.prev)) {
1475 				cpu_buffer->head_page = page;
1476 				return page;
1477 			}
1478 			rb_inc_page(&page);
1479 		} while (page != head);
1480 	}
1481 
1482 	RB_WARN_ON(cpu_buffer, 1);
1483 
1484 	return NULL;
1485 }
1486 
1487 static bool rb_head_page_replace(struct buffer_page *old,
1488 				struct buffer_page *new)
1489 {
1490 	unsigned long *ptr = (unsigned long *)&old->list.prev->next;
1491 	unsigned long val;
1492 
1493 	val = *ptr & ~RB_FLAG_MASK;
1494 	val |= RB_PAGE_HEAD;
1495 
1496 	return try_cmpxchg(ptr, &val, (unsigned long)&new->list);
1497 }
1498 
1499 /*
1500  * rb_tail_page_update - move the tail page forward
1501  */
1502 static void rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
1503 			       struct buffer_page *tail_page,
1504 			       struct buffer_page *next_page)
1505 {
1506 	unsigned long old_entries;
1507 	unsigned long old_write;
1508 
1509 	/*
1510 	 * The tail page now needs to be moved forward.
1511 	 *
1512 	 * We need to reset the tail page, but without messing
1513 	 * with possible erasing of data brought in by interrupts
1514 	 * that have moved the tail page and are currently on it.
1515 	 *
1516 	 * We add a counter to the write field to denote this.
1517 	 */
1518 	old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
1519 	old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
1520 
1521 	local_inc(&cpu_buffer->pages_touched);
1522 	/*
1523 	 * Just make sure we have seen our old_write and synchronize
1524 	 * with any interrupts that come in.
1525 	 */
1526 	barrier();
1527 
1528 	/*
1529 	 * If the tail page is still the same as what we think
1530 	 * it is, then it is up to us to update the tail
1531 	 * pointer.
1532 	 */
1533 	if (tail_page == READ_ONCE(cpu_buffer->tail_page)) {
1534 		/* Zero the write counter */
1535 		unsigned long val = old_write & ~RB_WRITE_MASK;
1536 		unsigned long eval = old_entries & ~RB_WRITE_MASK;
1537 
1538 		/*
1539 		 * This will only succeed if an interrupt did
1540 		 * not come in and change it. In which case, we
1541 		 * do not want to modify it.
1542 		 *
1543 		 * We add (void) to let the compiler know that we do not care
1544 		 * about the return value of these functions. We use the
1545 		 * cmpxchg to only update if an interrupt did not already
1546 		 * do it for us. If the cmpxchg fails, we don't care.
1547 		 */
1548 		(void)local_cmpxchg(&next_page->write, old_write, val);
1549 		(void)local_cmpxchg(&next_page->entries, old_entries, eval);
1550 
1551 		/*
1552 		 * No need to worry about races with clearing out the commit.
1553 		 * it only can increment when a commit takes place. But that
1554 		 * only happens in the outer most nested commit.
1555 		 */
1556 		local_set(&next_page->page->commit, 0);
1557 
1558 		/* Again, either we update tail_page or an interrupt does */
1559 		(void)cmpxchg(&cpu_buffer->tail_page, tail_page, next_page);
1560 	}
1561 }
1562 
1563 static void rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
1564 			  struct buffer_page *bpage)
1565 {
1566 	unsigned long val = (unsigned long)bpage;
1567 
1568 	RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK);
1569 }
1570 
1571 /**
1572  * rb_check_pages - integrity check of buffer pages
1573  * @cpu_buffer: CPU buffer with pages to test
1574  *
1575  * As a safety measure we check to make sure the data pages have not
1576  * been corrupted.
1577  */
1578 static void rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
1579 {
1580 	struct list_head *head = rb_list_head(cpu_buffer->pages);
1581 	struct list_head *tmp;
1582 
1583 	if (RB_WARN_ON(cpu_buffer,
1584 			rb_list_head(rb_list_head(head->next)->prev) != head))
1585 		return;
1586 
1587 	if (RB_WARN_ON(cpu_buffer,
1588 			rb_list_head(rb_list_head(head->prev)->next) != head))
1589 		return;
1590 
1591 	for (tmp = rb_list_head(head->next); tmp != head; tmp = rb_list_head(tmp->next)) {
1592 		if (RB_WARN_ON(cpu_buffer,
1593 				rb_list_head(rb_list_head(tmp->next)->prev) != tmp))
1594 			return;
1595 
1596 		if (RB_WARN_ON(cpu_buffer,
1597 				rb_list_head(rb_list_head(tmp->prev)->next) != tmp))
1598 			return;
1599 	}
1600 }
1601 
1602 static int __rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1603 		long nr_pages, struct list_head *pages)
1604 {
1605 	struct buffer_page *bpage, *tmp;
1606 	bool user_thread = current->mm != NULL;
1607 	gfp_t mflags;
1608 	long i;
1609 
1610 	/*
1611 	 * Check if the available memory is there first.
1612 	 * Note, si_mem_available() only gives us a rough estimate of available
1613 	 * memory. It may not be accurate. But we don't care, we just want
1614 	 * to prevent doing any allocation when it is obvious that it is
1615 	 * not going to succeed.
1616 	 */
1617 	i = si_mem_available();
1618 	if (i < nr_pages)
1619 		return -ENOMEM;
1620 
1621 	/*
1622 	 * __GFP_RETRY_MAYFAIL flag makes sure that the allocation fails
1623 	 * gracefully without invoking oom-killer and the system is not
1624 	 * destabilized.
1625 	 */
1626 	mflags = GFP_KERNEL | __GFP_RETRY_MAYFAIL;
1627 
1628 	/*
1629 	 * If a user thread allocates too much, and si_mem_available()
1630 	 * reports there's enough memory, even though there is not.
1631 	 * Make sure the OOM killer kills this thread. This can happen
1632 	 * even with RETRY_MAYFAIL because another task may be doing
1633 	 * an allocation after this task has taken all memory.
1634 	 * This is the task the OOM killer needs to take out during this
1635 	 * loop, even if it was triggered by an allocation somewhere else.
1636 	 */
1637 	if (user_thread)
1638 		set_current_oom_origin();
1639 	for (i = 0; i < nr_pages; i++) {
1640 		struct page *page;
1641 
1642 		bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1643 				    mflags, cpu_to_node(cpu_buffer->cpu));
1644 		if (!bpage)
1645 			goto free_pages;
1646 
1647 		rb_check_bpage(cpu_buffer, bpage);
1648 
1649 		list_add(&bpage->list, pages);
1650 
1651 		page = alloc_pages_node(cpu_to_node(cpu_buffer->cpu), mflags, 0);
1652 		if (!page)
1653 			goto free_pages;
1654 		bpage->page = page_address(page);
1655 		rb_init_page(bpage->page);
1656 
1657 		if (user_thread && fatal_signal_pending(current))
1658 			goto free_pages;
1659 	}
1660 	if (user_thread)
1661 		clear_current_oom_origin();
1662 
1663 	return 0;
1664 
1665 free_pages:
1666 	list_for_each_entry_safe(bpage, tmp, pages, list) {
1667 		list_del_init(&bpage->list);
1668 		free_buffer_page(bpage);
1669 	}
1670 	if (user_thread)
1671 		clear_current_oom_origin();
1672 
1673 	return -ENOMEM;
1674 }
1675 
1676 static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1677 			     unsigned long nr_pages)
1678 {
1679 	LIST_HEAD(pages);
1680 
1681 	WARN_ON(!nr_pages);
1682 
1683 	if (__rb_allocate_pages(cpu_buffer, nr_pages, &pages))
1684 		return -ENOMEM;
1685 
1686 	/*
1687 	 * The ring buffer page list is a circular list that does not
1688 	 * start and end with a list head. All page list items point to
1689 	 * other pages.
1690 	 */
1691 	cpu_buffer->pages = pages.next;
1692 	list_del(&pages);
1693 
1694 	cpu_buffer->nr_pages = nr_pages;
1695 
1696 	rb_check_pages(cpu_buffer);
1697 
1698 	return 0;
1699 }
1700 
1701 static struct ring_buffer_per_cpu *
1702 rb_allocate_cpu_buffer(struct trace_buffer *buffer, long nr_pages, int cpu)
1703 {
1704 	struct ring_buffer_per_cpu *cpu_buffer;
1705 	struct buffer_page *bpage;
1706 	struct page *page;
1707 	int ret;
1708 
1709 	cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1710 				  GFP_KERNEL, cpu_to_node(cpu));
1711 	if (!cpu_buffer)
1712 		return NULL;
1713 
1714 	cpu_buffer->cpu = cpu;
1715 	cpu_buffer->buffer = buffer;
1716 	raw_spin_lock_init(&cpu_buffer->reader_lock);
1717 	lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1718 	cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1719 	INIT_WORK(&cpu_buffer->update_pages_work, update_pages_handler);
1720 	init_completion(&cpu_buffer->update_done);
1721 	init_irq_work(&cpu_buffer->irq_work.work, rb_wake_up_waiters);
1722 	init_waitqueue_head(&cpu_buffer->irq_work.waiters);
1723 	init_waitqueue_head(&cpu_buffer->irq_work.full_waiters);
1724 
1725 	bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1726 			    GFP_KERNEL, cpu_to_node(cpu));
1727 	if (!bpage)
1728 		goto fail_free_buffer;
1729 
1730 	rb_check_bpage(cpu_buffer, bpage);
1731 
1732 	cpu_buffer->reader_page = bpage;
1733 	page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0);
1734 	if (!page)
1735 		goto fail_free_reader;
1736 	bpage->page = page_address(page);
1737 	rb_init_page(bpage->page);
1738 
1739 	INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1740 	INIT_LIST_HEAD(&cpu_buffer->new_pages);
1741 
1742 	ret = rb_allocate_pages(cpu_buffer, nr_pages);
1743 	if (ret < 0)
1744 		goto fail_free_reader;
1745 
1746 	cpu_buffer->head_page
1747 		= list_entry(cpu_buffer->pages, struct buffer_page, list);
1748 	cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1749 
1750 	rb_head_page_activate(cpu_buffer);
1751 
1752 	return cpu_buffer;
1753 
1754  fail_free_reader:
1755 	free_buffer_page(cpu_buffer->reader_page);
1756 
1757  fail_free_buffer:
1758 	kfree(cpu_buffer);
1759 	return NULL;
1760 }
1761 
1762 static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1763 {
1764 	struct list_head *head = cpu_buffer->pages;
1765 	struct buffer_page *bpage, *tmp;
1766 
1767 	irq_work_sync(&cpu_buffer->irq_work.work);
1768 
1769 	free_buffer_page(cpu_buffer->reader_page);
1770 
1771 	if (head) {
1772 		rb_head_page_deactivate(cpu_buffer);
1773 
1774 		list_for_each_entry_safe(bpage, tmp, head, list) {
1775 			list_del_init(&bpage->list);
1776 			free_buffer_page(bpage);
1777 		}
1778 		bpage = list_entry(head, struct buffer_page, list);
1779 		free_buffer_page(bpage);
1780 	}
1781 
1782 	kfree(cpu_buffer);
1783 }
1784 
1785 /**
1786  * __ring_buffer_alloc - allocate a new ring_buffer
1787  * @size: the size in bytes per cpu that is needed.
1788  * @flags: attributes to set for the ring buffer.
1789  * @key: ring buffer reader_lock_key.
1790  *
1791  * Currently the only flag that is available is the RB_FL_OVERWRITE
1792  * flag. This flag means that the buffer will overwrite old data
1793  * when the buffer wraps. If this flag is not set, the buffer will
1794  * drop data when the tail hits the head.
1795  */
1796 struct trace_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1797 					struct lock_class_key *key)
1798 {
1799 	struct trace_buffer *buffer;
1800 	long nr_pages;
1801 	int bsize;
1802 	int cpu;
1803 	int ret;
1804 
1805 	/* keep it in its own cache line */
1806 	buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1807 			 GFP_KERNEL);
1808 	if (!buffer)
1809 		return NULL;
1810 
1811 	if (!zalloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1812 		goto fail_free_buffer;
1813 
1814 	nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1815 	buffer->flags = flags;
1816 	buffer->clock = trace_clock_local;
1817 	buffer->reader_lock_key = key;
1818 
1819 	init_irq_work(&buffer->irq_work.work, rb_wake_up_waiters);
1820 	init_waitqueue_head(&buffer->irq_work.waiters);
1821 
1822 	/* need at least two pages */
1823 	if (nr_pages < 2)
1824 		nr_pages = 2;
1825 
1826 	buffer->cpus = nr_cpu_ids;
1827 
1828 	bsize = sizeof(void *) * nr_cpu_ids;
1829 	buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1830 				  GFP_KERNEL);
1831 	if (!buffer->buffers)
1832 		goto fail_free_cpumask;
1833 
1834 	cpu = raw_smp_processor_id();
1835 	cpumask_set_cpu(cpu, buffer->cpumask);
1836 	buffer->buffers[cpu] = rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
1837 	if (!buffer->buffers[cpu])
1838 		goto fail_free_buffers;
1839 
1840 	ret = cpuhp_state_add_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1841 	if (ret < 0)
1842 		goto fail_free_buffers;
1843 
1844 	mutex_init(&buffer->mutex);
1845 
1846 	return buffer;
1847 
1848  fail_free_buffers:
1849 	for_each_buffer_cpu(buffer, cpu) {
1850 		if (buffer->buffers[cpu])
1851 			rb_free_cpu_buffer(buffer->buffers[cpu]);
1852 	}
1853 	kfree(buffer->buffers);
1854 
1855  fail_free_cpumask:
1856 	free_cpumask_var(buffer->cpumask);
1857 
1858  fail_free_buffer:
1859 	kfree(buffer);
1860 	return NULL;
1861 }
1862 EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1863 
1864 /**
1865  * ring_buffer_free - free a ring buffer.
1866  * @buffer: the buffer to free.
1867  */
1868 void
1869 ring_buffer_free(struct trace_buffer *buffer)
1870 {
1871 	int cpu;
1872 
1873 	cpuhp_state_remove_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1874 
1875 	irq_work_sync(&buffer->irq_work.work);
1876 
1877 	for_each_buffer_cpu(buffer, cpu)
1878 		rb_free_cpu_buffer(buffer->buffers[cpu]);
1879 
1880 	kfree(buffer->buffers);
1881 	free_cpumask_var(buffer->cpumask);
1882 
1883 	kfree(buffer);
1884 }
1885 EXPORT_SYMBOL_GPL(ring_buffer_free);
1886 
1887 void ring_buffer_set_clock(struct trace_buffer *buffer,
1888 			   u64 (*clock)(void))
1889 {
1890 	buffer->clock = clock;
1891 }
1892 
1893 void ring_buffer_set_time_stamp_abs(struct trace_buffer *buffer, bool abs)
1894 {
1895 	buffer->time_stamp_abs = abs;
1896 }
1897 
1898 bool ring_buffer_time_stamp_abs(struct trace_buffer *buffer)
1899 {
1900 	return buffer->time_stamp_abs;
1901 }
1902 
1903 static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1904 
1905 static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1906 {
1907 	return local_read(&bpage->entries) & RB_WRITE_MASK;
1908 }
1909 
1910 static inline unsigned long rb_page_write(struct buffer_page *bpage)
1911 {
1912 	return local_read(&bpage->write) & RB_WRITE_MASK;
1913 }
1914 
1915 static bool
1916 rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned long nr_pages)
1917 {
1918 	struct list_head *tail_page, *to_remove, *next_page;
1919 	struct buffer_page *to_remove_page, *tmp_iter_page;
1920 	struct buffer_page *last_page, *first_page;
1921 	unsigned long nr_removed;
1922 	unsigned long head_bit;
1923 	int page_entries;
1924 
1925 	head_bit = 0;
1926 
1927 	raw_spin_lock_irq(&cpu_buffer->reader_lock);
1928 	atomic_inc(&cpu_buffer->record_disabled);
1929 	/*
1930 	 * We don't race with the readers since we have acquired the reader
1931 	 * lock. We also don't race with writers after disabling recording.
1932 	 * This makes it easy to figure out the first and the last page to be
1933 	 * removed from the list. We unlink all the pages in between including
1934 	 * the first and last pages. This is done in a busy loop so that we
1935 	 * lose the least number of traces.
1936 	 * The pages are freed after we restart recording and unlock readers.
1937 	 */
1938 	tail_page = &cpu_buffer->tail_page->list;
1939 
1940 	/*
1941 	 * tail page might be on reader page, we remove the next page
1942 	 * from the ring buffer
1943 	 */
1944 	if (cpu_buffer->tail_page == cpu_buffer->reader_page)
1945 		tail_page = rb_list_head(tail_page->next);
1946 	to_remove = tail_page;
1947 
1948 	/* start of pages to remove */
1949 	first_page = list_entry(rb_list_head(to_remove->next),
1950 				struct buffer_page, list);
1951 
1952 	for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
1953 		to_remove = rb_list_head(to_remove)->next;
1954 		head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
1955 	}
1956 	/* Read iterators need to reset themselves when some pages removed */
1957 	cpu_buffer->pages_removed += nr_removed;
1958 
1959 	next_page = rb_list_head(to_remove)->next;
1960 
1961 	/*
1962 	 * Now we remove all pages between tail_page and next_page.
1963 	 * Make sure that we have head_bit value preserved for the
1964 	 * next page
1965 	 */
1966 	tail_page->next = (struct list_head *)((unsigned long)next_page |
1967 						head_bit);
1968 	next_page = rb_list_head(next_page);
1969 	next_page->prev = tail_page;
1970 
1971 	/* make sure pages points to a valid page in the ring buffer */
1972 	cpu_buffer->pages = next_page;
1973 
1974 	/* update head page */
1975 	if (head_bit)
1976 		cpu_buffer->head_page = list_entry(next_page,
1977 						struct buffer_page, list);
1978 
1979 	/* pages are removed, resume tracing and then free the pages */
1980 	atomic_dec(&cpu_buffer->record_disabled);
1981 	raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1982 
1983 	RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
1984 
1985 	/* last buffer page to remove */
1986 	last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
1987 				list);
1988 	tmp_iter_page = first_page;
1989 
1990 	do {
1991 		cond_resched();
1992 
1993 		to_remove_page = tmp_iter_page;
1994 		rb_inc_page(&tmp_iter_page);
1995 
1996 		/* update the counters */
1997 		page_entries = rb_page_entries(to_remove_page);
1998 		if (page_entries) {
1999 			/*
2000 			 * If something was added to this page, it was full
2001 			 * since it is not the tail page. So we deduct the
2002 			 * bytes consumed in ring buffer from here.
2003 			 * Increment overrun to account for the lost events.
2004 			 */
2005 			local_add(page_entries, &cpu_buffer->overrun);
2006 			local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
2007 			local_inc(&cpu_buffer->pages_lost);
2008 		}
2009 
2010 		/*
2011 		 * We have already removed references to this list item, just
2012 		 * free up the buffer_page and its page
2013 		 */
2014 		free_buffer_page(to_remove_page);
2015 		nr_removed--;
2016 
2017 	} while (to_remove_page != last_page);
2018 
2019 	RB_WARN_ON(cpu_buffer, nr_removed);
2020 
2021 	return nr_removed == 0;
2022 }
2023 
2024 static bool
2025 rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
2026 {
2027 	struct list_head *pages = &cpu_buffer->new_pages;
2028 	unsigned long flags;
2029 	bool success;
2030 	int retries;
2031 
2032 	/* Can be called at early boot up, where interrupts must not been enabled */
2033 	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
2034 	/*
2035 	 * We are holding the reader lock, so the reader page won't be swapped
2036 	 * in the ring buffer. Now we are racing with the writer trying to
2037 	 * move head page and the tail page.
2038 	 * We are going to adapt the reader page update process where:
2039 	 * 1. We first splice the start and end of list of new pages between
2040 	 *    the head page and its previous page.
2041 	 * 2. We cmpxchg the prev_page->next to point from head page to the
2042 	 *    start of new pages list.
2043 	 * 3. Finally, we update the head->prev to the end of new list.
2044 	 *
2045 	 * We will try this process 10 times, to make sure that we don't keep
2046 	 * spinning.
2047 	 */
2048 	retries = 10;
2049 	success = false;
2050 	while (retries--) {
2051 		struct list_head *head_page, *prev_page, *r;
2052 		struct list_head *last_page, *first_page;
2053 		struct list_head *head_page_with_bit;
2054 		struct buffer_page *hpage = rb_set_head_page(cpu_buffer);
2055 
2056 		if (!hpage)
2057 			break;
2058 		head_page = &hpage->list;
2059 		prev_page = head_page->prev;
2060 
2061 		first_page = pages->next;
2062 		last_page  = pages->prev;
2063 
2064 		head_page_with_bit = (struct list_head *)
2065 				     ((unsigned long)head_page | RB_PAGE_HEAD);
2066 
2067 		last_page->next = head_page_with_bit;
2068 		first_page->prev = prev_page;
2069 
2070 		r = cmpxchg(&prev_page->next, head_page_with_bit, first_page);
2071 
2072 		if (r == head_page_with_bit) {
2073 			/*
2074 			 * yay, we replaced the page pointer to our new list,
2075 			 * now, we just have to update to head page's prev
2076 			 * pointer to point to end of list
2077 			 */
2078 			head_page->prev = last_page;
2079 			success = true;
2080 			break;
2081 		}
2082 	}
2083 
2084 	if (success)
2085 		INIT_LIST_HEAD(pages);
2086 	/*
2087 	 * If we weren't successful in adding in new pages, warn and stop
2088 	 * tracing
2089 	 */
2090 	RB_WARN_ON(cpu_buffer, !success);
2091 	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
2092 
2093 	/* free pages if they weren't inserted */
2094 	if (!success) {
2095 		struct buffer_page *bpage, *tmp;
2096 		list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
2097 					 list) {
2098 			list_del_init(&bpage->list);
2099 			free_buffer_page(bpage);
2100 		}
2101 	}
2102 	return success;
2103 }
2104 
2105 static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
2106 {
2107 	bool success;
2108 
2109 	if (cpu_buffer->nr_pages_to_update > 0)
2110 		success = rb_insert_pages(cpu_buffer);
2111 	else
2112 		success = rb_remove_pages(cpu_buffer,
2113 					-cpu_buffer->nr_pages_to_update);
2114 
2115 	if (success)
2116 		cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
2117 }
2118 
2119 static void update_pages_handler(struct work_struct *work)
2120 {
2121 	struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
2122 			struct ring_buffer_per_cpu, update_pages_work);
2123 	rb_update_pages(cpu_buffer);
2124 	complete(&cpu_buffer->update_done);
2125 }
2126 
2127 /**
2128  * ring_buffer_resize - resize the ring buffer
2129  * @buffer: the buffer to resize.
2130  * @size: the new size.
2131  * @cpu_id: the cpu buffer to resize
2132  *
2133  * Minimum size is 2 * BUF_PAGE_SIZE.
2134  *
2135  * Returns 0 on success and < 0 on failure.
2136  */
2137 int ring_buffer_resize(struct trace_buffer *buffer, unsigned long size,
2138 			int cpu_id)
2139 {
2140 	struct ring_buffer_per_cpu *cpu_buffer;
2141 	unsigned long nr_pages;
2142 	int cpu, err;
2143 
2144 	/*
2145 	 * Always succeed at resizing a non-existent buffer:
2146 	 */
2147 	if (!buffer)
2148 		return 0;
2149 
2150 	/* Make sure the requested buffer exists */
2151 	if (cpu_id != RING_BUFFER_ALL_CPUS &&
2152 	    !cpumask_test_cpu(cpu_id, buffer->cpumask))
2153 		return 0;
2154 
2155 	nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
2156 
2157 	/* we need a minimum of two pages */
2158 	if (nr_pages < 2)
2159 		nr_pages = 2;
2160 
2161 	/* prevent another thread from changing buffer sizes */
2162 	mutex_lock(&buffer->mutex);
2163 	atomic_inc(&buffer->resizing);
2164 
2165 	if (cpu_id == RING_BUFFER_ALL_CPUS) {
2166 		/*
2167 		 * Don't succeed if resizing is disabled, as a reader might be
2168 		 * manipulating the ring buffer and is expecting a sane state while
2169 		 * this is true.
2170 		 */
2171 		for_each_buffer_cpu(buffer, cpu) {
2172 			cpu_buffer = buffer->buffers[cpu];
2173 			if (atomic_read(&cpu_buffer->resize_disabled)) {
2174 				err = -EBUSY;
2175 				goto out_err_unlock;
2176 			}
2177 		}
2178 
2179 		/* calculate the pages to update */
2180 		for_each_buffer_cpu(buffer, cpu) {
2181 			cpu_buffer = buffer->buffers[cpu];
2182 
2183 			cpu_buffer->nr_pages_to_update = nr_pages -
2184 							cpu_buffer->nr_pages;
2185 			/*
2186 			 * nothing more to do for removing pages or no update
2187 			 */
2188 			if (cpu_buffer->nr_pages_to_update <= 0)
2189 				continue;
2190 			/*
2191 			 * to add pages, make sure all new pages can be
2192 			 * allocated without receiving ENOMEM
2193 			 */
2194 			INIT_LIST_HEAD(&cpu_buffer->new_pages);
2195 			if (__rb_allocate_pages(cpu_buffer, cpu_buffer->nr_pages_to_update,
2196 						&cpu_buffer->new_pages)) {
2197 				/* not enough memory for new pages */
2198 				err = -ENOMEM;
2199 				goto out_err;
2200 			}
2201 		}
2202 
2203 		cpus_read_lock();
2204 		/*
2205 		 * Fire off all the required work handlers
2206 		 * We can't schedule on offline CPUs, but it's not necessary
2207 		 * since we can change their buffer sizes without any race.
2208 		 */
2209 		for_each_buffer_cpu(buffer, cpu) {
2210 			cpu_buffer = buffer->buffers[cpu];
2211 			if (!cpu_buffer->nr_pages_to_update)
2212 				continue;
2213 
2214 			/* Can't run something on an offline CPU. */
2215 			if (!cpu_online(cpu)) {
2216 				rb_update_pages(cpu_buffer);
2217 				cpu_buffer->nr_pages_to_update = 0;
2218 			} else {
2219 				/* Run directly if possible. */
2220 				migrate_disable();
2221 				if (cpu != smp_processor_id()) {
2222 					migrate_enable();
2223 					schedule_work_on(cpu,
2224 							 &cpu_buffer->update_pages_work);
2225 				} else {
2226 					update_pages_handler(&cpu_buffer->update_pages_work);
2227 					migrate_enable();
2228 				}
2229 			}
2230 		}
2231 
2232 		/* wait for all the updates to complete */
2233 		for_each_buffer_cpu(buffer, cpu) {
2234 			cpu_buffer = buffer->buffers[cpu];
2235 			if (!cpu_buffer->nr_pages_to_update)
2236 				continue;
2237 
2238 			if (cpu_online(cpu))
2239 				wait_for_completion(&cpu_buffer->update_done);
2240 			cpu_buffer->nr_pages_to_update = 0;
2241 		}
2242 
2243 		cpus_read_unlock();
2244 	} else {
2245 		cpu_buffer = buffer->buffers[cpu_id];
2246 
2247 		if (nr_pages == cpu_buffer->nr_pages)
2248 			goto out;
2249 
2250 		/*
2251 		 * Don't succeed if resizing is disabled, as a reader might be
2252 		 * manipulating the ring buffer and is expecting a sane state while
2253 		 * this is true.
2254 		 */
2255 		if (atomic_read(&cpu_buffer->resize_disabled)) {
2256 			err = -EBUSY;
2257 			goto out_err_unlock;
2258 		}
2259 
2260 		cpu_buffer->nr_pages_to_update = nr_pages -
2261 						cpu_buffer->nr_pages;
2262 
2263 		INIT_LIST_HEAD(&cpu_buffer->new_pages);
2264 		if (cpu_buffer->nr_pages_to_update > 0 &&
2265 			__rb_allocate_pages(cpu_buffer, cpu_buffer->nr_pages_to_update,
2266 					    &cpu_buffer->new_pages)) {
2267 			err = -ENOMEM;
2268 			goto out_err;
2269 		}
2270 
2271 		cpus_read_lock();
2272 
2273 		/* Can't run something on an offline CPU. */
2274 		if (!cpu_online(cpu_id))
2275 			rb_update_pages(cpu_buffer);
2276 		else {
2277 			/* Run directly if possible. */
2278 			migrate_disable();
2279 			if (cpu_id == smp_processor_id()) {
2280 				rb_update_pages(cpu_buffer);
2281 				migrate_enable();
2282 			} else {
2283 				migrate_enable();
2284 				schedule_work_on(cpu_id,
2285 						 &cpu_buffer->update_pages_work);
2286 				wait_for_completion(&cpu_buffer->update_done);
2287 			}
2288 		}
2289 
2290 		cpu_buffer->nr_pages_to_update = 0;
2291 		cpus_read_unlock();
2292 	}
2293 
2294  out:
2295 	/*
2296 	 * The ring buffer resize can happen with the ring buffer
2297 	 * enabled, so that the update disturbs the tracing as little
2298 	 * as possible. But if the buffer is disabled, we do not need
2299 	 * to worry about that, and we can take the time to verify
2300 	 * that the buffer is not corrupt.
2301 	 */
2302 	if (atomic_read(&buffer->record_disabled)) {
2303 		atomic_inc(&buffer->record_disabled);
2304 		/*
2305 		 * Even though the buffer was disabled, we must make sure
2306 		 * that it is truly disabled before calling rb_check_pages.
2307 		 * There could have been a race between checking
2308 		 * record_disable and incrementing it.
2309 		 */
2310 		synchronize_rcu();
2311 		for_each_buffer_cpu(buffer, cpu) {
2312 			cpu_buffer = buffer->buffers[cpu];
2313 			rb_check_pages(cpu_buffer);
2314 		}
2315 		atomic_dec(&buffer->record_disabled);
2316 	}
2317 
2318 	atomic_dec(&buffer->resizing);
2319 	mutex_unlock(&buffer->mutex);
2320 	return 0;
2321 
2322  out_err:
2323 	for_each_buffer_cpu(buffer, cpu) {
2324 		struct buffer_page *bpage, *tmp;
2325 
2326 		cpu_buffer = buffer->buffers[cpu];
2327 		cpu_buffer->nr_pages_to_update = 0;
2328 
2329 		if (list_empty(&cpu_buffer->new_pages))
2330 			continue;
2331 
2332 		list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
2333 					list) {
2334 			list_del_init(&bpage->list);
2335 			free_buffer_page(bpage);
2336 		}
2337 	}
2338  out_err_unlock:
2339 	atomic_dec(&buffer->resizing);
2340 	mutex_unlock(&buffer->mutex);
2341 	return err;
2342 }
2343 EXPORT_SYMBOL_GPL(ring_buffer_resize);
2344 
2345 void ring_buffer_change_overwrite(struct trace_buffer *buffer, int val)
2346 {
2347 	mutex_lock(&buffer->mutex);
2348 	if (val)
2349 		buffer->flags |= RB_FL_OVERWRITE;
2350 	else
2351 		buffer->flags &= ~RB_FL_OVERWRITE;
2352 	mutex_unlock(&buffer->mutex);
2353 }
2354 EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
2355 
2356 static __always_inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
2357 {
2358 	return bpage->page->data + index;
2359 }
2360 
2361 static __always_inline struct ring_buffer_event *
2362 rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
2363 {
2364 	return __rb_page_index(cpu_buffer->reader_page,
2365 			       cpu_buffer->reader_page->read);
2366 }
2367 
2368 static __always_inline unsigned rb_page_commit(struct buffer_page *bpage)
2369 {
2370 	return local_read(&bpage->page->commit);
2371 }
2372 
2373 static struct ring_buffer_event *
2374 rb_iter_head_event(struct ring_buffer_iter *iter)
2375 {
2376 	struct ring_buffer_event *event;
2377 	struct buffer_page *iter_head_page = iter->head_page;
2378 	unsigned long commit;
2379 	unsigned length;
2380 
2381 	if (iter->head != iter->next_event)
2382 		return iter->event;
2383 
2384 	/*
2385 	 * When the writer goes across pages, it issues a cmpxchg which
2386 	 * is a mb(), which will synchronize with the rmb here.
2387 	 * (see rb_tail_page_update() and __rb_reserve_next())
2388 	 */
2389 	commit = rb_page_commit(iter_head_page);
2390 	smp_rmb();
2391 	event = __rb_page_index(iter_head_page, iter->head);
2392 	length = rb_event_length(event);
2393 
2394 	/*
2395 	 * READ_ONCE() doesn't work on functions and we don't want the
2396 	 * compiler doing any crazy optimizations with length.
2397 	 */
2398 	barrier();
2399 
2400 	if ((iter->head + length) > commit || length > BUF_MAX_DATA_SIZE)
2401 		/* Writer corrupted the read? */
2402 		goto reset;
2403 
2404 	memcpy(iter->event, event, length);
2405 	/*
2406 	 * If the page stamp is still the same after this rmb() then the
2407 	 * event was safely copied without the writer entering the page.
2408 	 */
2409 	smp_rmb();
2410 
2411 	/* Make sure the page didn't change since we read this */
2412 	if (iter->page_stamp != iter_head_page->page->time_stamp ||
2413 	    commit > rb_page_commit(iter_head_page))
2414 		goto reset;
2415 
2416 	iter->next_event = iter->head + length;
2417 	return iter->event;
2418  reset:
2419 	/* Reset to the beginning */
2420 	iter->page_stamp = iter->read_stamp = iter->head_page->page->time_stamp;
2421 	iter->head = 0;
2422 	iter->next_event = 0;
2423 	iter->missed_events = 1;
2424 	return NULL;
2425 }
2426 
2427 /* Size is determined by what has been committed */
2428 static __always_inline unsigned rb_page_size(struct buffer_page *bpage)
2429 {
2430 	return rb_page_commit(bpage);
2431 }
2432 
2433 static __always_inline unsigned
2434 rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
2435 {
2436 	return rb_page_commit(cpu_buffer->commit_page);
2437 }
2438 
2439 static __always_inline unsigned
2440 rb_event_index(struct ring_buffer_event *event)
2441 {
2442 	unsigned long addr = (unsigned long)event;
2443 
2444 	return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
2445 }
2446 
2447 static void rb_inc_iter(struct ring_buffer_iter *iter)
2448 {
2449 	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
2450 
2451 	/*
2452 	 * The iterator could be on the reader page (it starts there).
2453 	 * But the head could have moved, since the reader was
2454 	 * found. Check for this case and assign the iterator
2455 	 * to the head page instead of next.
2456 	 */
2457 	if (iter->head_page == cpu_buffer->reader_page)
2458 		iter->head_page = rb_set_head_page(cpu_buffer);
2459 	else
2460 		rb_inc_page(&iter->head_page);
2461 
2462 	iter->page_stamp = iter->read_stamp = iter->head_page->page->time_stamp;
2463 	iter->head = 0;
2464 	iter->next_event = 0;
2465 }
2466 
2467 /*
2468  * rb_handle_head_page - writer hit the head page
2469  *
2470  * Returns: +1 to retry page
2471  *           0 to continue
2472  *          -1 on error
2473  */
2474 static int
2475 rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
2476 		    struct buffer_page *tail_page,
2477 		    struct buffer_page *next_page)
2478 {
2479 	struct buffer_page *new_head;
2480 	int entries;
2481 	int type;
2482 	int ret;
2483 
2484 	entries = rb_page_entries(next_page);
2485 
2486 	/*
2487 	 * The hard part is here. We need to move the head
2488 	 * forward, and protect against both readers on
2489 	 * other CPUs and writers coming in via interrupts.
2490 	 */
2491 	type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
2492 				       RB_PAGE_HEAD);
2493 
2494 	/*
2495 	 * type can be one of four:
2496 	 *  NORMAL - an interrupt already moved it for us
2497 	 *  HEAD   - we are the first to get here.
2498 	 *  UPDATE - we are the interrupt interrupting
2499 	 *           a current move.
2500 	 *  MOVED  - a reader on another CPU moved the next
2501 	 *           pointer to its reader page. Give up
2502 	 *           and try again.
2503 	 */
2504 
2505 	switch (type) {
2506 	case RB_PAGE_HEAD:
2507 		/*
2508 		 * We changed the head to UPDATE, thus
2509 		 * it is our responsibility to update
2510 		 * the counters.
2511 		 */
2512 		local_add(entries, &cpu_buffer->overrun);
2513 		local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
2514 		local_inc(&cpu_buffer->pages_lost);
2515 
2516 		/*
2517 		 * The entries will be zeroed out when we move the
2518 		 * tail page.
2519 		 */
2520 
2521 		/* still more to do */
2522 		break;
2523 
2524 	case RB_PAGE_UPDATE:
2525 		/*
2526 		 * This is an interrupt that interrupt the
2527 		 * previous update. Still more to do.
2528 		 */
2529 		break;
2530 	case RB_PAGE_NORMAL:
2531 		/*
2532 		 * An interrupt came in before the update
2533 		 * and processed this for us.
2534 		 * Nothing left to do.
2535 		 */
2536 		return 1;
2537 	case RB_PAGE_MOVED:
2538 		/*
2539 		 * The reader is on another CPU and just did
2540 		 * a swap with our next_page.
2541 		 * Try again.
2542 		 */
2543 		return 1;
2544 	default:
2545 		RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
2546 		return -1;
2547 	}
2548 
2549 	/*
2550 	 * Now that we are here, the old head pointer is
2551 	 * set to UPDATE. This will keep the reader from
2552 	 * swapping the head page with the reader page.
2553 	 * The reader (on another CPU) will spin till
2554 	 * we are finished.
2555 	 *
2556 	 * We just need to protect against interrupts
2557 	 * doing the job. We will set the next pointer
2558 	 * to HEAD. After that, we set the old pointer
2559 	 * to NORMAL, but only if it was HEAD before.
2560 	 * otherwise we are an interrupt, and only
2561 	 * want the outer most commit to reset it.
2562 	 */
2563 	new_head = next_page;
2564 	rb_inc_page(&new_head);
2565 
2566 	ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
2567 				    RB_PAGE_NORMAL);
2568 
2569 	/*
2570 	 * Valid returns are:
2571 	 *  HEAD   - an interrupt came in and already set it.
2572 	 *  NORMAL - One of two things:
2573 	 *            1) We really set it.
2574 	 *            2) A bunch of interrupts came in and moved
2575 	 *               the page forward again.
2576 	 */
2577 	switch (ret) {
2578 	case RB_PAGE_HEAD:
2579 	case RB_PAGE_NORMAL:
2580 		/* OK */
2581 		break;
2582 	default:
2583 		RB_WARN_ON(cpu_buffer, 1);
2584 		return -1;
2585 	}
2586 
2587 	/*
2588 	 * It is possible that an interrupt came in,
2589 	 * set the head up, then more interrupts came in
2590 	 * and moved it again. When we get back here,
2591 	 * the page would have been set to NORMAL but we
2592 	 * just set it back to HEAD.
2593 	 *
2594 	 * How do you detect this? Well, if that happened
2595 	 * the tail page would have moved.
2596 	 */
2597 	if (ret == RB_PAGE_NORMAL) {
2598 		struct buffer_page *buffer_tail_page;
2599 
2600 		buffer_tail_page = READ_ONCE(cpu_buffer->tail_page);
2601 		/*
2602 		 * If the tail had moved passed next, then we need
2603 		 * to reset the pointer.
2604 		 */
2605 		if (buffer_tail_page != tail_page &&
2606 		    buffer_tail_page != next_page)
2607 			rb_head_page_set_normal(cpu_buffer, new_head,
2608 						next_page,
2609 						RB_PAGE_HEAD);
2610 	}
2611 
2612 	/*
2613 	 * If this was the outer most commit (the one that
2614 	 * changed the original pointer from HEAD to UPDATE),
2615 	 * then it is up to us to reset it to NORMAL.
2616 	 */
2617 	if (type == RB_PAGE_HEAD) {
2618 		ret = rb_head_page_set_normal(cpu_buffer, next_page,
2619 					      tail_page,
2620 					      RB_PAGE_UPDATE);
2621 		if (RB_WARN_ON(cpu_buffer,
2622 			       ret != RB_PAGE_UPDATE))
2623 			return -1;
2624 	}
2625 
2626 	return 0;
2627 }
2628 
2629 static inline void
2630 rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2631 	      unsigned long tail, struct rb_event_info *info)
2632 {
2633 	struct buffer_page *tail_page = info->tail_page;
2634 	struct ring_buffer_event *event;
2635 	unsigned long length = info->length;
2636 
2637 	/*
2638 	 * Only the event that crossed the page boundary
2639 	 * must fill the old tail_page with padding.
2640 	 */
2641 	if (tail >= BUF_PAGE_SIZE) {
2642 		/*
2643 		 * If the page was filled, then we still need
2644 		 * to update the real_end. Reset it to zero
2645 		 * and the reader will ignore it.
2646 		 */
2647 		if (tail == BUF_PAGE_SIZE)
2648 			tail_page->real_end = 0;
2649 
2650 		local_sub(length, &tail_page->write);
2651 		return;
2652 	}
2653 
2654 	event = __rb_page_index(tail_page, tail);
2655 
2656 	/* account for padding bytes */
2657 	local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes);
2658 
2659 	/*
2660 	 * Save the original length to the meta data.
2661 	 * This will be used by the reader to add lost event
2662 	 * counter.
2663 	 */
2664 	tail_page->real_end = tail;
2665 
2666 	/*
2667 	 * If this event is bigger than the minimum size, then
2668 	 * we need to be careful that we don't subtract the
2669 	 * write counter enough to allow another writer to slip
2670 	 * in on this page.
2671 	 * We put in a discarded commit instead, to make sure
2672 	 * that this space is not used again.
2673 	 *
2674 	 * If we are less than the minimum size, we don't need to
2675 	 * worry about it.
2676 	 */
2677 	if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
2678 		/* No room for any events */
2679 
2680 		/* Mark the rest of the page with padding */
2681 		rb_event_set_padding(event);
2682 
2683 		/* Make sure the padding is visible before the write update */
2684 		smp_wmb();
2685 
2686 		/* Set the write back to the previous setting */
2687 		local_sub(length, &tail_page->write);
2688 		return;
2689 	}
2690 
2691 	/* Put in a discarded event */
2692 	event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
2693 	event->type_len = RINGBUF_TYPE_PADDING;
2694 	/* time delta must be non zero */
2695 	event->time_delta = 1;
2696 
2697 	/* Make sure the padding is visible before the tail_page->write update */
2698 	smp_wmb();
2699 
2700 	/* Set write to end of buffer */
2701 	length = (tail + length) - BUF_PAGE_SIZE;
2702 	local_sub(length, &tail_page->write);
2703 }
2704 
2705 static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer);
2706 
2707 /*
2708  * This is the slow path, force gcc not to inline it.
2709  */
2710 static noinline struct ring_buffer_event *
2711 rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2712 	     unsigned long tail, struct rb_event_info *info)
2713 {
2714 	struct buffer_page *tail_page = info->tail_page;
2715 	struct buffer_page *commit_page = cpu_buffer->commit_page;
2716 	struct trace_buffer *buffer = cpu_buffer->buffer;
2717 	struct buffer_page *next_page;
2718 	int ret;
2719 
2720 	next_page = tail_page;
2721 
2722 	rb_inc_page(&next_page);
2723 
2724 	/*
2725 	 * If for some reason, we had an interrupt storm that made
2726 	 * it all the way around the buffer, bail, and warn
2727 	 * about it.
2728 	 */
2729 	if (unlikely(next_page == commit_page)) {
2730 		local_inc(&cpu_buffer->commit_overrun);
2731 		goto out_reset;
2732 	}
2733 
2734 	/*
2735 	 * This is where the fun begins!
2736 	 *
2737 	 * We are fighting against races between a reader that
2738 	 * could be on another CPU trying to swap its reader
2739 	 * page with the buffer head.
2740 	 *
2741 	 * We are also fighting against interrupts coming in and
2742 	 * moving the head or tail on us as well.
2743 	 *
2744 	 * If the next page is the head page then we have filled
2745 	 * the buffer, unless the commit page is still on the
2746 	 * reader page.
2747 	 */
2748 	if (rb_is_head_page(next_page, &tail_page->list)) {
2749 
2750 		/*
2751 		 * If the commit is not on the reader page, then
2752 		 * move the header page.
2753 		 */
2754 		if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2755 			/*
2756 			 * If we are not in overwrite mode,
2757 			 * this is easy, just stop here.
2758 			 */
2759 			if (!(buffer->flags & RB_FL_OVERWRITE)) {
2760 				local_inc(&cpu_buffer->dropped_events);
2761 				goto out_reset;
2762 			}
2763 
2764 			ret = rb_handle_head_page(cpu_buffer,
2765 						  tail_page,
2766 						  next_page);
2767 			if (ret < 0)
2768 				goto out_reset;
2769 			if (ret)
2770 				goto out_again;
2771 		} else {
2772 			/*
2773 			 * We need to be careful here too. The
2774 			 * commit page could still be on the reader
2775 			 * page. We could have a small buffer, and
2776 			 * have filled up the buffer with events
2777 			 * from interrupts and such, and wrapped.
2778 			 *
2779 			 * Note, if the tail page is also on the
2780 			 * reader_page, we let it move out.
2781 			 */
2782 			if (unlikely((cpu_buffer->commit_page !=
2783 				      cpu_buffer->tail_page) &&
2784 				     (cpu_buffer->commit_page ==
2785 				      cpu_buffer->reader_page))) {
2786 				local_inc(&cpu_buffer->commit_overrun);
2787 				goto out_reset;
2788 			}
2789 		}
2790 	}
2791 
2792 	rb_tail_page_update(cpu_buffer, tail_page, next_page);
2793 
2794  out_again:
2795 
2796 	rb_reset_tail(cpu_buffer, tail, info);
2797 
2798 	/* Commit what we have for now. */
2799 	rb_end_commit(cpu_buffer);
2800 	/* rb_end_commit() decs committing */
2801 	local_inc(&cpu_buffer->committing);
2802 
2803 	/* fail and let the caller try again */
2804 	return ERR_PTR(-EAGAIN);
2805 
2806  out_reset:
2807 	/* reset write */
2808 	rb_reset_tail(cpu_buffer, tail, info);
2809 
2810 	return NULL;
2811 }
2812 
2813 /* Slow path */
2814 static struct ring_buffer_event *
2815 rb_add_time_stamp(struct ring_buffer_event *event, u64 delta, bool abs)
2816 {
2817 	if (abs)
2818 		event->type_len = RINGBUF_TYPE_TIME_STAMP;
2819 	else
2820 		event->type_len = RINGBUF_TYPE_TIME_EXTEND;
2821 
2822 	/* Not the first event on the page, or not delta? */
2823 	if (abs || rb_event_index(event)) {
2824 		event->time_delta = delta & TS_MASK;
2825 		event->array[0] = delta >> TS_SHIFT;
2826 	} else {
2827 		/* nope, just zero it */
2828 		event->time_delta = 0;
2829 		event->array[0] = 0;
2830 	}
2831 
2832 	return skip_time_extend(event);
2833 }
2834 
2835 #ifndef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2836 static inline bool sched_clock_stable(void)
2837 {
2838 	return true;
2839 }
2840 #endif
2841 
2842 static void
2843 rb_check_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2844 		   struct rb_event_info *info)
2845 {
2846 	u64 write_stamp;
2847 
2848 	WARN_ONCE(1, "Delta way too big! %llu ts=%llu before=%llu after=%llu write stamp=%llu\n%s",
2849 		  (unsigned long long)info->delta,
2850 		  (unsigned long long)info->ts,
2851 		  (unsigned long long)info->before,
2852 		  (unsigned long long)info->after,
2853 		  (unsigned long long)(rb_time_read(&cpu_buffer->write_stamp, &write_stamp) ? write_stamp : 0),
2854 		  sched_clock_stable() ? "" :
2855 		  "If you just came from a suspend/resume,\n"
2856 		  "please switch to the trace global clock:\n"
2857 		  "  echo global > /sys/kernel/tracing/trace_clock\n"
2858 		  "or add trace_clock=global to the kernel command line\n");
2859 }
2860 
2861 static void rb_add_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2862 				      struct ring_buffer_event **event,
2863 				      struct rb_event_info *info,
2864 				      u64 *delta,
2865 				      unsigned int *length)
2866 {
2867 	bool abs = info->add_timestamp &
2868 		(RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE);
2869 
2870 	if (unlikely(info->delta > (1ULL << 59))) {
2871 		/*
2872 		 * Some timers can use more than 59 bits, and when a timestamp
2873 		 * is added to the buffer, it will lose those bits.
2874 		 */
2875 		if (abs && (info->ts & TS_MSB)) {
2876 			info->delta &= ABS_TS_MASK;
2877 
2878 		/* did the clock go backwards */
2879 		} else if (info->before == info->after && info->before > info->ts) {
2880 			/* not interrupted */
2881 			static int once;
2882 
2883 			/*
2884 			 * This is possible with a recalibrating of the TSC.
2885 			 * Do not produce a call stack, but just report it.
2886 			 */
2887 			if (!once) {
2888 				once++;
2889 				pr_warn("Ring buffer clock went backwards: %llu -> %llu\n",
2890 					info->before, info->ts);
2891 			}
2892 		} else
2893 			rb_check_timestamp(cpu_buffer, info);
2894 		if (!abs)
2895 			info->delta = 0;
2896 	}
2897 	*event = rb_add_time_stamp(*event, info->delta, abs);
2898 	*length -= RB_LEN_TIME_EXTEND;
2899 	*delta = 0;
2900 }
2901 
2902 /**
2903  * rb_update_event - update event type and data
2904  * @cpu_buffer: The per cpu buffer of the @event
2905  * @event: the event to update
2906  * @info: The info to update the @event with (contains length and delta)
2907  *
2908  * Update the type and data fields of the @event. The length
2909  * is the actual size that is written to the ring buffer,
2910  * and with this, we can determine what to place into the
2911  * data field.
2912  */
2913 static void
2914 rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
2915 		struct ring_buffer_event *event,
2916 		struct rb_event_info *info)
2917 {
2918 	unsigned length = info->length;
2919 	u64 delta = info->delta;
2920 	unsigned int nest = local_read(&cpu_buffer->committing) - 1;
2921 
2922 	if (!WARN_ON_ONCE(nest >= MAX_NEST))
2923 		cpu_buffer->event_stamp[nest] = info->ts;
2924 
2925 	/*
2926 	 * If we need to add a timestamp, then we
2927 	 * add it to the start of the reserved space.
2928 	 */
2929 	if (unlikely(info->add_timestamp))
2930 		rb_add_timestamp(cpu_buffer, &event, info, &delta, &length);
2931 
2932 	event->time_delta = delta;
2933 	length -= RB_EVNT_HDR_SIZE;
2934 	if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
2935 		event->type_len = 0;
2936 		event->array[0] = length;
2937 	} else
2938 		event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
2939 }
2940 
2941 static unsigned rb_calculate_event_length(unsigned length)
2942 {
2943 	struct ring_buffer_event event; /* Used only for sizeof array */
2944 
2945 	/* zero length can cause confusions */
2946 	if (!length)
2947 		length++;
2948 
2949 	if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
2950 		length += sizeof(event.array[0]);
2951 
2952 	length += RB_EVNT_HDR_SIZE;
2953 	length = ALIGN(length, RB_ARCH_ALIGNMENT);
2954 
2955 	/*
2956 	 * In case the time delta is larger than the 27 bits for it
2957 	 * in the header, we need to add a timestamp. If another
2958 	 * event comes in when trying to discard this one to increase
2959 	 * the length, then the timestamp will be added in the allocated
2960 	 * space of this event. If length is bigger than the size needed
2961 	 * for the TIME_EXTEND, then padding has to be used. The events
2962 	 * length must be either RB_LEN_TIME_EXTEND, or greater than or equal
2963 	 * to RB_LEN_TIME_EXTEND + 8, as 8 is the minimum size for padding.
2964 	 * As length is a multiple of 4, we only need to worry if it
2965 	 * is 12 (RB_LEN_TIME_EXTEND + 4).
2966 	 */
2967 	if (length == RB_LEN_TIME_EXTEND + RB_ALIGNMENT)
2968 		length += RB_ALIGNMENT;
2969 
2970 	return length;
2971 }
2972 
2973 static u64 rb_time_delta(struct ring_buffer_event *event)
2974 {
2975 	switch (event->type_len) {
2976 	case RINGBUF_TYPE_PADDING:
2977 		return 0;
2978 
2979 	case RINGBUF_TYPE_TIME_EXTEND:
2980 		return rb_event_time_stamp(event);
2981 
2982 	case RINGBUF_TYPE_TIME_STAMP:
2983 		return 0;
2984 
2985 	case RINGBUF_TYPE_DATA:
2986 		return event->time_delta;
2987 	default:
2988 		return 0;
2989 	}
2990 }
2991 
2992 static inline bool
2993 rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2994 		  struct ring_buffer_event *event)
2995 {
2996 	unsigned long new_index, old_index;
2997 	struct buffer_page *bpage;
2998 	unsigned long addr;
2999 	u64 write_stamp;
3000 	u64 delta;
3001 
3002 	new_index = rb_event_index(event);
3003 	old_index = new_index + rb_event_ts_length(event);
3004 	addr = (unsigned long)event;
3005 	addr &= PAGE_MASK;
3006 
3007 	bpage = READ_ONCE(cpu_buffer->tail_page);
3008 
3009 	delta = rb_time_delta(event);
3010 
3011 	if (!rb_time_read(&cpu_buffer->write_stamp, &write_stamp))
3012 		return false;
3013 
3014 	/* Make sure the write stamp is read before testing the location */
3015 	barrier();
3016 
3017 	if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
3018 		unsigned long write_mask =
3019 			local_read(&bpage->write) & ~RB_WRITE_MASK;
3020 		unsigned long event_length = rb_event_length(event);
3021 
3022 		/* Something came in, can't discard */
3023 		if (!rb_time_cmpxchg(&cpu_buffer->write_stamp,
3024 				       write_stamp, write_stamp - delta))
3025 			return false;
3026 
3027 		/*
3028 		 * It's possible that the event time delta is zero
3029 		 * (has the same time stamp as the previous event)
3030 		 * in which case write_stamp and before_stamp could
3031 		 * be the same. In such a case, force before_stamp
3032 		 * to be different than write_stamp. It doesn't
3033 		 * matter what it is, as long as its different.
3034 		 */
3035 		if (!delta)
3036 			rb_time_set(&cpu_buffer->before_stamp, 0);
3037 
3038 		/*
3039 		 * If an event were to come in now, it would see that the
3040 		 * write_stamp and the before_stamp are different, and assume
3041 		 * that this event just added itself before updating
3042 		 * the write stamp. The interrupting event will fix the
3043 		 * write stamp for us, and use the before stamp as its delta.
3044 		 */
3045 
3046 		/*
3047 		 * This is on the tail page. It is possible that
3048 		 * a write could come in and move the tail page
3049 		 * and write to the next page. That is fine
3050 		 * because we just shorten what is on this page.
3051 		 */
3052 		old_index += write_mask;
3053 		new_index += write_mask;
3054 
3055 		/* caution: old_index gets updated on cmpxchg failure */
3056 		if (local_try_cmpxchg(&bpage->write, &old_index, new_index)) {
3057 			/* update counters */
3058 			local_sub(event_length, &cpu_buffer->entries_bytes);
3059 			return true;
3060 		}
3061 	}
3062 
3063 	/* could not discard */
3064 	return false;
3065 }
3066 
3067 static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
3068 {
3069 	local_inc(&cpu_buffer->committing);
3070 	local_inc(&cpu_buffer->commits);
3071 }
3072 
3073 static __always_inline void
3074 rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
3075 {
3076 	unsigned long max_count;
3077 
3078 	/*
3079 	 * We only race with interrupts and NMIs on this CPU.
3080 	 * If we own the commit event, then we can commit
3081 	 * all others that interrupted us, since the interruptions
3082 	 * are in stack format (they finish before they come
3083 	 * back to us). This allows us to do a simple loop to
3084 	 * assign the commit to the tail.
3085 	 */
3086  again:
3087 	max_count = cpu_buffer->nr_pages * 100;
3088 
3089 	while (cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)) {
3090 		if (RB_WARN_ON(cpu_buffer, !(--max_count)))
3091 			return;
3092 		if (RB_WARN_ON(cpu_buffer,
3093 			       rb_is_reader_page(cpu_buffer->tail_page)))
3094 			return;
3095 		/*
3096 		 * No need for a memory barrier here, as the update
3097 		 * of the tail_page did it for this page.
3098 		 */
3099 		local_set(&cpu_buffer->commit_page->page->commit,
3100 			  rb_page_write(cpu_buffer->commit_page));
3101 		rb_inc_page(&cpu_buffer->commit_page);
3102 		/* add barrier to keep gcc from optimizing too much */
3103 		barrier();
3104 	}
3105 	while (rb_commit_index(cpu_buffer) !=
3106 	       rb_page_write(cpu_buffer->commit_page)) {
3107 
3108 		/* Make sure the readers see the content of what is committed. */
3109 		smp_wmb();
3110 		local_set(&cpu_buffer->commit_page->page->commit,
3111 			  rb_page_write(cpu_buffer->commit_page));
3112 		RB_WARN_ON(cpu_buffer,
3113 			   local_read(&cpu_buffer->commit_page->page->commit) &
3114 			   ~RB_WRITE_MASK);
3115 		barrier();
3116 	}
3117 
3118 	/* again, keep gcc from optimizing */
3119 	barrier();
3120 
3121 	/*
3122 	 * If an interrupt came in just after the first while loop
3123 	 * and pushed the tail page forward, we will be left with
3124 	 * a dangling commit that will never go forward.
3125 	 */
3126 	if (unlikely(cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)))
3127 		goto again;
3128 }
3129 
3130 static __always_inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
3131 {
3132 	unsigned long commits;
3133 
3134 	if (RB_WARN_ON(cpu_buffer,
3135 		       !local_read(&cpu_buffer->committing)))
3136 		return;
3137 
3138  again:
3139 	commits = local_read(&cpu_buffer->commits);
3140 	/* synchronize with interrupts */
3141 	barrier();
3142 	if (local_read(&cpu_buffer->committing) == 1)
3143 		rb_set_commit_to_write(cpu_buffer);
3144 
3145 	local_dec(&cpu_buffer->committing);
3146 
3147 	/* synchronize with interrupts */
3148 	barrier();
3149 
3150 	/*
3151 	 * Need to account for interrupts coming in between the
3152 	 * updating of the commit page and the clearing of the
3153 	 * committing counter.
3154 	 */
3155 	if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
3156 	    !local_read(&cpu_buffer->committing)) {
3157 		local_inc(&cpu_buffer->committing);
3158 		goto again;
3159 	}
3160 }
3161 
3162 static inline void rb_event_discard(struct ring_buffer_event *event)
3163 {
3164 	if (extended_time(event))
3165 		event = skip_time_extend(event);
3166 
3167 	/* array[0] holds the actual length for the discarded event */
3168 	event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
3169 	event->type_len = RINGBUF_TYPE_PADDING;
3170 	/* time delta must be non zero */
3171 	if (!event->time_delta)
3172 		event->time_delta = 1;
3173 }
3174 
3175 static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer)
3176 {
3177 	local_inc(&cpu_buffer->entries);
3178 	rb_end_commit(cpu_buffer);
3179 }
3180 
3181 static __always_inline void
3182 rb_wakeups(struct trace_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer)
3183 {
3184 	if (buffer->irq_work.waiters_pending) {
3185 		buffer->irq_work.waiters_pending = false;
3186 		/* irq_work_queue() supplies it's own memory barriers */
3187 		irq_work_queue(&buffer->irq_work.work);
3188 	}
3189 
3190 	if (cpu_buffer->irq_work.waiters_pending) {
3191 		cpu_buffer->irq_work.waiters_pending = false;
3192 		/* irq_work_queue() supplies it's own memory barriers */
3193 		irq_work_queue(&cpu_buffer->irq_work.work);
3194 	}
3195 
3196 	if (cpu_buffer->last_pages_touch == local_read(&cpu_buffer->pages_touched))
3197 		return;
3198 
3199 	if (cpu_buffer->reader_page == cpu_buffer->commit_page)
3200 		return;
3201 
3202 	if (!cpu_buffer->irq_work.full_waiters_pending)
3203 		return;
3204 
3205 	cpu_buffer->last_pages_touch = local_read(&cpu_buffer->pages_touched);
3206 
3207 	if (!full_hit(buffer, cpu_buffer->cpu, cpu_buffer->shortest_full))
3208 		return;
3209 
3210 	cpu_buffer->irq_work.wakeup_full = true;
3211 	cpu_buffer->irq_work.full_waiters_pending = false;
3212 	/* irq_work_queue() supplies it's own memory barriers */
3213 	irq_work_queue(&cpu_buffer->irq_work.work);
3214 }
3215 
3216 #ifdef CONFIG_RING_BUFFER_RECORD_RECURSION
3217 # define do_ring_buffer_record_recursion()	\
3218 	do_ftrace_record_recursion(_THIS_IP_, _RET_IP_)
3219 #else
3220 # define do_ring_buffer_record_recursion() do { } while (0)
3221 #endif
3222 
3223 /*
3224  * The lock and unlock are done within a preempt disable section.
3225  * The current_context per_cpu variable can only be modified
3226  * by the current task between lock and unlock. But it can
3227  * be modified more than once via an interrupt. To pass this
3228  * information from the lock to the unlock without having to
3229  * access the 'in_interrupt()' functions again (which do show
3230  * a bit of overhead in something as critical as function tracing,
3231  * we use a bitmask trick.
3232  *
3233  *  bit 1 =  NMI context
3234  *  bit 2 =  IRQ context
3235  *  bit 3 =  SoftIRQ context
3236  *  bit 4 =  normal context.
3237  *
3238  * This works because this is the order of contexts that can
3239  * preempt other contexts. A SoftIRQ never preempts an IRQ
3240  * context.
3241  *
3242  * When the context is determined, the corresponding bit is
3243  * checked and set (if it was set, then a recursion of that context
3244  * happened).
3245  *
3246  * On unlock, we need to clear this bit. To do so, just subtract
3247  * 1 from the current_context and AND it to itself.
3248  *
3249  * (binary)
3250  *  101 - 1 = 100
3251  *  101 & 100 = 100 (clearing bit zero)
3252  *
3253  *  1010 - 1 = 1001
3254  *  1010 & 1001 = 1000 (clearing bit 1)
3255  *
3256  * The least significant bit can be cleared this way, and it
3257  * just so happens that it is the same bit corresponding to
3258  * the current context.
3259  *
3260  * Now the TRANSITION bit breaks the above slightly. The TRANSITION bit
3261  * is set when a recursion is detected at the current context, and if
3262  * the TRANSITION bit is already set, it will fail the recursion.
3263  * This is needed because there's a lag between the changing of
3264  * interrupt context and updating the preempt count. In this case,
3265  * a false positive will be found. To handle this, one extra recursion
3266  * is allowed, and this is done by the TRANSITION bit. If the TRANSITION
3267  * bit is already set, then it is considered a recursion and the function
3268  * ends. Otherwise, the TRANSITION bit is set, and that bit is returned.
3269  *
3270  * On the trace_recursive_unlock(), the TRANSITION bit will be the first
3271  * to be cleared. Even if it wasn't the context that set it. That is,
3272  * if an interrupt comes in while NORMAL bit is set and the ring buffer
3273  * is called before preempt_count() is updated, since the check will
3274  * be on the NORMAL bit, the TRANSITION bit will then be set. If an
3275  * NMI then comes in, it will set the NMI bit, but when the NMI code
3276  * does the trace_recursive_unlock() it will clear the TRANSITION bit
3277  * and leave the NMI bit set. But this is fine, because the interrupt
3278  * code that set the TRANSITION bit will then clear the NMI bit when it
3279  * calls trace_recursive_unlock(). If another NMI comes in, it will
3280  * set the TRANSITION bit and continue.
3281  *
3282  * Note: The TRANSITION bit only handles a single transition between context.
3283  */
3284 
3285 static __always_inline bool
3286 trace_recursive_lock(struct ring_buffer_per_cpu *cpu_buffer)
3287 {
3288 	unsigned int val = cpu_buffer->current_context;
3289 	int bit = interrupt_context_level();
3290 
3291 	bit = RB_CTX_NORMAL - bit;
3292 
3293 	if (unlikely(val & (1 << (bit + cpu_buffer->nest)))) {
3294 		/*
3295 		 * It is possible that this was called by transitioning
3296 		 * between interrupt context, and preempt_count() has not
3297 		 * been updated yet. In this case, use the TRANSITION bit.
3298 		 */
3299 		bit = RB_CTX_TRANSITION;
3300 		if (val & (1 << (bit + cpu_buffer->nest))) {
3301 			do_ring_buffer_record_recursion();
3302 			return true;
3303 		}
3304 	}
3305 
3306 	val |= (1 << (bit + cpu_buffer->nest));
3307 	cpu_buffer->current_context = val;
3308 
3309 	return false;
3310 }
3311 
3312 static __always_inline void
3313 trace_recursive_unlock(struct ring_buffer_per_cpu *cpu_buffer)
3314 {
3315 	cpu_buffer->current_context &=
3316 		cpu_buffer->current_context - (1 << cpu_buffer->nest);
3317 }
3318 
3319 /* The recursive locking above uses 5 bits */
3320 #define NESTED_BITS 5
3321 
3322 /**
3323  * ring_buffer_nest_start - Allow to trace while nested
3324  * @buffer: The ring buffer to modify
3325  *
3326  * The ring buffer has a safety mechanism to prevent recursion.
3327  * But there may be a case where a trace needs to be done while
3328  * tracing something else. In this case, calling this function
3329  * will allow this function to nest within a currently active
3330  * ring_buffer_lock_reserve().
3331  *
3332  * Call this function before calling another ring_buffer_lock_reserve() and
3333  * call ring_buffer_nest_end() after the nested ring_buffer_unlock_commit().
3334  */
3335 void ring_buffer_nest_start(struct trace_buffer *buffer)
3336 {
3337 	struct ring_buffer_per_cpu *cpu_buffer;
3338 	int cpu;
3339 
3340 	/* Enabled by ring_buffer_nest_end() */
3341 	preempt_disable_notrace();
3342 	cpu = raw_smp_processor_id();
3343 	cpu_buffer = buffer->buffers[cpu];
3344 	/* This is the shift value for the above recursive locking */
3345 	cpu_buffer->nest += NESTED_BITS;
3346 }
3347 
3348 /**
3349  * ring_buffer_nest_end - Allow to trace while nested
3350  * @buffer: The ring buffer to modify
3351  *
3352  * Must be called after ring_buffer_nest_start() and after the
3353  * ring_buffer_unlock_commit().
3354  */
3355 void ring_buffer_nest_end(struct trace_buffer *buffer)
3356 {
3357 	struct ring_buffer_per_cpu *cpu_buffer;
3358 	int cpu;
3359 
3360 	/* disabled by ring_buffer_nest_start() */
3361 	cpu = raw_smp_processor_id();
3362 	cpu_buffer = buffer->buffers[cpu];
3363 	/* This is the shift value for the above recursive locking */
3364 	cpu_buffer->nest -= NESTED_BITS;
3365 	preempt_enable_notrace();
3366 }
3367 
3368 /**
3369  * ring_buffer_unlock_commit - commit a reserved
3370  * @buffer: The buffer to commit to
3371  *
3372  * This commits the data to the ring buffer, and releases any locks held.
3373  *
3374  * Must be paired with ring_buffer_lock_reserve.
3375  */
3376 int ring_buffer_unlock_commit(struct trace_buffer *buffer)
3377 {
3378 	struct ring_buffer_per_cpu *cpu_buffer;
3379 	int cpu = raw_smp_processor_id();
3380 
3381 	cpu_buffer = buffer->buffers[cpu];
3382 
3383 	rb_commit(cpu_buffer);
3384 
3385 	rb_wakeups(buffer, cpu_buffer);
3386 
3387 	trace_recursive_unlock(cpu_buffer);
3388 
3389 	preempt_enable_notrace();
3390 
3391 	return 0;
3392 }
3393 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
3394 
3395 /* Special value to validate all deltas on a page. */
3396 #define CHECK_FULL_PAGE		1L
3397 
3398 #ifdef CONFIG_RING_BUFFER_VALIDATE_TIME_DELTAS
3399 static void dump_buffer_page(struct buffer_data_page *bpage,
3400 			     struct rb_event_info *info,
3401 			     unsigned long tail)
3402 {
3403 	struct ring_buffer_event *event;
3404 	u64 ts, delta;
3405 	int e;
3406 
3407 	ts = bpage->time_stamp;
3408 	pr_warn("  [%lld] PAGE TIME STAMP\n", ts);
3409 
3410 	for (e = 0; e < tail; e += rb_event_length(event)) {
3411 
3412 		event = (struct ring_buffer_event *)(bpage->data + e);
3413 
3414 		switch (event->type_len) {
3415 
3416 		case RINGBUF_TYPE_TIME_EXTEND:
3417 			delta = rb_event_time_stamp(event);
3418 			ts += delta;
3419 			pr_warn("  [%lld] delta:%lld TIME EXTEND\n", ts, delta);
3420 			break;
3421 
3422 		case RINGBUF_TYPE_TIME_STAMP:
3423 			delta = rb_event_time_stamp(event);
3424 			ts = rb_fix_abs_ts(delta, ts);
3425 			pr_warn("  [%lld] absolute:%lld TIME STAMP\n", ts, delta);
3426 			break;
3427 
3428 		case RINGBUF_TYPE_PADDING:
3429 			ts += event->time_delta;
3430 			pr_warn("  [%lld] delta:%d PADDING\n", ts, event->time_delta);
3431 			break;
3432 
3433 		case RINGBUF_TYPE_DATA:
3434 			ts += event->time_delta;
3435 			pr_warn("  [%lld] delta:%d\n", ts, event->time_delta);
3436 			break;
3437 
3438 		default:
3439 			break;
3440 		}
3441 	}
3442 }
3443 
3444 static DEFINE_PER_CPU(atomic_t, checking);
3445 static atomic_t ts_dump;
3446 
3447 /*
3448  * Check if the current event time stamp matches the deltas on
3449  * the buffer page.
3450  */
3451 static void check_buffer(struct ring_buffer_per_cpu *cpu_buffer,
3452 			 struct rb_event_info *info,
3453 			 unsigned long tail)
3454 {
3455 	struct ring_buffer_event *event;
3456 	struct buffer_data_page *bpage;
3457 	u64 ts, delta;
3458 	bool full = false;
3459 	int e;
3460 
3461 	bpage = info->tail_page->page;
3462 
3463 	if (tail == CHECK_FULL_PAGE) {
3464 		full = true;
3465 		tail = local_read(&bpage->commit);
3466 	} else if (info->add_timestamp &
3467 		   (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE)) {
3468 		/* Ignore events with absolute time stamps */
3469 		return;
3470 	}
3471 
3472 	/*
3473 	 * Do not check the first event (skip possible extends too).
3474 	 * Also do not check if previous events have not been committed.
3475 	 */
3476 	if (tail <= 8 || tail > local_read(&bpage->commit))
3477 		return;
3478 
3479 	/*
3480 	 * If this interrupted another event,
3481 	 */
3482 	if (atomic_inc_return(this_cpu_ptr(&checking)) != 1)
3483 		goto out;
3484 
3485 	ts = bpage->time_stamp;
3486 
3487 	for (e = 0; e < tail; e += rb_event_length(event)) {
3488 
3489 		event = (struct ring_buffer_event *)(bpage->data + e);
3490 
3491 		switch (event->type_len) {
3492 
3493 		case RINGBUF_TYPE_TIME_EXTEND:
3494 			delta = rb_event_time_stamp(event);
3495 			ts += delta;
3496 			break;
3497 
3498 		case RINGBUF_TYPE_TIME_STAMP:
3499 			delta = rb_event_time_stamp(event);
3500 			ts = rb_fix_abs_ts(delta, ts);
3501 			break;
3502 
3503 		case RINGBUF_TYPE_PADDING:
3504 			if (event->time_delta == 1)
3505 				break;
3506 			fallthrough;
3507 		case RINGBUF_TYPE_DATA:
3508 			ts += event->time_delta;
3509 			break;
3510 
3511 		default:
3512 			RB_WARN_ON(cpu_buffer, 1);
3513 		}
3514 	}
3515 	if ((full && ts > info->ts) ||
3516 	    (!full && ts + info->delta != info->ts)) {
3517 		/* If another report is happening, ignore this one */
3518 		if (atomic_inc_return(&ts_dump) != 1) {
3519 			atomic_dec(&ts_dump);
3520 			goto out;
3521 		}
3522 		atomic_inc(&cpu_buffer->record_disabled);
3523 		/* There's some cases in boot up that this can happen */
3524 		WARN_ON_ONCE(system_state != SYSTEM_BOOTING);
3525 		pr_warn("[CPU: %d]TIME DOES NOT MATCH expected:%lld actual:%lld delta:%lld before:%lld after:%lld%s\n",
3526 			cpu_buffer->cpu,
3527 			ts + info->delta, info->ts, info->delta,
3528 			info->before, info->after,
3529 			full ? " (full)" : "");
3530 		dump_buffer_page(bpage, info, tail);
3531 		atomic_dec(&ts_dump);
3532 		/* Do not re-enable checking */
3533 		return;
3534 	}
3535 out:
3536 	atomic_dec(this_cpu_ptr(&checking));
3537 }
3538 #else
3539 static inline void check_buffer(struct ring_buffer_per_cpu *cpu_buffer,
3540 			 struct rb_event_info *info,
3541 			 unsigned long tail)
3542 {
3543 }
3544 #endif /* CONFIG_RING_BUFFER_VALIDATE_TIME_DELTAS */
3545 
3546 static struct ring_buffer_event *
3547 __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
3548 		  struct rb_event_info *info)
3549 {
3550 	struct ring_buffer_event *event;
3551 	struct buffer_page *tail_page;
3552 	unsigned long tail, write, w;
3553 	bool a_ok;
3554 	bool b_ok;
3555 
3556 	/* Don't let the compiler play games with cpu_buffer->tail_page */
3557 	tail_page = info->tail_page = READ_ONCE(cpu_buffer->tail_page);
3558 
3559  /*A*/	w = local_read(&tail_page->write) & RB_WRITE_MASK;
3560 	barrier();
3561 	b_ok = rb_time_read(&cpu_buffer->before_stamp, &info->before);
3562 	a_ok = rb_time_read(&cpu_buffer->write_stamp, &info->after);
3563 	barrier();
3564 	info->ts = rb_time_stamp(cpu_buffer->buffer);
3565 
3566 	if ((info->add_timestamp & RB_ADD_STAMP_ABSOLUTE)) {
3567 		info->delta = info->ts;
3568 	} else {
3569 		/*
3570 		 * If interrupting an event time update, we may need an
3571 		 * absolute timestamp.
3572 		 * Don't bother if this is the start of a new page (w == 0).
3573 		 */
3574 		if (unlikely(!a_ok || !b_ok || (info->before != info->after && w))) {
3575 			info->add_timestamp |= RB_ADD_STAMP_FORCE | RB_ADD_STAMP_EXTEND;
3576 			info->length += RB_LEN_TIME_EXTEND;
3577 		} else {
3578 			info->delta = info->ts - info->after;
3579 			if (unlikely(test_time_stamp(info->delta))) {
3580 				info->add_timestamp |= RB_ADD_STAMP_EXTEND;
3581 				info->length += RB_LEN_TIME_EXTEND;
3582 			}
3583 		}
3584 	}
3585 
3586  /*B*/	rb_time_set(&cpu_buffer->before_stamp, info->ts);
3587 
3588  /*C*/	write = local_add_return(info->length, &tail_page->write);
3589 
3590 	/* set write to only the index of the write */
3591 	write &= RB_WRITE_MASK;
3592 
3593 	tail = write - info->length;
3594 
3595 	/* See if we shot pass the end of this buffer page */
3596 	if (unlikely(write > BUF_PAGE_SIZE)) {
3597 		/* before and after may now different, fix it up*/
3598 		b_ok = rb_time_read(&cpu_buffer->before_stamp, &info->before);
3599 		a_ok = rb_time_read(&cpu_buffer->write_stamp, &info->after);
3600 		if (a_ok && b_ok && info->before != info->after)
3601 			(void)rb_time_cmpxchg(&cpu_buffer->before_stamp,
3602 					      info->before, info->after);
3603 		if (a_ok && b_ok)
3604 			check_buffer(cpu_buffer, info, CHECK_FULL_PAGE);
3605 		return rb_move_tail(cpu_buffer, tail, info);
3606 	}
3607 
3608 	if (likely(tail == w)) {
3609 		u64 save_before;
3610 		bool s_ok;
3611 
3612 		/* Nothing interrupted us between A and C */
3613  /*D*/		rb_time_set(&cpu_buffer->write_stamp, info->ts);
3614 		barrier();
3615  /*E*/		s_ok = rb_time_read(&cpu_buffer->before_stamp, &save_before);
3616 		RB_WARN_ON(cpu_buffer, !s_ok);
3617 		if (likely(!(info->add_timestamp &
3618 			     (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE))))
3619 			/* This did not interrupt any time update */
3620 			info->delta = info->ts - info->after;
3621 		else
3622 			/* Just use full timestamp for interrupting event */
3623 			info->delta = info->ts;
3624 		barrier();
3625 		check_buffer(cpu_buffer, info, tail);
3626 		if (unlikely(info->ts != save_before)) {
3627 			/* SLOW PATH - Interrupted between C and E */
3628 
3629 			a_ok = rb_time_read(&cpu_buffer->write_stamp, &info->after);
3630 			RB_WARN_ON(cpu_buffer, !a_ok);
3631 
3632 			/* Write stamp must only go forward */
3633 			if (save_before > info->after) {
3634 				/*
3635 				 * We do not care about the result, only that
3636 				 * it gets updated atomically.
3637 				 */
3638 				(void)rb_time_cmpxchg(&cpu_buffer->write_stamp,
3639 						      info->after, save_before);
3640 			}
3641 		}
3642 	} else {
3643 		u64 ts;
3644 		/* SLOW PATH - Interrupted between A and C */
3645 		a_ok = rb_time_read(&cpu_buffer->write_stamp, &info->after);
3646 		/* Was interrupted before here, write_stamp must be valid */
3647 		RB_WARN_ON(cpu_buffer, !a_ok);
3648 		ts = rb_time_stamp(cpu_buffer->buffer);
3649 		barrier();
3650  /*E*/		if (write == (local_read(&tail_page->write) & RB_WRITE_MASK) &&
3651 		    info->after < ts &&
3652 		    rb_time_cmpxchg(&cpu_buffer->write_stamp,
3653 				    info->after, ts)) {
3654 			/* Nothing came after this event between C and E */
3655 			info->delta = ts - info->after;
3656 		} else {
3657 			/*
3658 			 * Interrupted between C and E:
3659 			 * Lost the previous events time stamp. Just set the
3660 			 * delta to zero, and this will be the same time as
3661 			 * the event this event interrupted. And the events that
3662 			 * came after this will still be correct (as they would
3663 			 * have built their delta on the previous event.
3664 			 */
3665 			info->delta = 0;
3666 		}
3667 		info->ts = ts;
3668 		info->add_timestamp &= ~RB_ADD_STAMP_FORCE;
3669 	}
3670 
3671 	/*
3672 	 * If this is the first commit on the page, then it has the same
3673 	 * timestamp as the page itself.
3674 	 */
3675 	if (unlikely(!tail && !(info->add_timestamp &
3676 				(RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE))))
3677 		info->delta = 0;
3678 
3679 	/* We reserved something on the buffer */
3680 
3681 	event = __rb_page_index(tail_page, tail);
3682 	rb_update_event(cpu_buffer, event, info);
3683 
3684 	local_inc(&tail_page->entries);
3685 
3686 	/*
3687 	 * If this is the first commit on the page, then update
3688 	 * its timestamp.
3689 	 */
3690 	if (unlikely(!tail))
3691 		tail_page->page->time_stamp = info->ts;
3692 
3693 	/* account for these added bytes */
3694 	local_add(info->length, &cpu_buffer->entries_bytes);
3695 
3696 	return event;
3697 }
3698 
3699 static __always_inline struct ring_buffer_event *
3700 rb_reserve_next_event(struct trace_buffer *buffer,
3701 		      struct ring_buffer_per_cpu *cpu_buffer,
3702 		      unsigned long length)
3703 {
3704 	struct ring_buffer_event *event;
3705 	struct rb_event_info info;
3706 	int nr_loops = 0;
3707 	int add_ts_default;
3708 
3709 	rb_start_commit(cpu_buffer);
3710 	/* The commit page can not change after this */
3711 
3712 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
3713 	/*
3714 	 * Due to the ability to swap a cpu buffer from a buffer
3715 	 * it is possible it was swapped before we committed.
3716 	 * (committing stops a swap). We check for it here and
3717 	 * if it happened, we have to fail the write.
3718 	 */
3719 	barrier();
3720 	if (unlikely(READ_ONCE(cpu_buffer->buffer) != buffer)) {
3721 		local_dec(&cpu_buffer->committing);
3722 		local_dec(&cpu_buffer->commits);
3723 		return NULL;
3724 	}
3725 #endif
3726 
3727 	info.length = rb_calculate_event_length(length);
3728 
3729 	if (ring_buffer_time_stamp_abs(cpu_buffer->buffer)) {
3730 		add_ts_default = RB_ADD_STAMP_ABSOLUTE;
3731 		info.length += RB_LEN_TIME_EXTEND;
3732 	} else {
3733 		add_ts_default = RB_ADD_STAMP_NONE;
3734 	}
3735 
3736  again:
3737 	info.add_timestamp = add_ts_default;
3738 	info.delta = 0;
3739 
3740 	/*
3741 	 * We allow for interrupts to reenter here and do a trace.
3742 	 * If one does, it will cause this original code to loop
3743 	 * back here. Even with heavy interrupts happening, this
3744 	 * should only happen a few times in a row. If this happens
3745 	 * 1000 times in a row, there must be either an interrupt
3746 	 * storm or we have something buggy.
3747 	 * Bail!
3748 	 */
3749 	if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
3750 		goto out_fail;
3751 
3752 	event = __rb_reserve_next(cpu_buffer, &info);
3753 
3754 	if (unlikely(PTR_ERR(event) == -EAGAIN)) {
3755 		if (info.add_timestamp & (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_EXTEND))
3756 			info.length -= RB_LEN_TIME_EXTEND;
3757 		goto again;
3758 	}
3759 
3760 	if (likely(event))
3761 		return event;
3762  out_fail:
3763 	rb_end_commit(cpu_buffer);
3764 	return NULL;
3765 }
3766 
3767 /**
3768  * ring_buffer_lock_reserve - reserve a part of the buffer
3769  * @buffer: the ring buffer to reserve from
3770  * @length: the length of the data to reserve (excluding event header)
3771  *
3772  * Returns a reserved event on the ring buffer to copy directly to.
3773  * The user of this interface will need to get the body to write into
3774  * and can use the ring_buffer_event_data() interface.
3775  *
3776  * The length is the length of the data needed, not the event length
3777  * which also includes the event header.
3778  *
3779  * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
3780  * If NULL is returned, then nothing has been allocated or locked.
3781  */
3782 struct ring_buffer_event *
3783 ring_buffer_lock_reserve(struct trace_buffer *buffer, unsigned long length)
3784 {
3785 	struct ring_buffer_per_cpu *cpu_buffer;
3786 	struct ring_buffer_event *event;
3787 	int cpu;
3788 
3789 	/* If we are tracing schedule, we don't want to recurse */
3790 	preempt_disable_notrace();
3791 
3792 	if (unlikely(atomic_read(&buffer->record_disabled)))
3793 		goto out;
3794 
3795 	cpu = raw_smp_processor_id();
3796 
3797 	if (unlikely(!cpumask_test_cpu(cpu, buffer->cpumask)))
3798 		goto out;
3799 
3800 	cpu_buffer = buffer->buffers[cpu];
3801 
3802 	if (unlikely(atomic_read(&cpu_buffer->record_disabled)))
3803 		goto out;
3804 
3805 	if (unlikely(length > BUF_MAX_DATA_SIZE))
3806 		goto out;
3807 
3808 	if (unlikely(trace_recursive_lock(cpu_buffer)))
3809 		goto out;
3810 
3811 	event = rb_reserve_next_event(buffer, cpu_buffer, length);
3812 	if (!event)
3813 		goto out_unlock;
3814 
3815 	return event;
3816 
3817  out_unlock:
3818 	trace_recursive_unlock(cpu_buffer);
3819  out:
3820 	preempt_enable_notrace();
3821 	return NULL;
3822 }
3823 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
3824 
3825 /*
3826  * Decrement the entries to the page that an event is on.
3827  * The event does not even need to exist, only the pointer
3828  * to the page it is on. This may only be called before the commit
3829  * takes place.
3830  */
3831 static inline void
3832 rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
3833 		   struct ring_buffer_event *event)
3834 {
3835 	unsigned long addr = (unsigned long)event;
3836 	struct buffer_page *bpage = cpu_buffer->commit_page;
3837 	struct buffer_page *start;
3838 
3839 	addr &= PAGE_MASK;
3840 
3841 	/* Do the likely case first */
3842 	if (likely(bpage->page == (void *)addr)) {
3843 		local_dec(&bpage->entries);
3844 		return;
3845 	}
3846 
3847 	/*
3848 	 * Because the commit page may be on the reader page we
3849 	 * start with the next page and check the end loop there.
3850 	 */
3851 	rb_inc_page(&bpage);
3852 	start = bpage;
3853 	do {
3854 		if (bpage->page == (void *)addr) {
3855 			local_dec(&bpage->entries);
3856 			return;
3857 		}
3858 		rb_inc_page(&bpage);
3859 	} while (bpage != start);
3860 
3861 	/* commit not part of this buffer?? */
3862 	RB_WARN_ON(cpu_buffer, 1);
3863 }
3864 
3865 /**
3866  * ring_buffer_discard_commit - discard an event that has not been committed
3867  * @buffer: the ring buffer
3868  * @event: non committed event to discard
3869  *
3870  * Sometimes an event that is in the ring buffer needs to be ignored.
3871  * This function lets the user discard an event in the ring buffer
3872  * and then that event will not be read later.
3873  *
3874  * This function only works if it is called before the item has been
3875  * committed. It will try to free the event from the ring buffer
3876  * if another event has not been added behind it.
3877  *
3878  * If another event has been added behind it, it will set the event
3879  * up as discarded, and perform the commit.
3880  *
3881  * If this function is called, do not call ring_buffer_unlock_commit on
3882  * the event.
3883  */
3884 void ring_buffer_discard_commit(struct trace_buffer *buffer,
3885 				struct ring_buffer_event *event)
3886 {
3887 	struct ring_buffer_per_cpu *cpu_buffer;
3888 	int cpu;
3889 
3890 	/* The event is discarded regardless */
3891 	rb_event_discard(event);
3892 
3893 	cpu = smp_processor_id();
3894 	cpu_buffer = buffer->buffers[cpu];
3895 
3896 	/*
3897 	 * This must only be called if the event has not been
3898 	 * committed yet. Thus we can assume that preemption
3899 	 * is still disabled.
3900 	 */
3901 	RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
3902 
3903 	rb_decrement_entry(cpu_buffer, event);
3904 	if (rb_try_to_discard(cpu_buffer, event))
3905 		goto out;
3906 
3907  out:
3908 	rb_end_commit(cpu_buffer);
3909 
3910 	trace_recursive_unlock(cpu_buffer);
3911 
3912 	preempt_enable_notrace();
3913 
3914 }
3915 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
3916 
3917 /**
3918  * ring_buffer_write - write data to the buffer without reserving
3919  * @buffer: The ring buffer to write to.
3920  * @length: The length of the data being written (excluding the event header)
3921  * @data: The data to write to the buffer.
3922  *
3923  * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
3924  * one function. If you already have the data to write to the buffer, it
3925  * may be easier to simply call this function.
3926  *
3927  * Note, like ring_buffer_lock_reserve, the length is the length of the data
3928  * and not the length of the event which would hold the header.
3929  */
3930 int ring_buffer_write(struct trace_buffer *buffer,
3931 		      unsigned long length,
3932 		      void *data)
3933 {
3934 	struct ring_buffer_per_cpu *cpu_buffer;
3935 	struct ring_buffer_event *event;
3936 	void *body;
3937 	int ret = -EBUSY;
3938 	int cpu;
3939 
3940 	preempt_disable_notrace();
3941 
3942 	if (atomic_read(&buffer->record_disabled))
3943 		goto out;
3944 
3945 	cpu = raw_smp_processor_id();
3946 
3947 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3948 		goto out;
3949 
3950 	cpu_buffer = buffer->buffers[cpu];
3951 
3952 	if (atomic_read(&cpu_buffer->record_disabled))
3953 		goto out;
3954 
3955 	if (length > BUF_MAX_DATA_SIZE)
3956 		goto out;
3957 
3958 	if (unlikely(trace_recursive_lock(cpu_buffer)))
3959 		goto out;
3960 
3961 	event = rb_reserve_next_event(buffer, cpu_buffer, length);
3962 	if (!event)
3963 		goto out_unlock;
3964 
3965 	body = rb_event_data(event);
3966 
3967 	memcpy(body, data, length);
3968 
3969 	rb_commit(cpu_buffer);
3970 
3971 	rb_wakeups(buffer, cpu_buffer);
3972 
3973 	ret = 0;
3974 
3975  out_unlock:
3976 	trace_recursive_unlock(cpu_buffer);
3977 
3978  out:
3979 	preempt_enable_notrace();
3980 
3981 	return ret;
3982 }
3983 EXPORT_SYMBOL_GPL(ring_buffer_write);
3984 
3985 static bool rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
3986 {
3987 	struct buffer_page *reader = cpu_buffer->reader_page;
3988 	struct buffer_page *head = rb_set_head_page(cpu_buffer);
3989 	struct buffer_page *commit = cpu_buffer->commit_page;
3990 
3991 	/* In case of error, head will be NULL */
3992 	if (unlikely(!head))
3993 		return true;
3994 
3995 	/* Reader should exhaust content in reader page */
3996 	if (reader->read != rb_page_commit(reader))
3997 		return false;
3998 
3999 	/*
4000 	 * If writers are committing on the reader page, knowing all
4001 	 * committed content has been read, the ring buffer is empty.
4002 	 */
4003 	if (commit == reader)
4004 		return true;
4005 
4006 	/*
4007 	 * If writers are committing on a page other than reader page
4008 	 * and head page, there should always be content to read.
4009 	 */
4010 	if (commit != head)
4011 		return false;
4012 
4013 	/*
4014 	 * Writers are committing on the head page, we just need
4015 	 * to care about there're committed data, and the reader will
4016 	 * swap reader page with head page when it is to read data.
4017 	 */
4018 	return rb_page_commit(commit) == 0;
4019 }
4020 
4021 /**
4022  * ring_buffer_record_disable - stop all writes into the buffer
4023  * @buffer: The ring buffer to stop writes to.
4024  *
4025  * This prevents all writes to the buffer. Any attempt to write
4026  * to the buffer after this will fail and return NULL.
4027  *
4028  * The caller should call synchronize_rcu() after this.
4029  */
4030 void ring_buffer_record_disable(struct trace_buffer *buffer)
4031 {
4032 	atomic_inc(&buffer->record_disabled);
4033 }
4034 EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
4035 
4036 /**
4037  * ring_buffer_record_enable - enable writes to the buffer
4038  * @buffer: The ring buffer to enable writes
4039  *
4040  * Note, multiple disables will need the same number of enables
4041  * to truly enable the writing (much like preempt_disable).
4042  */
4043 void ring_buffer_record_enable(struct trace_buffer *buffer)
4044 {
4045 	atomic_dec(&buffer->record_disabled);
4046 }
4047 EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
4048 
4049 /**
4050  * ring_buffer_record_off - stop all writes into the buffer
4051  * @buffer: The ring buffer to stop writes to.
4052  *
4053  * This prevents all writes to the buffer. Any attempt to write
4054  * to the buffer after this will fail and return NULL.
4055  *
4056  * This is different than ring_buffer_record_disable() as
4057  * it works like an on/off switch, where as the disable() version
4058  * must be paired with a enable().
4059  */
4060 void ring_buffer_record_off(struct trace_buffer *buffer)
4061 {
4062 	unsigned int rd;
4063 	unsigned int new_rd;
4064 
4065 	rd = atomic_read(&buffer->record_disabled);
4066 	do {
4067 		new_rd = rd | RB_BUFFER_OFF;
4068 	} while (!atomic_try_cmpxchg(&buffer->record_disabled, &rd, new_rd));
4069 }
4070 EXPORT_SYMBOL_GPL(ring_buffer_record_off);
4071 
4072 /**
4073  * ring_buffer_record_on - restart writes into the buffer
4074  * @buffer: The ring buffer to start writes to.
4075  *
4076  * This enables all writes to the buffer that was disabled by
4077  * ring_buffer_record_off().
4078  *
4079  * This is different than ring_buffer_record_enable() as
4080  * it works like an on/off switch, where as the enable() version
4081  * must be paired with a disable().
4082  */
4083 void ring_buffer_record_on(struct trace_buffer *buffer)
4084 {
4085 	unsigned int rd;
4086 	unsigned int new_rd;
4087 
4088 	rd = atomic_read(&buffer->record_disabled);
4089 	do {
4090 		new_rd = rd & ~RB_BUFFER_OFF;
4091 	} while (!atomic_try_cmpxchg(&buffer->record_disabled, &rd, new_rd));
4092 }
4093 EXPORT_SYMBOL_GPL(ring_buffer_record_on);
4094 
4095 /**
4096  * ring_buffer_record_is_on - return true if the ring buffer can write
4097  * @buffer: The ring buffer to see if write is enabled
4098  *
4099  * Returns true if the ring buffer is in a state that it accepts writes.
4100  */
4101 bool ring_buffer_record_is_on(struct trace_buffer *buffer)
4102 {
4103 	return !atomic_read(&buffer->record_disabled);
4104 }
4105 
4106 /**
4107  * ring_buffer_record_is_set_on - return true if the ring buffer is set writable
4108  * @buffer: The ring buffer to see if write is set enabled
4109  *
4110  * Returns true if the ring buffer is set writable by ring_buffer_record_on().
4111  * Note that this does NOT mean it is in a writable state.
4112  *
4113  * It may return true when the ring buffer has been disabled by
4114  * ring_buffer_record_disable(), as that is a temporary disabling of
4115  * the ring buffer.
4116  */
4117 bool ring_buffer_record_is_set_on(struct trace_buffer *buffer)
4118 {
4119 	return !(atomic_read(&buffer->record_disabled) & RB_BUFFER_OFF);
4120 }
4121 
4122 /**
4123  * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
4124  * @buffer: The ring buffer to stop writes to.
4125  * @cpu: The CPU buffer to stop
4126  *
4127  * This prevents all writes to the buffer. Any attempt to write
4128  * to the buffer after this will fail and return NULL.
4129  *
4130  * The caller should call synchronize_rcu() after this.
4131  */
4132 void ring_buffer_record_disable_cpu(struct trace_buffer *buffer, int cpu)
4133 {
4134 	struct ring_buffer_per_cpu *cpu_buffer;
4135 
4136 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4137 		return;
4138 
4139 	cpu_buffer = buffer->buffers[cpu];
4140 	atomic_inc(&cpu_buffer->record_disabled);
4141 }
4142 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
4143 
4144 /**
4145  * ring_buffer_record_enable_cpu - enable writes to the buffer
4146  * @buffer: The ring buffer to enable writes
4147  * @cpu: The CPU to enable.
4148  *
4149  * Note, multiple disables will need the same number of enables
4150  * to truly enable the writing (much like preempt_disable).
4151  */
4152 void ring_buffer_record_enable_cpu(struct trace_buffer *buffer, int cpu)
4153 {
4154 	struct ring_buffer_per_cpu *cpu_buffer;
4155 
4156 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4157 		return;
4158 
4159 	cpu_buffer = buffer->buffers[cpu];
4160 	atomic_dec(&cpu_buffer->record_disabled);
4161 }
4162 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
4163 
4164 /*
4165  * The total entries in the ring buffer is the running counter
4166  * of entries entered into the ring buffer, minus the sum of
4167  * the entries read from the ring buffer and the number of
4168  * entries that were overwritten.
4169  */
4170 static inline unsigned long
4171 rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
4172 {
4173 	return local_read(&cpu_buffer->entries) -
4174 		(local_read(&cpu_buffer->overrun) + cpu_buffer->read);
4175 }
4176 
4177 /**
4178  * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
4179  * @buffer: The ring buffer
4180  * @cpu: The per CPU buffer to read from.
4181  */
4182 u64 ring_buffer_oldest_event_ts(struct trace_buffer *buffer, int cpu)
4183 {
4184 	unsigned long flags;
4185 	struct ring_buffer_per_cpu *cpu_buffer;
4186 	struct buffer_page *bpage;
4187 	u64 ret = 0;
4188 
4189 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4190 		return 0;
4191 
4192 	cpu_buffer = buffer->buffers[cpu];
4193 	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4194 	/*
4195 	 * if the tail is on reader_page, oldest time stamp is on the reader
4196 	 * page
4197 	 */
4198 	if (cpu_buffer->tail_page == cpu_buffer->reader_page)
4199 		bpage = cpu_buffer->reader_page;
4200 	else
4201 		bpage = rb_set_head_page(cpu_buffer);
4202 	if (bpage)
4203 		ret = bpage->page->time_stamp;
4204 	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4205 
4206 	return ret;
4207 }
4208 EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
4209 
4210 /**
4211  * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
4212  * @buffer: The ring buffer
4213  * @cpu: The per CPU buffer to read from.
4214  */
4215 unsigned long ring_buffer_bytes_cpu(struct trace_buffer *buffer, int cpu)
4216 {
4217 	struct ring_buffer_per_cpu *cpu_buffer;
4218 	unsigned long ret;
4219 
4220 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4221 		return 0;
4222 
4223 	cpu_buffer = buffer->buffers[cpu];
4224 	ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
4225 
4226 	return ret;
4227 }
4228 EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
4229 
4230 /**
4231  * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
4232  * @buffer: The ring buffer
4233  * @cpu: The per CPU buffer to get the entries from.
4234  */
4235 unsigned long ring_buffer_entries_cpu(struct trace_buffer *buffer, int cpu)
4236 {
4237 	struct ring_buffer_per_cpu *cpu_buffer;
4238 
4239 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4240 		return 0;
4241 
4242 	cpu_buffer = buffer->buffers[cpu];
4243 
4244 	return rb_num_of_entries(cpu_buffer);
4245 }
4246 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
4247 
4248 /**
4249  * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
4250  * buffer wrapping around (only if RB_FL_OVERWRITE is on).
4251  * @buffer: The ring buffer
4252  * @cpu: The per CPU buffer to get the number of overruns from
4253  */
4254 unsigned long ring_buffer_overrun_cpu(struct trace_buffer *buffer, int cpu)
4255 {
4256 	struct ring_buffer_per_cpu *cpu_buffer;
4257 	unsigned long ret;
4258 
4259 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4260 		return 0;
4261 
4262 	cpu_buffer = buffer->buffers[cpu];
4263 	ret = local_read(&cpu_buffer->overrun);
4264 
4265 	return ret;
4266 }
4267 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
4268 
4269 /**
4270  * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
4271  * commits failing due to the buffer wrapping around while there are uncommitted
4272  * events, such as during an interrupt storm.
4273  * @buffer: The ring buffer
4274  * @cpu: The per CPU buffer to get the number of overruns from
4275  */
4276 unsigned long
4277 ring_buffer_commit_overrun_cpu(struct trace_buffer *buffer, int cpu)
4278 {
4279 	struct ring_buffer_per_cpu *cpu_buffer;
4280 	unsigned long ret;
4281 
4282 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4283 		return 0;
4284 
4285 	cpu_buffer = buffer->buffers[cpu];
4286 	ret = local_read(&cpu_buffer->commit_overrun);
4287 
4288 	return ret;
4289 }
4290 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
4291 
4292 /**
4293  * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
4294  * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
4295  * @buffer: The ring buffer
4296  * @cpu: The per CPU buffer to get the number of overruns from
4297  */
4298 unsigned long
4299 ring_buffer_dropped_events_cpu(struct trace_buffer *buffer, int cpu)
4300 {
4301 	struct ring_buffer_per_cpu *cpu_buffer;
4302 	unsigned long ret;
4303 
4304 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4305 		return 0;
4306 
4307 	cpu_buffer = buffer->buffers[cpu];
4308 	ret = local_read(&cpu_buffer->dropped_events);
4309 
4310 	return ret;
4311 }
4312 EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
4313 
4314 /**
4315  * ring_buffer_read_events_cpu - get the number of events successfully read
4316  * @buffer: The ring buffer
4317  * @cpu: The per CPU buffer to get the number of events read
4318  */
4319 unsigned long
4320 ring_buffer_read_events_cpu(struct trace_buffer *buffer, int cpu)
4321 {
4322 	struct ring_buffer_per_cpu *cpu_buffer;
4323 
4324 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4325 		return 0;
4326 
4327 	cpu_buffer = buffer->buffers[cpu];
4328 	return cpu_buffer->read;
4329 }
4330 EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu);
4331 
4332 /**
4333  * ring_buffer_entries - get the number of entries in a buffer
4334  * @buffer: The ring buffer
4335  *
4336  * Returns the total number of entries in the ring buffer
4337  * (all CPU entries)
4338  */
4339 unsigned long ring_buffer_entries(struct trace_buffer *buffer)
4340 {
4341 	struct ring_buffer_per_cpu *cpu_buffer;
4342 	unsigned long entries = 0;
4343 	int cpu;
4344 
4345 	/* if you care about this being correct, lock the buffer */
4346 	for_each_buffer_cpu(buffer, cpu) {
4347 		cpu_buffer = buffer->buffers[cpu];
4348 		entries += rb_num_of_entries(cpu_buffer);
4349 	}
4350 
4351 	return entries;
4352 }
4353 EXPORT_SYMBOL_GPL(ring_buffer_entries);
4354 
4355 /**
4356  * ring_buffer_overruns - get the number of overruns in buffer
4357  * @buffer: The ring buffer
4358  *
4359  * Returns the total number of overruns in the ring buffer
4360  * (all CPU entries)
4361  */
4362 unsigned long ring_buffer_overruns(struct trace_buffer *buffer)
4363 {
4364 	struct ring_buffer_per_cpu *cpu_buffer;
4365 	unsigned long overruns = 0;
4366 	int cpu;
4367 
4368 	/* if you care about this being correct, lock the buffer */
4369 	for_each_buffer_cpu(buffer, cpu) {
4370 		cpu_buffer = buffer->buffers[cpu];
4371 		overruns += local_read(&cpu_buffer->overrun);
4372 	}
4373 
4374 	return overruns;
4375 }
4376 EXPORT_SYMBOL_GPL(ring_buffer_overruns);
4377 
4378 static void rb_iter_reset(struct ring_buffer_iter *iter)
4379 {
4380 	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4381 
4382 	/* Iterator usage is expected to have record disabled */
4383 	iter->head_page = cpu_buffer->reader_page;
4384 	iter->head = cpu_buffer->reader_page->read;
4385 	iter->next_event = iter->head;
4386 
4387 	iter->cache_reader_page = iter->head_page;
4388 	iter->cache_read = cpu_buffer->read;
4389 	iter->cache_pages_removed = cpu_buffer->pages_removed;
4390 
4391 	if (iter->head) {
4392 		iter->read_stamp = cpu_buffer->read_stamp;
4393 		iter->page_stamp = cpu_buffer->reader_page->page->time_stamp;
4394 	} else {
4395 		iter->read_stamp = iter->head_page->page->time_stamp;
4396 		iter->page_stamp = iter->read_stamp;
4397 	}
4398 }
4399 
4400 /**
4401  * ring_buffer_iter_reset - reset an iterator
4402  * @iter: The iterator to reset
4403  *
4404  * Resets the iterator, so that it will start from the beginning
4405  * again.
4406  */
4407 void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
4408 {
4409 	struct ring_buffer_per_cpu *cpu_buffer;
4410 	unsigned long flags;
4411 
4412 	if (!iter)
4413 		return;
4414 
4415 	cpu_buffer = iter->cpu_buffer;
4416 
4417 	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4418 	rb_iter_reset(iter);
4419 	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4420 }
4421 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
4422 
4423 /**
4424  * ring_buffer_iter_empty - check if an iterator has no more to read
4425  * @iter: The iterator to check
4426  */
4427 int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
4428 {
4429 	struct ring_buffer_per_cpu *cpu_buffer;
4430 	struct buffer_page *reader;
4431 	struct buffer_page *head_page;
4432 	struct buffer_page *commit_page;
4433 	struct buffer_page *curr_commit_page;
4434 	unsigned commit;
4435 	u64 curr_commit_ts;
4436 	u64 commit_ts;
4437 
4438 	cpu_buffer = iter->cpu_buffer;
4439 	reader = cpu_buffer->reader_page;
4440 	head_page = cpu_buffer->head_page;
4441 	commit_page = cpu_buffer->commit_page;
4442 	commit_ts = commit_page->page->time_stamp;
4443 
4444 	/*
4445 	 * When the writer goes across pages, it issues a cmpxchg which
4446 	 * is a mb(), which will synchronize with the rmb here.
4447 	 * (see rb_tail_page_update())
4448 	 */
4449 	smp_rmb();
4450 	commit = rb_page_commit(commit_page);
4451 	/* We want to make sure that the commit page doesn't change */
4452 	smp_rmb();
4453 
4454 	/* Make sure commit page didn't change */
4455 	curr_commit_page = READ_ONCE(cpu_buffer->commit_page);
4456 	curr_commit_ts = READ_ONCE(curr_commit_page->page->time_stamp);
4457 
4458 	/* If the commit page changed, then there's more data */
4459 	if (curr_commit_page != commit_page ||
4460 	    curr_commit_ts != commit_ts)
4461 		return 0;
4462 
4463 	/* Still racy, as it may return a false positive, but that's OK */
4464 	return ((iter->head_page == commit_page && iter->head >= commit) ||
4465 		(iter->head_page == reader && commit_page == head_page &&
4466 		 head_page->read == commit &&
4467 		 iter->head == rb_page_commit(cpu_buffer->reader_page)));
4468 }
4469 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
4470 
4471 static void
4472 rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
4473 		     struct ring_buffer_event *event)
4474 {
4475 	u64 delta;
4476 
4477 	switch (event->type_len) {
4478 	case RINGBUF_TYPE_PADDING:
4479 		return;
4480 
4481 	case RINGBUF_TYPE_TIME_EXTEND:
4482 		delta = rb_event_time_stamp(event);
4483 		cpu_buffer->read_stamp += delta;
4484 		return;
4485 
4486 	case RINGBUF_TYPE_TIME_STAMP:
4487 		delta = rb_event_time_stamp(event);
4488 		delta = rb_fix_abs_ts(delta, cpu_buffer->read_stamp);
4489 		cpu_buffer->read_stamp = delta;
4490 		return;
4491 
4492 	case RINGBUF_TYPE_DATA:
4493 		cpu_buffer->read_stamp += event->time_delta;
4494 		return;
4495 
4496 	default:
4497 		RB_WARN_ON(cpu_buffer, 1);
4498 	}
4499 }
4500 
4501 static void
4502 rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
4503 			  struct ring_buffer_event *event)
4504 {
4505 	u64 delta;
4506 
4507 	switch (event->type_len) {
4508 	case RINGBUF_TYPE_PADDING:
4509 		return;
4510 
4511 	case RINGBUF_TYPE_TIME_EXTEND:
4512 		delta = rb_event_time_stamp(event);
4513 		iter->read_stamp += delta;
4514 		return;
4515 
4516 	case RINGBUF_TYPE_TIME_STAMP:
4517 		delta = rb_event_time_stamp(event);
4518 		delta = rb_fix_abs_ts(delta, iter->read_stamp);
4519 		iter->read_stamp = delta;
4520 		return;
4521 
4522 	case RINGBUF_TYPE_DATA:
4523 		iter->read_stamp += event->time_delta;
4524 		return;
4525 
4526 	default:
4527 		RB_WARN_ON(iter->cpu_buffer, 1);
4528 	}
4529 }
4530 
4531 static struct buffer_page *
4532 rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
4533 {
4534 	struct buffer_page *reader = NULL;
4535 	unsigned long overwrite;
4536 	unsigned long flags;
4537 	int nr_loops = 0;
4538 	bool ret;
4539 
4540 	local_irq_save(flags);
4541 	arch_spin_lock(&cpu_buffer->lock);
4542 
4543  again:
4544 	/*
4545 	 * This should normally only loop twice. But because the
4546 	 * start of the reader inserts an empty page, it causes
4547 	 * a case where we will loop three times. There should be no
4548 	 * reason to loop four times (that I know of).
4549 	 */
4550 	if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
4551 		reader = NULL;
4552 		goto out;
4553 	}
4554 
4555 	reader = cpu_buffer->reader_page;
4556 
4557 	/* If there's more to read, return this page */
4558 	if (cpu_buffer->reader_page->read < rb_page_size(reader))
4559 		goto out;
4560 
4561 	/* Never should we have an index greater than the size */
4562 	if (RB_WARN_ON(cpu_buffer,
4563 		       cpu_buffer->reader_page->read > rb_page_size(reader)))
4564 		goto out;
4565 
4566 	/* check if we caught up to the tail */
4567 	reader = NULL;
4568 	if (cpu_buffer->commit_page == cpu_buffer->reader_page)
4569 		goto out;
4570 
4571 	/* Don't bother swapping if the ring buffer is empty */
4572 	if (rb_num_of_entries(cpu_buffer) == 0)
4573 		goto out;
4574 
4575 	/*
4576 	 * Reset the reader page to size zero.
4577 	 */
4578 	local_set(&cpu_buffer->reader_page->write, 0);
4579 	local_set(&cpu_buffer->reader_page->entries, 0);
4580 	local_set(&cpu_buffer->reader_page->page->commit, 0);
4581 	cpu_buffer->reader_page->real_end = 0;
4582 
4583  spin:
4584 	/*
4585 	 * Splice the empty reader page into the list around the head.
4586 	 */
4587 	reader = rb_set_head_page(cpu_buffer);
4588 	if (!reader)
4589 		goto out;
4590 	cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
4591 	cpu_buffer->reader_page->list.prev = reader->list.prev;
4592 
4593 	/*
4594 	 * cpu_buffer->pages just needs to point to the buffer, it
4595 	 *  has no specific buffer page to point to. Lets move it out
4596 	 *  of our way so we don't accidentally swap it.
4597 	 */
4598 	cpu_buffer->pages = reader->list.prev;
4599 
4600 	/* The reader page will be pointing to the new head */
4601 	rb_set_list_to_head(&cpu_buffer->reader_page->list);
4602 
4603 	/*
4604 	 * We want to make sure we read the overruns after we set up our
4605 	 * pointers to the next object. The writer side does a
4606 	 * cmpxchg to cross pages which acts as the mb on the writer
4607 	 * side. Note, the reader will constantly fail the swap
4608 	 * while the writer is updating the pointers, so this
4609 	 * guarantees that the overwrite recorded here is the one we
4610 	 * want to compare with the last_overrun.
4611 	 */
4612 	smp_mb();
4613 	overwrite = local_read(&(cpu_buffer->overrun));
4614 
4615 	/*
4616 	 * Here's the tricky part.
4617 	 *
4618 	 * We need to move the pointer past the header page.
4619 	 * But we can only do that if a writer is not currently
4620 	 * moving it. The page before the header page has the
4621 	 * flag bit '1' set if it is pointing to the page we want.
4622 	 * but if the writer is in the process of moving it
4623 	 * than it will be '2' or already moved '0'.
4624 	 */
4625 
4626 	ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
4627 
4628 	/*
4629 	 * If we did not convert it, then we must try again.
4630 	 */
4631 	if (!ret)
4632 		goto spin;
4633 
4634 	/*
4635 	 * Yay! We succeeded in replacing the page.
4636 	 *
4637 	 * Now make the new head point back to the reader page.
4638 	 */
4639 	rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
4640 	rb_inc_page(&cpu_buffer->head_page);
4641 
4642 	local_inc(&cpu_buffer->pages_read);
4643 
4644 	/* Finally update the reader page to the new head */
4645 	cpu_buffer->reader_page = reader;
4646 	cpu_buffer->reader_page->read = 0;
4647 
4648 	if (overwrite != cpu_buffer->last_overrun) {
4649 		cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
4650 		cpu_buffer->last_overrun = overwrite;
4651 	}
4652 
4653 	goto again;
4654 
4655  out:
4656 	/* Update the read_stamp on the first event */
4657 	if (reader && reader->read == 0)
4658 		cpu_buffer->read_stamp = reader->page->time_stamp;
4659 
4660 	arch_spin_unlock(&cpu_buffer->lock);
4661 	local_irq_restore(flags);
4662 
4663 	/*
4664 	 * The writer has preempt disable, wait for it. But not forever
4665 	 * Although, 1 second is pretty much "forever"
4666 	 */
4667 #define USECS_WAIT	1000000
4668         for (nr_loops = 0; nr_loops < USECS_WAIT; nr_loops++) {
4669 		/* If the write is past the end of page, a writer is still updating it */
4670 		if (likely(!reader || rb_page_write(reader) <= BUF_PAGE_SIZE))
4671 			break;
4672 
4673 		udelay(1);
4674 
4675 		/* Get the latest version of the reader write value */
4676 		smp_rmb();
4677 	}
4678 
4679 	/* The writer is not moving forward? Something is wrong */
4680 	if (RB_WARN_ON(cpu_buffer, nr_loops == USECS_WAIT))
4681 		reader = NULL;
4682 
4683 	/*
4684 	 * Make sure we see any padding after the write update
4685 	 * (see rb_reset_tail()).
4686 	 *
4687 	 * In addition, a writer may be writing on the reader page
4688 	 * if the page has not been fully filled, so the read barrier
4689 	 * is also needed to make sure we see the content of what is
4690 	 * committed by the writer (see rb_set_commit_to_write()).
4691 	 */
4692 	smp_rmb();
4693 
4694 
4695 	return reader;
4696 }
4697 
4698 static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
4699 {
4700 	struct ring_buffer_event *event;
4701 	struct buffer_page *reader;
4702 	unsigned length;
4703 
4704 	reader = rb_get_reader_page(cpu_buffer);
4705 
4706 	/* This function should not be called when buffer is empty */
4707 	if (RB_WARN_ON(cpu_buffer, !reader))
4708 		return;
4709 
4710 	event = rb_reader_event(cpu_buffer);
4711 
4712 	if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
4713 		cpu_buffer->read++;
4714 
4715 	rb_update_read_stamp(cpu_buffer, event);
4716 
4717 	length = rb_event_length(event);
4718 	cpu_buffer->reader_page->read += length;
4719 }
4720 
4721 static void rb_advance_iter(struct ring_buffer_iter *iter)
4722 {
4723 	struct ring_buffer_per_cpu *cpu_buffer;
4724 
4725 	cpu_buffer = iter->cpu_buffer;
4726 
4727 	/* If head == next_event then we need to jump to the next event */
4728 	if (iter->head == iter->next_event) {
4729 		/* If the event gets overwritten again, there's nothing to do */
4730 		if (rb_iter_head_event(iter) == NULL)
4731 			return;
4732 	}
4733 
4734 	iter->head = iter->next_event;
4735 
4736 	/*
4737 	 * Check if we are at the end of the buffer.
4738 	 */
4739 	if (iter->next_event >= rb_page_size(iter->head_page)) {
4740 		/* discarded commits can make the page empty */
4741 		if (iter->head_page == cpu_buffer->commit_page)
4742 			return;
4743 		rb_inc_iter(iter);
4744 		return;
4745 	}
4746 
4747 	rb_update_iter_read_stamp(iter, iter->event);
4748 }
4749 
4750 static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
4751 {
4752 	return cpu_buffer->lost_events;
4753 }
4754 
4755 static struct ring_buffer_event *
4756 rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
4757 	       unsigned long *lost_events)
4758 {
4759 	struct ring_buffer_event *event;
4760 	struct buffer_page *reader;
4761 	int nr_loops = 0;
4762 
4763 	if (ts)
4764 		*ts = 0;
4765  again:
4766 	/*
4767 	 * We repeat when a time extend is encountered.
4768 	 * Since the time extend is always attached to a data event,
4769 	 * we should never loop more than once.
4770 	 * (We never hit the following condition more than twice).
4771 	 */
4772 	if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
4773 		return NULL;
4774 
4775 	reader = rb_get_reader_page(cpu_buffer);
4776 	if (!reader)
4777 		return NULL;
4778 
4779 	event = rb_reader_event(cpu_buffer);
4780 
4781 	switch (event->type_len) {
4782 	case RINGBUF_TYPE_PADDING:
4783 		if (rb_null_event(event))
4784 			RB_WARN_ON(cpu_buffer, 1);
4785 		/*
4786 		 * Because the writer could be discarding every
4787 		 * event it creates (which would probably be bad)
4788 		 * if we were to go back to "again" then we may never
4789 		 * catch up, and will trigger the warn on, or lock
4790 		 * the box. Return the padding, and we will release
4791 		 * the current locks, and try again.
4792 		 */
4793 		return event;
4794 
4795 	case RINGBUF_TYPE_TIME_EXTEND:
4796 		/* Internal data, OK to advance */
4797 		rb_advance_reader(cpu_buffer);
4798 		goto again;
4799 
4800 	case RINGBUF_TYPE_TIME_STAMP:
4801 		if (ts) {
4802 			*ts = rb_event_time_stamp(event);
4803 			*ts = rb_fix_abs_ts(*ts, reader->page->time_stamp);
4804 			ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
4805 							 cpu_buffer->cpu, ts);
4806 		}
4807 		/* Internal data, OK to advance */
4808 		rb_advance_reader(cpu_buffer);
4809 		goto again;
4810 
4811 	case RINGBUF_TYPE_DATA:
4812 		if (ts && !(*ts)) {
4813 			*ts = cpu_buffer->read_stamp + event->time_delta;
4814 			ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
4815 							 cpu_buffer->cpu, ts);
4816 		}
4817 		if (lost_events)
4818 			*lost_events = rb_lost_events(cpu_buffer);
4819 		return event;
4820 
4821 	default:
4822 		RB_WARN_ON(cpu_buffer, 1);
4823 	}
4824 
4825 	return NULL;
4826 }
4827 EXPORT_SYMBOL_GPL(ring_buffer_peek);
4828 
4829 static struct ring_buffer_event *
4830 rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
4831 {
4832 	struct trace_buffer *buffer;
4833 	struct ring_buffer_per_cpu *cpu_buffer;
4834 	struct ring_buffer_event *event;
4835 	int nr_loops = 0;
4836 
4837 	if (ts)
4838 		*ts = 0;
4839 
4840 	cpu_buffer = iter->cpu_buffer;
4841 	buffer = cpu_buffer->buffer;
4842 
4843 	/*
4844 	 * Check if someone performed a consuming read to the buffer
4845 	 * or removed some pages from the buffer. In these cases,
4846 	 * iterator was invalidated and we need to reset it.
4847 	 */
4848 	if (unlikely(iter->cache_read != cpu_buffer->read ||
4849 		     iter->cache_reader_page != cpu_buffer->reader_page ||
4850 		     iter->cache_pages_removed != cpu_buffer->pages_removed))
4851 		rb_iter_reset(iter);
4852 
4853  again:
4854 	if (ring_buffer_iter_empty(iter))
4855 		return NULL;
4856 
4857 	/*
4858 	 * As the writer can mess with what the iterator is trying
4859 	 * to read, just give up if we fail to get an event after
4860 	 * three tries. The iterator is not as reliable when reading
4861 	 * the ring buffer with an active write as the consumer is.
4862 	 * Do not warn if the three failures is reached.
4863 	 */
4864 	if (++nr_loops > 3)
4865 		return NULL;
4866 
4867 	if (rb_per_cpu_empty(cpu_buffer))
4868 		return NULL;
4869 
4870 	if (iter->head >= rb_page_size(iter->head_page)) {
4871 		rb_inc_iter(iter);
4872 		goto again;
4873 	}
4874 
4875 	event = rb_iter_head_event(iter);
4876 	if (!event)
4877 		goto again;
4878 
4879 	switch (event->type_len) {
4880 	case RINGBUF_TYPE_PADDING:
4881 		if (rb_null_event(event)) {
4882 			rb_inc_iter(iter);
4883 			goto again;
4884 		}
4885 		rb_advance_iter(iter);
4886 		return event;
4887 
4888 	case RINGBUF_TYPE_TIME_EXTEND:
4889 		/* Internal data, OK to advance */
4890 		rb_advance_iter(iter);
4891 		goto again;
4892 
4893 	case RINGBUF_TYPE_TIME_STAMP:
4894 		if (ts) {
4895 			*ts = rb_event_time_stamp(event);
4896 			*ts = rb_fix_abs_ts(*ts, iter->head_page->page->time_stamp);
4897 			ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
4898 							 cpu_buffer->cpu, ts);
4899 		}
4900 		/* Internal data, OK to advance */
4901 		rb_advance_iter(iter);
4902 		goto again;
4903 
4904 	case RINGBUF_TYPE_DATA:
4905 		if (ts && !(*ts)) {
4906 			*ts = iter->read_stamp + event->time_delta;
4907 			ring_buffer_normalize_time_stamp(buffer,
4908 							 cpu_buffer->cpu, ts);
4909 		}
4910 		return event;
4911 
4912 	default:
4913 		RB_WARN_ON(cpu_buffer, 1);
4914 	}
4915 
4916 	return NULL;
4917 }
4918 EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
4919 
4920 static inline bool rb_reader_lock(struct ring_buffer_per_cpu *cpu_buffer)
4921 {
4922 	if (likely(!in_nmi())) {
4923 		raw_spin_lock(&cpu_buffer->reader_lock);
4924 		return true;
4925 	}
4926 
4927 	/*
4928 	 * If an NMI die dumps out the content of the ring buffer
4929 	 * trylock must be used to prevent a deadlock if the NMI
4930 	 * preempted a task that holds the ring buffer locks. If
4931 	 * we get the lock then all is fine, if not, then continue
4932 	 * to do the read, but this can corrupt the ring buffer,
4933 	 * so it must be permanently disabled from future writes.
4934 	 * Reading from NMI is a oneshot deal.
4935 	 */
4936 	if (raw_spin_trylock(&cpu_buffer->reader_lock))
4937 		return true;
4938 
4939 	/* Continue without locking, but disable the ring buffer */
4940 	atomic_inc(&cpu_buffer->record_disabled);
4941 	return false;
4942 }
4943 
4944 static inline void
4945 rb_reader_unlock(struct ring_buffer_per_cpu *cpu_buffer, bool locked)
4946 {
4947 	if (likely(locked))
4948 		raw_spin_unlock(&cpu_buffer->reader_lock);
4949 }
4950 
4951 /**
4952  * ring_buffer_peek - peek at the next event to be read
4953  * @buffer: The ring buffer to read
4954  * @cpu: The cpu to peak at
4955  * @ts: The timestamp counter of this event.
4956  * @lost_events: a variable to store if events were lost (may be NULL)
4957  *
4958  * This will return the event that will be read next, but does
4959  * not consume the data.
4960  */
4961 struct ring_buffer_event *
4962 ring_buffer_peek(struct trace_buffer *buffer, int cpu, u64 *ts,
4963 		 unsigned long *lost_events)
4964 {
4965 	struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4966 	struct ring_buffer_event *event;
4967 	unsigned long flags;
4968 	bool dolock;
4969 
4970 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4971 		return NULL;
4972 
4973  again:
4974 	local_irq_save(flags);
4975 	dolock = rb_reader_lock(cpu_buffer);
4976 	event = rb_buffer_peek(cpu_buffer, ts, lost_events);
4977 	if (event && event->type_len == RINGBUF_TYPE_PADDING)
4978 		rb_advance_reader(cpu_buffer);
4979 	rb_reader_unlock(cpu_buffer, dolock);
4980 	local_irq_restore(flags);
4981 
4982 	if (event && event->type_len == RINGBUF_TYPE_PADDING)
4983 		goto again;
4984 
4985 	return event;
4986 }
4987 
4988 /** ring_buffer_iter_dropped - report if there are dropped events
4989  * @iter: The ring buffer iterator
4990  *
4991  * Returns true if there was dropped events since the last peek.
4992  */
4993 bool ring_buffer_iter_dropped(struct ring_buffer_iter *iter)
4994 {
4995 	bool ret = iter->missed_events != 0;
4996 
4997 	iter->missed_events = 0;
4998 	return ret;
4999 }
5000 EXPORT_SYMBOL_GPL(ring_buffer_iter_dropped);
5001 
5002 /**
5003  * ring_buffer_iter_peek - peek at the next event to be read
5004  * @iter: The ring buffer iterator
5005  * @ts: The timestamp counter of this event.
5006  *
5007  * This will return the event that will be read next, but does
5008  * not increment the iterator.
5009  */
5010 struct ring_buffer_event *
5011 ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
5012 {
5013 	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
5014 	struct ring_buffer_event *event;
5015 	unsigned long flags;
5016 
5017  again:
5018 	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5019 	event = rb_iter_peek(iter, ts);
5020 	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5021 
5022 	if (event && event->type_len == RINGBUF_TYPE_PADDING)
5023 		goto again;
5024 
5025 	return event;
5026 }
5027 
5028 /**
5029  * ring_buffer_consume - return an event and consume it
5030  * @buffer: The ring buffer to get the next event from
5031  * @cpu: the cpu to read the buffer from
5032  * @ts: a variable to store the timestamp (may be NULL)
5033  * @lost_events: a variable to store if events were lost (may be NULL)
5034  *
5035  * Returns the next event in the ring buffer, and that event is consumed.
5036  * Meaning, that sequential reads will keep returning a different event,
5037  * and eventually empty the ring buffer if the producer is slower.
5038  */
5039 struct ring_buffer_event *
5040 ring_buffer_consume(struct trace_buffer *buffer, int cpu, u64 *ts,
5041 		    unsigned long *lost_events)
5042 {
5043 	struct ring_buffer_per_cpu *cpu_buffer;
5044 	struct ring_buffer_event *event = NULL;
5045 	unsigned long flags;
5046 	bool dolock;
5047 
5048  again:
5049 	/* might be called in atomic */
5050 	preempt_disable();
5051 
5052 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
5053 		goto out;
5054 
5055 	cpu_buffer = buffer->buffers[cpu];
5056 	local_irq_save(flags);
5057 	dolock = rb_reader_lock(cpu_buffer);
5058 
5059 	event = rb_buffer_peek(cpu_buffer, ts, lost_events);
5060 	if (event) {
5061 		cpu_buffer->lost_events = 0;
5062 		rb_advance_reader(cpu_buffer);
5063 	}
5064 
5065 	rb_reader_unlock(cpu_buffer, dolock);
5066 	local_irq_restore(flags);
5067 
5068  out:
5069 	preempt_enable();
5070 
5071 	if (event && event->type_len == RINGBUF_TYPE_PADDING)
5072 		goto again;
5073 
5074 	return event;
5075 }
5076 EXPORT_SYMBOL_GPL(ring_buffer_consume);
5077 
5078 /**
5079  * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
5080  * @buffer: The ring buffer to read from
5081  * @cpu: The cpu buffer to iterate over
5082  * @flags: gfp flags to use for memory allocation
5083  *
5084  * This performs the initial preparations necessary to iterate
5085  * through the buffer.  Memory is allocated, buffer recording
5086  * is disabled, and the iterator pointer is returned to the caller.
5087  *
5088  * Disabling buffer recording prevents the reading from being
5089  * corrupted. This is not a consuming read, so a producer is not
5090  * expected.
5091  *
5092  * After a sequence of ring_buffer_read_prepare calls, the user is
5093  * expected to make at least one call to ring_buffer_read_prepare_sync.
5094  * Afterwards, ring_buffer_read_start is invoked to get things going
5095  * for real.
5096  *
5097  * This overall must be paired with ring_buffer_read_finish.
5098  */
5099 struct ring_buffer_iter *
5100 ring_buffer_read_prepare(struct trace_buffer *buffer, int cpu, gfp_t flags)
5101 {
5102 	struct ring_buffer_per_cpu *cpu_buffer;
5103 	struct ring_buffer_iter *iter;
5104 
5105 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
5106 		return NULL;
5107 
5108 	iter = kzalloc(sizeof(*iter), flags);
5109 	if (!iter)
5110 		return NULL;
5111 
5112 	iter->event = kmalloc(BUF_MAX_DATA_SIZE, flags);
5113 	if (!iter->event) {
5114 		kfree(iter);
5115 		return NULL;
5116 	}
5117 
5118 	cpu_buffer = buffer->buffers[cpu];
5119 
5120 	iter->cpu_buffer = cpu_buffer;
5121 
5122 	atomic_inc(&cpu_buffer->resize_disabled);
5123 
5124 	return iter;
5125 }
5126 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
5127 
5128 /**
5129  * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
5130  *
5131  * All previously invoked ring_buffer_read_prepare calls to prepare
5132  * iterators will be synchronized.  Afterwards, read_buffer_read_start
5133  * calls on those iterators are allowed.
5134  */
5135 void
5136 ring_buffer_read_prepare_sync(void)
5137 {
5138 	synchronize_rcu();
5139 }
5140 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
5141 
5142 /**
5143  * ring_buffer_read_start - start a non consuming read of the buffer
5144  * @iter: The iterator returned by ring_buffer_read_prepare
5145  *
5146  * This finalizes the startup of an iteration through the buffer.
5147  * The iterator comes from a call to ring_buffer_read_prepare and
5148  * an intervening ring_buffer_read_prepare_sync must have been
5149  * performed.
5150  *
5151  * Must be paired with ring_buffer_read_finish.
5152  */
5153 void
5154 ring_buffer_read_start(struct ring_buffer_iter *iter)
5155 {
5156 	struct ring_buffer_per_cpu *cpu_buffer;
5157 	unsigned long flags;
5158 
5159 	if (!iter)
5160 		return;
5161 
5162 	cpu_buffer = iter->cpu_buffer;
5163 
5164 	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5165 	arch_spin_lock(&cpu_buffer->lock);
5166 	rb_iter_reset(iter);
5167 	arch_spin_unlock(&cpu_buffer->lock);
5168 	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5169 }
5170 EXPORT_SYMBOL_GPL(ring_buffer_read_start);
5171 
5172 /**
5173  * ring_buffer_read_finish - finish reading the iterator of the buffer
5174  * @iter: The iterator retrieved by ring_buffer_start
5175  *
5176  * This re-enables the recording to the buffer, and frees the
5177  * iterator.
5178  */
5179 void
5180 ring_buffer_read_finish(struct ring_buffer_iter *iter)
5181 {
5182 	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
5183 	unsigned long flags;
5184 
5185 	/*
5186 	 * Ring buffer is disabled from recording, here's a good place
5187 	 * to check the integrity of the ring buffer.
5188 	 * Must prevent readers from trying to read, as the check
5189 	 * clears the HEAD page and readers require it.
5190 	 */
5191 	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5192 	rb_check_pages(cpu_buffer);
5193 	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5194 
5195 	atomic_dec(&cpu_buffer->resize_disabled);
5196 	kfree(iter->event);
5197 	kfree(iter);
5198 }
5199 EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
5200 
5201 /**
5202  * ring_buffer_iter_advance - advance the iterator to the next location
5203  * @iter: The ring buffer iterator
5204  *
5205  * Move the location of the iterator such that the next read will
5206  * be the next location of the iterator.
5207  */
5208 void ring_buffer_iter_advance(struct ring_buffer_iter *iter)
5209 {
5210 	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
5211 	unsigned long flags;
5212 
5213 	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5214 
5215 	rb_advance_iter(iter);
5216 
5217 	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5218 }
5219 EXPORT_SYMBOL_GPL(ring_buffer_iter_advance);
5220 
5221 /**
5222  * ring_buffer_size - return the size of the ring buffer (in bytes)
5223  * @buffer: The ring buffer.
5224  * @cpu: The CPU to get ring buffer size from.
5225  */
5226 unsigned long ring_buffer_size(struct trace_buffer *buffer, int cpu)
5227 {
5228 	/*
5229 	 * Earlier, this method returned
5230 	 *	BUF_PAGE_SIZE * buffer->nr_pages
5231 	 * Since the nr_pages field is now removed, we have converted this to
5232 	 * return the per cpu buffer value.
5233 	 */
5234 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
5235 		return 0;
5236 
5237 	return BUF_PAGE_SIZE * buffer->buffers[cpu]->nr_pages;
5238 }
5239 EXPORT_SYMBOL_GPL(ring_buffer_size);
5240 
5241 static void rb_clear_buffer_page(struct buffer_page *page)
5242 {
5243 	local_set(&page->write, 0);
5244 	local_set(&page->entries, 0);
5245 	rb_init_page(page->page);
5246 	page->read = 0;
5247 }
5248 
5249 static void
5250 rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
5251 {
5252 	struct buffer_page *page;
5253 
5254 	rb_head_page_deactivate(cpu_buffer);
5255 
5256 	cpu_buffer->head_page
5257 		= list_entry(cpu_buffer->pages, struct buffer_page, list);
5258 	rb_clear_buffer_page(cpu_buffer->head_page);
5259 	list_for_each_entry(page, cpu_buffer->pages, list) {
5260 		rb_clear_buffer_page(page);
5261 	}
5262 
5263 	cpu_buffer->tail_page = cpu_buffer->head_page;
5264 	cpu_buffer->commit_page = cpu_buffer->head_page;
5265 
5266 	INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
5267 	INIT_LIST_HEAD(&cpu_buffer->new_pages);
5268 	rb_clear_buffer_page(cpu_buffer->reader_page);
5269 
5270 	local_set(&cpu_buffer->entries_bytes, 0);
5271 	local_set(&cpu_buffer->overrun, 0);
5272 	local_set(&cpu_buffer->commit_overrun, 0);
5273 	local_set(&cpu_buffer->dropped_events, 0);
5274 	local_set(&cpu_buffer->entries, 0);
5275 	local_set(&cpu_buffer->committing, 0);
5276 	local_set(&cpu_buffer->commits, 0);
5277 	local_set(&cpu_buffer->pages_touched, 0);
5278 	local_set(&cpu_buffer->pages_lost, 0);
5279 	local_set(&cpu_buffer->pages_read, 0);
5280 	cpu_buffer->last_pages_touch = 0;
5281 	cpu_buffer->shortest_full = 0;
5282 	cpu_buffer->read = 0;
5283 	cpu_buffer->read_bytes = 0;
5284 
5285 	rb_time_set(&cpu_buffer->write_stamp, 0);
5286 	rb_time_set(&cpu_buffer->before_stamp, 0);
5287 
5288 	memset(cpu_buffer->event_stamp, 0, sizeof(cpu_buffer->event_stamp));
5289 
5290 	cpu_buffer->lost_events = 0;
5291 	cpu_buffer->last_overrun = 0;
5292 
5293 	rb_head_page_activate(cpu_buffer);
5294 	cpu_buffer->pages_removed = 0;
5295 }
5296 
5297 /* Must have disabled the cpu buffer then done a synchronize_rcu */
5298 static void reset_disabled_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
5299 {
5300 	unsigned long flags;
5301 
5302 	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5303 
5304 	if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
5305 		goto out;
5306 
5307 	arch_spin_lock(&cpu_buffer->lock);
5308 
5309 	rb_reset_cpu(cpu_buffer);
5310 
5311 	arch_spin_unlock(&cpu_buffer->lock);
5312 
5313  out:
5314 	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5315 }
5316 
5317 /**
5318  * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
5319  * @buffer: The ring buffer to reset a per cpu buffer of
5320  * @cpu: The CPU buffer to be reset
5321  */
5322 void ring_buffer_reset_cpu(struct trace_buffer *buffer, int cpu)
5323 {
5324 	struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
5325 
5326 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
5327 		return;
5328 
5329 	/* prevent another thread from changing buffer sizes */
5330 	mutex_lock(&buffer->mutex);
5331 
5332 	atomic_inc(&cpu_buffer->resize_disabled);
5333 	atomic_inc(&cpu_buffer->record_disabled);
5334 
5335 	/* Make sure all commits have finished */
5336 	synchronize_rcu();
5337 
5338 	reset_disabled_cpu_buffer(cpu_buffer);
5339 
5340 	atomic_dec(&cpu_buffer->record_disabled);
5341 	atomic_dec(&cpu_buffer->resize_disabled);
5342 
5343 	mutex_unlock(&buffer->mutex);
5344 }
5345 EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
5346 
5347 /* Flag to ensure proper resetting of atomic variables */
5348 #define RESET_BIT	(1 << 30)
5349 
5350 /**
5351  * ring_buffer_reset_online_cpus - reset a ring buffer per CPU buffer
5352  * @buffer: The ring buffer to reset a per cpu buffer of
5353  */
5354 void ring_buffer_reset_online_cpus(struct trace_buffer *buffer)
5355 {
5356 	struct ring_buffer_per_cpu *cpu_buffer;
5357 	int cpu;
5358 
5359 	/* prevent another thread from changing buffer sizes */
5360 	mutex_lock(&buffer->mutex);
5361 
5362 	for_each_online_buffer_cpu(buffer, cpu) {
5363 		cpu_buffer = buffer->buffers[cpu];
5364 
5365 		atomic_add(RESET_BIT, &cpu_buffer->resize_disabled);
5366 		atomic_inc(&cpu_buffer->record_disabled);
5367 	}
5368 
5369 	/* Make sure all commits have finished */
5370 	synchronize_rcu();
5371 
5372 	for_each_buffer_cpu(buffer, cpu) {
5373 		cpu_buffer = buffer->buffers[cpu];
5374 
5375 		/*
5376 		 * If a CPU came online during the synchronize_rcu(), then
5377 		 * ignore it.
5378 		 */
5379 		if (!(atomic_read(&cpu_buffer->resize_disabled) & RESET_BIT))
5380 			continue;
5381 
5382 		reset_disabled_cpu_buffer(cpu_buffer);
5383 
5384 		atomic_dec(&cpu_buffer->record_disabled);
5385 		atomic_sub(RESET_BIT, &cpu_buffer->resize_disabled);
5386 	}
5387 
5388 	mutex_unlock(&buffer->mutex);
5389 }
5390 
5391 /**
5392  * ring_buffer_reset - reset a ring buffer
5393  * @buffer: The ring buffer to reset all cpu buffers
5394  */
5395 void ring_buffer_reset(struct trace_buffer *buffer)
5396 {
5397 	struct ring_buffer_per_cpu *cpu_buffer;
5398 	int cpu;
5399 
5400 	/* prevent another thread from changing buffer sizes */
5401 	mutex_lock(&buffer->mutex);
5402 
5403 	for_each_buffer_cpu(buffer, cpu) {
5404 		cpu_buffer = buffer->buffers[cpu];
5405 
5406 		atomic_inc(&cpu_buffer->resize_disabled);
5407 		atomic_inc(&cpu_buffer->record_disabled);
5408 	}
5409 
5410 	/* Make sure all commits have finished */
5411 	synchronize_rcu();
5412 
5413 	for_each_buffer_cpu(buffer, cpu) {
5414 		cpu_buffer = buffer->buffers[cpu];
5415 
5416 		reset_disabled_cpu_buffer(cpu_buffer);
5417 
5418 		atomic_dec(&cpu_buffer->record_disabled);
5419 		atomic_dec(&cpu_buffer->resize_disabled);
5420 	}
5421 
5422 	mutex_unlock(&buffer->mutex);
5423 }
5424 EXPORT_SYMBOL_GPL(ring_buffer_reset);
5425 
5426 /**
5427  * ring_buffer_empty - is the ring buffer empty?
5428  * @buffer: The ring buffer to test
5429  */
5430 bool ring_buffer_empty(struct trace_buffer *buffer)
5431 {
5432 	struct ring_buffer_per_cpu *cpu_buffer;
5433 	unsigned long flags;
5434 	bool dolock;
5435 	bool ret;
5436 	int cpu;
5437 
5438 	/* yes this is racy, but if you don't like the race, lock the buffer */
5439 	for_each_buffer_cpu(buffer, cpu) {
5440 		cpu_buffer = buffer->buffers[cpu];
5441 		local_irq_save(flags);
5442 		dolock = rb_reader_lock(cpu_buffer);
5443 		ret = rb_per_cpu_empty(cpu_buffer);
5444 		rb_reader_unlock(cpu_buffer, dolock);
5445 		local_irq_restore(flags);
5446 
5447 		if (!ret)
5448 			return false;
5449 	}
5450 
5451 	return true;
5452 }
5453 EXPORT_SYMBOL_GPL(ring_buffer_empty);
5454 
5455 /**
5456  * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
5457  * @buffer: The ring buffer
5458  * @cpu: The CPU buffer to test
5459  */
5460 bool ring_buffer_empty_cpu(struct trace_buffer *buffer, int cpu)
5461 {
5462 	struct ring_buffer_per_cpu *cpu_buffer;
5463 	unsigned long flags;
5464 	bool dolock;
5465 	bool ret;
5466 
5467 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
5468 		return true;
5469 
5470 	cpu_buffer = buffer->buffers[cpu];
5471 	local_irq_save(flags);
5472 	dolock = rb_reader_lock(cpu_buffer);
5473 	ret = rb_per_cpu_empty(cpu_buffer);
5474 	rb_reader_unlock(cpu_buffer, dolock);
5475 	local_irq_restore(flags);
5476 
5477 	return ret;
5478 }
5479 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
5480 
5481 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
5482 /**
5483  * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
5484  * @buffer_a: One buffer to swap with
5485  * @buffer_b: The other buffer to swap with
5486  * @cpu: the CPU of the buffers to swap
5487  *
5488  * This function is useful for tracers that want to take a "snapshot"
5489  * of a CPU buffer and has another back up buffer lying around.
5490  * it is expected that the tracer handles the cpu buffer not being
5491  * used at the moment.
5492  */
5493 int ring_buffer_swap_cpu(struct trace_buffer *buffer_a,
5494 			 struct trace_buffer *buffer_b, int cpu)
5495 {
5496 	struct ring_buffer_per_cpu *cpu_buffer_a;
5497 	struct ring_buffer_per_cpu *cpu_buffer_b;
5498 	int ret = -EINVAL;
5499 
5500 	if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
5501 	    !cpumask_test_cpu(cpu, buffer_b->cpumask))
5502 		goto out;
5503 
5504 	cpu_buffer_a = buffer_a->buffers[cpu];
5505 	cpu_buffer_b = buffer_b->buffers[cpu];
5506 
5507 	/* At least make sure the two buffers are somewhat the same */
5508 	if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
5509 		goto out;
5510 
5511 	ret = -EAGAIN;
5512 
5513 	if (atomic_read(&buffer_a->record_disabled))
5514 		goto out;
5515 
5516 	if (atomic_read(&buffer_b->record_disabled))
5517 		goto out;
5518 
5519 	if (atomic_read(&cpu_buffer_a->record_disabled))
5520 		goto out;
5521 
5522 	if (atomic_read(&cpu_buffer_b->record_disabled))
5523 		goto out;
5524 
5525 	/*
5526 	 * We can't do a synchronize_rcu here because this
5527 	 * function can be called in atomic context.
5528 	 * Normally this will be called from the same CPU as cpu.
5529 	 * If not it's up to the caller to protect this.
5530 	 */
5531 	atomic_inc(&cpu_buffer_a->record_disabled);
5532 	atomic_inc(&cpu_buffer_b->record_disabled);
5533 
5534 	ret = -EBUSY;
5535 	if (local_read(&cpu_buffer_a->committing))
5536 		goto out_dec;
5537 	if (local_read(&cpu_buffer_b->committing))
5538 		goto out_dec;
5539 
5540 	/*
5541 	 * When resize is in progress, we cannot swap it because
5542 	 * it will mess the state of the cpu buffer.
5543 	 */
5544 	if (atomic_read(&buffer_a->resizing))
5545 		goto out_dec;
5546 	if (atomic_read(&buffer_b->resizing))
5547 		goto out_dec;
5548 
5549 	buffer_a->buffers[cpu] = cpu_buffer_b;
5550 	buffer_b->buffers[cpu] = cpu_buffer_a;
5551 
5552 	cpu_buffer_b->buffer = buffer_a;
5553 	cpu_buffer_a->buffer = buffer_b;
5554 
5555 	ret = 0;
5556 
5557 out_dec:
5558 	atomic_dec(&cpu_buffer_a->record_disabled);
5559 	atomic_dec(&cpu_buffer_b->record_disabled);
5560 out:
5561 	return ret;
5562 }
5563 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
5564 #endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
5565 
5566 /**
5567  * ring_buffer_alloc_read_page - allocate a page to read from buffer
5568  * @buffer: the buffer to allocate for.
5569  * @cpu: the cpu buffer to allocate.
5570  *
5571  * This function is used in conjunction with ring_buffer_read_page.
5572  * When reading a full page from the ring buffer, these functions
5573  * can be used to speed up the process. The calling function should
5574  * allocate a few pages first with this function. Then when it
5575  * needs to get pages from the ring buffer, it passes the result
5576  * of this function into ring_buffer_read_page, which will swap
5577  * the page that was allocated, with the read page of the buffer.
5578  *
5579  * Returns:
5580  *  The page allocated, or ERR_PTR
5581  */
5582 void *ring_buffer_alloc_read_page(struct trace_buffer *buffer, int cpu)
5583 {
5584 	struct ring_buffer_per_cpu *cpu_buffer;
5585 	struct buffer_data_page *bpage = NULL;
5586 	unsigned long flags;
5587 	struct page *page;
5588 
5589 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
5590 		return ERR_PTR(-ENODEV);
5591 
5592 	cpu_buffer = buffer->buffers[cpu];
5593 	local_irq_save(flags);
5594 	arch_spin_lock(&cpu_buffer->lock);
5595 
5596 	if (cpu_buffer->free_page) {
5597 		bpage = cpu_buffer->free_page;
5598 		cpu_buffer->free_page = NULL;
5599 	}
5600 
5601 	arch_spin_unlock(&cpu_buffer->lock);
5602 	local_irq_restore(flags);
5603 
5604 	if (bpage)
5605 		goto out;
5606 
5607 	page = alloc_pages_node(cpu_to_node(cpu),
5608 				GFP_KERNEL | __GFP_NORETRY, 0);
5609 	if (!page)
5610 		return ERR_PTR(-ENOMEM);
5611 
5612 	bpage = page_address(page);
5613 
5614  out:
5615 	rb_init_page(bpage);
5616 
5617 	return bpage;
5618 }
5619 EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
5620 
5621 /**
5622  * ring_buffer_free_read_page - free an allocated read page
5623  * @buffer: the buffer the page was allocate for
5624  * @cpu: the cpu buffer the page came from
5625  * @data: the page to free
5626  *
5627  * Free a page allocated from ring_buffer_alloc_read_page.
5628  */
5629 void ring_buffer_free_read_page(struct trace_buffer *buffer, int cpu, void *data)
5630 {
5631 	struct ring_buffer_per_cpu *cpu_buffer;
5632 	struct buffer_data_page *bpage = data;
5633 	struct page *page = virt_to_page(bpage);
5634 	unsigned long flags;
5635 
5636 	if (!buffer || !buffer->buffers || !buffer->buffers[cpu])
5637 		return;
5638 
5639 	cpu_buffer = buffer->buffers[cpu];
5640 
5641 	/* If the page is still in use someplace else, we can't reuse it */
5642 	if (page_ref_count(page) > 1)
5643 		goto out;
5644 
5645 	local_irq_save(flags);
5646 	arch_spin_lock(&cpu_buffer->lock);
5647 
5648 	if (!cpu_buffer->free_page) {
5649 		cpu_buffer->free_page = bpage;
5650 		bpage = NULL;
5651 	}
5652 
5653 	arch_spin_unlock(&cpu_buffer->lock);
5654 	local_irq_restore(flags);
5655 
5656  out:
5657 	free_page((unsigned long)bpage);
5658 }
5659 EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
5660 
5661 /**
5662  * ring_buffer_read_page - extract a page from the ring buffer
5663  * @buffer: buffer to extract from
5664  * @data_page: the page to use allocated from ring_buffer_alloc_read_page
5665  * @len: amount to extract
5666  * @cpu: the cpu of the buffer to extract
5667  * @full: should the extraction only happen when the page is full.
5668  *
5669  * This function will pull out a page from the ring buffer and consume it.
5670  * @data_page must be the address of the variable that was returned
5671  * from ring_buffer_alloc_read_page. This is because the page might be used
5672  * to swap with a page in the ring buffer.
5673  *
5674  * for example:
5675  *	rpage = ring_buffer_alloc_read_page(buffer, cpu);
5676  *	if (IS_ERR(rpage))
5677  *		return PTR_ERR(rpage);
5678  *	ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
5679  *	if (ret >= 0)
5680  *		process_page(rpage, ret);
5681  *
5682  * When @full is set, the function will not return true unless
5683  * the writer is off the reader page.
5684  *
5685  * Note: it is up to the calling functions to handle sleeps and wakeups.
5686  *  The ring buffer can be used anywhere in the kernel and can not
5687  *  blindly call wake_up. The layer that uses the ring buffer must be
5688  *  responsible for that.
5689  *
5690  * Returns:
5691  *  >=0 if data has been transferred, returns the offset of consumed data.
5692  *  <0 if no data has been transferred.
5693  */
5694 int ring_buffer_read_page(struct trace_buffer *buffer,
5695 			  void **data_page, size_t len, int cpu, int full)
5696 {
5697 	struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
5698 	struct ring_buffer_event *event;
5699 	struct buffer_data_page *bpage;
5700 	struct buffer_page *reader;
5701 	unsigned long missed_events;
5702 	unsigned long flags;
5703 	unsigned int commit;
5704 	unsigned int read;
5705 	u64 save_timestamp;
5706 	int ret = -1;
5707 
5708 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
5709 		goto out;
5710 
5711 	/*
5712 	 * If len is not big enough to hold the page header, then
5713 	 * we can not copy anything.
5714 	 */
5715 	if (len <= BUF_PAGE_HDR_SIZE)
5716 		goto out;
5717 
5718 	len -= BUF_PAGE_HDR_SIZE;
5719 
5720 	if (!data_page)
5721 		goto out;
5722 
5723 	bpage = *data_page;
5724 	if (!bpage)
5725 		goto out;
5726 
5727 	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5728 
5729 	reader = rb_get_reader_page(cpu_buffer);
5730 	if (!reader)
5731 		goto out_unlock;
5732 
5733 	event = rb_reader_event(cpu_buffer);
5734 
5735 	read = reader->read;
5736 	commit = rb_page_commit(reader);
5737 
5738 	/* Check if any events were dropped */
5739 	missed_events = cpu_buffer->lost_events;
5740 
5741 	/*
5742 	 * If this page has been partially read or
5743 	 * if len is not big enough to read the rest of the page or
5744 	 * a writer is still on the page, then
5745 	 * we must copy the data from the page to the buffer.
5746 	 * Otherwise, we can simply swap the page with the one passed in.
5747 	 */
5748 	if (read || (len < (commit - read)) ||
5749 	    cpu_buffer->reader_page == cpu_buffer->commit_page) {
5750 		struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
5751 		unsigned int rpos = read;
5752 		unsigned int pos = 0;
5753 		unsigned int size;
5754 
5755 		/*
5756 		 * If a full page is expected, this can still be returned
5757 		 * if there's been a previous partial read and the
5758 		 * rest of the page can be read and the commit page is off
5759 		 * the reader page.
5760 		 */
5761 		if (full &&
5762 		    (!read || (len < (commit - read)) ||
5763 		     cpu_buffer->reader_page == cpu_buffer->commit_page))
5764 			goto out_unlock;
5765 
5766 		if (len > (commit - read))
5767 			len = (commit - read);
5768 
5769 		/* Always keep the time extend and data together */
5770 		size = rb_event_ts_length(event);
5771 
5772 		if (len < size)
5773 			goto out_unlock;
5774 
5775 		/* save the current timestamp, since the user will need it */
5776 		save_timestamp = cpu_buffer->read_stamp;
5777 
5778 		/* Need to copy one event at a time */
5779 		do {
5780 			/* We need the size of one event, because
5781 			 * rb_advance_reader only advances by one event,
5782 			 * whereas rb_event_ts_length may include the size of
5783 			 * one or two events.
5784 			 * We have already ensured there's enough space if this
5785 			 * is a time extend. */
5786 			size = rb_event_length(event);
5787 			memcpy(bpage->data + pos, rpage->data + rpos, size);
5788 
5789 			len -= size;
5790 
5791 			rb_advance_reader(cpu_buffer);
5792 			rpos = reader->read;
5793 			pos += size;
5794 
5795 			if (rpos >= commit)
5796 				break;
5797 
5798 			event = rb_reader_event(cpu_buffer);
5799 			/* Always keep the time extend and data together */
5800 			size = rb_event_ts_length(event);
5801 		} while (len >= size);
5802 
5803 		/* update bpage */
5804 		local_set(&bpage->commit, pos);
5805 		bpage->time_stamp = save_timestamp;
5806 
5807 		/* we copied everything to the beginning */
5808 		read = 0;
5809 	} else {
5810 		/* update the entry counter */
5811 		cpu_buffer->read += rb_page_entries(reader);
5812 		cpu_buffer->read_bytes += BUF_PAGE_SIZE;
5813 
5814 		/* swap the pages */
5815 		rb_init_page(bpage);
5816 		bpage = reader->page;
5817 		reader->page = *data_page;
5818 		local_set(&reader->write, 0);
5819 		local_set(&reader->entries, 0);
5820 		reader->read = 0;
5821 		*data_page = bpage;
5822 
5823 		/*
5824 		 * Use the real_end for the data size,
5825 		 * This gives us a chance to store the lost events
5826 		 * on the page.
5827 		 */
5828 		if (reader->real_end)
5829 			local_set(&bpage->commit, reader->real_end);
5830 	}
5831 	ret = read;
5832 
5833 	cpu_buffer->lost_events = 0;
5834 
5835 	commit = local_read(&bpage->commit);
5836 	/*
5837 	 * Set a flag in the commit field if we lost events
5838 	 */
5839 	if (missed_events) {
5840 		/* If there is room at the end of the page to save the
5841 		 * missed events, then record it there.
5842 		 */
5843 		if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
5844 			memcpy(&bpage->data[commit], &missed_events,
5845 			       sizeof(missed_events));
5846 			local_add(RB_MISSED_STORED, &bpage->commit);
5847 			commit += sizeof(missed_events);
5848 		}
5849 		local_add(RB_MISSED_EVENTS, &bpage->commit);
5850 	}
5851 
5852 	/*
5853 	 * This page may be off to user land. Zero it out here.
5854 	 */
5855 	if (commit < BUF_PAGE_SIZE)
5856 		memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
5857 
5858  out_unlock:
5859 	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5860 
5861  out:
5862 	return ret;
5863 }
5864 EXPORT_SYMBOL_GPL(ring_buffer_read_page);
5865 
5866 /*
5867  * We only allocate new buffers, never free them if the CPU goes down.
5868  * If we were to free the buffer, then the user would lose any trace that was in
5869  * the buffer.
5870  */
5871 int trace_rb_cpu_prepare(unsigned int cpu, struct hlist_node *node)
5872 {
5873 	struct trace_buffer *buffer;
5874 	long nr_pages_same;
5875 	int cpu_i;
5876 	unsigned long nr_pages;
5877 
5878 	buffer = container_of(node, struct trace_buffer, node);
5879 	if (cpumask_test_cpu(cpu, buffer->cpumask))
5880 		return 0;
5881 
5882 	nr_pages = 0;
5883 	nr_pages_same = 1;
5884 	/* check if all cpu sizes are same */
5885 	for_each_buffer_cpu(buffer, cpu_i) {
5886 		/* fill in the size from first enabled cpu */
5887 		if (nr_pages == 0)
5888 			nr_pages = buffer->buffers[cpu_i]->nr_pages;
5889 		if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
5890 			nr_pages_same = 0;
5891 			break;
5892 		}
5893 	}
5894 	/* allocate minimum pages, user can later expand it */
5895 	if (!nr_pages_same)
5896 		nr_pages = 2;
5897 	buffer->buffers[cpu] =
5898 		rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
5899 	if (!buffer->buffers[cpu]) {
5900 		WARN(1, "failed to allocate ring buffer on CPU %u\n",
5901 		     cpu);
5902 		return -ENOMEM;
5903 	}
5904 	smp_wmb();
5905 	cpumask_set_cpu(cpu, buffer->cpumask);
5906 	return 0;
5907 }
5908 
5909 #ifdef CONFIG_RING_BUFFER_STARTUP_TEST
5910 /*
5911  * This is a basic integrity check of the ring buffer.
5912  * Late in the boot cycle this test will run when configured in.
5913  * It will kick off a thread per CPU that will go into a loop
5914  * writing to the per cpu ring buffer various sizes of data.
5915  * Some of the data will be large items, some small.
5916  *
5917  * Another thread is created that goes into a spin, sending out
5918  * IPIs to the other CPUs to also write into the ring buffer.
5919  * this is to test the nesting ability of the buffer.
5920  *
5921  * Basic stats are recorded and reported. If something in the
5922  * ring buffer should happen that's not expected, a big warning
5923  * is displayed and all ring buffers are disabled.
5924  */
5925 static struct task_struct *rb_threads[NR_CPUS] __initdata;
5926 
5927 struct rb_test_data {
5928 	struct trace_buffer *buffer;
5929 	unsigned long		events;
5930 	unsigned long		bytes_written;
5931 	unsigned long		bytes_alloc;
5932 	unsigned long		bytes_dropped;
5933 	unsigned long		events_nested;
5934 	unsigned long		bytes_written_nested;
5935 	unsigned long		bytes_alloc_nested;
5936 	unsigned long		bytes_dropped_nested;
5937 	int			min_size_nested;
5938 	int			max_size_nested;
5939 	int			max_size;
5940 	int			min_size;
5941 	int			cpu;
5942 	int			cnt;
5943 };
5944 
5945 static struct rb_test_data rb_data[NR_CPUS] __initdata;
5946 
5947 /* 1 meg per cpu */
5948 #define RB_TEST_BUFFER_SIZE	1048576
5949 
5950 static char rb_string[] __initdata =
5951 	"abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
5952 	"?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
5953 	"!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
5954 
5955 static bool rb_test_started __initdata;
5956 
5957 struct rb_item {
5958 	int size;
5959 	char str[];
5960 };
5961 
5962 static __init int rb_write_something(struct rb_test_data *data, bool nested)
5963 {
5964 	struct ring_buffer_event *event;
5965 	struct rb_item *item;
5966 	bool started;
5967 	int event_len;
5968 	int size;
5969 	int len;
5970 	int cnt;
5971 
5972 	/* Have nested writes different that what is written */
5973 	cnt = data->cnt + (nested ? 27 : 0);
5974 
5975 	/* Multiply cnt by ~e, to make some unique increment */
5976 	size = (cnt * 68 / 25) % (sizeof(rb_string) - 1);
5977 
5978 	len = size + sizeof(struct rb_item);
5979 
5980 	started = rb_test_started;
5981 	/* read rb_test_started before checking buffer enabled */
5982 	smp_rmb();
5983 
5984 	event = ring_buffer_lock_reserve(data->buffer, len);
5985 	if (!event) {
5986 		/* Ignore dropped events before test starts. */
5987 		if (started) {
5988 			if (nested)
5989 				data->bytes_dropped += len;
5990 			else
5991 				data->bytes_dropped_nested += len;
5992 		}
5993 		return len;
5994 	}
5995 
5996 	event_len = ring_buffer_event_length(event);
5997 
5998 	if (RB_WARN_ON(data->buffer, event_len < len))
5999 		goto out;
6000 
6001 	item = ring_buffer_event_data(event);
6002 	item->size = size;
6003 	memcpy(item->str, rb_string, size);
6004 
6005 	if (nested) {
6006 		data->bytes_alloc_nested += event_len;
6007 		data->bytes_written_nested += len;
6008 		data->events_nested++;
6009 		if (!data->min_size_nested || len < data->min_size_nested)
6010 			data->min_size_nested = len;
6011 		if (len > data->max_size_nested)
6012 			data->max_size_nested = len;
6013 	} else {
6014 		data->bytes_alloc += event_len;
6015 		data->bytes_written += len;
6016 		data->events++;
6017 		if (!data->min_size || len < data->min_size)
6018 			data->max_size = len;
6019 		if (len > data->max_size)
6020 			data->max_size = len;
6021 	}
6022 
6023  out:
6024 	ring_buffer_unlock_commit(data->buffer);
6025 
6026 	return 0;
6027 }
6028 
6029 static __init int rb_test(void *arg)
6030 {
6031 	struct rb_test_data *data = arg;
6032 
6033 	while (!kthread_should_stop()) {
6034 		rb_write_something(data, false);
6035 		data->cnt++;
6036 
6037 		set_current_state(TASK_INTERRUPTIBLE);
6038 		/* Now sleep between a min of 100-300us and a max of 1ms */
6039 		usleep_range(((data->cnt % 3) + 1) * 100, 1000);
6040 	}
6041 
6042 	return 0;
6043 }
6044 
6045 static __init void rb_ipi(void *ignore)
6046 {
6047 	struct rb_test_data *data;
6048 	int cpu = smp_processor_id();
6049 
6050 	data = &rb_data[cpu];
6051 	rb_write_something(data, true);
6052 }
6053 
6054 static __init int rb_hammer_test(void *arg)
6055 {
6056 	while (!kthread_should_stop()) {
6057 
6058 		/* Send an IPI to all cpus to write data! */
6059 		smp_call_function(rb_ipi, NULL, 1);
6060 		/* No sleep, but for non preempt, let others run */
6061 		schedule();
6062 	}
6063 
6064 	return 0;
6065 }
6066 
6067 static __init int test_ringbuffer(void)
6068 {
6069 	struct task_struct *rb_hammer;
6070 	struct trace_buffer *buffer;
6071 	int cpu;
6072 	int ret = 0;
6073 
6074 	if (security_locked_down(LOCKDOWN_TRACEFS)) {
6075 		pr_warn("Lockdown is enabled, skipping ring buffer tests\n");
6076 		return 0;
6077 	}
6078 
6079 	pr_info("Running ring buffer tests...\n");
6080 
6081 	buffer = ring_buffer_alloc(RB_TEST_BUFFER_SIZE, RB_FL_OVERWRITE);
6082 	if (WARN_ON(!buffer))
6083 		return 0;
6084 
6085 	/* Disable buffer so that threads can't write to it yet */
6086 	ring_buffer_record_off(buffer);
6087 
6088 	for_each_online_cpu(cpu) {
6089 		rb_data[cpu].buffer = buffer;
6090 		rb_data[cpu].cpu = cpu;
6091 		rb_data[cpu].cnt = cpu;
6092 		rb_threads[cpu] = kthread_run_on_cpu(rb_test, &rb_data[cpu],
6093 						     cpu, "rbtester/%u");
6094 		if (WARN_ON(IS_ERR(rb_threads[cpu]))) {
6095 			pr_cont("FAILED\n");
6096 			ret = PTR_ERR(rb_threads[cpu]);
6097 			goto out_free;
6098 		}
6099 	}
6100 
6101 	/* Now create the rb hammer! */
6102 	rb_hammer = kthread_run(rb_hammer_test, NULL, "rbhammer");
6103 	if (WARN_ON(IS_ERR(rb_hammer))) {
6104 		pr_cont("FAILED\n");
6105 		ret = PTR_ERR(rb_hammer);
6106 		goto out_free;
6107 	}
6108 
6109 	ring_buffer_record_on(buffer);
6110 	/*
6111 	 * Show buffer is enabled before setting rb_test_started.
6112 	 * Yes there's a small race window where events could be
6113 	 * dropped and the thread wont catch it. But when a ring
6114 	 * buffer gets enabled, there will always be some kind of
6115 	 * delay before other CPUs see it. Thus, we don't care about
6116 	 * those dropped events. We care about events dropped after
6117 	 * the threads see that the buffer is active.
6118 	 */
6119 	smp_wmb();
6120 	rb_test_started = true;
6121 
6122 	set_current_state(TASK_INTERRUPTIBLE);
6123 	/* Just run for 10 seconds */;
6124 	schedule_timeout(10 * HZ);
6125 
6126 	kthread_stop(rb_hammer);
6127 
6128  out_free:
6129 	for_each_online_cpu(cpu) {
6130 		if (!rb_threads[cpu])
6131 			break;
6132 		kthread_stop(rb_threads[cpu]);
6133 	}
6134 	if (ret) {
6135 		ring_buffer_free(buffer);
6136 		return ret;
6137 	}
6138 
6139 	/* Report! */
6140 	pr_info("finished\n");
6141 	for_each_online_cpu(cpu) {
6142 		struct ring_buffer_event *event;
6143 		struct rb_test_data *data = &rb_data[cpu];
6144 		struct rb_item *item;
6145 		unsigned long total_events;
6146 		unsigned long total_dropped;
6147 		unsigned long total_written;
6148 		unsigned long total_alloc;
6149 		unsigned long total_read = 0;
6150 		unsigned long total_size = 0;
6151 		unsigned long total_len = 0;
6152 		unsigned long total_lost = 0;
6153 		unsigned long lost;
6154 		int big_event_size;
6155 		int small_event_size;
6156 
6157 		ret = -1;
6158 
6159 		total_events = data->events + data->events_nested;
6160 		total_written = data->bytes_written + data->bytes_written_nested;
6161 		total_alloc = data->bytes_alloc + data->bytes_alloc_nested;
6162 		total_dropped = data->bytes_dropped + data->bytes_dropped_nested;
6163 
6164 		big_event_size = data->max_size + data->max_size_nested;
6165 		small_event_size = data->min_size + data->min_size_nested;
6166 
6167 		pr_info("CPU %d:\n", cpu);
6168 		pr_info("              events:    %ld\n", total_events);
6169 		pr_info("       dropped bytes:    %ld\n", total_dropped);
6170 		pr_info("       alloced bytes:    %ld\n", total_alloc);
6171 		pr_info("       written bytes:    %ld\n", total_written);
6172 		pr_info("       biggest event:    %d\n", big_event_size);
6173 		pr_info("      smallest event:    %d\n", small_event_size);
6174 
6175 		if (RB_WARN_ON(buffer, total_dropped))
6176 			break;
6177 
6178 		ret = 0;
6179 
6180 		while ((event = ring_buffer_consume(buffer, cpu, NULL, &lost))) {
6181 			total_lost += lost;
6182 			item = ring_buffer_event_data(event);
6183 			total_len += ring_buffer_event_length(event);
6184 			total_size += item->size + sizeof(struct rb_item);
6185 			if (memcmp(&item->str[0], rb_string, item->size) != 0) {
6186 				pr_info("FAILED!\n");
6187 				pr_info("buffer had: %.*s\n", item->size, item->str);
6188 				pr_info("expected:   %.*s\n", item->size, rb_string);
6189 				RB_WARN_ON(buffer, 1);
6190 				ret = -1;
6191 				break;
6192 			}
6193 			total_read++;
6194 		}
6195 		if (ret)
6196 			break;
6197 
6198 		ret = -1;
6199 
6200 		pr_info("         read events:   %ld\n", total_read);
6201 		pr_info("         lost events:   %ld\n", total_lost);
6202 		pr_info("        total events:   %ld\n", total_lost + total_read);
6203 		pr_info("  recorded len bytes:   %ld\n", total_len);
6204 		pr_info(" recorded size bytes:   %ld\n", total_size);
6205 		if (total_lost) {
6206 			pr_info(" With dropped events, record len and size may not match\n"
6207 				" alloced and written from above\n");
6208 		} else {
6209 			if (RB_WARN_ON(buffer, total_len != total_alloc ||
6210 				       total_size != total_written))
6211 				break;
6212 		}
6213 		if (RB_WARN_ON(buffer, total_lost + total_read != total_events))
6214 			break;
6215 
6216 		ret = 0;
6217 	}
6218 	if (!ret)
6219 		pr_info("Ring buffer PASSED!\n");
6220 
6221 	ring_buffer_free(buffer);
6222 	return 0;
6223 }
6224 
6225 late_initcall(test_ringbuffer);
6226 #endif /* CONFIG_RING_BUFFER_STARTUP_TEST */
6227