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