• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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_events.h>
8 #include <linux/ring_buffer.h>
9 #include <linux/trace_clock.h>
10 #include <linux/sched/clock.h>
11 #include <linux/trace_seq.h>
12 #include <linux/spinlock.h>
13 #include <linux/irq_work.h>
14 #include <linux/security.h>
15 #include <linux/uaccess.h>
16 #include <linux/hardirq.h>
17 #include <linux/kthread.h>	/* for self test */
18 #include <linux/module.h>
19 #include <linux/percpu.h>
20 #include <linux/mutex.h>
21 #include <linux/delay.h>
22 #include <linux/slab.h>
23 #include <linux/init.h>
24 #include <linux/hash.h>
25 #include <linux/list.h>
26 #include <linux/cpu.h>
27 #include <linux/oom.h>
28 
29 #include <asm/local.h>
30 
31 static void update_pages_handler(struct work_struct *work);
32 
33 /*
34  * The ring buffer header is special. We must manually up keep it.
35  */
ring_buffer_print_entry_header(struct trace_seq * s)36 int ring_buffer_print_entry_header(struct trace_seq *s)
37 {
38 	trace_seq_puts(s, "# compressed entry header\n");
39 	trace_seq_puts(s, "\ttype_len    :    5 bits\n");
40 	trace_seq_puts(s, "\ttime_delta  :   27 bits\n");
41 	trace_seq_puts(s, "\tarray       :   32 bits\n");
42 	trace_seq_putc(s, '\n');
43 	trace_seq_printf(s, "\tpadding     : type == %d\n",
44 			 RINGBUF_TYPE_PADDING);
45 	trace_seq_printf(s, "\ttime_extend : type == %d\n",
46 			 RINGBUF_TYPE_TIME_EXTEND);
47 	trace_seq_printf(s, "\ttime_stamp : type == %d\n",
48 			 RINGBUF_TYPE_TIME_STAMP);
49 	trace_seq_printf(s, "\tdata max type_len  == %d\n",
50 			 RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
51 
52 	return !trace_seq_has_overflowed(s);
53 }
54 
55 /*
56  * The ring buffer is made up of a list of pages. A separate list of pages is
57  * allocated for each CPU. A writer may only write to a buffer that is
58  * associated with the CPU it is currently executing on.  A reader may read
59  * from any per cpu buffer.
60  *
61  * The reader is special. For each per cpu buffer, the reader has its own
62  * reader page. When a reader has read the entire reader page, this reader
63  * page is swapped with another page in the ring buffer.
64  *
65  * Now, as long as the writer is off the reader page, the reader can do what
66  * ever it wants with that page. The writer will never write to that page
67  * again (as long as it is out of the ring buffer).
68  *
69  * Here's some silly ASCII art.
70  *
71  *   +------+
72  *   |reader|          RING BUFFER
73  *   |page  |
74  *   +------+        +---+   +---+   +---+
75  *                   |   |-->|   |-->|   |
76  *                   +---+   +---+   +---+
77  *                     ^               |
78  *                     |               |
79  *                     +---------------+
80  *
81  *
82  *   +------+
83  *   |reader|          RING BUFFER
84  *   |page  |------------------v
85  *   +------+        +---+   +---+   +---+
86  *                   |   |-->|   |-->|   |
87  *                   +---+   +---+   +---+
88  *                     ^               |
89  *                     |               |
90  *                     +---------------+
91  *
92  *
93  *   +------+
94  *   |reader|          RING BUFFER
95  *   |page  |------------------v
96  *   +------+        +---+   +---+   +---+
97  *      ^            |   |-->|   |-->|   |
98  *      |            +---+   +---+   +---+
99  *      |                              |
100  *      |                              |
101  *      +------------------------------+
102  *
103  *
104  *   +------+
105  *   |buffer|          RING BUFFER
106  *   |page  |------------------v
107  *   +------+        +---+   +---+   +---+
108  *      ^            |   |   |   |-->|   |
109  *      |   New      +---+   +---+   +---+
110  *      |  Reader------^               |
111  *      |   page                       |
112  *      +------------------------------+
113  *
114  *
115  * After we make this swap, the reader can hand this page off to the splice
116  * code and be done with it. It can even allocate a new page if it needs to
117  * and swap that into the ring buffer.
118  *
119  * We will be using cmpxchg soon to make all this lockless.
120  *
121  */
122 
123 /* Used for individual buffers (after the counter) */
124 #define RB_BUFFER_OFF		(1 << 20)
125 
126 #define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
127 
128 #define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
129 #define RB_ALIGNMENT		4U
130 #define RB_MAX_SMALL_DATA	(RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
131 #define RB_EVNT_MIN_SIZE	8U	/* two 32bit words */
132 
133 #ifndef CONFIG_HAVE_64BIT_ALIGNED_ACCESS
134 # define RB_FORCE_8BYTE_ALIGNMENT	0
135 # define RB_ARCH_ALIGNMENT		RB_ALIGNMENT
136 #else
137 # define RB_FORCE_8BYTE_ALIGNMENT	1
138 # define RB_ARCH_ALIGNMENT		8U
139 #endif
140 
141 #define RB_ALIGN_DATA		__aligned(RB_ARCH_ALIGNMENT)
142 
143 /* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
144 #define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
145 
146 enum {
147 	RB_LEN_TIME_EXTEND = 8,
148 	RB_LEN_TIME_STAMP =  8,
149 };
150 
151 #define skip_time_extend(event) \
152 	((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
153 
154 #define extended_time(event) \
155 	(event->type_len >= RINGBUF_TYPE_TIME_EXTEND)
156 
rb_null_event(struct ring_buffer_event * event)157 static inline int rb_null_event(struct ring_buffer_event *event)
158 {
159 	return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
160 }
161 
rb_event_set_padding(struct ring_buffer_event * event)162 static void rb_event_set_padding(struct ring_buffer_event *event)
163 {
164 	/* padding has a NULL time_delta */
165 	event->type_len = RINGBUF_TYPE_PADDING;
166 	event->time_delta = 0;
167 }
168 
169 static unsigned
rb_event_data_length(struct ring_buffer_event * event)170 rb_event_data_length(struct ring_buffer_event *event)
171 {
172 	unsigned length;
173 
174 	if (event->type_len)
175 		length = event->type_len * RB_ALIGNMENT;
176 	else
177 		length = event->array[0];
178 	return length + RB_EVNT_HDR_SIZE;
179 }
180 
181 /*
182  * Return the length of the given event. Will return
183  * the length of the time extend if the event is a
184  * time extend.
185  */
186 static inline unsigned
rb_event_length(struct ring_buffer_event * event)187 rb_event_length(struct ring_buffer_event *event)
188 {
189 	switch (event->type_len) {
190 	case RINGBUF_TYPE_PADDING:
191 		if (rb_null_event(event))
192 			/* undefined */
193 			return -1;
194 		return  event->array[0] + RB_EVNT_HDR_SIZE;
195 
196 	case RINGBUF_TYPE_TIME_EXTEND:
197 		return RB_LEN_TIME_EXTEND;
198 
199 	case RINGBUF_TYPE_TIME_STAMP:
200 		return RB_LEN_TIME_STAMP;
201 
202 	case RINGBUF_TYPE_DATA:
203 		return rb_event_data_length(event);
204 	default:
205 		WARN_ON_ONCE(1);
206 	}
207 	/* not hit */
208 	return 0;
209 }
210 
211 /*
212  * Return total length of time extend and data,
213  *   or just the event length for all other events.
214  */
215 static inline unsigned
rb_event_ts_length(struct ring_buffer_event * event)216 rb_event_ts_length(struct ring_buffer_event *event)
217 {
218 	unsigned len = 0;
219 
220 	if (extended_time(event)) {
221 		/* time extends include the data event after it */
222 		len = RB_LEN_TIME_EXTEND;
223 		event = skip_time_extend(event);
224 	}
225 	return len + rb_event_length(event);
226 }
227 
228 /**
229  * ring_buffer_event_length - return the length of the event
230  * @event: the event to get the length of
231  *
232  * Returns the size of the data load of a data event.
233  * If the event is something other than a data event, it
234  * returns the size of the event itself. With the exception
235  * of a TIME EXTEND, where it still returns the size of the
236  * data load of the data event after it.
237  */
ring_buffer_event_length(struct ring_buffer_event * event)238 unsigned ring_buffer_event_length(struct ring_buffer_event *event)
239 {
240 	unsigned length;
241 
242 	if (extended_time(event))
243 		event = skip_time_extend(event);
244 
245 	length = rb_event_length(event);
246 	if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
247 		return length;
248 	length -= RB_EVNT_HDR_SIZE;
249 	if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
250                 length -= sizeof(event->array[0]);
251 	return length;
252 }
253 EXPORT_SYMBOL_GPL(ring_buffer_event_length);
254 
255 /* inline for ring buffer fast paths */
256 static __always_inline void *
rb_event_data(struct ring_buffer_event * event)257 rb_event_data(struct ring_buffer_event *event)
258 {
259 	if (extended_time(event))
260 		event = skip_time_extend(event);
261 	WARN_ON_ONCE(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
262 	/* If length is in len field, then array[0] has the data */
263 	if (event->type_len)
264 		return (void *)&event->array[0];
265 	/* Otherwise length is in array[0] and array[1] has the data */
266 	return (void *)&event->array[1];
267 }
268 
269 /**
270  * ring_buffer_event_data - return the data of the event
271  * @event: the event to get the data from
272  */
ring_buffer_event_data(struct ring_buffer_event * event)273 void *ring_buffer_event_data(struct ring_buffer_event *event)
274 {
275 	return rb_event_data(event);
276 }
277 EXPORT_SYMBOL_GPL(ring_buffer_event_data);
278 
279 #define for_each_buffer_cpu(buffer, cpu)		\
280 	for_each_cpu(cpu, buffer->cpumask)
281 
282 #define for_each_online_buffer_cpu(buffer, cpu)		\
283 	for_each_cpu_and(cpu, buffer->cpumask, cpu_online_mask)
284 
285 #define TS_SHIFT	27
286 #define TS_MASK		((1ULL << TS_SHIFT) - 1)
287 #define TS_DELTA_TEST	(~TS_MASK)
288 
289 /**
290  * ring_buffer_event_time_stamp - return the event's extended timestamp
291  * @event: the event to get the timestamp of
292  *
293  * Returns the extended timestamp associated with a data event.
294  * An extended time_stamp is a 64-bit timestamp represented
295  * internally in a special way that makes the best use of space
296  * contained within a ring buffer event.  This function decodes
297  * it and maps it to a straight u64 value.
298  */
ring_buffer_event_time_stamp(struct ring_buffer_event * event)299 u64 ring_buffer_event_time_stamp(struct ring_buffer_event *event)
300 {
301 	u64 ts;
302 
303 	ts = event->array[0];
304 	ts <<= TS_SHIFT;
305 	ts += event->time_delta;
306 
307 	return ts;
308 }
309 
310 /* Flag when events were overwritten */
311 #define RB_MISSED_EVENTS	(1 << 31)
312 /* Missed count stored at end */
313 #define RB_MISSED_STORED	(1 << 30)
314 
315 struct buffer_data_page {
316 	u64		 time_stamp;	/* page time stamp */
317 	local_t		 commit;	/* write committed index */
318 	unsigned char	 data[] RB_ALIGN_DATA;	/* data of buffer page */
319 };
320 
321 /*
322  * Note, the buffer_page list must be first. The buffer pages
323  * are allocated in cache lines, which means that each buffer
324  * page will be at the beginning of a cache line, and thus
325  * the least significant bits will be zero. We use this to
326  * add flags in the list struct pointers, to make the ring buffer
327  * lockless.
328  */
329 struct buffer_page {
330 	struct list_head list;		/* list of buffer pages */
331 	local_t		 write;		/* index for next write */
332 	unsigned	 read;		/* index for next read */
333 	local_t		 entries;	/* entries on this page */
334 	unsigned long	 real_end;	/* real end of data */
335 	struct buffer_data_page *page;	/* Actual data page */
336 };
337 
338 /*
339  * The buffer page counters, write and entries, must be reset
340  * atomically when crossing page boundaries. To synchronize this
341  * update, two counters are inserted into the number. One is
342  * the actual counter for the write position or count on the page.
343  *
344  * The other is a counter of updaters. Before an update happens
345  * the update partition of the counter is incremented. This will
346  * allow the updater to update the counter atomically.
347  *
348  * The counter is 20 bits, and the state data is 12.
349  */
350 #define RB_WRITE_MASK		0xfffff
351 #define RB_WRITE_INTCNT		(1 << 20)
352 
rb_init_page(struct buffer_data_page * bpage)353 static void rb_init_page(struct buffer_data_page *bpage)
354 {
355 	local_set(&bpage->commit, 0);
356 }
357 
358 /*
359  * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
360  * this issue out.
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 int test_time_stamp(u64 delta)
372 {
373 	if (delta & TS_DELTA_TEST)
374 		return 1;
375 	return 0;
376 }
377 
378 #define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
379 
380 /* Max payload is BUF_PAGE_SIZE - header (8bytes) */
381 #define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
382 
ring_buffer_print_page_header(struct trace_seq * s)383 int ring_buffer_print_page_header(struct trace_seq *s)
384 {
385 	struct buffer_data_page field;
386 
387 	trace_seq_printf(s, "\tfield: u64 timestamp;\t"
388 			 "offset:0;\tsize:%u;\tsigned:%u;\n",
389 			 (unsigned int)sizeof(field.time_stamp),
390 			 (unsigned int)is_signed_type(u64));
391 
392 	trace_seq_printf(s, "\tfield: local_t commit;\t"
393 			 "offset:%u;\tsize:%u;\tsigned:%u;\n",
394 			 (unsigned int)offsetof(typeof(field), commit),
395 			 (unsigned int)sizeof(field.commit),
396 			 (unsigned int)is_signed_type(long));
397 
398 	trace_seq_printf(s, "\tfield: int overwrite;\t"
399 			 "offset:%u;\tsize:%u;\tsigned:%u;\n",
400 			 (unsigned int)offsetof(typeof(field), commit),
401 			 1,
402 			 (unsigned int)is_signed_type(long));
403 
404 	trace_seq_printf(s, "\tfield: char data;\t"
405 			 "offset:%u;\tsize:%u;\tsigned:%u;\n",
406 			 (unsigned int)offsetof(typeof(field), data),
407 			 (unsigned int)BUF_PAGE_SIZE,
408 			 (unsigned int)is_signed_type(char));
409 
410 	return !trace_seq_has_overflowed(s);
411 }
412 
413 struct rb_irq_work {
414 	struct irq_work			work;
415 	wait_queue_head_t		waiters;
416 	wait_queue_head_t		full_waiters;
417 	long				wait_index;
418 	bool				waiters_pending;
419 	bool				full_waiters_pending;
420 	bool				wakeup_full;
421 };
422 
423 /*
424  * Structure to hold event state and handle nested events.
425  */
426 struct rb_event_info {
427 	u64			ts;
428 	u64			delta;
429 	u64			before;
430 	u64			after;
431 	unsigned long		length;
432 	struct buffer_page	*tail_page;
433 	int			add_timestamp;
434 };
435 
436 /*
437  * Used for the add_timestamp
438  *  NONE
439  *  EXTEND - wants a time extend
440  *  ABSOLUTE - the buffer requests all events to have absolute time stamps
441  *  FORCE - force a full time stamp.
442  */
443 enum {
444 	RB_ADD_STAMP_NONE		= 0,
445 	RB_ADD_STAMP_EXTEND		= BIT(1),
446 	RB_ADD_STAMP_ABSOLUTE		= BIT(2),
447 	RB_ADD_STAMP_FORCE		= BIT(3)
448 };
449 /*
450  * Used for which event context the event is in.
451  *  TRANSITION = 0
452  *  NMI     = 1
453  *  IRQ     = 2
454  *  SOFTIRQ = 3
455  *  NORMAL  = 4
456  *
457  * See trace_recursive_lock() comment below for more details.
458  */
459 enum {
460 	RB_CTX_TRANSITION,
461 	RB_CTX_NMI,
462 	RB_CTX_IRQ,
463 	RB_CTX_SOFTIRQ,
464 	RB_CTX_NORMAL,
465 	RB_CTX_MAX
466 };
467 
468 #if BITS_PER_LONG == 32
469 #define RB_TIME_32
470 #endif
471 
472 /* To test on 64 bit machines */
473 //#define RB_TIME_32
474 
475 #ifdef RB_TIME_32
476 
477 struct rb_time_struct {
478 	local_t		cnt;
479 	local_t		top;
480 	local_t		bottom;
481 };
482 #else
483 #include <asm/local64.h>
484 struct rb_time_struct {
485 	local64_t	time;
486 };
487 #endif
488 typedef struct rb_time_struct rb_time_t;
489 
490 /*
491  * head_page == tail_page && head == tail then buffer is empty.
492  */
493 struct ring_buffer_per_cpu {
494 	int				cpu;
495 	atomic_t			record_disabled;
496 	atomic_t			resize_disabled;
497 	struct trace_buffer	*buffer;
498 	raw_spinlock_t			reader_lock;	/* serialize readers */
499 	arch_spinlock_t			lock;
500 	struct lock_class_key		lock_key;
501 	struct buffer_data_page		*free_page;
502 	unsigned long			nr_pages;
503 	unsigned int			current_context;
504 	struct list_head		*pages;
505 	struct buffer_page		*head_page;	/* read from head */
506 	struct buffer_page		*tail_page;	/* write to tail */
507 	struct buffer_page		*commit_page;	/* committed pages */
508 	struct buffer_page		*reader_page;
509 	unsigned long			lost_events;
510 	unsigned long			last_overrun;
511 	unsigned long			nest;
512 	local_t				entries_bytes;
513 	local_t				entries;
514 	local_t				overrun;
515 	local_t				commit_overrun;
516 	local_t				dropped_events;
517 	local_t				committing;
518 	local_t				commits;
519 	local_t				pages_touched;
520 	local_t				pages_lost;
521 	local_t				pages_read;
522 	long				last_pages_touch;
523 	size_t				shortest_full;
524 	unsigned long			read;
525 	unsigned long			read_bytes;
526 	rb_time_t			write_stamp;
527 	rb_time_t			before_stamp;
528 	u64				read_stamp;
529 	/* ring buffer pages to update, > 0 to add, < 0 to remove */
530 	long				nr_pages_to_update;
531 	struct list_head		new_pages; /* new pages to add */
532 	struct work_struct		update_pages_work;
533 	struct completion		update_done;
534 
535 	struct rb_irq_work		irq_work;
536 };
537 
538 struct trace_buffer {
539 	unsigned			flags;
540 	int				cpus;
541 	atomic_t			record_disabled;
542 	cpumask_var_t			cpumask;
543 
544 	struct lock_class_key		*reader_lock_key;
545 
546 	struct mutex			mutex;
547 
548 	struct ring_buffer_per_cpu	**buffers;
549 
550 	struct hlist_node		node;
551 	u64				(*clock)(void);
552 
553 	struct rb_irq_work		irq_work;
554 	bool				time_stamp_abs;
555 };
556 
557 struct ring_buffer_iter {
558 	struct ring_buffer_per_cpu	*cpu_buffer;
559 	unsigned long			head;
560 	unsigned long			next_event;
561 	struct buffer_page		*head_page;
562 	struct buffer_page		*cache_reader_page;
563 	unsigned long			cache_read;
564 	u64				read_stamp;
565 	u64				page_stamp;
566 	struct ring_buffer_event	*event;
567 	int				missed_events;
568 };
569 
570 #ifdef RB_TIME_32
571 
572 /*
573  * On 32 bit machines, local64_t is very expensive. As the ring
574  * buffer doesn't need all the features of a true 64 bit atomic,
575  * on 32 bit, it uses these functions (64 still uses local64_t).
576  *
577  * For the ring buffer, 64 bit required operations for the time is
578  * the following:
579  *
580  *  - Only need 59 bits (uses 60 to make it even).
581  *  - Reads may fail if it interrupted a modification of the time stamp.
582  *      It will succeed if it did not interrupt another write even if
583  *      the read itself is interrupted by a write.
584  *      It returns whether it was successful or not.
585  *
586  *  - Writes always succeed and will overwrite other writes and writes
587  *      that were done by events interrupting the current write.
588  *
589  *  - A write followed by a read of the same time stamp will always succeed,
590  *      but may not contain the same value.
591  *
592  *  - A cmpxchg will fail if it interrupted another write or cmpxchg.
593  *      Other than that, it acts like a normal cmpxchg.
594  *
595  * The 60 bit time stamp is broken up by 30 bits in a top and bottom half
596  *  (bottom being the least significant 30 bits of the 60 bit time stamp).
597  *
598  * The two most significant bits of each half holds a 2 bit counter (0-3).
599  * Each update will increment this counter by one.
600  * When reading the top and bottom, if the two counter bits match then the
601  *  top and bottom together make a valid 60 bit number.
602  */
603 #define RB_TIME_SHIFT	30
604 #define RB_TIME_VAL_MASK ((1 << RB_TIME_SHIFT) - 1)
605 
rb_time_cnt(unsigned long val)606 static inline int rb_time_cnt(unsigned long val)
607 {
608 	return (val >> RB_TIME_SHIFT) & 3;
609 }
610 
rb_time_val(unsigned long top,unsigned long bottom)611 static inline u64 rb_time_val(unsigned long top, unsigned long bottom)
612 {
613 	u64 val;
614 
615 	val = top & RB_TIME_VAL_MASK;
616 	val <<= RB_TIME_SHIFT;
617 	val |= bottom & RB_TIME_VAL_MASK;
618 
619 	return val;
620 }
621 
__rb_time_read(rb_time_t * t,u64 * ret,unsigned long * cnt)622 static inline bool __rb_time_read(rb_time_t *t, u64 *ret, unsigned long *cnt)
623 {
624 	unsigned long top, bottom;
625 	unsigned long c;
626 
627 	/*
628 	 * If the read is interrupted by a write, then the cnt will
629 	 * be different. Loop until both top and bottom have been read
630 	 * without interruption.
631 	 */
632 	do {
633 		c = local_read(&t->cnt);
634 		top = local_read(&t->top);
635 		bottom = local_read(&t->bottom);
636 	} while (c != local_read(&t->cnt));
637 
638 	*cnt = rb_time_cnt(top);
639 
640 	/* If top and bottom counts don't match, this interrupted a write */
641 	if (*cnt != rb_time_cnt(bottom))
642 		return false;
643 
644 	*ret = rb_time_val(top, bottom);
645 	return true;
646 }
647 
rb_time_read(rb_time_t * t,u64 * ret)648 static bool rb_time_read(rb_time_t *t, u64 *ret)
649 {
650 	unsigned long cnt;
651 
652 	return __rb_time_read(t, ret, &cnt);
653 }
654 
rb_time_val_cnt(unsigned long val,unsigned long cnt)655 static inline unsigned long rb_time_val_cnt(unsigned long val, unsigned long cnt)
656 {
657 	return (val & RB_TIME_VAL_MASK) | ((cnt & 3) << RB_TIME_SHIFT);
658 }
659 
rb_time_split(u64 val,unsigned long * top,unsigned long * bottom)660 static inline void rb_time_split(u64 val, unsigned long *top, unsigned long *bottom)
661 {
662 	*top = (unsigned long)((val >> RB_TIME_SHIFT) & RB_TIME_VAL_MASK);
663 	*bottom = (unsigned long)(val & RB_TIME_VAL_MASK);
664 }
665 
rb_time_val_set(local_t * t,unsigned long val,unsigned long cnt)666 static inline void rb_time_val_set(local_t *t, unsigned long val, unsigned long cnt)
667 {
668 	val = rb_time_val_cnt(val, cnt);
669 	local_set(t, val);
670 }
671 
rb_time_set(rb_time_t * t,u64 val)672 static void rb_time_set(rb_time_t *t, u64 val)
673 {
674 	unsigned long cnt, top, bottom;
675 
676 	rb_time_split(val, &top, &bottom);
677 
678 	/* Writes always succeed with a valid number even if it gets interrupted. */
679 	do {
680 		cnt = local_inc_return(&t->cnt);
681 		rb_time_val_set(&t->top, top, cnt);
682 		rb_time_val_set(&t->bottom, bottom, cnt);
683 	} while (cnt != local_read(&t->cnt));
684 }
685 
686 static inline bool
rb_time_read_cmpxchg(local_t * l,unsigned long expect,unsigned long set)687 rb_time_read_cmpxchg(local_t *l, unsigned long expect, unsigned long set)
688 {
689 	unsigned long ret;
690 
691 	ret = local_cmpxchg(l, expect, set);
692 	return ret == expect;
693 }
694 
rb_time_cmpxchg(rb_time_t * t,u64 expect,u64 set)695 static int rb_time_cmpxchg(rb_time_t *t, u64 expect, u64 set)
696 {
697 	unsigned long cnt, top, bottom;
698 	unsigned long cnt2, top2, bottom2;
699 	u64 val;
700 
701 	/* The cmpxchg always fails if it interrupted an update */
702 	 if (!__rb_time_read(t, &val, &cnt2))
703 		 return false;
704 
705 	 if (val != expect)
706 		 return false;
707 
708 	 cnt = local_read(&t->cnt);
709 	 if ((cnt & 3) != cnt2)
710 		 return false;
711 
712 	 cnt2 = cnt + 1;
713 
714 	 rb_time_split(val, &top, &bottom);
715 	 top = rb_time_val_cnt(top, cnt);
716 	 bottom = rb_time_val_cnt(bottom, cnt);
717 
718 	 rb_time_split(set, &top2, &bottom2);
719 	 top2 = rb_time_val_cnt(top2, cnt2);
720 	 bottom2 = rb_time_val_cnt(bottom2, cnt2);
721 
722 	if (!rb_time_read_cmpxchg(&t->cnt, cnt, cnt2))
723 		return false;
724 	if (!rb_time_read_cmpxchg(&t->top, top, top2))
725 		return false;
726 	if (!rb_time_read_cmpxchg(&t->bottom, bottom, bottom2))
727 		return false;
728 	return true;
729 }
730 
731 #else /* 64 bits */
732 
733 /* local64_t always succeeds */
734 
rb_time_read(rb_time_t * t,u64 * ret)735 static inline bool rb_time_read(rb_time_t *t, u64 *ret)
736 {
737 	*ret = local64_read(&t->time);
738 	return true;
739 }
rb_time_set(rb_time_t * t,u64 val)740 static void rb_time_set(rb_time_t *t, u64 val)
741 {
742 	local64_set(&t->time, val);
743 }
744 
rb_time_cmpxchg(rb_time_t * t,u64 expect,u64 set)745 static bool rb_time_cmpxchg(rb_time_t *t, u64 expect, u64 set)
746 {
747 	u64 val;
748 	val = local64_cmpxchg(&t->time, expect, set);
749 	return val == expect;
750 }
751 #endif
752 
753 /**
754  * ring_buffer_nr_pages - get the number of buffer pages in the ring buffer
755  * @buffer: The ring_buffer to get the number of pages from
756  * @cpu: The cpu of the ring_buffer to get the number of pages from
757  *
758  * Returns the number of pages used by a per_cpu buffer of the ring buffer.
759  */
ring_buffer_nr_pages(struct trace_buffer * buffer,int cpu)760 size_t ring_buffer_nr_pages(struct trace_buffer *buffer, int cpu)
761 {
762 	return buffer->buffers[cpu]->nr_pages;
763 }
764 
765 /**
766  * ring_buffer_nr_pages_dirty - get the number of used pages in the ring buffer
767  * @buffer: The ring_buffer to get the number of pages from
768  * @cpu: The cpu of the ring_buffer to get the number of pages from
769  *
770  * Returns the number of pages that have content in the ring buffer.
771  */
ring_buffer_nr_dirty_pages(struct trace_buffer * buffer,int cpu)772 size_t ring_buffer_nr_dirty_pages(struct trace_buffer *buffer, int cpu)
773 {
774 	size_t read;
775 	size_t lost;
776 	size_t cnt;
777 
778 	read = local_read(&buffer->buffers[cpu]->pages_read);
779 	lost = local_read(&buffer->buffers[cpu]->pages_lost);
780 	cnt = local_read(&buffer->buffers[cpu]->pages_touched);
781 
782 	if (WARN_ON_ONCE(cnt < lost))
783 		return 0;
784 
785 	cnt -= lost;
786 
787 	/* The reader can read an empty page, but not more than that */
788 	if (cnt < read) {
789 		WARN_ON_ONCE(read > cnt + 1);
790 		return 0;
791 	}
792 
793 	return cnt - read;
794 }
795 
full_hit(struct trace_buffer * buffer,int cpu,int full)796 static __always_inline bool full_hit(struct trace_buffer *buffer, int cpu, int full)
797 {
798 	struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
799 	size_t nr_pages;
800 	size_t dirty;
801 
802 	nr_pages = cpu_buffer->nr_pages;
803 	if (!nr_pages || !full)
804 		return true;
805 
806 	dirty = ring_buffer_nr_dirty_pages(buffer, cpu);
807 
808 	return (dirty * 100) > (full * nr_pages);
809 }
810 
811 /*
812  * rb_wake_up_waiters - wake up tasks waiting for ring buffer input
813  *
814  * Schedules a delayed work to wake up any task that is blocked on the
815  * ring buffer waiters queue.
816  */
rb_wake_up_waiters(struct irq_work * work)817 static void rb_wake_up_waiters(struct irq_work *work)
818 {
819 	struct rb_irq_work *rbwork = container_of(work, struct rb_irq_work, work);
820 
821 	wake_up_all(&rbwork->waiters);
822 	if (rbwork->full_waiters_pending || rbwork->wakeup_full) {
823 		rbwork->wakeup_full = false;
824 		rbwork->full_waiters_pending = false;
825 		wake_up_all(&rbwork->full_waiters);
826 	}
827 }
828 
829 /**
830  * ring_buffer_wake_waiters - wake up any waiters on this ring buffer
831  * @buffer: The ring buffer to wake waiters on
832  *
833  * In the case of a file that represents a ring buffer is closing,
834  * it is prudent to wake up any waiters that are on this.
835  */
ring_buffer_wake_waiters(struct trace_buffer * buffer,int cpu)836 void ring_buffer_wake_waiters(struct trace_buffer *buffer, int cpu)
837 {
838 	struct ring_buffer_per_cpu *cpu_buffer;
839 	struct rb_irq_work *rbwork;
840 
841 	if (cpu == RING_BUFFER_ALL_CPUS) {
842 
843 		/* Wake up individual ones too. One level recursion */
844 		for_each_buffer_cpu(buffer, cpu)
845 			ring_buffer_wake_waiters(buffer, cpu);
846 
847 		rbwork = &buffer->irq_work;
848 	} else {
849 		cpu_buffer = buffer->buffers[cpu];
850 		rbwork = &cpu_buffer->irq_work;
851 	}
852 
853 	rbwork->wait_index++;
854 	/* make sure the waiters see the new index */
855 	smp_wmb();
856 
857 	rb_wake_up_waiters(&rbwork->work);
858 }
859 
860 /**
861  * ring_buffer_wait - wait for input to the ring buffer
862  * @buffer: buffer to wait on
863  * @cpu: the cpu buffer to wait on
864  * @full: wait until the percentage of pages are available, if @cpu != RING_BUFFER_ALL_CPUS
865  *
866  * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
867  * as data is added to any of the @buffer's cpu buffers. Otherwise
868  * it will wait for data to be added to a specific cpu buffer.
869  */
ring_buffer_wait(struct trace_buffer * buffer,int cpu,int full)870 int ring_buffer_wait(struct trace_buffer *buffer, int cpu, int full)
871 {
872 	struct ring_buffer_per_cpu *cpu_buffer;
873 	DEFINE_WAIT(wait);
874 	struct rb_irq_work *work;
875 	long wait_index;
876 	int ret = 0;
877 
878 	/*
879 	 * Depending on what the caller is waiting for, either any
880 	 * data in any cpu buffer, or a specific buffer, put the
881 	 * caller on the appropriate wait queue.
882 	 */
883 	if (cpu == RING_BUFFER_ALL_CPUS) {
884 		work = &buffer->irq_work;
885 		/* Full only makes sense on per cpu reads */
886 		full = 0;
887 	} else {
888 		if (!cpumask_test_cpu(cpu, buffer->cpumask))
889 			return -ENODEV;
890 		cpu_buffer = buffer->buffers[cpu];
891 		work = &cpu_buffer->irq_work;
892 	}
893 
894 	wait_index = READ_ONCE(work->wait_index);
895 
896 	while (true) {
897 		if (full)
898 			prepare_to_wait(&work->full_waiters, &wait, TASK_INTERRUPTIBLE);
899 		else
900 			prepare_to_wait(&work->waiters, &wait, TASK_INTERRUPTIBLE);
901 
902 		/*
903 		 * The events can happen in critical sections where
904 		 * checking a work queue can cause deadlocks.
905 		 * After adding a task to the queue, this flag is set
906 		 * only to notify events to try to wake up the queue
907 		 * using irq_work.
908 		 *
909 		 * We don't clear it even if the buffer is no longer
910 		 * empty. The flag only causes the next event to run
911 		 * irq_work to do the work queue wake up. The worse
912 		 * that can happen if we race with !trace_empty() is that
913 		 * an event will cause an irq_work to try to wake up
914 		 * an empty queue.
915 		 *
916 		 * There's no reason to protect this flag either, as
917 		 * the work queue and irq_work logic will do the necessary
918 		 * synchronization for the wake ups. The only thing
919 		 * that is necessary is that the wake up happens after
920 		 * a task has been queued. It's OK for spurious wake ups.
921 		 */
922 		if (full)
923 			work->full_waiters_pending = true;
924 		else
925 			work->waiters_pending = true;
926 
927 		if (signal_pending(current)) {
928 			ret = -EINTR;
929 			break;
930 		}
931 
932 		if (cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer))
933 			break;
934 
935 		if (cpu != RING_BUFFER_ALL_CPUS &&
936 		    !ring_buffer_empty_cpu(buffer, cpu)) {
937 			unsigned long flags;
938 			bool pagebusy;
939 			bool done;
940 
941 			if (!full)
942 				break;
943 
944 			raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
945 			pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
946 			done = !pagebusy && full_hit(buffer, cpu, full);
947 
948 			if (!cpu_buffer->shortest_full ||
949 			    cpu_buffer->shortest_full > full)
950 				cpu_buffer->shortest_full = full;
951 			raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
952 			if (done)
953 				break;
954 		}
955 
956 		schedule();
957 
958 		/* Make sure to see the new wait index */
959 		smp_rmb();
960 		if (wait_index != work->wait_index)
961 			break;
962 	}
963 
964 	if (full)
965 		finish_wait(&work->full_waiters, &wait);
966 	else
967 		finish_wait(&work->waiters, &wait);
968 
969 	return ret;
970 }
971 
972 /**
973  * ring_buffer_poll_wait - poll on buffer input
974  * @buffer: buffer to wait on
975  * @cpu: the cpu buffer to wait on
976  * @filp: the file descriptor
977  * @poll_table: The poll descriptor
978  * @full: wait until the percentage of pages are available, if @cpu != RING_BUFFER_ALL_CPUS
979  *
980  * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
981  * as data is added to any of the @buffer's cpu buffers. Otherwise
982  * it will wait for data to be added to a specific cpu buffer.
983  *
984  * Returns EPOLLIN | EPOLLRDNORM if data exists in the buffers,
985  * zero otherwise.
986  */
ring_buffer_poll_wait(struct trace_buffer * buffer,int cpu,struct file * filp,poll_table * poll_table,int full)987 __poll_t ring_buffer_poll_wait(struct trace_buffer *buffer, int cpu,
988 			  struct file *filp, poll_table *poll_table, int full)
989 {
990 	struct ring_buffer_per_cpu *cpu_buffer;
991 	struct rb_irq_work *work;
992 
993 	if (cpu == RING_BUFFER_ALL_CPUS) {
994 		work = &buffer->irq_work;
995 		full = 0;
996 	} else {
997 		if (!cpumask_test_cpu(cpu, buffer->cpumask))
998 			return -EINVAL;
999 
1000 		cpu_buffer = buffer->buffers[cpu];
1001 		work = &cpu_buffer->irq_work;
1002 	}
1003 
1004 	if (full) {
1005 		poll_wait(filp, &work->full_waiters, poll_table);
1006 		work->full_waiters_pending = true;
1007 	} else {
1008 		poll_wait(filp, &work->waiters, poll_table);
1009 		work->waiters_pending = true;
1010 	}
1011 
1012 	/*
1013 	 * There's a tight race between setting the waiters_pending and
1014 	 * checking if the ring buffer is empty.  Once the waiters_pending bit
1015 	 * is set, the next event will wake the task up, but we can get stuck
1016 	 * if there's only a single event in.
1017 	 *
1018 	 * FIXME: Ideally, we need a memory barrier on the writer side as well,
1019 	 * but adding a memory barrier to all events will cause too much of a
1020 	 * performance hit in the fast path.  We only need a memory barrier when
1021 	 * the buffer goes from empty to having content.  But as this race is
1022 	 * extremely small, and it's not a problem if another event comes in, we
1023 	 * will fix it later.
1024 	 */
1025 	smp_mb();
1026 
1027 	if (full)
1028 		return full_hit(buffer, cpu, full) ? EPOLLIN | EPOLLRDNORM : 0;
1029 
1030 	if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
1031 	    (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
1032 		return EPOLLIN | EPOLLRDNORM;
1033 	return 0;
1034 }
1035 
1036 /* buffer may be either ring_buffer or ring_buffer_per_cpu */
1037 #define RB_WARN_ON(b, cond)						\
1038 	({								\
1039 		int _____ret = unlikely(cond);				\
1040 		if (_____ret) {						\
1041 			if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
1042 				struct ring_buffer_per_cpu *__b =	\
1043 					(void *)b;			\
1044 				atomic_inc(&__b->buffer->record_disabled); \
1045 			} else						\
1046 				atomic_inc(&b->record_disabled);	\
1047 			WARN_ON(1);					\
1048 		}							\
1049 		_____ret;						\
1050 	})
1051 
1052 /* Up this if you want to test the TIME_EXTENTS and normalization */
1053 #define DEBUG_SHIFT 0
1054 
rb_time_stamp(struct trace_buffer * buffer)1055 static inline u64 rb_time_stamp(struct trace_buffer *buffer)
1056 {
1057 	u64 ts;
1058 
1059 	/* Skip retpolines :-( */
1060 	if (IS_ENABLED(CONFIG_RETPOLINE) && likely(buffer->clock == trace_clock_local))
1061 		ts = trace_clock_local();
1062 	else
1063 		ts = buffer->clock();
1064 
1065 	/* shift to debug/test normalization and TIME_EXTENTS */
1066 	return ts << DEBUG_SHIFT;
1067 }
1068 
ring_buffer_time_stamp(struct trace_buffer * buffer,int cpu)1069 u64 ring_buffer_time_stamp(struct trace_buffer *buffer, int cpu)
1070 {
1071 	u64 time;
1072 
1073 	preempt_disable_notrace();
1074 	time = rb_time_stamp(buffer);
1075 	preempt_enable_notrace();
1076 
1077 	return time;
1078 }
1079 EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
1080 
ring_buffer_normalize_time_stamp(struct trace_buffer * buffer,int cpu,u64 * ts)1081 void ring_buffer_normalize_time_stamp(struct trace_buffer *buffer,
1082 				      int cpu, u64 *ts)
1083 {
1084 	/* Just stupid testing the normalize function and deltas */
1085 	*ts >>= DEBUG_SHIFT;
1086 }
1087 EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
1088 
1089 /*
1090  * Making the ring buffer lockless makes things tricky.
1091  * Although writes only happen on the CPU that they are on,
1092  * and they only need to worry about interrupts. Reads can
1093  * happen on any CPU.
1094  *
1095  * The reader page is always off the ring buffer, but when the
1096  * reader finishes with a page, it needs to swap its page with
1097  * a new one from the buffer. The reader needs to take from
1098  * the head (writes go to the tail). But if a writer is in overwrite
1099  * mode and wraps, it must push the head page forward.
1100  *
1101  * Here lies the problem.
1102  *
1103  * The reader must be careful to replace only the head page, and
1104  * not another one. As described at the top of the file in the
1105  * ASCII art, the reader sets its old page to point to the next
1106  * page after head. It then sets the page after head to point to
1107  * the old reader page. But if the writer moves the head page
1108  * during this operation, the reader could end up with the tail.
1109  *
1110  * We use cmpxchg to help prevent this race. We also do something
1111  * special with the page before head. We set the LSB to 1.
1112  *
1113  * When the writer must push the page forward, it will clear the
1114  * bit that points to the head page, move the head, and then set
1115  * the bit that points to the new head page.
1116  *
1117  * We also don't want an interrupt coming in and moving the head
1118  * page on another writer. Thus we use the second LSB to catch
1119  * that too. Thus:
1120  *
1121  * head->list->prev->next        bit 1          bit 0
1122  *                              -------        -------
1123  * Normal page                     0              0
1124  * Points to head page             0              1
1125  * New head page                   1              0
1126  *
1127  * Note we can not trust the prev pointer of the head page, because:
1128  *
1129  * +----+       +-----+        +-----+
1130  * |    |------>|  T  |---X--->|  N  |
1131  * |    |<------|     |        |     |
1132  * +----+       +-----+        +-----+
1133  *   ^                           ^ |
1134  *   |          +-----+          | |
1135  *   +----------|  R  |----------+ |
1136  *              |     |<-----------+
1137  *              +-----+
1138  *
1139  * Key:  ---X-->  HEAD flag set in pointer
1140  *         T      Tail page
1141  *         R      Reader page
1142  *         N      Next page
1143  *
1144  * (see __rb_reserve_next() to see where this happens)
1145  *
1146  *  What the above shows is that the reader just swapped out
1147  *  the reader page with a page in the buffer, but before it
1148  *  could make the new header point back to the new page added
1149  *  it was preempted by a writer. The writer moved forward onto
1150  *  the new page added by the reader and is about to move forward
1151  *  again.
1152  *
1153  *  You can see, it is legitimate for the previous pointer of
1154  *  the head (or any page) not to point back to itself. But only
1155  *  temporarily.
1156  */
1157 
1158 #define RB_PAGE_NORMAL		0UL
1159 #define RB_PAGE_HEAD		1UL
1160 #define RB_PAGE_UPDATE		2UL
1161 
1162 
1163 #define RB_FLAG_MASK		3UL
1164 
1165 /* PAGE_MOVED is not part of the mask */
1166 #define RB_PAGE_MOVED		4UL
1167 
1168 /*
1169  * rb_list_head - remove any bit
1170  */
rb_list_head(struct list_head * list)1171 static struct list_head *rb_list_head(struct list_head *list)
1172 {
1173 	unsigned long val = (unsigned long)list;
1174 
1175 	return (struct list_head *)(val & ~RB_FLAG_MASK);
1176 }
1177 
1178 /*
1179  * rb_is_head_page - test if the given page is the head page
1180  *
1181  * Because the reader may move the head_page pointer, we can
1182  * not trust what the head page is (it may be pointing to
1183  * the reader page). But if the next page is a header page,
1184  * its flags will be non zero.
1185  */
1186 static inline int
rb_is_head_page(struct ring_buffer_per_cpu * cpu_buffer,struct buffer_page * page,struct list_head * list)1187 rb_is_head_page(struct ring_buffer_per_cpu *cpu_buffer,
1188 		struct buffer_page *page, struct list_head *list)
1189 {
1190 	unsigned long val;
1191 
1192 	val = (unsigned long)list->next;
1193 
1194 	if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
1195 		return RB_PAGE_MOVED;
1196 
1197 	return val & RB_FLAG_MASK;
1198 }
1199 
1200 /*
1201  * rb_is_reader_page
1202  *
1203  * The unique thing about the reader page, is that, if the
1204  * writer is ever on it, the previous pointer never points
1205  * back to the reader page.
1206  */
rb_is_reader_page(struct buffer_page * page)1207 static bool rb_is_reader_page(struct buffer_page *page)
1208 {
1209 	struct list_head *list = page->list.prev;
1210 
1211 	return rb_list_head(list->next) != &page->list;
1212 }
1213 
1214 /*
1215  * rb_set_list_to_head - set a list_head to be pointing to head.
1216  */
rb_set_list_to_head(struct ring_buffer_per_cpu * cpu_buffer,struct list_head * list)1217 static void rb_set_list_to_head(struct ring_buffer_per_cpu *cpu_buffer,
1218 				struct list_head *list)
1219 {
1220 	unsigned long *ptr;
1221 
1222 	ptr = (unsigned long *)&list->next;
1223 	*ptr |= RB_PAGE_HEAD;
1224 	*ptr &= ~RB_PAGE_UPDATE;
1225 }
1226 
1227 /*
1228  * rb_head_page_activate - sets up head page
1229  */
rb_head_page_activate(struct ring_buffer_per_cpu * cpu_buffer)1230 static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
1231 {
1232 	struct buffer_page *head;
1233 
1234 	head = cpu_buffer->head_page;
1235 	if (!head)
1236 		return;
1237 
1238 	/*
1239 	 * Set the previous list pointer to have the HEAD flag.
1240 	 */
1241 	rb_set_list_to_head(cpu_buffer, head->list.prev);
1242 }
1243 
rb_list_head_clear(struct list_head * list)1244 static void rb_list_head_clear(struct list_head *list)
1245 {
1246 	unsigned long *ptr = (unsigned long *)&list->next;
1247 
1248 	*ptr &= ~RB_FLAG_MASK;
1249 }
1250 
1251 /*
1252  * rb_head_page_deactivate - clears head page ptr (for free list)
1253  */
1254 static void
rb_head_page_deactivate(struct ring_buffer_per_cpu * cpu_buffer)1255 rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
1256 {
1257 	struct list_head *hd;
1258 
1259 	/* Go through the whole list and clear any pointers found. */
1260 	rb_list_head_clear(cpu_buffer->pages);
1261 
1262 	list_for_each(hd, cpu_buffer->pages)
1263 		rb_list_head_clear(hd);
1264 }
1265 
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)1266 static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
1267 			    struct buffer_page *head,
1268 			    struct buffer_page *prev,
1269 			    int old_flag, int new_flag)
1270 {
1271 	struct list_head *list;
1272 	unsigned long val = (unsigned long)&head->list;
1273 	unsigned long ret;
1274 
1275 	list = &prev->list;
1276 
1277 	val &= ~RB_FLAG_MASK;
1278 
1279 	ret = cmpxchg((unsigned long *)&list->next,
1280 		      val | old_flag, val | new_flag);
1281 
1282 	/* check if the reader took the page */
1283 	if ((ret & ~RB_FLAG_MASK) != val)
1284 		return RB_PAGE_MOVED;
1285 
1286 	return ret & RB_FLAG_MASK;
1287 }
1288 
rb_head_page_set_update(struct ring_buffer_per_cpu * cpu_buffer,struct buffer_page * head,struct buffer_page * prev,int old_flag)1289 static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
1290 				   struct buffer_page *head,
1291 				   struct buffer_page *prev,
1292 				   int old_flag)
1293 {
1294 	return rb_head_page_set(cpu_buffer, head, prev,
1295 				old_flag, RB_PAGE_UPDATE);
1296 }
1297 
rb_head_page_set_head(struct ring_buffer_per_cpu * cpu_buffer,struct buffer_page * head,struct buffer_page * prev,int old_flag)1298 static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
1299 				 struct buffer_page *head,
1300 				 struct buffer_page *prev,
1301 				 int old_flag)
1302 {
1303 	return rb_head_page_set(cpu_buffer, head, prev,
1304 				old_flag, RB_PAGE_HEAD);
1305 }
1306 
rb_head_page_set_normal(struct ring_buffer_per_cpu * cpu_buffer,struct buffer_page * head,struct buffer_page * prev,int old_flag)1307 static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
1308 				   struct buffer_page *head,
1309 				   struct buffer_page *prev,
1310 				   int old_flag)
1311 {
1312 	return rb_head_page_set(cpu_buffer, head, prev,
1313 				old_flag, RB_PAGE_NORMAL);
1314 }
1315 
rb_inc_page(struct ring_buffer_per_cpu * cpu_buffer,struct buffer_page ** bpage)1316 static inline void rb_inc_page(struct ring_buffer_per_cpu *cpu_buffer,
1317 			       struct buffer_page **bpage)
1318 {
1319 	struct list_head *p = rb_list_head((*bpage)->list.next);
1320 
1321 	*bpage = list_entry(p, struct buffer_page, list);
1322 }
1323 
1324 static struct buffer_page *
rb_set_head_page(struct ring_buffer_per_cpu * cpu_buffer)1325 rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
1326 {
1327 	struct buffer_page *head;
1328 	struct buffer_page *page;
1329 	struct list_head *list;
1330 	int i;
1331 
1332 	if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
1333 		return NULL;
1334 
1335 	/* sanity check */
1336 	list = cpu_buffer->pages;
1337 	if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
1338 		return NULL;
1339 
1340 	page = head = cpu_buffer->head_page;
1341 	/*
1342 	 * It is possible that the writer moves the header behind
1343 	 * where we started, and we miss in one loop.
1344 	 * A second loop should grab the header, but we'll do
1345 	 * three loops just because I'm paranoid.
1346 	 */
1347 	for (i = 0; i < 3; i++) {
1348 		do {
1349 			if (rb_is_head_page(cpu_buffer, page, page->list.prev)) {
1350 				cpu_buffer->head_page = page;
1351 				return page;
1352 			}
1353 			rb_inc_page(cpu_buffer, &page);
1354 		} while (page != head);
1355 	}
1356 
1357 	RB_WARN_ON(cpu_buffer, 1);
1358 
1359 	return NULL;
1360 }
1361 
rb_head_page_replace(struct buffer_page * old,struct buffer_page * new)1362 static int rb_head_page_replace(struct buffer_page *old,
1363 				struct buffer_page *new)
1364 {
1365 	unsigned long *ptr = (unsigned long *)&old->list.prev->next;
1366 	unsigned long val;
1367 	unsigned long ret;
1368 
1369 	val = *ptr & ~RB_FLAG_MASK;
1370 	val |= RB_PAGE_HEAD;
1371 
1372 	ret = cmpxchg(ptr, val, (unsigned long)&new->list);
1373 
1374 	return ret == val;
1375 }
1376 
1377 /*
1378  * rb_tail_page_update - move the tail page forward
1379  */
rb_tail_page_update(struct ring_buffer_per_cpu * cpu_buffer,struct buffer_page * tail_page,struct buffer_page * next_page)1380 static void rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
1381 			       struct buffer_page *tail_page,
1382 			       struct buffer_page *next_page)
1383 {
1384 	unsigned long old_entries;
1385 	unsigned long old_write;
1386 
1387 	/*
1388 	 * The tail page now needs to be moved forward.
1389 	 *
1390 	 * We need to reset the tail page, but without messing
1391 	 * with possible erasing of data brought in by interrupts
1392 	 * that have moved the tail page and are currently on it.
1393 	 *
1394 	 * We add a counter to the write field to denote this.
1395 	 */
1396 	old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
1397 	old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
1398 
1399 	local_inc(&cpu_buffer->pages_touched);
1400 	/*
1401 	 * Just make sure we have seen our old_write and synchronize
1402 	 * with any interrupts that come in.
1403 	 */
1404 	barrier();
1405 
1406 	/*
1407 	 * If the tail page is still the same as what we think
1408 	 * it is, then it is up to us to update the tail
1409 	 * pointer.
1410 	 */
1411 	if (tail_page == READ_ONCE(cpu_buffer->tail_page)) {
1412 		/* Zero the write counter */
1413 		unsigned long val = old_write & ~RB_WRITE_MASK;
1414 		unsigned long eval = old_entries & ~RB_WRITE_MASK;
1415 
1416 		/*
1417 		 * This will only succeed if an interrupt did
1418 		 * not come in and change it. In which case, we
1419 		 * do not want to modify it.
1420 		 *
1421 		 * We add (void) to let the compiler know that we do not care
1422 		 * about the return value of these functions. We use the
1423 		 * cmpxchg to only update if an interrupt did not already
1424 		 * do it for us. If the cmpxchg fails, we don't care.
1425 		 */
1426 		(void)local_cmpxchg(&next_page->write, old_write, val);
1427 		(void)local_cmpxchg(&next_page->entries, old_entries, eval);
1428 
1429 		/*
1430 		 * No need to worry about races with clearing out the commit.
1431 		 * it only can increment when a commit takes place. But that
1432 		 * only happens in the outer most nested commit.
1433 		 */
1434 		local_set(&next_page->page->commit, 0);
1435 
1436 		/* Again, either we update tail_page or an interrupt does */
1437 		(void)cmpxchg(&cpu_buffer->tail_page, tail_page, next_page);
1438 	}
1439 }
1440 
rb_check_bpage(struct ring_buffer_per_cpu * cpu_buffer,struct buffer_page * bpage)1441 static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
1442 			  struct buffer_page *bpage)
1443 {
1444 	unsigned long val = (unsigned long)bpage;
1445 
1446 	if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK))
1447 		return 1;
1448 
1449 	return 0;
1450 }
1451 
1452 /**
1453  * rb_check_list - make sure a pointer to a list has the last bits zero
1454  */
rb_check_list(struct ring_buffer_per_cpu * cpu_buffer,struct list_head * list)1455 static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer,
1456 			 struct list_head *list)
1457 {
1458 	if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev))
1459 		return 1;
1460 	if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next))
1461 		return 1;
1462 	return 0;
1463 }
1464 
1465 /**
1466  * rb_check_pages - integrity check of buffer pages
1467  * @cpu_buffer: CPU buffer with pages to test
1468  *
1469  * As a safety measure we check to make sure the data pages have not
1470  * been corrupted.
1471  */
rb_check_pages(struct ring_buffer_per_cpu * cpu_buffer)1472 static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
1473 {
1474 	struct list_head *head = cpu_buffer->pages;
1475 	struct buffer_page *bpage, *tmp;
1476 
1477 	/* Reset the head page if it exists */
1478 	if (cpu_buffer->head_page)
1479 		rb_set_head_page(cpu_buffer);
1480 
1481 	rb_head_page_deactivate(cpu_buffer);
1482 
1483 	if (RB_WARN_ON(cpu_buffer, head->next->prev != head))
1484 		return -1;
1485 	if (RB_WARN_ON(cpu_buffer, head->prev->next != head))
1486 		return -1;
1487 
1488 	if (rb_check_list(cpu_buffer, head))
1489 		return -1;
1490 
1491 	list_for_each_entry_safe(bpage, tmp, head, list) {
1492 		if (RB_WARN_ON(cpu_buffer,
1493 			       bpage->list.next->prev != &bpage->list))
1494 			return -1;
1495 		if (RB_WARN_ON(cpu_buffer,
1496 			       bpage->list.prev->next != &bpage->list))
1497 			return -1;
1498 		if (rb_check_list(cpu_buffer, &bpage->list))
1499 			return -1;
1500 	}
1501 
1502 	rb_head_page_activate(cpu_buffer);
1503 
1504 	return 0;
1505 }
1506 
__rb_allocate_pages(long nr_pages,struct list_head * pages,int cpu)1507 static int __rb_allocate_pages(long nr_pages, struct list_head *pages, int cpu)
1508 {
1509 	struct buffer_page *bpage, *tmp;
1510 	bool user_thread = current->mm != NULL;
1511 	gfp_t mflags;
1512 	long i;
1513 
1514 	/*
1515 	 * Check if the available memory is there first.
1516 	 * Note, si_mem_available() only gives us a rough estimate of available
1517 	 * memory. It may not be accurate. But we don't care, we just want
1518 	 * to prevent doing any allocation when it is obvious that it is
1519 	 * not going to succeed.
1520 	 */
1521 	i = si_mem_available();
1522 	if (i < nr_pages)
1523 		return -ENOMEM;
1524 
1525 	/*
1526 	 * __GFP_RETRY_MAYFAIL flag makes sure that the allocation fails
1527 	 * gracefully without invoking oom-killer and the system is not
1528 	 * destabilized.
1529 	 */
1530 	mflags = GFP_KERNEL | __GFP_RETRY_MAYFAIL;
1531 
1532 	/*
1533 	 * If a user thread allocates too much, and si_mem_available()
1534 	 * reports there's enough memory, even though there is not.
1535 	 * Make sure the OOM killer kills this thread. This can happen
1536 	 * even with RETRY_MAYFAIL because another task may be doing
1537 	 * an allocation after this task has taken all memory.
1538 	 * This is the task the OOM killer needs to take out during this
1539 	 * loop, even if it was triggered by an allocation somewhere else.
1540 	 */
1541 	if (user_thread)
1542 		set_current_oom_origin();
1543 	for (i = 0; i < nr_pages; i++) {
1544 		struct page *page;
1545 
1546 		bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1547 				    mflags, cpu_to_node(cpu));
1548 		if (!bpage)
1549 			goto free_pages;
1550 
1551 		list_add(&bpage->list, pages);
1552 
1553 		page = alloc_pages_node(cpu_to_node(cpu), mflags, 0);
1554 		if (!page)
1555 			goto free_pages;
1556 		bpage->page = page_address(page);
1557 		rb_init_page(bpage->page);
1558 
1559 		if (user_thread && fatal_signal_pending(current))
1560 			goto free_pages;
1561 	}
1562 	if (user_thread)
1563 		clear_current_oom_origin();
1564 
1565 	return 0;
1566 
1567 free_pages:
1568 	list_for_each_entry_safe(bpage, tmp, pages, list) {
1569 		list_del_init(&bpage->list);
1570 		free_buffer_page(bpage);
1571 	}
1572 	if (user_thread)
1573 		clear_current_oom_origin();
1574 
1575 	return -ENOMEM;
1576 }
1577 
rb_allocate_pages(struct ring_buffer_per_cpu * cpu_buffer,unsigned long nr_pages)1578 static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1579 			     unsigned long nr_pages)
1580 {
1581 	LIST_HEAD(pages);
1582 
1583 	WARN_ON(!nr_pages);
1584 
1585 	if (__rb_allocate_pages(nr_pages, &pages, cpu_buffer->cpu))
1586 		return -ENOMEM;
1587 
1588 	/*
1589 	 * The ring buffer page list is a circular list that does not
1590 	 * start and end with a list head. All page list items point to
1591 	 * other pages.
1592 	 */
1593 	cpu_buffer->pages = pages.next;
1594 	list_del(&pages);
1595 
1596 	cpu_buffer->nr_pages = nr_pages;
1597 
1598 	rb_check_pages(cpu_buffer);
1599 
1600 	return 0;
1601 }
1602 
1603 static struct ring_buffer_per_cpu *
rb_allocate_cpu_buffer(struct trace_buffer * buffer,long nr_pages,int cpu)1604 rb_allocate_cpu_buffer(struct trace_buffer *buffer, long nr_pages, int cpu)
1605 {
1606 	struct ring_buffer_per_cpu *cpu_buffer;
1607 	struct buffer_page *bpage;
1608 	struct page *page;
1609 	int ret;
1610 
1611 	cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1612 				  GFP_KERNEL, cpu_to_node(cpu));
1613 	if (!cpu_buffer)
1614 		return NULL;
1615 
1616 	cpu_buffer->cpu = cpu;
1617 	cpu_buffer->buffer = buffer;
1618 	raw_spin_lock_init(&cpu_buffer->reader_lock);
1619 	lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1620 	cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1621 	INIT_WORK(&cpu_buffer->update_pages_work, update_pages_handler);
1622 	init_completion(&cpu_buffer->update_done);
1623 	init_irq_work(&cpu_buffer->irq_work.work, rb_wake_up_waiters);
1624 	init_waitqueue_head(&cpu_buffer->irq_work.waiters);
1625 	init_waitqueue_head(&cpu_buffer->irq_work.full_waiters);
1626 
1627 	bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1628 			    GFP_KERNEL, cpu_to_node(cpu));
1629 	if (!bpage)
1630 		goto fail_free_buffer;
1631 
1632 	rb_check_bpage(cpu_buffer, bpage);
1633 
1634 	cpu_buffer->reader_page = bpage;
1635 	page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0);
1636 	if (!page)
1637 		goto fail_free_reader;
1638 	bpage->page = page_address(page);
1639 	rb_init_page(bpage->page);
1640 
1641 	INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1642 	INIT_LIST_HEAD(&cpu_buffer->new_pages);
1643 
1644 	ret = rb_allocate_pages(cpu_buffer, nr_pages);
1645 	if (ret < 0)
1646 		goto fail_free_reader;
1647 
1648 	cpu_buffer->head_page
1649 		= list_entry(cpu_buffer->pages, struct buffer_page, list);
1650 	cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1651 
1652 	rb_head_page_activate(cpu_buffer);
1653 
1654 	return cpu_buffer;
1655 
1656  fail_free_reader:
1657 	free_buffer_page(cpu_buffer->reader_page);
1658 
1659  fail_free_buffer:
1660 	kfree(cpu_buffer);
1661 	return NULL;
1662 }
1663 
rb_free_cpu_buffer(struct ring_buffer_per_cpu * cpu_buffer)1664 static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1665 {
1666 	struct list_head *head = cpu_buffer->pages;
1667 	struct buffer_page *bpage, *tmp;
1668 
1669 	free_buffer_page(cpu_buffer->reader_page);
1670 
1671 	if (head) {
1672 		rb_head_page_deactivate(cpu_buffer);
1673 
1674 		list_for_each_entry_safe(bpage, tmp, head, list) {
1675 			list_del_init(&bpage->list);
1676 			free_buffer_page(bpage);
1677 		}
1678 		bpage = list_entry(head, struct buffer_page, list);
1679 		free_buffer_page(bpage);
1680 	}
1681 
1682 	kfree(cpu_buffer);
1683 }
1684 
1685 /**
1686  * __ring_buffer_alloc - allocate a new ring_buffer
1687  * @size: the size in bytes per cpu that is needed.
1688  * @flags: attributes to set for the ring buffer.
1689  * @key: ring buffer reader_lock_key.
1690  *
1691  * Currently the only flag that is available is the RB_FL_OVERWRITE
1692  * flag. This flag means that the buffer will overwrite old data
1693  * when the buffer wraps. If this flag is not set, the buffer will
1694  * drop data when the tail hits the head.
1695  */
__ring_buffer_alloc(unsigned long size,unsigned flags,struct lock_class_key * key)1696 struct trace_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1697 					struct lock_class_key *key)
1698 {
1699 	struct trace_buffer *buffer;
1700 	long nr_pages;
1701 	int bsize;
1702 	int cpu;
1703 	int ret;
1704 
1705 	/* keep it in its own cache line */
1706 	buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1707 			 GFP_KERNEL);
1708 	if (!buffer)
1709 		return NULL;
1710 
1711 	if (!zalloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1712 		goto fail_free_buffer;
1713 
1714 	nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1715 	buffer->flags = flags;
1716 	buffer->clock = trace_clock_local;
1717 	buffer->reader_lock_key = key;
1718 
1719 	init_irq_work(&buffer->irq_work.work, rb_wake_up_waiters);
1720 	init_waitqueue_head(&buffer->irq_work.waiters);
1721 
1722 	/* need at least two pages */
1723 	if (nr_pages < 2)
1724 		nr_pages = 2;
1725 
1726 	buffer->cpus = nr_cpu_ids;
1727 
1728 	bsize = sizeof(void *) * nr_cpu_ids;
1729 	buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1730 				  GFP_KERNEL);
1731 	if (!buffer->buffers)
1732 		goto fail_free_cpumask;
1733 
1734 	cpu = raw_smp_processor_id();
1735 	cpumask_set_cpu(cpu, buffer->cpumask);
1736 	buffer->buffers[cpu] = rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
1737 	if (!buffer->buffers[cpu])
1738 		goto fail_free_buffers;
1739 
1740 	ret = cpuhp_state_add_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1741 	if (ret < 0)
1742 		goto fail_free_buffers;
1743 
1744 	mutex_init(&buffer->mutex);
1745 
1746 	return buffer;
1747 
1748  fail_free_buffers:
1749 	for_each_buffer_cpu(buffer, cpu) {
1750 		if (buffer->buffers[cpu])
1751 			rb_free_cpu_buffer(buffer->buffers[cpu]);
1752 	}
1753 	kfree(buffer->buffers);
1754 
1755  fail_free_cpumask:
1756 	free_cpumask_var(buffer->cpumask);
1757 
1758  fail_free_buffer:
1759 	kfree(buffer);
1760 	return NULL;
1761 }
1762 EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1763 
1764 /**
1765  * ring_buffer_free - free a ring buffer.
1766  * @buffer: the buffer to free.
1767  */
1768 void
ring_buffer_free(struct trace_buffer * buffer)1769 ring_buffer_free(struct trace_buffer *buffer)
1770 {
1771 	int cpu;
1772 
1773 	cpuhp_state_remove_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1774 
1775 	for_each_buffer_cpu(buffer, cpu)
1776 		rb_free_cpu_buffer(buffer->buffers[cpu]);
1777 
1778 	kfree(buffer->buffers);
1779 	free_cpumask_var(buffer->cpumask);
1780 
1781 	kfree(buffer);
1782 }
1783 EXPORT_SYMBOL_GPL(ring_buffer_free);
1784 
ring_buffer_set_clock(struct trace_buffer * buffer,u64 (* clock)(void))1785 void ring_buffer_set_clock(struct trace_buffer *buffer,
1786 			   u64 (*clock)(void))
1787 {
1788 	buffer->clock = clock;
1789 }
1790 
ring_buffer_set_time_stamp_abs(struct trace_buffer * buffer,bool abs)1791 void ring_buffer_set_time_stamp_abs(struct trace_buffer *buffer, bool abs)
1792 {
1793 	buffer->time_stamp_abs = abs;
1794 }
1795 
ring_buffer_time_stamp_abs(struct trace_buffer * buffer)1796 bool ring_buffer_time_stamp_abs(struct trace_buffer *buffer)
1797 {
1798 	return buffer->time_stamp_abs;
1799 }
1800 
1801 static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1802 
rb_page_entries(struct buffer_page * bpage)1803 static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1804 {
1805 	return local_read(&bpage->entries) & RB_WRITE_MASK;
1806 }
1807 
rb_page_write(struct buffer_page * bpage)1808 static inline unsigned long rb_page_write(struct buffer_page *bpage)
1809 {
1810 	return local_read(&bpage->write) & RB_WRITE_MASK;
1811 }
1812 
1813 static int
rb_remove_pages(struct ring_buffer_per_cpu * cpu_buffer,unsigned long nr_pages)1814 rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned long nr_pages)
1815 {
1816 	struct list_head *tail_page, *to_remove, *next_page;
1817 	struct buffer_page *to_remove_page, *tmp_iter_page;
1818 	struct buffer_page *last_page, *first_page;
1819 	unsigned long nr_removed;
1820 	unsigned long head_bit;
1821 	int page_entries;
1822 
1823 	head_bit = 0;
1824 
1825 	raw_spin_lock_irq(&cpu_buffer->reader_lock);
1826 	atomic_inc(&cpu_buffer->record_disabled);
1827 	/*
1828 	 * We don't race with the readers since we have acquired the reader
1829 	 * lock. We also don't race with writers after disabling recording.
1830 	 * This makes it easy to figure out the first and the last page to be
1831 	 * removed from the list. We unlink all the pages in between including
1832 	 * the first and last pages. This is done in a busy loop so that we
1833 	 * lose the least number of traces.
1834 	 * The pages are freed after we restart recording and unlock readers.
1835 	 */
1836 	tail_page = &cpu_buffer->tail_page->list;
1837 
1838 	/*
1839 	 * tail page might be on reader page, we remove the next page
1840 	 * from the ring buffer
1841 	 */
1842 	if (cpu_buffer->tail_page == cpu_buffer->reader_page)
1843 		tail_page = rb_list_head(tail_page->next);
1844 	to_remove = tail_page;
1845 
1846 	/* start of pages to remove */
1847 	first_page = list_entry(rb_list_head(to_remove->next),
1848 				struct buffer_page, list);
1849 
1850 	for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
1851 		to_remove = rb_list_head(to_remove)->next;
1852 		head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
1853 	}
1854 
1855 	next_page = rb_list_head(to_remove)->next;
1856 
1857 	/*
1858 	 * Now we remove all pages between tail_page and next_page.
1859 	 * Make sure that we have head_bit value preserved for the
1860 	 * next page
1861 	 */
1862 	tail_page->next = (struct list_head *)((unsigned long)next_page |
1863 						head_bit);
1864 	next_page = rb_list_head(next_page);
1865 	next_page->prev = tail_page;
1866 
1867 	/* make sure pages points to a valid page in the ring buffer */
1868 	cpu_buffer->pages = next_page;
1869 
1870 	/* update head page */
1871 	if (head_bit)
1872 		cpu_buffer->head_page = list_entry(next_page,
1873 						struct buffer_page, list);
1874 
1875 	/*
1876 	 * change read pointer to make sure any read iterators reset
1877 	 * themselves
1878 	 */
1879 	cpu_buffer->read = 0;
1880 
1881 	/* pages are removed, resume tracing and then free the pages */
1882 	atomic_dec(&cpu_buffer->record_disabled);
1883 	raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1884 
1885 	RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
1886 
1887 	/* last buffer page to remove */
1888 	last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
1889 				list);
1890 	tmp_iter_page = first_page;
1891 
1892 	do {
1893 		cond_resched();
1894 
1895 		to_remove_page = tmp_iter_page;
1896 		rb_inc_page(cpu_buffer, &tmp_iter_page);
1897 
1898 		/* update the counters */
1899 		page_entries = rb_page_entries(to_remove_page);
1900 		if (page_entries) {
1901 			/*
1902 			 * If something was added to this page, it was full
1903 			 * since it is not the tail page. So we deduct the
1904 			 * bytes consumed in ring buffer from here.
1905 			 * Increment overrun to account for the lost events.
1906 			 */
1907 			local_add(page_entries, &cpu_buffer->overrun);
1908 			local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1909 			local_inc(&cpu_buffer->pages_lost);
1910 		}
1911 
1912 		/*
1913 		 * We have already removed references to this list item, just
1914 		 * free up the buffer_page and its page
1915 		 */
1916 		free_buffer_page(to_remove_page);
1917 		nr_removed--;
1918 
1919 	} while (to_remove_page != last_page);
1920 
1921 	RB_WARN_ON(cpu_buffer, nr_removed);
1922 
1923 	return nr_removed == 0;
1924 }
1925 
1926 static int
rb_insert_pages(struct ring_buffer_per_cpu * cpu_buffer)1927 rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
1928 {
1929 	struct list_head *pages = &cpu_buffer->new_pages;
1930 	int retries, success;
1931 
1932 	raw_spin_lock_irq(&cpu_buffer->reader_lock);
1933 	/*
1934 	 * We are holding the reader lock, so the reader page won't be swapped
1935 	 * in the ring buffer. Now we are racing with the writer trying to
1936 	 * move head page and the tail page.
1937 	 * We are going to adapt the reader page update process where:
1938 	 * 1. We first splice the start and end of list of new pages between
1939 	 *    the head page and its previous page.
1940 	 * 2. We cmpxchg the prev_page->next to point from head page to the
1941 	 *    start of new pages list.
1942 	 * 3. Finally, we update the head->prev to the end of new list.
1943 	 *
1944 	 * We will try this process 10 times, to make sure that we don't keep
1945 	 * spinning.
1946 	 */
1947 	retries = 10;
1948 	success = 0;
1949 	while (retries--) {
1950 		struct list_head *head_page, *prev_page, *r;
1951 		struct list_head *last_page, *first_page;
1952 		struct list_head *head_page_with_bit;
1953 
1954 		head_page = &rb_set_head_page(cpu_buffer)->list;
1955 		if (!head_page)
1956 			break;
1957 		prev_page = head_page->prev;
1958 
1959 		first_page = pages->next;
1960 		last_page  = pages->prev;
1961 
1962 		head_page_with_bit = (struct list_head *)
1963 				     ((unsigned long)head_page | RB_PAGE_HEAD);
1964 
1965 		last_page->next = head_page_with_bit;
1966 		first_page->prev = prev_page;
1967 
1968 		r = cmpxchg(&prev_page->next, head_page_with_bit, first_page);
1969 
1970 		if (r == head_page_with_bit) {
1971 			/*
1972 			 * yay, we replaced the page pointer to our new list,
1973 			 * now, we just have to update to head page's prev
1974 			 * pointer to point to end of list
1975 			 */
1976 			head_page->prev = last_page;
1977 			success = 1;
1978 			break;
1979 		}
1980 	}
1981 
1982 	if (success)
1983 		INIT_LIST_HEAD(pages);
1984 	/*
1985 	 * If we weren't successful in adding in new pages, warn and stop
1986 	 * tracing
1987 	 */
1988 	RB_WARN_ON(cpu_buffer, !success);
1989 	raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1990 
1991 	/* free pages if they weren't inserted */
1992 	if (!success) {
1993 		struct buffer_page *bpage, *tmp;
1994 		list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1995 					 list) {
1996 			list_del_init(&bpage->list);
1997 			free_buffer_page(bpage);
1998 		}
1999 	}
2000 	return success;
2001 }
2002 
rb_update_pages(struct ring_buffer_per_cpu * cpu_buffer)2003 static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
2004 {
2005 	int success;
2006 
2007 	if (cpu_buffer->nr_pages_to_update > 0)
2008 		success = rb_insert_pages(cpu_buffer);
2009 	else
2010 		success = rb_remove_pages(cpu_buffer,
2011 					-cpu_buffer->nr_pages_to_update);
2012 
2013 	if (success)
2014 		cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
2015 }
2016 
update_pages_handler(struct work_struct * work)2017 static void update_pages_handler(struct work_struct *work)
2018 {
2019 	struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
2020 			struct ring_buffer_per_cpu, update_pages_work);
2021 	rb_update_pages(cpu_buffer);
2022 	complete(&cpu_buffer->update_done);
2023 }
2024 
2025 /**
2026  * ring_buffer_resize - resize the ring buffer
2027  * @buffer: the buffer to resize.
2028  * @size: the new size.
2029  * @cpu_id: the cpu buffer to resize
2030  *
2031  * Minimum size is 2 * BUF_PAGE_SIZE.
2032  *
2033  * Returns 0 on success and < 0 on failure.
2034  */
ring_buffer_resize(struct trace_buffer * buffer,unsigned long size,int cpu_id)2035 int ring_buffer_resize(struct trace_buffer *buffer, unsigned long size,
2036 			int cpu_id)
2037 {
2038 	struct ring_buffer_per_cpu *cpu_buffer;
2039 	unsigned long nr_pages;
2040 	int cpu, err;
2041 
2042 	/*
2043 	 * Always succeed at resizing a non-existent buffer:
2044 	 */
2045 	if (!buffer)
2046 		return 0;
2047 
2048 	/* Make sure the requested buffer exists */
2049 	if (cpu_id != RING_BUFFER_ALL_CPUS &&
2050 	    !cpumask_test_cpu(cpu_id, buffer->cpumask))
2051 		return 0;
2052 
2053 	nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
2054 
2055 	/* we need a minimum of two pages */
2056 	if (nr_pages < 2)
2057 		nr_pages = 2;
2058 
2059 	size = nr_pages * BUF_PAGE_SIZE;
2060 
2061 	/* prevent another thread from changing buffer sizes */
2062 	mutex_lock(&buffer->mutex);
2063 
2064 
2065 	if (cpu_id == RING_BUFFER_ALL_CPUS) {
2066 		/*
2067 		 * Don't succeed if resizing is disabled, as a reader might be
2068 		 * manipulating the ring buffer and is expecting a sane state while
2069 		 * this is true.
2070 		 */
2071 		for_each_buffer_cpu(buffer, cpu) {
2072 			cpu_buffer = buffer->buffers[cpu];
2073 			if (atomic_read(&cpu_buffer->resize_disabled)) {
2074 				err = -EBUSY;
2075 				goto out_err_unlock;
2076 			}
2077 		}
2078 
2079 		/* calculate the pages to update */
2080 		for_each_buffer_cpu(buffer, cpu) {
2081 			cpu_buffer = buffer->buffers[cpu];
2082 
2083 			cpu_buffer->nr_pages_to_update = nr_pages -
2084 							cpu_buffer->nr_pages;
2085 			/*
2086 			 * nothing more to do for removing pages or no update
2087 			 */
2088 			if (cpu_buffer->nr_pages_to_update <= 0)
2089 				continue;
2090 			/*
2091 			 * to add pages, make sure all new pages can be
2092 			 * allocated without receiving ENOMEM
2093 			 */
2094 			INIT_LIST_HEAD(&cpu_buffer->new_pages);
2095 			if (__rb_allocate_pages(cpu_buffer->nr_pages_to_update,
2096 						&cpu_buffer->new_pages, cpu)) {
2097 				/* not enough memory for new pages */
2098 				err = -ENOMEM;
2099 				goto out_err;
2100 			}
2101 		}
2102 
2103 		get_online_cpus();
2104 		/*
2105 		 * Fire off all the required work handlers
2106 		 * We can't schedule on offline CPUs, but it's not necessary
2107 		 * since we can change their buffer sizes without any race.
2108 		 */
2109 		for_each_buffer_cpu(buffer, cpu) {
2110 			cpu_buffer = buffer->buffers[cpu];
2111 			if (!cpu_buffer->nr_pages_to_update)
2112 				continue;
2113 
2114 			/* Can't run something on an offline CPU. */
2115 			if (!cpu_online(cpu)) {
2116 				rb_update_pages(cpu_buffer);
2117 				cpu_buffer->nr_pages_to_update = 0;
2118 			} else {
2119 				schedule_work_on(cpu,
2120 						&cpu_buffer->update_pages_work);
2121 			}
2122 		}
2123 
2124 		/* wait for all the updates to complete */
2125 		for_each_buffer_cpu(buffer, cpu) {
2126 			cpu_buffer = buffer->buffers[cpu];
2127 			if (!cpu_buffer->nr_pages_to_update)
2128 				continue;
2129 
2130 			if (cpu_online(cpu))
2131 				wait_for_completion(&cpu_buffer->update_done);
2132 			cpu_buffer->nr_pages_to_update = 0;
2133 		}
2134 
2135 		put_online_cpus();
2136 	} else {
2137 		/* Make sure this CPU has been initialized */
2138 		if (!cpumask_test_cpu(cpu_id, buffer->cpumask))
2139 			goto out;
2140 
2141 		cpu_buffer = buffer->buffers[cpu_id];
2142 
2143 		if (nr_pages == cpu_buffer->nr_pages)
2144 			goto out;
2145 
2146 		/*
2147 		 * Don't succeed if resizing is disabled, as a reader might be
2148 		 * manipulating the ring buffer and is expecting a sane state while
2149 		 * this is true.
2150 		 */
2151 		if (atomic_read(&cpu_buffer->resize_disabled)) {
2152 			err = -EBUSY;
2153 			goto out_err_unlock;
2154 		}
2155 
2156 		cpu_buffer->nr_pages_to_update = nr_pages -
2157 						cpu_buffer->nr_pages;
2158 
2159 		INIT_LIST_HEAD(&cpu_buffer->new_pages);
2160 		if (cpu_buffer->nr_pages_to_update > 0 &&
2161 			__rb_allocate_pages(cpu_buffer->nr_pages_to_update,
2162 					    &cpu_buffer->new_pages, cpu_id)) {
2163 			err = -ENOMEM;
2164 			goto out_err;
2165 		}
2166 
2167 		get_online_cpus();
2168 
2169 		/* Can't run something on an offline CPU. */
2170 		if (!cpu_online(cpu_id))
2171 			rb_update_pages(cpu_buffer);
2172 		else {
2173 			schedule_work_on(cpu_id,
2174 					 &cpu_buffer->update_pages_work);
2175 			wait_for_completion(&cpu_buffer->update_done);
2176 		}
2177 
2178 		cpu_buffer->nr_pages_to_update = 0;
2179 		put_online_cpus();
2180 	}
2181 
2182  out:
2183 	/*
2184 	 * The ring buffer resize can happen with the ring buffer
2185 	 * enabled, so that the update disturbs the tracing as little
2186 	 * as possible. But if the buffer is disabled, we do not need
2187 	 * to worry about that, and we can take the time to verify
2188 	 * that the buffer is not corrupt.
2189 	 */
2190 	if (atomic_read(&buffer->record_disabled)) {
2191 		atomic_inc(&buffer->record_disabled);
2192 		/*
2193 		 * Even though the buffer was disabled, we must make sure
2194 		 * that it is truly disabled before calling rb_check_pages.
2195 		 * There could have been a race between checking
2196 		 * record_disable and incrementing it.
2197 		 */
2198 		synchronize_rcu();
2199 		for_each_buffer_cpu(buffer, cpu) {
2200 			cpu_buffer = buffer->buffers[cpu];
2201 			rb_check_pages(cpu_buffer);
2202 		}
2203 		atomic_dec(&buffer->record_disabled);
2204 	}
2205 
2206 	mutex_unlock(&buffer->mutex);
2207 	return 0;
2208 
2209  out_err:
2210 	for_each_buffer_cpu(buffer, cpu) {
2211 		struct buffer_page *bpage, *tmp;
2212 
2213 		cpu_buffer = buffer->buffers[cpu];
2214 		cpu_buffer->nr_pages_to_update = 0;
2215 
2216 		if (list_empty(&cpu_buffer->new_pages))
2217 			continue;
2218 
2219 		list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
2220 					list) {
2221 			list_del_init(&bpage->list);
2222 			free_buffer_page(bpage);
2223 		}
2224 	}
2225  out_err_unlock:
2226 	mutex_unlock(&buffer->mutex);
2227 	return err;
2228 }
2229 EXPORT_SYMBOL_GPL(ring_buffer_resize);
2230 
ring_buffer_change_overwrite(struct trace_buffer * buffer,int val)2231 void ring_buffer_change_overwrite(struct trace_buffer *buffer, int val)
2232 {
2233 	mutex_lock(&buffer->mutex);
2234 	if (val)
2235 		buffer->flags |= RB_FL_OVERWRITE;
2236 	else
2237 		buffer->flags &= ~RB_FL_OVERWRITE;
2238 	mutex_unlock(&buffer->mutex);
2239 }
2240 EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
2241 
__rb_page_index(struct buffer_page * bpage,unsigned index)2242 static __always_inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
2243 {
2244 	return bpage->page->data + index;
2245 }
2246 
2247 static __always_inline struct ring_buffer_event *
rb_reader_event(struct ring_buffer_per_cpu * cpu_buffer)2248 rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
2249 {
2250 	return __rb_page_index(cpu_buffer->reader_page,
2251 			       cpu_buffer->reader_page->read);
2252 }
2253 
rb_page_commit(struct buffer_page * bpage)2254 static __always_inline unsigned rb_page_commit(struct buffer_page *bpage)
2255 {
2256 	return local_read(&bpage->page->commit);
2257 }
2258 
2259 static struct ring_buffer_event *
rb_iter_head_event(struct ring_buffer_iter * iter)2260 rb_iter_head_event(struct ring_buffer_iter *iter)
2261 {
2262 	struct ring_buffer_event *event;
2263 	struct buffer_page *iter_head_page = iter->head_page;
2264 	unsigned long commit;
2265 	unsigned length;
2266 
2267 	if (iter->head != iter->next_event)
2268 		return iter->event;
2269 
2270 	/*
2271 	 * When the writer goes across pages, it issues a cmpxchg which
2272 	 * is a mb(), which will synchronize with the rmb here.
2273 	 * (see rb_tail_page_update() and __rb_reserve_next())
2274 	 */
2275 	commit = rb_page_commit(iter_head_page);
2276 	smp_rmb();
2277 	event = __rb_page_index(iter_head_page, iter->head);
2278 	length = rb_event_length(event);
2279 
2280 	/*
2281 	 * READ_ONCE() doesn't work on functions and we don't want the
2282 	 * compiler doing any crazy optimizations with length.
2283 	 */
2284 	barrier();
2285 
2286 	if ((iter->head + length) > commit || length > BUF_MAX_DATA_SIZE)
2287 		/* Writer corrupted the read? */
2288 		goto reset;
2289 
2290 	memcpy(iter->event, event, length);
2291 	/*
2292 	 * If the page stamp is still the same after this rmb() then the
2293 	 * event was safely copied without the writer entering the page.
2294 	 */
2295 	smp_rmb();
2296 
2297 	/* Make sure the page didn't change since we read this */
2298 	if (iter->page_stamp != iter_head_page->page->time_stamp ||
2299 	    commit > rb_page_commit(iter_head_page))
2300 		goto reset;
2301 
2302 	iter->next_event = iter->head + length;
2303 	return iter->event;
2304  reset:
2305 	/* Reset to the beginning */
2306 	iter->page_stamp = iter->read_stamp = iter->head_page->page->time_stamp;
2307 	iter->head = 0;
2308 	iter->next_event = 0;
2309 	iter->missed_events = 1;
2310 	return NULL;
2311 }
2312 
2313 /* Size is determined by what has been committed */
rb_page_size(struct buffer_page * bpage)2314 static __always_inline unsigned rb_page_size(struct buffer_page *bpage)
2315 {
2316 	return rb_page_commit(bpage);
2317 }
2318 
2319 static __always_inline unsigned
rb_commit_index(struct ring_buffer_per_cpu * cpu_buffer)2320 rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
2321 {
2322 	return rb_page_commit(cpu_buffer->commit_page);
2323 }
2324 
2325 static __always_inline unsigned
rb_event_index(struct ring_buffer_event * event)2326 rb_event_index(struct ring_buffer_event *event)
2327 {
2328 	unsigned long addr = (unsigned long)event;
2329 
2330 	return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
2331 }
2332 
rb_inc_iter(struct ring_buffer_iter * iter)2333 static void rb_inc_iter(struct ring_buffer_iter *iter)
2334 {
2335 	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
2336 
2337 	/*
2338 	 * The iterator could be on the reader page (it starts there).
2339 	 * But the head could have moved, since the reader was
2340 	 * found. Check for this case and assign the iterator
2341 	 * to the head page instead of next.
2342 	 */
2343 	if (iter->head_page == cpu_buffer->reader_page)
2344 		iter->head_page = rb_set_head_page(cpu_buffer);
2345 	else
2346 		rb_inc_page(cpu_buffer, &iter->head_page);
2347 
2348 	iter->page_stamp = iter->read_stamp = iter->head_page->page->time_stamp;
2349 	iter->head = 0;
2350 	iter->next_event = 0;
2351 }
2352 
2353 /*
2354  * rb_handle_head_page - writer hit the head page
2355  *
2356  * Returns: +1 to retry page
2357  *           0 to continue
2358  *          -1 on error
2359  */
2360 static int
rb_handle_head_page(struct ring_buffer_per_cpu * cpu_buffer,struct buffer_page * tail_page,struct buffer_page * next_page)2361 rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
2362 		    struct buffer_page *tail_page,
2363 		    struct buffer_page *next_page)
2364 {
2365 	struct buffer_page *new_head;
2366 	int entries;
2367 	int type;
2368 	int ret;
2369 
2370 	entries = rb_page_entries(next_page);
2371 
2372 	/*
2373 	 * The hard part is here. We need to move the head
2374 	 * forward, and protect against both readers on
2375 	 * other CPUs and writers coming in via interrupts.
2376 	 */
2377 	type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
2378 				       RB_PAGE_HEAD);
2379 
2380 	/*
2381 	 * type can be one of four:
2382 	 *  NORMAL - an interrupt already moved it for us
2383 	 *  HEAD   - we are the first to get here.
2384 	 *  UPDATE - we are the interrupt interrupting
2385 	 *           a current move.
2386 	 *  MOVED  - a reader on another CPU moved the next
2387 	 *           pointer to its reader page. Give up
2388 	 *           and try again.
2389 	 */
2390 
2391 	switch (type) {
2392 	case RB_PAGE_HEAD:
2393 		/*
2394 		 * We changed the head to UPDATE, thus
2395 		 * it is our responsibility to update
2396 		 * the counters.
2397 		 */
2398 		local_add(entries, &cpu_buffer->overrun);
2399 		local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
2400 		local_inc(&cpu_buffer->pages_lost);
2401 
2402 		/*
2403 		 * The entries will be zeroed out when we move the
2404 		 * tail page.
2405 		 */
2406 
2407 		/* still more to do */
2408 		break;
2409 
2410 	case RB_PAGE_UPDATE:
2411 		/*
2412 		 * This is an interrupt that interrupt the
2413 		 * previous update. Still more to do.
2414 		 */
2415 		break;
2416 	case RB_PAGE_NORMAL:
2417 		/*
2418 		 * An interrupt came in before the update
2419 		 * and processed this for us.
2420 		 * Nothing left to do.
2421 		 */
2422 		return 1;
2423 	case RB_PAGE_MOVED:
2424 		/*
2425 		 * The reader is on another CPU and just did
2426 		 * a swap with our next_page.
2427 		 * Try again.
2428 		 */
2429 		return 1;
2430 	default:
2431 		RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
2432 		return -1;
2433 	}
2434 
2435 	/*
2436 	 * Now that we are here, the old head pointer is
2437 	 * set to UPDATE. This will keep the reader from
2438 	 * swapping the head page with the reader page.
2439 	 * The reader (on another CPU) will spin till
2440 	 * we are finished.
2441 	 *
2442 	 * We just need to protect against interrupts
2443 	 * doing the job. We will set the next pointer
2444 	 * to HEAD. After that, we set the old pointer
2445 	 * to NORMAL, but only if it was HEAD before.
2446 	 * otherwise we are an interrupt, and only
2447 	 * want the outer most commit to reset it.
2448 	 */
2449 	new_head = next_page;
2450 	rb_inc_page(cpu_buffer, &new_head);
2451 
2452 	ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
2453 				    RB_PAGE_NORMAL);
2454 
2455 	/*
2456 	 * Valid returns are:
2457 	 *  HEAD   - an interrupt came in and already set it.
2458 	 *  NORMAL - One of two things:
2459 	 *            1) We really set it.
2460 	 *            2) A bunch of interrupts came in and moved
2461 	 *               the page forward again.
2462 	 */
2463 	switch (ret) {
2464 	case RB_PAGE_HEAD:
2465 	case RB_PAGE_NORMAL:
2466 		/* OK */
2467 		break;
2468 	default:
2469 		RB_WARN_ON(cpu_buffer, 1);
2470 		return -1;
2471 	}
2472 
2473 	/*
2474 	 * It is possible that an interrupt came in,
2475 	 * set the head up, then more interrupts came in
2476 	 * and moved it again. When we get back here,
2477 	 * the page would have been set to NORMAL but we
2478 	 * just set it back to HEAD.
2479 	 *
2480 	 * How do you detect this? Well, if that happened
2481 	 * the tail page would have moved.
2482 	 */
2483 	if (ret == RB_PAGE_NORMAL) {
2484 		struct buffer_page *buffer_tail_page;
2485 
2486 		buffer_tail_page = READ_ONCE(cpu_buffer->tail_page);
2487 		/*
2488 		 * If the tail had moved passed next, then we need
2489 		 * to reset the pointer.
2490 		 */
2491 		if (buffer_tail_page != tail_page &&
2492 		    buffer_tail_page != next_page)
2493 			rb_head_page_set_normal(cpu_buffer, new_head,
2494 						next_page,
2495 						RB_PAGE_HEAD);
2496 	}
2497 
2498 	/*
2499 	 * If this was the outer most commit (the one that
2500 	 * changed the original pointer from HEAD to UPDATE),
2501 	 * then it is up to us to reset it to NORMAL.
2502 	 */
2503 	if (type == RB_PAGE_HEAD) {
2504 		ret = rb_head_page_set_normal(cpu_buffer, next_page,
2505 					      tail_page,
2506 					      RB_PAGE_UPDATE);
2507 		if (RB_WARN_ON(cpu_buffer,
2508 			       ret != RB_PAGE_UPDATE))
2509 			return -1;
2510 	}
2511 
2512 	return 0;
2513 }
2514 
2515 static inline void
rb_reset_tail(struct ring_buffer_per_cpu * cpu_buffer,unsigned long tail,struct rb_event_info * info)2516 rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2517 	      unsigned long tail, struct rb_event_info *info)
2518 {
2519 	struct buffer_page *tail_page = info->tail_page;
2520 	struct ring_buffer_event *event;
2521 	unsigned long length = info->length;
2522 
2523 	/*
2524 	 * Only the event that crossed the page boundary
2525 	 * must fill the old tail_page with padding.
2526 	 */
2527 	if (tail >= BUF_PAGE_SIZE) {
2528 		/*
2529 		 * If the page was filled, then we still need
2530 		 * to update the real_end. Reset it to zero
2531 		 * and the reader will ignore it.
2532 		 */
2533 		if (tail == BUF_PAGE_SIZE)
2534 			tail_page->real_end = 0;
2535 
2536 		local_sub(length, &tail_page->write);
2537 		return;
2538 	}
2539 
2540 	event = __rb_page_index(tail_page, tail);
2541 
2542 	/* account for padding bytes */
2543 	local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes);
2544 
2545 	/*
2546 	 * Save the original length to the meta data.
2547 	 * This will be used by the reader to add lost event
2548 	 * counter.
2549 	 */
2550 	tail_page->real_end = tail;
2551 
2552 	/*
2553 	 * If this event is bigger than the minimum size, then
2554 	 * we need to be careful that we don't subtract the
2555 	 * write counter enough to allow another writer to slip
2556 	 * in on this page.
2557 	 * We put in a discarded commit instead, to make sure
2558 	 * that this space is not used again.
2559 	 *
2560 	 * If we are less than the minimum size, we don't need to
2561 	 * worry about it.
2562 	 */
2563 	if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
2564 		/* No room for any events */
2565 
2566 		/* Mark the rest of the page with padding */
2567 		rb_event_set_padding(event);
2568 
2569 		/* Make sure the padding is visible before the write update */
2570 		smp_wmb();
2571 
2572 		/* Set the write back to the previous setting */
2573 		local_sub(length, &tail_page->write);
2574 		return;
2575 	}
2576 
2577 	/* Put in a discarded event */
2578 	event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
2579 	event->type_len = RINGBUF_TYPE_PADDING;
2580 	/* time delta must be non zero */
2581 	event->time_delta = 1;
2582 
2583 	/* Make sure the padding is visible before the tail_page->write update */
2584 	smp_wmb();
2585 
2586 	/* Set write to end of buffer */
2587 	length = (tail + length) - BUF_PAGE_SIZE;
2588 	local_sub(length, &tail_page->write);
2589 }
2590 
2591 static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer);
2592 
2593 /*
2594  * This is the slow path, force gcc not to inline it.
2595  */
2596 static noinline struct ring_buffer_event *
rb_move_tail(struct ring_buffer_per_cpu * cpu_buffer,unsigned long tail,struct rb_event_info * info)2597 rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2598 	     unsigned long tail, struct rb_event_info *info)
2599 {
2600 	struct buffer_page *tail_page = info->tail_page;
2601 	struct buffer_page *commit_page = cpu_buffer->commit_page;
2602 	struct trace_buffer *buffer = cpu_buffer->buffer;
2603 	struct buffer_page *next_page;
2604 	int ret;
2605 
2606 	next_page = tail_page;
2607 
2608 	rb_inc_page(cpu_buffer, &next_page);
2609 
2610 	/*
2611 	 * If for some reason, we had an interrupt storm that made
2612 	 * it all the way around the buffer, bail, and warn
2613 	 * about it.
2614 	 */
2615 	if (unlikely(next_page == commit_page)) {
2616 		local_inc(&cpu_buffer->commit_overrun);
2617 		goto out_reset;
2618 	}
2619 
2620 	/*
2621 	 * This is where the fun begins!
2622 	 *
2623 	 * We are fighting against races between a reader that
2624 	 * could be on another CPU trying to swap its reader
2625 	 * page with the buffer head.
2626 	 *
2627 	 * We are also fighting against interrupts coming in and
2628 	 * moving the head or tail on us as well.
2629 	 *
2630 	 * If the next page is the head page then we have filled
2631 	 * the buffer, unless the commit page is still on the
2632 	 * reader page.
2633 	 */
2634 	if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) {
2635 
2636 		/*
2637 		 * If the commit is not on the reader page, then
2638 		 * move the header page.
2639 		 */
2640 		if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2641 			/*
2642 			 * If we are not in overwrite mode,
2643 			 * this is easy, just stop here.
2644 			 */
2645 			if (!(buffer->flags & RB_FL_OVERWRITE)) {
2646 				local_inc(&cpu_buffer->dropped_events);
2647 				goto out_reset;
2648 			}
2649 
2650 			ret = rb_handle_head_page(cpu_buffer,
2651 						  tail_page,
2652 						  next_page);
2653 			if (ret < 0)
2654 				goto out_reset;
2655 			if (ret)
2656 				goto out_again;
2657 		} else {
2658 			/*
2659 			 * We need to be careful here too. The
2660 			 * commit page could still be on the reader
2661 			 * page. We could have a small buffer, and
2662 			 * have filled up the buffer with events
2663 			 * from interrupts and such, and wrapped.
2664 			 *
2665 			 * Note, if the tail page is also the on the
2666 			 * reader_page, we let it move out.
2667 			 */
2668 			if (unlikely((cpu_buffer->commit_page !=
2669 				      cpu_buffer->tail_page) &&
2670 				     (cpu_buffer->commit_page ==
2671 				      cpu_buffer->reader_page))) {
2672 				local_inc(&cpu_buffer->commit_overrun);
2673 				goto out_reset;
2674 			}
2675 		}
2676 	}
2677 
2678 	rb_tail_page_update(cpu_buffer, tail_page, next_page);
2679 
2680  out_again:
2681 
2682 	rb_reset_tail(cpu_buffer, tail, info);
2683 
2684 	/* Commit what we have for now. */
2685 	rb_end_commit(cpu_buffer);
2686 	/* rb_end_commit() decs committing */
2687 	local_inc(&cpu_buffer->committing);
2688 
2689 	/* fail and let the caller try again */
2690 	return ERR_PTR(-EAGAIN);
2691 
2692  out_reset:
2693 	/* reset write */
2694 	rb_reset_tail(cpu_buffer, tail, info);
2695 
2696 	return NULL;
2697 }
2698 
2699 /* Slow path */
2700 static struct ring_buffer_event *
rb_add_time_stamp(struct ring_buffer_event * event,u64 delta,bool abs)2701 rb_add_time_stamp(struct ring_buffer_event *event, u64 delta, bool abs)
2702 {
2703 	if (abs)
2704 		event->type_len = RINGBUF_TYPE_TIME_STAMP;
2705 	else
2706 		event->type_len = RINGBUF_TYPE_TIME_EXTEND;
2707 
2708 	/* Not the first event on the page, or not delta? */
2709 	if (abs || rb_event_index(event)) {
2710 		event->time_delta = delta & TS_MASK;
2711 		event->array[0] = delta >> TS_SHIFT;
2712 	} else {
2713 		/* nope, just zero it */
2714 		event->time_delta = 0;
2715 		event->array[0] = 0;
2716 	}
2717 
2718 	return skip_time_extend(event);
2719 }
2720 
2721 static inline bool rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2722 				     struct ring_buffer_event *event);
2723 
2724 #ifndef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
sched_clock_stable(void)2725 static inline bool sched_clock_stable(void)
2726 {
2727 	return true;
2728 }
2729 #endif
2730 
2731 static void
rb_check_timestamp(struct ring_buffer_per_cpu * cpu_buffer,struct rb_event_info * info)2732 rb_check_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2733 		   struct rb_event_info *info)
2734 {
2735 	u64 write_stamp;
2736 
2737 	WARN_ONCE(1, "Delta way too big! %llu ts=%llu before=%llu after=%llu write stamp=%llu\n%s",
2738 		  (unsigned long long)info->delta,
2739 		  (unsigned long long)info->ts,
2740 		  (unsigned long long)info->before,
2741 		  (unsigned long long)info->after,
2742 		  (unsigned long long)(rb_time_read(&cpu_buffer->write_stamp, &write_stamp) ? write_stamp : 0),
2743 		  sched_clock_stable() ? "" :
2744 		  "If you just came from a suspend/resume,\n"
2745 		  "please switch to the trace global clock:\n"
2746 		  "  echo global > /sys/kernel/debug/tracing/trace_clock\n"
2747 		  "or add trace_clock=global to the kernel command line\n");
2748 }
2749 
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)2750 static void rb_add_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2751 				      struct ring_buffer_event **event,
2752 				      struct rb_event_info *info,
2753 				      u64 *delta,
2754 				      unsigned int *length)
2755 {
2756 	bool abs = info->add_timestamp &
2757 		(RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE);
2758 
2759 	if (unlikely(info->delta > (1ULL << 59))) {
2760 		/* did the clock go backwards */
2761 		if (info->before == info->after && info->before > info->ts) {
2762 			/* not interrupted */
2763 			static int once;
2764 
2765 			/*
2766 			 * This is possible with a recalibrating of the TSC.
2767 			 * Do not produce a call stack, but just report it.
2768 			 */
2769 			if (!once) {
2770 				once++;
2771 				pr_warn("Ring buffer clock went backwards: %llu -> %llu\n",
2772 					info->before, info->ts);
2773 			}
2774 		} else
2775 			rb_check_timestamp(cpu_buffer, info);
2776 		if (!abs)
2777 			info->delta = 0;
2778 	}
2779 	*event = rb_add_time_stamp(*event, info->delta, abs);
2780 	*length -= RB_LEN_TIME_EXTEND;
2781 	*delta = 0;
2782 }
2783 
2784 /**
2785  * rb_update_event - update event type and data
2786  * @cpu_buffer: The per cpu buffer of the @event
2787  * @event: the event to update
2788  * @info: The info to update the @event with (contains length and delta)
2789  *
2790  * Update the type and data fields of the @event. The length
2791  * is the actual size that is written to the ring buffer,
2792  * and with this, we can determine what to place into the
2793  * data field.
2794  */
2795 static void
rb_update_event(struct ring_buffer_per_cpu * cpu_buffer,struct ring_buffer_event * event,struct rb_event_info * info)2796 rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
2797 		struct ring_buffer_event *event,
2798 		struct rb_event_info *info)
2799 {
2800 	unsigned length = info->length;
2801 	u64 delta = info->delta;
2802 
2803 	/*
2804 	 * If we need to add a timestamp, then we
2805 	 * add it to the start of the reserved space.
2806 	 */
2807 	if (unlikely(info->add_timestamp))
2808 		rb_add_timestamp(cpu_buffer, &event, info, &delta, &length);
2809 
2810 	event->time_delta = delta;
2811 	length -= RB_EVNT_HDR_SIZE;
2812 	if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
2813 		event->type_len = 0;
2814 		event->array[0] = length;
2815 	} else
2816 		event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
2817 }
2818 
rb_calculate_event_length(unsigned length)2819 static unsigned rb_calculate_event_length(unsigned length)
2820 {
2821 	struct ring_buffer_event event; /* Used only for sizeof array */
2822 
2823 	/* zero length can cause confusions */
2824 	if (!length)
2825 		length++;
2826 
2827 	if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
2828 		length += sizeof(event.array[0]);
2829 
2830 	length += RB_EVNT_HDR_SIZE;
2831 	length = ALIGN(length, RB_ARCH_ALIGNMENT);
2832 
2833 	/*
2834 	 * In case the time delta is larger than the 27 bits for it
2835 	 * in the header, we need to add a timestamp. If another
2836 	 * event comes in when trying to discard this one to increase
2837 	 * the length, then the timestamp will be added in the allocated
2838 	 * space of this event. If length is bigger than the size needed
2839 	 * for the TIME_EXTEND, then padding has to be used. The events
2840 	 * length must be either RB_LEN_TIME_EXTEND, or greater than or equal
2841 	 * to RB_LEN_TIME_EXTEND + 8, as 8 is the minimum size for padding.
2842 	 * As length is a multiple of 4, we only need to worry if it
2843 	 * is 12 (RB_LEN_TIME_EXTEND + 4).
2844 	 */
2845 	if (length == RB_LEN_TIME_EXTEND + RB_ALIGNMENT)
2846 		length += RB_ALIGNMENT;
2847 
2848 	return length;
2849 }
2850 
2851 static __always_inline bool
rb_event_is_commit(struct ring_buffer_per_cpu * cpu_buffer,struct ring_buffer_event * event)2852 rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2853 		   struct ring_buffer_event *event)
2854 {
2855 	unsigned long addr = (unsigned long)event;
2856 	unsigned long index;
2857 
2858 	index = rb_event_index(event);
2859 	addr &= PAGE_MASK;
2860 
2861 	return cpu_buffer->commit_page->page == (void *)addr &&
2862 		rb_commit_index(cpu_buffer) == index;
2863 }
2864 
rb_time_delta(struct ring_buffer_event * event)2865 static u64 rb_time_delta(struct ring_buffer_event *event)
2866 {
2867 	switch (event->type_len) {
2868 	case RINGBUF_TYPE_PADDING:
2869 		return 0;
2870 
2871 	case RINGBUF_TYPE_TIME_EXTEND:
2872 		return ring_buffer_event_time_stamp(event);
2873 
2874 	case RINGBUF_TYPE_TIME_STAMP:
2875 		return 0;
2876 
2877 	case RINGBUF_TYPE_DATA:
2878 		return event->time_delta;
2879 	default:
2880 		return 0;
2881 	}
2882 }
2883 
2884 static inline int
rb_try_to_discard(struct ring_buffer_per_cpu * cpu_buffer,struct ring_buffer_event * event)2885 rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2886 		  struct ring_buffer_event *event)
2887 {
2888 	unsigned long new_index, old_index;
2889 	struct buffer_page *bpage;
2890 	unsigned long index;
2891 	unsigned long addr;
2892 	u64 write_stamp;
2893 	u64 delta;
2894 
2895 	new_index = rb_event_index(event);
2896 	old_index = new_index + rb_event_ts_length(event);
2897 	addr = (unsigned long)event;
2898 	addr &= PAGE_MASK;
2899 
2900 	bpage = READ_ONCE(cpu_buffer->tail_page);
2901 
2902 	delta = rb_time_delta(event);
2903 
2904 	if (!rb_time_read(&cpu_buffer->write_stamp, &write_stamp))
2905 		return 0;
2906 
2907 	/* Make sure the write stamp is read before testing the location */
2908 	barrier();
2909 
2910 	if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2911 		unsigned long write_mask =
2912 			local_read(&bpage->write) & ~RB_WRITE_MASK;
2913 		unsigned long event_length = rb_event_length(event);
2914 
2915 		/* Something came in, can't discard */
2916 		if (!rb_time_cmpxchg(&cpu_buffer->write_stamp,
2917 				       write_stamp, write_stamp - delta))
2918 			return 0;
2919 
2920 		/*
2921 		 * It's possible that the event time delta is zero
2922 		 * (has the same time stamp as the previous event)
2923 		 * in which case write_stamp and before_stamp could
2924 		 * be the same. In such a case, force before_stamp
2925 		 * to be different than write_stamp. It doesn't
2926 		 * matter what it is, as long as its different.
2927 		 */
2928 		if (!delta)
2929 			rb_time_set(&cpu_buffer->before_stamp, 0);
2930 
2931 		/*
2932 		 * If an event were to come in now, it would see that the
2933 		 * write_stamp and the before_stamp are different, and assume
2934 		 * that this event just added itself before updating
2935 		 * the write stamp. The interrupting event will fix the
2936 		 * write stamp for us, and use the before stamp as its delta.
2937 		 */
2938 
2939 		/*
2940 		 * This is on the tail page. It is possible that
2941 		 * a write could come in and move the tail page
2942 		 * and write to the next page. That is fine
2943 		 * because we just shorten what is on this page.
2944 		 */
2945 		old_index += write_mask;
2946 		new_index += write_mask;
2947 		index = local_cmpxchg(&bpage->write, old_index, new_index);
2948 		if (index == old_index) {
2949 			/* update counters */
2950 			local_sub(event_length, &cpu_buffer->entries_bytes);
2951 			return 1;
2952 		}
2953 	}
2954 
2955 	/* could not discard */
2956 	return 0;
2957 }
2958 
rb_start_commit(struct ring_buffer_per_cpu * cpu_buffer)2959 static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2960 {
2961 	local_inc(&cpu_buffer->committing);
2962 	local_inc(&cpu_buffer->commits);
2963 }
2964 
2965 static __always_inline void
rb_set_commit_to_write(struct ring_buffer_per_cpu * cpu_buffer)2966 rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
2967 {
2968 	unsigned long max_count;
2969 
2970 	/*
2971 	 * We only race with interrupts and NMIs on this CPU.
2972 	 * If we own the commit event, then we can commit
2973 	 * all others that interrupted us, since the interruptions
2974 	 * are in stack format (they finish before they come
2975 	 * back to us). This allows us to do a simple loop to
2976 	 * assign the commit to the tail.
2977 	 */
2978  again:
2979 	max_count = cpu_buffer->nr_pages * 100;
2980 
2981 	while (cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)) {
2982 		if (RB_WARN_ON(cpu_buffer, !(--max_count)))
2983 			return;
2984 		if (RB_WARN_ON(cpu_buffer,
2985 			       rb_is_reader_page(cpu_buffer->tail_page)))
2986 			return;
2987 		local_set(&cpu_buffer->commit_page->page->commit,
2988 			  rb_page_write(cpu_buffer->commit_page));
2989 		rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
2990 		/* add barrier to keep gcc from optimizing too much */
2991 		barrier();
2992 	}
2993 	while (rb_commit_index(cpu_buffer) !=
2994 	       rb_page_write(cpu_buffer->commit_page)) {
2995 
2996 		local_set(&cpu_buffer->commit_page->page->commit,
2997 			  rb_page_write(cpu_buffer->commit_page));
2998 		RB_WARN_ON(cpu_buffer,
2999 			   local_read(&cpu_buffer->commit_page->page->commit) &
3000 			   ~RB_WRITE_MASK);
3001 		barrier();
3002 	}
3003 
3004 	/* again, keep gcc from optimizing */
3005 	barrier();
3006 
3007 	/*
3008 	 * If an interrupt came in just after the first while loop
3009 	 * and pushed the tail page forward, we will be left with
3010 	 * a dangling commit that will never go forward.
3011 	 */
3012 	if (unlikely(cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)))
3013 		goto again;
3014 }
3015 
rb_end_commit(struct ring_buffer_per_cpu * cpu_buffer)3016 static __always_inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
3017 {
3018 	unsigned long commits;
3019 
3020 	if (RB_WARN_ON(cpu_buffer,
3021 		       !local_read(&cpu_buffer->committing)))
3022 		return;
3023 
3024  again:
3025 	commits = local_read(&cpu_buffer->commits);
3026 	/* synchronize with interrupts */
3027 	barrier();
3028 	if (local_read(&cpu_buffer->committing) == 1)
3029 		rb_set_commit_to_write(cpu_buffer);
3030 
3031 	local_dec(&cpu_buffer->committing);
3032 
3033 	/* synchronize with interrupts */
3034 	barrier();
3035 
3036 	/*
3037 	 * Need to account for interrupts coming in between the
3038 	 * updating of the commit page and the clearing of the
3039 	 * committing counter.
3040 	 */
3041 	if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
3042 	    !local_read(&cpu_buffer->committing)) {
3043 		local_inc(&cpu_buffer->committing);
3044 		goto again;
3045 	}
3046 }
3047 
rb_event_discard(struct ring_buffer_event * event)3048 static inline void rb_event_discard(struct ring_buffer_event *event)
3049 {
3050 	if (extended_time(event))
3051 		event = skip_time_extend(event);
3052 
3053 	/* array[0] holds the actual length for the discarded event */
3054 	event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
3055 	event->type_len = RINGBUF_TYPE_PADDING;
3056 	/* time delta must be non zero */
3057 	if (!event->time_delta)
3058 		event->time_delta = 1;
3059 }
3060 
rb_commit(struct ring_buffer_per_cpu * cpu_buffer,struct ring_buffer_event * event)3061 static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
3062 		      struct ring_buffer_event *event)
3063 {
3064 	local_inc(&cpu_buffer->entries);
3065 	rb_end_commit(cpu_buffer);
3066 }
3067 
3068 static __always_inline void
rb_wakeups(struct trace_buffer * buffer,struct ring_buffer_per_cpu * cpu_buffer)3069 rb_wakeups(struct trace_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer)
3070 {
3071 	if (buffer->irq_work.waiters_pending) {
3072 		buffer->irq_work.waiters_pending = false;
3073 		/* irq_work_queue() supplies it's own memory barriers */
3074 		irq_work_queue(&buffer->irq_work.work);
3075 	}
3076 
3077 	if (cpu_buffer->irq_work.waiters_pending) {
3078 		cpu_buffer->irq_work.waiters_pending = false;
3079 		/* irq_work_queue() supplies it's own memory barriers */
3080 		irq_work_queue(&cpu_buffer->irq_work.work);
3081 	}
3082 
3083 	if (cpu_buffer->last_pages_touch == local_read(&cpu_buffer->pages_touched))
3084 		return;
3085 
3086 	if (cpu_buffer->reader_page == cpu_buffer->commit_page)
3087 		return;
3088 
3089 	if (!cpu_buffer->irq_work.full_waiters_pending)
3090 		return;
3091 
3092 	cpu_buffer->last_pages_touch = local_read(&cpu_buffer->pages_touched);
3093 
3094 	if (!full_hit(buffer, cpu_buffer->cpu, cpu_buffer->shortest_full))
3095 		return;
3096 
3097 	cpu_buffer->irq_work.wakeup_full = true;
3098 	cpu_buffer->irq_work.full_waiters_pending = false;
3099 	/* irq_work_queue() supplies it's own memory barriers */
3100 	irq_work_queue(&cpu_buffer->irq_work.work);
3101 }
3102 
3103 /*
3104  * The lock and unlock are done within a preempt disable section.
3105  * The current_context per_cpu variable can only be modified
3106  * by the current task between lock and unlock. But it can
3107  * be modified more than once via an interrupt. To pass this
3108  * information from the lock to the unlock without having to
3109  * access the 'in_interrupt()' functions again (which do show
3110  * a bit of overhead in something as critical as function tracing,
3111  * we use a bitmask trick.
3112  *
3113  *  bit 1 =  NMI context
3114  *  bit 2 =  IRQ context
3115  *  bit 3 =  SoftIRQ context
3116  *  bit 4 =  normal context.
3117  *
3118  * This works because this is the order of contexts that can
3119  * preempt other contexts. A SoftIRQ never preempts an IRQ
3120  * context.
3121  *
3122  * When the context is determined, the corresponding bit is
3123  * checked and set (if it was set, then a recursion of that context
3124  * happened).
3125  *
3126  * On unlock, we need to clear this bit. To do so, just subtract
3127  * 1 from the current_context and AND it to itself.
3128  *
3129  * (binary)
3130  *  101 - 1 = 100
3131  *  101 & 100 = 100 (clearing bit zero)
3132  *
3133  *  1010 - 1 = 1001
3134  *  1010 & 1001 = 1000 (clearing bit 1)
3135  *
3136  * The least significant bit can be cleared this way, and it
3137  * just so happens that it is the same bit corresponding to
3138  * the current context.
3139  *
3140  * Now the TRANSITION bit breaks the above slightly. The TRANSITION bit
3141  * is set when a recursion is detected at the current context, and if
3142  * the TRANSITION bit is already set, it will fail the recursion.
3143  * This is needed because there's a lag between the changing of
3144  * interrupt context and updating the preempt count. In this case,
3145  * a false positive will be found. To handle this, one extra recursion
3146  * is allowed, and this is done by the TRANSITION bit. If the TRANSITION
3147  * bit is already set, then it is considered a recursion and the function
3148  * ends. Otherwise, the TRANSITION bit is set, and that bit is returned.
3149  *
3150  * On the trace_recursive_unlock(), the TRANSITION bit will be the first
3151  * to be cleared. Even if it wasn't the context that set it. That is,
3152  * if an interrupt comes in while NORMAL bit is set and the ring buffer
3153  * is called before preempt_count() is updated, since the check will
3154  * be on the NORMAL bit, the TRANSITION bit will then be set. If an
3155  * NMI then comes in, it will set the NMI bit, but when the NMI code
3156  * does the trace_recursive_unlock() it will clear the TRANSTION bit
3157  * and leave the NMI bit set. But this is fine, because the interrupt
3158  * code that set the TRANSITION bit will then clear the NMI bit when it
3159  * calls trace_recursive_unlock(). If another NMI comes in, it will
3160  * set the TRANSITION bit and continue.
3161  *
3162  * Note: The TRANSITION bit only handles a single transition between context.
3163  */
3164 
3165 static __always_inline int
trace_recursive_lock(struct ring_buffer_per_cpu * cpu_buffer)3166 trace_recursive_lock(struct ring_buffer_per_cpu *cpu_buffer)
3167 {
3168 	unsigned int val = cpu_buffer->current_context;
3169 	unsigned long pc = preempt_count();
3170 	int bit;
3171 
3172 	if (!(pc & (NMI_MASK | HARDIRQ_MASK | SOFTIRQ_OFFSET)))
3173 		bit = RB_CTX_NORMAL;
3174 	else
3175 		bit = pc & NMI_MASK ? RB_CTX_NMI :
3176 			pc & HARDIRQ_MASK ? RB_CTX_IRQ : RB_CTX_SOFTIRQ;
3177 
3178 	if (unlikely(val & (1 << (bit + cpu_buffer->nest)))) {
3179 		/*
3180 		 * It is possible that this was called by transitioning
3181 		 * between interrupt context, and preempt_count() has not
3182 		 * been updated yet. In this case, use the TRANSITION bit.
3183 		 */
3184 		bit = RB_CTX_TRANSITION;
3185 		if (val & (1 << (bit + cpu_buffer->nest)))
3186 			return 1;
3187 	}
3188 
3189 	val |= (1 << (bit + cpu_buffer->nest));
3190 	cpu_buffer->current_context = val;
3191 
3192 	return 0;
3193 }
3194 
3195 static __always_inline void
trace_recursive_unlock(struct ring_buffer_per_cpu * cpu_buffer)3196 trace_recursive_unlock(struct ring_buffer_per_cpu *cpu_buffer)
3197 {
3198 	cpu_buffer->current_context &=
3199 		cpu_buffer->current_context - (1 << cpu_buffer->nest);
3200 }
3201 
3202 /* The recursive locking above uses 5 bits */
3203 #define NESTED_BITS 5
3204 
3205 /**
3206  * ring_buffer_nest_start - Allow to trace while nested
3207  * @buffer: The ring buffer to modify
3208  *
3209  * The ring buffer has a safety mechanism to prevent recursion.
3210  * But there may be a case where a trace needs to be done while
3211  * tracing something else. In this case, calling this function
3212  * will allow this function to nest within a currently active
3213  * ring_buffer_lock_reserve().
3214  *
3215  * Call this function before calling another ring_buffer_lock_reserve() and
3216  * call ring_buffer_nest_end() after the nested ring_buffer_unlock_commit().
3217  */
ring_buffer_nest_start(struct trace_buffer * buffer)3218 void ring_buffer_nest_start(struct trace_buffer *buffer)
3219 {
3220 	struct ring_buffer_per_cpu *cpu_buffer;
3221 	int cpu;
3222 
3223 	/* Enabled by ring_buffer_nest_end() */
3224 	preempt_disable_notrace();
3225 	cpu = raw_smp_processor_id();
3226 	cpu_buffer = buffer->buffers[cpu];
3227 	/* This is the shift value for the above recursive locking */
3228 	cpu_buffer->nest += NESTED_BITS;
3229 }
3230 
3231 /**
3232  * ring_buffer_nest_end - Allow to trace while nested
3233  * @buffer: The ring buffer to modify
3234  *
3235  * Must be called after ring_buffer_nest_start() and after the
3236  * ring_buffer_unlock_commit().
3237  */
ring_buffer_nest_end(struct trace_buffer * buffer)3238 void ring_buffer_nest_end(struct trace_buffer *buffer)
3239 {
3240 	struct ring_buffer_per_cpu *cpu_buffer;
3241 	int cpu;
3242 
3243 	/* disabled by ring_buffer_nest_start() */
3244 	cpu = raw_smp_processor_id();
3245 	cpu_buffer = buffer->buffers[cpu];
3246 	/* This is the shift value for the above recursive locking */
3247 	cpu_buffer->nest -= NESTED_BITS;
3248 	preempt_enable_notrace();
3249 }
3250 
3251 /**
3252  * ring_buffer_unlock_commit - commit a reserved
3253  * @buffer: The buffer to commit to
3254  * @event: The event pointer to commit.
3255  *
3256  * This commits the data to the ring buffer, and releases any locks held.
3257  *
3258  * Must be paired with ring_buffer_lock_reserve.
3259  */
ring_buffer_unlock_commit(struct trace_buffer * buffer,struct ring_buffer_event * event)3260 int ring_buffer_unlock_commit(struct trace_buffer *buffer,
3261 			      struct ring_buffer_event *event)
3262 {
3263 	struct ring_buffer_per_cpu *cpu_buffer;
3264 	int cpu = raw_smp_processor_id();
3265 
3266 	cpu_buffer = buffer->buffers[cpu];
3267 
3268 	rb_commit(cpu_buffer, event);
3269 
3270 	rb_wakeups(buffer, cpu_buffer);
3271 
3272 	trace_recursive_unlock(cpu_buffer);
3273 
3274 	preempt_enable_notrace();
3275 
3276 	return 0;
3277 }
3278 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
3279 
3280 static struct ring_buffer_event *
__rb_reserve_next(struct ring_buffer_per_cpu * cpu_buffer,struct rb_event_info * info)3281 __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
3282 		  struct rb_event_info *info)
3283 {
3284 	struct ring_buffer_event *event;
3285 	struct buffer_page *tail_page;
3286 	unsigned long tail, write, w;
3287 	bool a_ok;
3288 	bool b_ok;
3289 
3290 	/* Don't let the compiler play games with cpu_buffer->tail_page */
3291 	tail_page = info->tail_page = READ_ONCE(cpu_buffer->tail_page);
3292 
3293  /*A*/	w = local_read(&tail_page->write) & RB_WRITE_MASK;
3294 	barrier();
3295 	b_ok = rb_time_read(&cpu_buffer->before_stamp, &info->before);
3296 	a_ok = rb_time_read(&cpu_buffer->write_stamp, &info->after);
3297 	barrier();
3298 	info->ts = rb_time_stamp(cpu_buffer->buffer);
3299 
3300 	if ((info->add_timestamp & RB_ADD_STAMP_ABSOLUTE)) {
3301 		info->delta = info->ts;
3302 	} else {
3303 		/*
3304 		 * If interrupting an event time update, we may need an
3305 		 * absolute timestamp.
3306 		 * Don't bother if this is the start of a new page (w == 0).
3307 		 */
3308 		if (unlikely(!a_ok || !b_ok || (info->before != info->after && w))) {
3309 			info->add_timestamp |= RB_ADD_STAMP_FORCE | RB_ADD_STAMP_EXTEND;
3310 			info->length += RB_LEN_TIME_EXTEND;
3311 		} else {
3312 			info->delta = info->ts - info->after;
3313 			if (unlikely(test_time_stamp(info->delta))) {
3314 				info->add_timestamp |= RB_ADD_STAMP_EXTEND;
3315 				info->length += RB_LEN_TIME_EXTEND;
3316 			}
3317 		}
3318 	}
3319 
3320  /*B*/	rb_time_set(&cpu_buffer->before_stamp, info->ts);
3321 
3322  /*C*/	write = local_add_return(info->length, &tail_page->write);
3323 
3324 	/* set write to only the index of the write */
3325 	write &= RB_WRITE_MASK;
3326 
3327 	tail = write - info->length;
3328 
3329 	/* See if we shot pass the end of this buffer page */
3330 	if (unlikely(write > BUF_PAGE_SIZE)) {
3331 		/* before and after may now different, fix it up*/
3332 		b_ok = rb_time_read(&cpu_buffer->before_stamp, &info->before);
3333 		a_ok = rb_time_read(&cpu_buffer->write_stamp, &info->after);
3334 		if (a_ok && b_ok && info->before != info->after)
3335 			(void)rb_time_cmpxchg(&cpu_buffer->before_stamp,
3336 					      info->before, info->after);
3337 		return rb_move_tail(cpu_buffer, tail, info);
3338 	}
3339 
3340 	if (likely(tail == w)) {
3341 		u64 save_before;
3342 		bool s_ok;
3343 
3344 		/* Nothing interrupted us between A and C */
3345  /*D*/		rb_time_set(&cpu_buffer->write_stamp, info->ts);
3346 		barrier();
3347  /*E*/		s_ok = rb_time_read(&cpu_buffer->before_stamp, &save_before);
3348 		RB_WARN_ON(cpu_buffer, !s_ok);
3349 		if (likely(!(info->add_timestamp &
3350 			     (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE))))
3351 			/* This did not interrupt any time update */
3352 			info->delta = info->ts - info->after;
3353 		else
3354 			/* Just use full timestamp for inerrupting event */
3355 			info->delta = info->ts;
3356 		barrier();
3357 		if (unlikely(info->ts != save_before)) {
3358 			/* SLOW PATH - Interrupted between C and E */
3359 
3360 			a_ok = rb_time_read(&cpu_buffer->write_stamp, &info->after);
3361 			RB_WARN_ON(cpu_buffer, !a_ok);
3362 
3363 			/* Write stamp must only go forward */
3364 			if (save_before > info->after) {
3365 				/*
3366 				 * We do not care about the result, only that
3367 				 * it gets updated atomically.
3368 				 */
3369 				(void)rb_time_cmpxchg(&cpu_buffer->write_stamp,
3370 						      info->after, save_before);
3371 			}
3372 		}
3373 	} else {
3374 		u64 ts;
3375 		/* SLOW PATH - Interrupted between A and C */
3376 		a_ok = rb_time_read(&cpu_buffer->write_stamp, &info->after);
3377 		/* Was interrupted before here, write_stamp must be valid */
3378 		RB_WARN_ON(cpu_buffer, !a_ok);
3379 		ts = rb_time_stamp(cpu_buffer->buffer);
3380 		barrier();
3381  /*E*/		if (write == (local_read(&tail_page->write) & RB_WRITE_MASK) &&
3382 		    info->after < ts &&
3383 		    rb_time_cmpxchg(&cpu_buffer->write_stamp,
3384 				    info->after, ts)) {
3385 			/* Nothing came after this event between C and E */
3386 			info->delta = ts - info->after;
3387 			info->ts = ts;
3388 		} else {
3389 			/*
3390 			 * Interrupted beween C and E:
3391 			 * Lost the previous events time stamp. Just set the
3392 			 * delta to zero, and this will be the same time as
3393 			 * the event this event interrupted. And the events that
3394 			 * came after this will still be correct (as they would
3395 			 * have built their delta on the previous event.
3396 			 */
3397 			info->delta = 0;
3398 		}
3399 		info->add_timestamp &= ~RB_ADD_STAMP_FORCE;
3400 	}
3401 
3402 	/*
3403 	 * If this is the first commit on the page, then it has the same
3404 	 * timestamp as the page itself.
3405 	 */
3406 	if (unlikely(!tail && !(info->add_timestamp &
3407 				(RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE))))
3408 		info->delta = 0;
3409 
3410 	/* We reserved something on the buffer */
3411 
3412 	event = __rb_page_index(tail_page, tail);
3413 	rb_update_event(cpu_buffer, event, info);
3414 
3415 	local_inc(&tail_page->entries);
3416 
3417 	/*
3418 	 * If this is the first commit on the page, then update
3419 	 * its timestamp.
3420 	 */
3421 	if (unlikely(!tail))
3422 		tail_page->page->time_stamp = info->ts;
3423 
3424 	/* account for these added bytes */
3425 	local_add(info->length, &cpu_buffer->entries_bytes);
3426 
3427 	return event;
3428 }
3429 
3430 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)3431 rb_reserve_next_event(struct trace_buffer *buffer,
3432 		      struct ring_buffer_per_cpu *cpu_buffer,
3433 		      unsigned long length)
3434 {
3435 	struct ring_buffer_event *event;
3436 	struct rb_event_info info;
3437 	int nr_loops = 0;
3438 	int add_ts_default;
3439 
3440 	rb_start_commit(cpu_buffer);
3441 	/* The commit page can not change after this */
3442 
3443 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
3444 	/*
3445 	 * Due to the ability to swap a cpu buffer from a buffer
3446 	 * it is possible it was swapped before we committed.
3447 	 * (committing stops a swap). We check for it here and
3448 	 * if it happened, we have to fail the write.
3449 	 */
3450 	barrier();
3451 	if (unlikely(READ_ONCE(cpu_buffer->buffer) != buffer)) {
3452 		local_dec(&cpu_buffer->committing);
3453 		local_dec(&cpu_buffer->commits);
3454 		return NULL;
3455 	}
3456 #endif
3457 
3458 	info.length = rb_calculate_event_length(length);
3459 
3460 	if (ring_buffer_time_stamp_abs(cpu_buffer->buffer)) {
3461 		add_ts_default = RB_ADD_STAMP_ABSOLUTE;
3462 		info.length += RB_LEN_TIME_EXTEND;
3463 	} else {
3464 		add_ts_default = RB_ADD_STAMP_NONE;
3465 	}
3466 
3467  again:
3468 	info.add_timestamp = add_ts_default;
3469 	info.delta = 0;
3470 
3471 	/*
3472 	 * We allow for interrupts to reenter here and do a trace.
3473 	 * If one does, it will cause this original code to loop
3474 	 * back here. Even with heavy interrupts happening, this
3475 	 * should only happen a few times in a row. If this happens
3476 	 * 1000 times in a row, there must be either an interrupt
3477 	 * storm or we have something buggy.
3478 	 * Bail!
3479 	 */
3480 	if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
3481 		goto out_fail;
3482 
3483 	event = __rb_reserve_next(cpu_buffer, &info);
3484 
3485 	if (unlikely(PTR_ERR(event) == -EAGAIN)) {
3486 		if (info.add_timestamp & (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_EXTEND))
3487 			info.length -= RB_LEN_TIME_EXTEND;
3488 		goto again;
3489 	}
3490 
3491 	if (likely(event))
3492 		return event;
3493  out_fail:
3494 	rb_end_commit(cpu_buffer);
3495 	return NULL;
3496 }
3497 
3498 /**
3499  * ring_buffer_lock_reserve - reserve a part of the buffer
3500  * @buffer: the ring buffer to reserve from
3501  * @length: the length of the data to reserve (excluding event header)
3502  *
3503  * Returns a reserved event on the ring buffer to copy directly to.
3504  * The user of this interface will need to get the body to write into
3505  * and can use the ring_buffer_event_data() interface.
3506  *
3507  * The length is the length of the data needed, not the event length
3508  * which also includes the event header.
3509  *
3510  * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
3511  * If NULL is returned, then nothing has been allocated or locked.
3512  */
3513 struct ring_buffer_event *
ring_buffer_lock_reserve(struct trace_buffer * buffer,unsigned long length)3514 ring_buffer_lock_reserve(struct trace_buffer *buffer, unsigned long length)
3515 {
3516 	struct ring_buffer_per_cpu *cpu_buffer;
3517 	struct ring_buffer_event *event;
3518 	int cpu;
3519 
3520 	/* If we are tracing schedule, we don't want to recurse */
3521 	preempt_disable_notrace();
3522 
3523 	if (unlikely(atomic_read(&buffer->record_disabled)))
3524 		goto out;
3525 
3526 	cpu = raw_smp_processor_id();
3527 
3528 	if (unlikely(!cpumask_test_cpu(cpu, buffer->cpumask)))
3529 		goto out;
3530 
3531 	cpu_buffer = buffer->buffers[cpu];
3532 
3533 	if (unlikely(atomic_read(&cpu_buffer->record_disabled)))
3534 		goto out;
3535 
3536 	if (unlikely(length > BUF_MAX_DATA_SIZE))
3537 		goto out;
3538 
3539 	if (unlikely(trace_recursive_lock(cpu_buffer)))
3540 		goto out;
3541 
3542 	event = rb_reserve_next_event(buffer, cpu_buffer, length);
3543 	if (!event)
3544 		goto out_unlock;
3545 
3546 	return event;
3547 
3548  out_unlock:
3549 	trace_recursive_unlock(cpu_buffer);
3550  out:
3551 	preempt_enable_notrace();
3552 	return NULL;
3553 }
3554 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
3555 
3556 /*
3557  * Decrement the entries to the page that an event is on.
3558  * The event does not even need to exist, only the pointer
3559  * to the page it is on. This may only be called before the commit
3560  * takes place.
3561  */
3562 static inline void
rb_decrement_entry(struct ring_buffer_per_cpu * cpu_buffer,struct ring_buffer_event * event)3563 rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
3564 		   struct ring_buffer_event *event)
3565 {
3566 	unsigned long addr = (unsigned long)event;
3567 	struct buffer_page *bpage = cpu_buffer->commit_page;
3568 	struct buffer_page *start;
3569 
3570 	addr &= PAGE_MASK;
3571 
3572 	/* Do the likely case first */
3573 	if (likely(bpage->page == (void *)addr)) {
3574 		local_dec(&bpage->entries);
3575 		return;
3576 	}
3577 
3578 	/*
3579 	 * Because the commit page may be on the reader page we
3580 	 * start with the next page and check the end loop there.
3581 	 */
3582 	rb_inc_page(cpu_buffer, &bpage);
3583 	start = bpage;
3584 	do {
3585 		if (bpage->page == (void *)addr) {
3586 			local_dec(&bpage->entries);
3587 			return;
3588 		}
3589 		rb_inc_page(cpu_buffer, &bpage);
3590 	} while (bpage != start);
3591 
3592 	/* commit not part of this buffer?? */
3593 	RB_WARN_ON(cpu_buffer, 1);
3594 }
3595 
3596 /**
3597  * ring_buffer_commit_discard - discard an event that has not been committed
3598  * @buffer: the ring buffer
3599  * @event: non committed event to discard
3600  *
3601  * Sometimes an event that is in the ring buffer needs to be ignored.
3602  * This function lets the user discard an event in the ring buffer
3603  * and then that event will not be read later.
3604  *
3605  * This function only works if it is called before the item has been
3606  * committed. It will try to free the event from the ring buffer
3607  * if another event has not been added behind it.
3608  *
3609  * If another event has been added behind it, it will set the event
3610  * up as discarded, and perform the commit.
3611  *
3612  * If this function is called, do not call ring_buffer_unlock_commit on
3613  * the event.
3614  */
ring_buffer_discard_commit(struct trace_buffer * buffer,struct ring_buffer_event * event)3615 void ring_buffer_discard_commit(struct trace_buffer *buffer,
3616 				struct ring_buffer_event *event)
3617 {
3618 	struct ring_buffer_per_cpu *cpu_buffer;
3619 	int cpu;
3620 
3621 	/* The event is discarded regardless */
3622 	rb_event_discard(event);
3623 
3624 	cpu = smp_processor_id();
3625 	cpu_buffer = buffer->buffers[cpu];
3626 
3627 	/*
3628 	 * This must only be called if the event has not been
3629 	 * committed yet. Thus we can assume that preemption
3630 	 * is still disabled.
3631 	 */
3632 	RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
3633 
3634 	rb_decrement_entry(cpu_buffer, event);
3635 	if (rb_try_to_discard(cpu_buffer, event))
3636 		goto out;
3637 
3638  out:
3639 	rb_end_commit(cpu_buffer);
3640 
3641 	trace_recursive_unlock(cpu_buffer);
3642 
3643 	preempt_enable_notrace();
3644 
3645 }
3646 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
3647 
3648 /**
3649  * ring_buffer_write - write data to the buffer without reserving
3650  * @buffer: The ring buffer to write to.
3651  * @length: The length of the data being written (excluding the event header)
3652  * @data: The data to write to the buffer.
3653  *
3654  * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
3655  * one function. If you already have the data to write to the buffer, it
3656  * may be easier to simply call this function.
3657  *
3658  * Note, like ring_buffer_lock_reserve, the length is the length of the data
3659  * and not the length of the event which would hold the header.
3660  */
ring_buffer_write(struct trace_buffer * buffer,unsigned long length,void * data)3661 int ring_buffer_write(struct trace_buffer *buffer,
3662 		      unsigned long length,
3663 		      void *data)
3664 {
3665 	struct ring_buffer_per_cpu *cpu_buffer;
3666 	struct ring_buffer_event *event;
3667 	void *body;
3668 	int ret = -EBUSY;
3669 	int cpu;
3670 
3671 	preempt_disable_notrace();
3672 
3673 	if (atomic_read(&buffer->record_disabled))
3674 		goto out;
3675 
3676 	cpu = raw_smp_processor_id();
3677 
3678 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3679 		goto out;
3680 
3681 	cpu_buffer = buffer->buffers[cpu];
3682 
3683 	if (atomic_read(&cpu_buffer->record_disabled))
3684 		goto out;
3685 
3686 	if (length > BUF_MAX_DATA_SIZE)
3687 		goto out;
3688 
3689 	if (unlikely(trace_recursive_lock(cpu_buffer)))
3690 		goto out;
3691 
3692 	event = rb_reserve_next_event(buffer, cpu_buffer, length);
3693 	if (!event)
3694 		goto out_unlock;
3695 
3696 	body = rb_event_data(event);
3697 
3698 	memcpy(body, data, length);
3699 
3700 	rb_commit(cpu_buffer, event);
3701 
3702 	rb_wakeups(buffer, cpu_buffer);
3703 
3704 	ret = 0;
3705 
3706  out_unlock:
3707 	trace_recursive_unlock(cpu_buffer);
3708 
3709  out:
3710 	preempt_enable_notrace();
3711 
3712 	return ret;
3713 }
3714 EXPORT_SYMBOL_GPL(ring_buffer_write);
3715 
rb_per_cpu_empty(struct ring_buffer_per_cpu * cpu_buffer)3716 static bool rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
3717 {
3718 	struct buffer_page *reader = cpu_buffer->reader_page;
3719 	struct buffer_page *head = rb_set_head_page(cpu_buffer);
3720 	struct buffer_page *commit = cpu_buffer->commit_page;
3721 
3722 	/* In case of error, head will be NULL */
3723 	if (unlikely(!head))
3724 		return true;
3725 
3726 	/* Reader should exhaust content in reader page */
3727 	if (reader->read != rb_page_commit(reader))
3728 		return false;
3729 
3730 	/*
3731 	 * If writers are committing on the reader page, knowing all
3732 	 * committed content has been read, the ring buffer is empty.
3733 	 */
3734 	if (commit == reader)
3735 		return true;
3736 
3737 	/*
3738 	 * If writers are committing on a page other than reader page
3739 	 * and head page, there should always be content to read.
3740 	 */
3741 	if (commit != head)
3742 		return false;
3743 
3744 	/*
3745 	 * Writers are committing on the head page, we just need
3746 	 * to care about there're committed data, and the reader will
3747 	 * swap reader page with head page when it is to read data.
3748 	 */
3749 	return rb_page_commit(commit) == 0;
3750 }
3751 
3752 /**
3753  * ring_buffer_record_disable - stop all writes into the buffer
3754  * @buffer: The ring buffer to stop writes to.
3755  *
3756  * This prevents all writes to the buffer. Any attempt to write
3757  * to the buffer after this will fail and return NULL.
3758  *
3759  * The caller should call synchronize_rcu() after this.
3760  */
ring_buffer_record_disable(struct trace_buffer * buffer)3761 void ring_buffer_record_disable(struct trace_buffer *buffer)
3762 {
3763 	atomic_inc(&buffer->record_disabled);
3764 }
3765 EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
3766 
3767 /**
3768  * ring_buffer_record_enable - enable writes to the buffer
3769  * @buffer: The ring buffer to enable writes
3770  *
3771  * Note, multiple disables will need the same number of enables
3772  * to truly enable the writing (much like preempt_disable).
3773  */
ring_buffer_record_enable(struct trace_buffer * buffer)3774 void ring_buffer_record_enable(struct trace_buffer *buffer)
3775 {
3776 	atomic_dec(&buffer->record_disabled);
3777 }
3778 EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
3779 
3780 /**
3781  * ring_buffer_record_off - stop all writes into the buffer
3782  * @buffer: The ring buffer to stop writes to.
3783  *
3784  * This prevents all writes to the buffer. Any attempt to write
3785  * to the buffer after this will fail and return NULL.
3786  *
3787  * This is different than ring_buffer_record_disable() as
3788  * it works like an on/off switch, where as the disable() version
3789  * must be paired with a enable().
3790  */
ring_buffer_record_off(struct trace_buffer * buffer)3791 void ring_buffer_record_off(struct trace_buffer *buffer)
3792 {
3793 	unsigned int rd;
3794 	unsigned int new_rd;
3795 
3796 	do {
3797 		rd = atomic_read(&buffer->record_disabled);
3798 		new_rd = rd | RB_BUFFER_OFF;
3799 	} while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3800 }
3801 EXPORT_SYMBOL_GPL(ring_buffer_record_off);
3802 
3803 /**
3804  * ring_buffer_record_on - restart writes into the buffer
3805  * @buffer: The ring buffer to start writes to.
3806  *
3807  * This enables all writes to the buffer that was disabled by
3808  * ring_buffer_record_off().
3809  *
3810  * This is different than ring_buffer_record_enable() as
3811  * it works like an on/off switch, where as the enable() version
3812  * must be paired with a disable().
3813  */
ring_buffer_record_on(struct trace_buffer * buffer)3814 void ring_buffer_record_on(struct trace_buffer *buffer)
3815 {
3816 	unsigned int rd;
3817 	unsigned int new_rd;
3818 
3819 	do {
3820 		rd = atomic_read(&buffer->record_disabled);
3821 		new_rd = rd & ~RB_BUFFER_OFF;
3822 	} while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3823 }
3824 EXPORT_SYMBOL_GPL(ring_buffer_record_on);
3825 
3826 /**
3827  * ring_buffer_record_is_on - return true if the ring buffer can write
3828  * @buffer: The ring buffer to see if write is enabled
3829  *
3830  * Returns true if the ring buffer is in a state that it accepts writes.
3831  */
ring_buffer_record_is_on(struct trace_buffer * buffer)3832 bool ring_buffer_record_is_on(struct trace_buffer *buffer)
3833 {
3834 	return !atomic_read(&buffer->record_disabled);
3835 }
3836 
3837 /**
3838  * ring_buffer_record_is_set_on - return true if the ring buffer is set writable
3839  * @buffer: The ring buffer to see if write is set enabled
3840  *
3841  * Returns true if the ring buffer is set writable by ring_buffer_record_on().
3842  * Note that this does NOT mean it is in a writable state.
3843  *
3844  * It may return true when the ring buffer has been disabled by
3845  * ring_buffer_record_disable(), as that is a temporary disabling of
3846  * the ring buffer.
3847  */
ring_buffer_record_is_set_on(struct trace_buffer * buffer)3848 bool ring_buffer_record_is_set_on(struct trace_buffer *buffer)
3849 {
3850 	return !(atomic_read(&buffer->record_disabled) & RB_BUFFER_OFF);
3851 }
3852 
3853 /**
3854  * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
3855  * @buffer: The ring buffer to stop writes to.
3856  * @cpu: The CPU buffer to stop
3857  *
3858  * This prevents all writes to the buffer. Any attempt to write
3859  * to the buffer after this will fail and return NULL.
3860  *
3861  * The caller should call synchronize_rcu() after this.
3862  */
ring_buffer_record_disable_cpu(struct trace_buffer * buffer,int cpu)3863 void ring_buffer_record_disable_cpu(struct trace_buffer *buffer, int cpu)
3864 {
3865 	struct ring_buffer_per_cpu *cpu_buffer;
3866 
3867 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3868 		return;
3869 
3870 	cpu_buffer = buffer->buffers[cpu];
3871 	atomic_inc(&cpu_buffer->record_disabled);
3872 }
3873 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
3874 
3875 /**
3876  * ring_buffer_record_enable_cpu - enable writes to the buffer
3877  * @buffer: The ring buffer to enable writes
3878  * @cpu: The CPU to enable.
3879  *
3880  * Note, multiple disables will need the same number of enables
3881  * to truly enable the writing (much like preempt_disable).
3882  */
ring_buffer_record_enable_cpu(struct trace_buffer * buffer,int cpu)3883 void ring_buffer_record_enable_cpu(struct trace_buffer *buffer, int cpu)
3884 {
3885 	struct ring_buffer_per_cpu *cpu_buffer;
3886 
3887 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3888 		return;
3889 
3890 	cpu_buffer = buffer->buffers[cpu];
3891 	atomic_dec(&cpu_buffer->record_disabled);
3892 }
3893 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
3894 
3895 /*
3896  * The total entries in the ring buffer is the running counter
3897  * of entries entered into the ring buffer, minus the sum of
3898  * the entries read from the ring buffer and the number of
3899  * entries that were overwritten.
3900  */
3901 static inline unsigned long
rb_num_of_entries(struct ring_buffer_per_cpu * cpu_buffer)3902 rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
3903 {
3904 	return local_read(&cpu_buffer->entries) -
3905 		(local_read(&cpu_buffer->overrun) + cpu_buffer->read);
3906 }
3907 
3908 /**
3909  * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
3910  * @buffer: The ring buffer
3911  * @cpu: The per CPU buffer to read from.
3912  */
ring_buffer_oldest_event_ts(struct trace_buffer * buffer,int cpu)3913 u64 ring_buffer_oldest_event_ts(struct trace_buffer *buffer, int cpu)
3914 {
3915 	unsigned long flags;
3916 	struct ring_buffer_per_cpu *cpu_buffer;
3917 	struct buffer_page *bpage;
3918 	u64 ret = 0;
3919 
3920 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3921 		return 0;
3922 
3923 	cpu_buffer = buffer->buffers[cpu];
3924 	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3925 	/*
3926 	 * if the tail is on reader_page, oldest time stamp is on the reader
3927 	 * page
3928 	 */
3929 	if (cpu_buffer->tail_page == cpu_buffer->reader_page)
3930 		bpage = cpu_buffer->reader_page;
3931 	else
3932 		bpage = rb_set_head_page(cpu_buffer);
3933 	if (bpage)
3934 		ret = bpage->page->time_stamp;
3935 	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3936 
3937 	return ret;
3938 }
3939 EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
3940 
3941 /**
3942  * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
3943  * @buffer: The ring buffer
3944  * @cpu: The per CPU buffer to read from.
3945  */
ring_buffer_bytes_cpu(struct trace_buffer * buffer,int cpu)3946 unsigned long ring_buffer_bytes_cpu(struct trace_buffer *buffer, int cpu)
3947 {
3948 	struct ring_buffer_per_cpu *cpu_buffer;
3949 	unsigned long ret;
3950 
3951 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3952 		return 0;
3953 
3954 	cpu_buffer = buffer->buffers[cpu];
3955 	ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
3956 
3957 	return ret;
3958 }
3959 EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
3960 
3961 /**
3962  * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
3963  * @buffer: The ring buffer
3964  * @cpu: The per CPU buffer to get the entries from.
3965  */
ring_buffer_entries_cpu(struct trace_buffer * buffer,int cpu)3966 unsigned long ring_buffer_entries_cpu(struct trace_buffer *buffer, int cpu)
3967 {
3968 	struct ring_buffer_per_cpu *cpu_buffer;
3969 
3970 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3971 		return 0;
3972 
3973 	cpu_buffer = buffer->buffers[cpu];
3974 
3975 	return rb_num_of_entries(cpu_buffer);
3976 }
3977 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
3978 
3979 /**
3980  * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
3981  * buffer wrapping around (only if RB_FL_OVERWRITE is on).
3982  * @buffer: The ring buffer
3983  * @cpu: The per CPU buffer to get the number of overruns from
3984  */
ring_buffer_overrun_cpu(struct trace_buffer * buffer,int cpu)3985 unsigned long ring_buffer_overrun_cpu(struct trace_buffer *buffer, int cpu)
3986 {
3987 	struct ring_buffer_per_cpu *cpu_buffer;
3988 	unsigned long ret;
3989 
3990 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3991 		return 0;
3992 
3993 	cpu_buffer = buffer->buffers[cpu];
3994 	ret = local_read(&cpu_buffer->overrun);
3995 
3996 	return ret;
3997 }
3998 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
3999 
4000 /**
4001  * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
4002  * commits failing due to the buffer wrapping around while there are uncommitted
4003  * events, such as during an interrupt storm.
4004  * @buffer: The ring buffer
4005  * @cpu: The per CPU buffer to get the number of overruns from
4006  */
4007 unsigned long
ring_buffer_commit_overrun_cpu(struct trace_buffer * buffer,int cpu)4008 ring_buffer_commit_overrun_cpu(struct trace_buffer *buffer, int cpu)
4009 {
4010 	struct ring_buffer_per_cpu *cpu_buffer;
4011 	unsigned long ret;
4012 
4013 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4014 		return 0;
4015 
4016 	cpu_buffer = buffer->buffers[cpu];
4017 	ret = local_read(&cpu_buffer->commit_overrun);
4018 
4019 	return ret;
4020 }
4021 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
4022 
4023 /**
4024  * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
4025  * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
4026  * @buffer: The ring buffer
4027  * @cpu: The per CPU buffer to get the number of overruns from
4028  */
4029 unsigned long
ring_buffer_dropped_events_cpu(struct trace_buffer * buffer,int cpu)4030 ring_buffer_dropped_events_cpu(struct trace_buffer *buffer, int cpu)
4031 {
4032 	struct ring_buffer_per_cpu *cpu_buffer;
4033 	unsigned long ret;
4034 
4035 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4036 		return 0;
4037 
4038 	cpu_buffer = buffer->buffers[cpu];
4039 	ret = local_read(&cpu_buffer->dropped_events);
4040 
4041 	return ret;
4042 }
4043 EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
4044 
4045 /**
4046  * ring_buffer_read_events_cpu - get the number of events successfully read
4047  * @buffer: The ring buffer
4048  * @cpu: The per CPU buffer to get the number of events read
4049  */
4050 unsigned long
ring_buffer_read_events_cpu(struct trace_buffer * buffer,int cpu)4051 ring_buffer_read_events_cpu(struct trace_buffer *buffer, int cpu)
4052 {
4053 	struct ring_buffer_per_cpu *cpu_buffer;
4054 
4055 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4056 		return 0;
4057 
4058 	cpu_buffer = buffer->buffers[cpu];
4059 	return cpu_buffer->read;
4060 }
4061 EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu);
4062 
4063 /**
4064  * ring_buffer_entries - get the number of entries in a buffer
4065  * @buffer: The ring buffer
4066  *
4067  * Returns the total number of entries in the ring buffer
4068  * (all CPU entries)
4069  */
ring_buffer_entries(struct trace_buffer * buffer)4070 unsigned long ring_buffer_entries(struct trace_buffer *buffer)
4071 {
4072 	struct ring_buffer_per_cpu *cpu_buffer;
4073 	unsigned long entries = 0;
4074 	int cpu;
4075 
4076 	/* if you care about this being correct, lock the buffer */
4077 	for_each_buffer_cpu(buffer, cpu) {
4078 		cpu_buffer = buffer->buffers[cpu];
4079 		entries += rb_num_of_entries(cpu_buffer);
4080 	}
4081 
4082 	return entries;
4083 }
4084 EXPORT_SYMBOL_GPL(ring_buffer_entries);
4085 
4086 /**
4087  * ring_buffer_overruns - get the number of overruns in buffer
4088  * @buffer: The ring buffer
4089  *
4090  * Returns the total number of overruns in the ring buffer
4091  * (all CPU entries)
4092  */
ring_buffer_overruns(struct trace_buffer * buffer)4093 unsigned long ring_buffer_overruns(struct trace_buffer *buffer)
4094 {
4095 	struct ring_buffer_per_cpu *cpu_buffer;
4096 	unsigned long overruns = 0;
4097 	int cpu;
4098 
4099 	/* if you care about this being correct, lock the buffer */
4100 	for_each_buffer_cpu(buffer, cpu) {
4101 		cpu_buffer = buffer->buffers[cpu];
4102 		overruns += local_read(&cpu_buffer->overrun);
4103 	}
4104 
4105 	return overruns;
4106 }
4107 EXPORT_SYMBOL_GPL(ring_buffer_overruns);
4108 
rb_iter_reset(struct ring_buffer_iter * iter)4109 static void rb_iter_reset(struct ring_buffer_iter *iter)
4110 {
4111 	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4112 
4113 	/* Iterator usage is expected to have record disabled */
4114 	iter->head_page = cpu_buffer->reader_page;
4115 	iter->head = cpu_buffer->reader_page->read;
4116 	iter->next_event = iter->head;
4117 
4118 	iter->cache_reader_page = iter->head_page;
4119 	iter->cache_read = cpu_buffer->read;
4120 
4121 	if (iter->head) {
4122 		iter->read_stamp = cpu_buffer->read_stamp;
4123 		iter->page_stamp = cpu_buffer->reader_page->page->time_stamp;
4124 	} else {
4125 		iter->read_stamp = iter->head_page->page->time_stamp;
4126 		iter->page_stamp = iter->read_stamp;
4127 	}
4128 }
4129 
4130 /**
4131  * ring_buffer_iter_reset - reset an iterator
4132  * @iter: The iterator to reset
4133  *
4134  * Resets the iterator, so that it will start from the beginning
4135  * again.
4136  */
ring_buffer_iter_reset(struct ring_buffer_iter * iter)4137 void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
4138 {
4139 	struct ring_buffer_per_cpu *cpu_buffer;
4140 	unsigned long flags;
4141 
4142 	if (!iter)
4143 		return;
4144 
4145 	cpu_buffer = iter->cpu_buffer;
4146 
4147 	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4148 	rb_iter_reset(iter);
4149 	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4150 }
4151 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
4152 
4153 /**
4154  * ring_buffer_iter_empty - check if an iterator has no more to read
4155  * @iter: The iterator to check
4156  */
ring_buffer_iter_empty(struct ring_buffer_iter * iter)4157 int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
4158 {
4159 	struct ring_buffer_per_cpu *cpu_buffer;
4160 	struct buffer_page *reader;
4161 	struct buffer_page *head_page;
4162 	struct buffer_page *commit_page;
4163 	struct buffer_page *curr_commit_page;
4164 	unsigned commit;
4165 	u64 curr_commit_ts;
4166 	u64 commit_ts;
4167 
4168 	cpu_buffer = iter->cpu_buffer;
4169 	reader = cpu_buffer->reader_page;
4170 	head_page = cpu_buffer->head_page;
4171 	commit_page = cpu_buffer->commit_page;
4172 	commit_ts = commit_page->page->time_stamp;
4173 
4174 	/*
4175 	 * When the writer goes across pages, it issues a cmpxchg which
4176 	 * is a mb(), which will synchronize with the rmb here.
4177 	 * (see rb_tail_page_update())
4178 	 */
4179 	smp_rmb();
4180 	commit = rb_page_commit(commit_page);
4181 	/* We want to make sure that the commit page doesn't change */
4182 	smp_rmb();
4183 
4184 	/* Make sure commit page didn't change */
4185 	curr_commit_page = READ_ONCE(cpu_buffer->commit_page);
4186 	curr_commit_ts = READ_ONCE(curr_commit_page->page->time_stamp);
4187 
4188 	/* If the commit page changed, then there's more data */
4189 	if (curr_commit_page != commit_page ||
4190 	    curr_commit_ts != commit_ts)
4191 		return 0;
4192 
4193 	/* Still racy, as it may return a false positive, but that's OK */
4194 	return ((iter->head_page == commit_page && iter->head >= commit) ||
4195 		(iter->head_page == reader && commit_page == head_page &&
4196 		 head_page->read == commit &&
4197 		 iter->head == rb_page_commit(cpu_buffer->reader_page)));
4198 }
4199 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
4200 
4201 static void
rb_update_read_stamp(struct ring_buffer_per_cpu * cpu_buffer,struct ring_buffer_event * event)4202 rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
4203 		     struct ring_buffer_event *event)
4204 {
4205 	u64 delta;
4206 
4207 	switch (event->type_len) {
4208 	case RINGBUF_TYPE_PADDING:
4209 		return;
4210 
4211 	case RINGBUF_TYPE_TIME_EXTEND:
4212 		delta = ring_buffer_event_time_stamp(event);
4213 		cpu_buffer->read_stamp += delta;
4214 		return;
4215 
4216 	case RINGBUF_TYPE_TIME_STAMP:
4217 		delta = ring_buffer_event_time_stamp(event);
4218 		cpu_buffer->read_stamp = delta;
4219 		return;
4220 
4221 	case RINGBUF_TYPE_DATA:
4222 		cpu_buffer->read_stamp += event->time_delta;
4223 		return;
4224 
4225 	default:
4226 		RB_WARN_ON(cpu_buffer, 1);
4227 	}
4228 	return;
4229 }
4230 
4231 static void
rb_update_iter_read_stamp(struct ring_buffer_iter * iter,struct ring_buffer_event * event)4232 rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
4233 			  struct ring_buffer_event *event)
4234 {
4235 	u64 delta;
4236 
4237 	switch (event->type_len) {
4238 	case RINGBUF_TYPE_PADDING:
4239 		return;
4240 
4241 	case RINGBUF_TYPE_TIME_EXTEND:
4242 		delta = ring_buffer_event_time_stamp(event);
4243 		iter->read_stamp += delta;
4244 		return;
4245 
4246 	case RINGBUF_TYPE_TIME_STAMP:
4247 		delta = ring_buffer_event_time_stamp(event);
4248 		iter->read_stamp = delta;
4249 		return;
4250 
4251 	case RINGBUF_TYPE_DATA:
4252 		iter->read_stamp += event->time_delta;
4253 		return;
4254 
4255 	default:
4256 		RB_WARN_ON(iter->cpu_buffer, 1);
4257 	}
4258 	return;
4259 }
4260 
4261 static struct buffer_page *
rb_get_reader_page(struct ring_buffer_per_cpu * cpu_buffer)4262 rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
4263 {
4264 	struct buffer_page *reader = NULL;
4265 	unsigned long overwrite;
4266 	unsigned long flags;
4267 	int nr_loops = 0;
4268 	int ret;
4269 
4270 	local_irq_save(flags);
4271 	arch_spin_lock(&cpu_buffer->lock);
4272 
4273  again:
4274 	/*
4275 	 * This should normally only loop twice. But because the
4276 	 * start of the reader inserts an empty page, it causes
4277 	 * a case where we will loop three times. There should be no
4278 	 * reason to loop four times (that I know of).
4279 	 */
4280 	if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
4281 		reader = NULL;
4282 		goto out;
4283 	}
4284 
4285 	reader = cpu_buffer->reader_page;
4286 
4287 	/* If there's more to read, return this page */
4288 	if (cpu_buffer->reader_page->read < rb_page_size(reader))
4289 		goto out;
4290 
4291 	/* Never should we have an index greater than the size */
4292 	if (RB_WARN_ON(cpu_buffer,
4293 		       cpu_buffer->reader_page->read > rb_page_size(reader)))
4294 		goto out;
4295 
4296 	/* check if we caught up to the tail */
4297 	reader = NULL;
4298 	if (cpu_buffer->commit_page == cpu_buffer->reader_page)
4299 		goto out;
4300 
4301 	/* Don't bother swapping if the ring buffer is empty */
4302 	if (rb_num_of_entries(cpu_buffer) == 0)
4303 		goto out;
4304 
4305 	/*
4306 	 * Reset the reader page to size zero.
4307 	 */
4308 	local_set(&cpu_buffer->reader_page->write, 0);
4309 	local_set(&cpu_buffer->reader_page->entries, 0);
4310 	local_set(&cpu_buffer->reader_page->page->commit, 0);
4311 	cpu_buffer->reader_page->real_end = 0;
4312 
4313  spin:
4314 	/*
4315 	 * Splice the empty reader page into the list around the head.
4316 	 */
4317 	reader = rb_set_head_page(cpu_buffer);
4318 	if (!reader)
4319 		goto out;
4320 	cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
4321 	cpu_buffer->reader_page->list.prev = reader->list.prev;
4322 
4323 	/*
4324 	 * cpu_buffer->pages just needs to point to the buffer, it
4325 	 *  has no specific buffer page to point to. Lets move it out
4326 	 *  of our way so we don't accidentally swap it.
4327 	 */
4328 	cpu_buffer->pages = reader->list.prev;
4329 
4330 	/* The reader page will be pointing to the new head */
4331 	rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list);
4332 
4333 	/*
4334 	 * We want to make sure we read the overruns after we set up our
4335 	 * pointers to the next object. The writer side does a
4336 	 * cmpxchg to cross pages which acts as the mb on the writer
4337 	 * side. Note, the reader will constantly fail the swap
4338 	 * while the writer is updating the pointers, so this
4339 	 * guarantees that the overwrite recorded here is the one we
4340 	 * want to compare with the last_overrun.
4341 	 */
4342 	smp_mb();
4343 	overwrite = local_read(&(cpu_buffer->overrun));
4344 
4345 	/*
4346 	 * Here's the tricky part.
4347 	 *
4348 	 * We need to move the pointer past the header page.
4349 	 * But we can only do that if a writer is not currently
4350 	 * moving it. The page before the header page has the
4351 	 * flag bit '1' set if it is pointing to the page we want.
4352 	 * but if the writer is in the process of moving it
4353 	 * than it will be '2' or already moved '0'.
4354 	 */
4355 
4356 	ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
4357 
4358 	/*
4359 	 * If we did not convert it, then we must try again.
4360 	 */
4361 	if (!ret)
4362 		goto spin;
4363 
4364 	/*
4365 	 * Yay! We succeeded in replacing the page.
4366 	 *
4367 	 * Now make the new head point back to the reader page.
4368 	 */
4369 	rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
4370 	rb_inc_page(cpu_buffer, &cpu_buffer->head_page);
4371 
4372 	local_inc(&cpu_buffer->pages_read);
4373 
4374 	/* Finally update the reader page to the new head */
4375 	cpu_buffer->reader_page = reader;
4376 	cpu_buffer->reader_page->read = 0;
4377 
4378 	if (overwrite != cpu_buffer->last_overrun) {
4379 		cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
4380 		cpu_buffer->last_overrun = overwrite;
4381 	}
4382 
4383 	goto again;
4384 
4385  out:
4386 	/* Update the read_stamp on the first event */
4387 	if (reader && reader->read == 0)
4388 		cpu_buffer->read_stamp = reader->page->time_stamp;
4389 
4390 	arch_spin_unlock(&cpu_buffer->lock);
4391 	local_irq_restore(flags);
4392 
4393 	/*
4394 	 * The writer has preempt disable, wait for it. But not forever
4395 	 * Although, 1 second is pretty much "forever"
4396 	 */
4397 #define USECS_WAIT	1000000
4398         for (nr_loops = 0; nr_loops < USECS_WAIT; nr_loops++) {
4399 		/* If the write is past the end of page, a writer is still updating it */
4400 		if (likely(!reader || rb_page_write(reader) <= BUF_PAGE_SIZE))
4401 			break;
4402 
4403 		udelay(1);
4404 
4405 		/* Get the latest version of the reader write value */
4406 		smp_rmb();
4407 	}
4408 
4409 	/* The writer is not moving forward? Something is wrong */
4410 	if (RB_WARN_ON(cpu_buffer, nr_loops == USECS_WAIT))
4411 		reader = NULL;
4412 
4413 	/*
4414 	 * Make sure we see any padding after the write update
4415 	 * (see rb_reset_tail())
4416 	 */
4417 	smp_rmb();
4418 
4419 
4420 	return reader;
4421 }
4422 
rb_advance_reader(struct ring_buffer_per_cpu * cpu_buffer)4423 static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
4424 {
4425 	struct ring_buffer_event *event;
4426 	struct buffer_page *reader;
4427 	unsigned length;
4428 
4429 	reader = rb_get_reader_page(cpu_buffer);
4430 
4431 	/* This function should not be called when buffer is empty */
4432 	if (RB_WARN_ON(cpu_buffer, !reader))
4433 		return;
4434 
4435 	event = rb_reader_event(cpu_buffer);
4436 
4437 	if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
4438 		cpu_buffer->read++;
4439 
4440 	rb_update_read_stamp(cpu_buffer, event);
4441 
4442 	length = rb_event_length(event);
4443 	cpu_buffer->reader_page->read += length;
4444 }
4445 
rb_advance_iter(struct ring_buffer_iter * iter)4446 static void rb_advance_iter(struct ring_buffer_iter *iter)
4447 {
4448 	struct ring_buffer_per_cpu *cpu_buffer;
4449 
4450 	cpu_buffer = iter->cpu_buffer;
4451 
4452 	/* If head == next_event then we need to jump to the next event */
4453 	if (iter->head == iter->next_event) {
4454 		/* If the event gets overwritten again, there's nothing to do */
4455 		if (rb_iter_head_event(iter) == NULL)
4456 			return;
4457 	}
4458 
4459 	iter->head = iter->next_event;
4460 
4461 	/*
4462 	 * Check if we are at the end of the buffer.
4463 	 */
4464 	if (iter->next_event >= rb_page_size(iter->head_page)) {
4465 		/* discarded commits can make the page empty */
4466 		if (iter->head_page == cpu_buffer->commit_page)
4467 			return;
4468 		rb_inc_iter(iter);
4469 		return;
4470 	}
4471 
4472 	rb_update_iter_read_stamp(iter, iter->event);
4473 }
4474 
rb_lost_events(struct ring_buffer_per_cpu * cpu_buffer)4475 static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
4476 {
4477 	return cpu_buffer->lost_events;
4478 }
4479 
4480 static struct ring_buffer_event *
rb_buffer_peek(struct ring_buffer_per_cpu * cpu_buffer,u64 * ts,unsigned long * lost_events)4481 rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
4482 	       unsigned long *lost_events)
4483 {
4484 	struct ring_buffer_event *event;
4485 	struct buffer_page *reader;
4486 	int nr_loops = 0;
4487 
4488 	if (ts)
4489 		*ts = 0;
4490  again:
4491 	/*
4492 	 * We repeat when a time extend is encountered.
4493 	 * Since the time extend is always attached to a data event,
4494 	 * we should never loop more than once.
4495 	 * (We never hit the following condition more than twice).
4496 	 */
4497 	if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
4498 		return NULL;
4499 
4500 	reader = rb_get_reader_page(cpu_buffer);
4501 	if (!reader)
4502 		return NULL;
4503 
4504 	event = rb_reader_event(cpu_buffer);
4505 
4506 	switch (event->type_len) {
4507 	case RINGBUF_TYPE_PADDING:
4508 		if (rb_null_event(event))
4509 			RB_WARN_ON(cpu_buffer, 1);
4510 		/*
4511 		 * Because the writer could be discarding every
4512 		 * event it creates (which would probably be bad)
4513 		 * if we were to go back to "again" then we may never
4514 		 * catch up, and will trigger the warn on, or lock
4515 		 * the box. Return the padding, and we will release
4516 		 * the current locks, and try again.
4517 		 */
4518 		return event;
4519 
4520 	case RINGBUF_TYPE_TIME_EXTEND:
4521 		/* Internal data, OK to advance */
4522 		rb_advance_reader(cpu_buffer);
4523 		goto again;
4524 
4525 	case RINGBUF_TYPE_TIME_STAMP:
4526 		if (ts) {
4527 			*ts = ring_buffer_event_time_stamp(event);
4528 			ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
4529 							 cpu_buffer->cpu, ts);
4530 		}
4531 		/* Internal data, OK to advance */
4532 		rb_advance_reader(cpu_buffer);
4533 		goto again;
4534 
4535 	case RINGBUF_TYPE_DATA:
4536 		if (ts && !(*ts)) {
4537 			*ts = cpu_buffer->read_stamp + event->time_delta;
4538 			ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
4539 							 cpu_buffer->cpu, ts);
4540 		}
4541 		if (lost_events)
4542 			*lost_events = rb_lost_events(cpu_buffer);
4543 		return event;
4544 
4545 	default:
4546 		RB_WARN_ON(cpu_buffer, 1);
4547 	}
4548 
4549 	return NULL;
4550 }
4551 EXPORT_SYMBOL_GPL(ring_buffer_peek);
4552 
4553 static struct ring_buffer_event *
rb_iter_peek(struct ring_buffer_iter * iter,u64 * ts)4554 rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
4555 {
4556 	struct trace_buffer *buffer;
4557 	struct ring_buffer_per_cpu *cpu_buffer;
4558 	struct ring_buffer_event *event;
4559 	int nr_loops = 0;
4560 
4561 	if (ts)
4562 		*ts = 0;
4563 
4564 	cpu_buffer = iter->cpu_buffer;
4565 	buffer = cpu_buffer->buffer;
4566 
4567 	/*
4568 	 * Check if someone performed a consuming read to
4569 	 * the buffer. A consuming read invalidates the iterator
4570 	 * and we need to reset the iterator in this case.
4571 	 */
4572 	if (unlikely(iter->cache_read != cpu_buffer->read ||
4573 		     iter->cache_reader_page != cpu_buffer->reader_page))
4574 		rb_iter_reset(iter);
4575 
4576  again:
4577 	if (ring_buffer_iter_empty(iter))
4578 		return NULL;
4579 
4580 	/*
4581 	 * As the writer can mess with what the iterator is trying
4582 	 * to read, just give up if we fail to get an event after
4583 	 * three tries. The iterator is not as reliable when reading
4584 	 * the ring buffer with an active write as the consumer is.
4585 	 * Do not warn if the three failures is reached.
4586 	 */
4587 	if (++nr_loops > 3)
4588 		return NULL;
4589 
4590 	if (rb_per_cpu_empty(cpu_buffer))
4591 		return NULL;
4592 
4593 	if (iter->head >= rb_page_size(iter->head_page)) {
4594 		rb_inc_iter(iter);
4595 		goto again;
4596 	}
4597 
4598 	event = rb_iter_head_event(iter);
4599 	if (!event)
4600 		goto again;
4601 
4602 	switch (event->type_len) {
4603 	case RINGBUF_TYPE_PADDING:
4604 		if (rb_null_event(event)) {
4605 			rb_inc_iter(iter);
4606 			goto again;
4607 		}
4608 		rb_advance_iter(iter);
4609 		return event;
4610 
4611 	case RINGBUF_TYPE_TIME_EXTEND:
4612 		/* Internal data, OK to advance */
4613 		rb_advance_iter(iter);
4614 		goto again;
4615 
4616 	case RINGBUF_TYPE_TIME_STAMP:
4617 		if (ts) {
4618 			*ts = ring_buffer_event_time_stamp(event);
4619 			ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
4620 							 cpu_buffer->cpu, ts);
4621 		}
4622 		/* Internal data, OK to advance */
4623 		rb_advance_iter(iter);
4624 		goto again;
4625 
4626 	case RINGBUF_TYPE_DATA:
4627 		if (ts && !(*ts)) {
4628 			*ts = iter->read_stamp + event->time_delta;
4629 			ring_buffer_normalize_time_stamp(buffer,
4630 							 cpu_buffer->cpu, ts);
4631 		}
4632 		return event;
4633 
4634 	default:
4635 		RB_WARN_ON(cpu_buffer, 1);
4636 	}
4637 
4638 	return NULL;
4639 }
4640 EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
4641 
rb_reader_lock(struct ring_buffer_per_cpu * cpu_buffer)4642 static inline bool rb_reader_lock(struct ring_buffer_per_cpu *cpu_buffer)
4643 {
4644 	if (likely(!in_nmi())) {
4645 		raw_spin_lock(&cpu_buffer->reader_lock);
4646 		return true;
4647 	}
4648 
4649 	/*
4650 	 * If an NMI die dumps out the content of the ring buffer
4651 	 * trylock must be used to prevent a deadlock if the NMI
4652 	 * preempted a task that holds the ring buffer locks. If
4653 	 * we get the lock then all is fine, if not, then continue
4654 	 * to do the read, but this can corrupt the ring buffer,
4655 	 * so it must be permanently disabled from future writes.
4656 	 * Reading from NMI is a oneshot deal.
4657 	 */
4658 	if (raw_spin_trylock(&cpu_buffer->reader_lock))
4659 		return true;
4660 
4661 	/* Continue without locking, but disable the ring buffer */
4662 	atomic_inc(&cpu_buffer->record_disabled);
4663 	return false;
4664 }
4665 
4666 static inline void
rb_reader_unlock(struct ring_buffer_per_cpu * cpu_buffer,bool locked)4667 rb_reader_unlock(struct ring_buffer_per_cpu *cpu_buffer, bool locked)
4668 {
4669 	if (likely(locked))
4670 		raw_spin_unlock(&cpu_buffer->reader_lock);
4671 	return;
4672 }
4673 
4674 /**
4675  * ring_buffer_peek - peek at the next event to be read
4676  * @buffer: The ring buffer to read
4677  * @cpu: The cpu to peak at
4678  * @ts: The timestamp counter of this event.
4679  * @lost_events: a variable to store if events were lost (may be NULL)
4680  *
4681  * This will return the event that will be read next, but does
4682  * not consume the data.
4683  */
4684 struct ring_buffer_event *
ring_buffer_peek(struct trace_buffer * buffer,int cpu,u64 * ts,unsigned long * lost_events)4685 ring_buffer_peek(struct trace_buffer *buffer, int cpu, u64 *ts,
4686 		 unsigned long *lost_events)
4687 {
4688 	struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4689 	struct ring_buffer_event *event;
4690 	unsigned long flags;
4691 	bool dolock;
4692 
4693 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4694 		return NULL;
4695 
4696  again:
4697 	local_irq_save(flags);
4698 	dolock = rb_reader_lock(cpu_buffer);
4699 	event = rb_buffer_peek(cpu_buffer, ts, lost_events);
4700 	if (event && event->type_len == RINGBUF_TYPE_PADDING)
4701 		rb_advance_reader(cpu_buffer);
4702 	rb_reader_unlock(cpu_buffer, dolock);
4703 	local_irq_restore(flags);
4704 
4705 	if (event && event->type_len == RINGBUF_TYPE_PADDING)
4706 		goto again;
4707 
4708 	return event;
4709 }
4710 
4711 /** ring_buffer_iter_dropped - report if there are dropped events
4712  * @iter: The ring buffer iterator
4713  *
4714  * Returns true if there was dropped events since the last peek.
4715  */
ring_buffer_iter_dropped(struct ring_buffer_iter * iter)4716 bool ring_buffer_iter_dropped(struct ring_buffer_iter *iter)
4717 {
4718 	bool ret = iter->missed_events != 0;
4719 
4720 	iter->missed_events = 0;
4721 	return ret;
4722 }
4723 EXPORT_SYMBOL_GPL(ring_buffer_iter_dropped);
4724 
4725 /**
4726  * ring_buffer_iter_peek - peek at the next event to be read
4727  * @iter: The ring buffer iterator
4728  * @ts: The timestamp counter of this event.
4729  *
4730  * This will return the event that will be read next, but does
4731  * not increment the iterator.
4732  */
4733 struct ring_buffer_event *
ring_buffer_iter_peek(struct ring_buffer_iter * iter,u64 * ts)4734 ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
4735 {
4736 	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4737 	struct ring_buffer_event *event;
4738 	unsigned long flags;
4739 
4740  again:
4741 	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4742 	event = rb_iter_peek(iter, ts);
4743 	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4744 
4745 	if (event && event->type_len == RINGBUF_TYPE_PADDING)
4746 		goto again;
4747 
4748 	return event;
4749 }
4750 
4751 /**
4752  * ring_buffer_consume - return an event and consume it
4753  * @buffer: The ring buffer to get the next event from
4754  * @cpu: the cpu to read the buffer from
4755  * @ts: a variable to store the timestamp (may be NULL)
4756  * @lost_events: a variable to store if events were lost (may be NULL)
4757  *
4758  * Returns the next event in the ring buffer, and that event is consumed.
4759  * Meaning, that sequential reads will keep returning a different event,
4760  * and eventually empty the ring buffer if the producer is slower.
4761  */
4762 struct ring_buffer_event *
ring_buffer_consume(struct trace_buffer * buffer,int cpu,u64 * ts,unsigned long * lost_events)4763 ring_buffer_consume(struct trace_buffer *buffer, int cpu, u64 *ts,
4764 		    unsigned long *lost_events)
4765 {
4766 	struct ring_buffer_per_cpu *cpu_buffer;
4767 	struct ring_buffer_event *event = NULL;
4768 	unsigned long flags;
4769 	bool dolock;
4770 
4771  again:
4772 	/* might be called in atomic */
4773 	preempt_disable();
4774 
4775 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4776 		goto out;
4777 
4778 	cpu_buffer = buffer->buffers[cpu];
4779 	local_irq_save(flags);
4780 	dolock = rb_reader_lock(cpu_buffer);
4781 
4782 	event = rb_buffer_peek(cpu_buffer, ts, lost_events);
4783 	if (event) {
4784 		cpu_buffer->lost_events = 0;
4785 		rb_advance_reader(cpu_buffer);
4786 	}
4787 
4788 	rb_reader_unlock(cpu_buffer, dolock);
4789 	local_irq_restore(flags);
4790 
4791  out:
4792 	preempt_enable();
4793 
4794 	if (event && event->type_len == RINGBUF_TYPE_PADDING)
4795 		goto again;
4796 
4797 	return event;
4798 }
4799 EXPORT_SYMBOL_GPL(ring_buffer_consume);
4800 
4801 /**
4802  * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
4803  * @buffer: The ring buffer to read from
4804  * @cpu: The cpu buffer to iterate over
4805  * @flags: gfp flags to use for memory allocation
4806  *
4807  * This performs the initial preparations necessary to iterate
4808  * through the buffer.  Memory is allocated, buffer recording
4809  * is disabled, and the iterator pointer is returned to the caller.
4810  *
4811  * Disabling buffer recording prevents the reading from being
4812  * corrupted. This is not a consuming read, so a producer is not
4813  * expected.
4814  *
4815  * After a sequence of ring_buffer_read_prepare calls, the user is
4816  * expected to make at least one call to ring_buffer_read_prepare_sync.
4817  * Afterwards, ring_buffer_read_start is invoked to get things going
4818  * for real.
4819  *
4820  * This overall must be paired with ring_buffer_read_finish.
4821  */
4822 struct ring_buffer_iter *
ring_buffer_read_prepare(struct trace_buffer * buffer,int cpu,gfp_t flags)4823 ring_buffer_read_prepare(struct trace_buffer *buffer, int cpu, gfp_t flags)
4824 {
4825 	struct ring_buffer_per_cpu *cpu_buffer;
4826 	struct ring_buffer_iter *iter;
4827 
4828 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4829 		return NULL;
4830 
4831 	iter = kzalloc(sizeof(*iter), flags);
4832 	if (!iter)
4833 		return NULL;
4834 
4835 	iter->event = kmalloc(BUF_MAX_DATA_SIZE, flags);
4836 	if (!iter->event) {
4837 		kfree(iter);
4838 		return NULL;
4839 	}
4840 
4841 	cpu_buffer = buffer->buffers[cpu];
4842 
4843 	iter->cpu_buffer = cpu_buffer;
4844 
4845 	atomic_inc(&cpu_buffer->resize_disabled);
4846 
4847 	return iter;
4848 }
4849 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
4850 
4851 /**
4852  * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
4853  *
4854  * All previously invoked ring_buffer_read_prepare calls to prepare
4855  * iterators will be synchronized.  Afterwards, read_buffer_read_start
4856  * calls on those iterators are allowed.
4857  */
4858 void
ring_buffer_read_prepare_sync(void)4859 ring_buffer_read_prepare_sync(void)
4860 {
4861 	synchronize_rcu();
4862 }
4863 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
4864 
4865 /**
4866  * ring_buffer_read_start - start a non consuming read of the buffer
4867  * @iter: The iterator returned by ring_buffer_read_prepare
4868  *
4869  * This finalizes the startup of an iteration through the buffer.
4870  * The iterator comes from a call to ring_buffer_read_prepare and
4871  * an intervening ring_buffer_read_prepare_sync must have been
4872  * performed.
4873  *
4874  * Must be paired with ring_buffer_read_finish.
4875  */
4876 void
ring_buffer_read_start(struct ring_buffer_iter * iter)4877 ring_buffer_read_start(struct ring_buffer_iter *iter)
4878 {
4879 	struct ring_buffer_per_cpu *cpu_buffer;
4880 	unsigned long flags;
4881 
4882 	if (!iter)
4883 		return;
4884 
4885 	cpu_buffer = iter->cpu_buffer;
4886 
4887 	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4888 	arch_spin_lock(&cpu_buffer->lock);
4889 	rb_iter_reset(iter);
4890 	arch_spin_unlock(&cpu_buffer->lock);
4891 	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4892 }
4893 EXPORT_SYMBOL_GPL(ring_buffer_read_start);
4894 
4895 /**
4896  * ring_buffer_read_finish - finish reading the iterator of the buffer
4897  * @iter: The iterator retrieved by ring_buffer_start
4898  *
4899  * This re-enables the recording to the buffer, and frees the
4900  * iterator.
4901  */
4902 void
ring_buffer_read_finish(struct ring_buffer_iter * iter)4903 ring_buffer_read_finish(struct ring_buffer_iter *iter)
4904 {
4905 	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4906 	unsigned long flags;
4907 
4908 	/*
4909 	 * Ring buffer is disabled from recording, here's a good place
4910 	 * to check the integrity of the ring buffer.
4911 	 * Must prevent readers from trying to read, as the check
4912 	 * clears the HEAD page and readers require it.
4913 	 */
4914 	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4915 	rb_check_pages(cpu_buffer);
4916 	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4917 
4918 	atomic_dec(&cpu_buffer->resize_disabled);
4919 	kfree(iter->event);
4920 	kfree(iter);
4921 }
4922 EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
4923 
4924 /**
4925  * ring_buffer_iter_advance - advance the iterator to the next location
4926  * @iter: The ring buffer iterator
4927  *
4928  * Move the location of the iterator such that the next read will
4929  * be the next location of the iterator.
4930  */
ring_buffer_iter_advance(struct ring_buffer_iter * iter)4931 void ring_buffer_iter_advance(struct ring_buffer_iter *iter)
4932 {
4933 	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4934 	unsigned long flags;
4935 
4936 	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4937 
4938 	rb_advance_iter(iter);
4939 
4940 	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4941 }
4942 EXPORT_SYMBOL_GPL(ring_buffer_iter_advance);
4943 
4944 /**
4945  * ring_buffer_size - return the size of the ring buffer (in bytes)
4946  * @buffer: The ring buffer.
4947  * @cpu: The CPU to get ring buffer size from.
4948  */
ring_buffer_size(struct trace_buffer * buffer,int cpu)4949 unsigned long ring_buffer_size(struct trace_buffer *buffer, int cpu)
4950 {
4951 	/*
4952 	 * Earlier, this method returned
4953 	 *	BUF_PAGE_SIZE * buffer->nr_pages
4954 	 * Since the nr_pages field is now removed, we have converted this to
4955 	 * return the per cpu buffer value.
4956 	 */
4957 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4958 		return 0;
4959 
4960 	return BUF_PAGE_SIZE * buffer->buffers[cpu]->nr_pages;
4961 }
4962 EXPORT_SYMBOL_GPL(ring_buffer_size);
4963 
4964 static void
rb_reset_cpu(struct ring_buffer_per_cpu * cpu_buffer)4965 rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
4966 {
4967 	rb_head_page_deactivate(cpu_buffer);
4968 
4969 	cpu_buffer->head_page
4970 		= list_entry(cpu_buffer->pages, struct buffer_page, list);
4971 	local_set(&cpu_buffer->head_page->write, 0);
4972 	local_set(&cpu_buffer->head_page->entries, 0);
4973 	local_set(&cpu_buffer->head_page->page->commit, 0);
4974 
4975 	cpu_buffer->head_page->read = 0;
4976 
4977 	cpu_buffer->tail_page = cpu_buffer->head_page;
4978 	cpu_buffer->commit_page = cpu_buffer->head_page;
4979 
4980 	INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
4981 	INIT_LIST_HEAD(&cpu_buffer->new_pages);
4982 	local_set(&cpu_buffer->reader_page->write, 0);
4983 	local_set(&cpu_buffer->reader_page->entries, 0);
4984 	local_set(&cpu_buffer->reader_page->page->commit, 0);
4985 	cpu_buffer->reader_page->read = 0;
4986 
4987 	local_set(&cpu_buffer->entries_bytes, 0);
4988 	local_set(&cpu_buffer->overrun, 0);
4989 	local_set(&cpu_buffer->commit_overrun, 0);
4990 	local_set(&cpu_buffer->dropped_events, 0);
4991 	local_set(&cpu_buffer->entries, 0);
4992 	local_set(&cpu_buffer->committing, 0);
4993 	local_set(&cpu_buffer->commits, 0);
4994 	local_set(&cpu_buffer->pages_touched, 0);
4995 	local_set(&cpu_buffer->pages_lost, 0);
4996 	local_set(&cpu_buffer->pages_read, 0);
4997 	cpu_buffer->last_pages_touch = 0;
4998 	cpu_buffer->shortest_full = 0;
4999 	cpu_buffer->read = 0;
5000 	cpu_buffer->read_bytes = 0;
5001 
5002 	rb_time_set(&cpu_buffer->write_stamp, 0);
5003 	rb_time_set(&cpu_buffer->before_stamp, 0);
5004 
5005 	cpu_buffer->lost_events = 0;
5006 	cpu_buffer->last_overrun = 0;
5007 
5008 	rb_head_page_activate(cpu_buffer);
5009 }
5010 
5011 /* Must have disabled the cpu buffer then done a synchronize_rcu */
reset_disabled_cpu_buffer(struct ring_buffer_per_cpu * cpu_buffer)5012 static void reset_disabled_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
5013 {
5014 	unsigned long flags;
5015 
5016 	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5017 
5018 	if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
5019 		goto out;
5020 
5021 	arch_spin_lock(&cpu_buffer->lock);
5022 
5023 	rb_reset_cpu(cpu_buffer);
5024 
5025 	arch_spin_unlock(&cpu_buffer->lock);
5026 
5027  out:
5028 	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5029 }
5030 
5031 /**
5032  * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
5033  * @buffer: The ring buffer to reset a per cpu buffer of
5034  * @cpu: The CPU buffer to be reset
5035  */
ring_buffer_reset_cpu(struct trace_buffer * buffer,int cpu)5036 void ring_buffer_reset_cpu(struct trace_buffer *buffer, int cpu)
5037 {
5038 	struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
5039 
5040 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
5041 		return;
5042 
5043 	/* prevent another thread from changing buffer sizes */
5044 	mutex_lock(&buffer->mutex);
5045 
5046 	atomic_inc(&cpu_buffer->resize_disabled);
5047 	atomic_inc(&cpu_buffer->record_disabled);
5048 
5049 	/* Make sure all commits have finished */
5050 	synchronize_rcu();
5051 
5052 	reset_disabled_cpu_buffer(cpu_buffer);
5053 
5054 	atomic_dec(&cpu_buffer->record_disabled);
5055 	atomic_dec(&cpu_buffer->resize_disabled);
5056 
5057 	mutex_unlock(&buffer->mutex);
5058 }
5059 EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
5060 
5061 /**
5062  * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
5063  * @buffer: The ring buffer to reset a per cpu buffer of
5064  * @cpu: The CPU buffer to be reset
5065  */
ring_buffer_reset_online_cpus(struct trace_buffer * buffer)5066 void ring_buffer_reset_online_cpus(struct trace_buffer *buffer)
5067 {
5068 	struct ring_buffer_per_cpu *cpu_buffer;
5069 	int cpu;
5070 
5071 	/* prevent another thread from changing buffer sizes */
5072 	mutex_lock(&buffer->mutex);
5073 
5074 	for_each_online_buffer_cpu(buffer, cpu) {
5075 		cpu_buffer = buffer->buffers[cpu];
5076 
5077 		atomic_inc(&cpu_buffer->resize_disabled);
5078 		atomic_inc(&cpu_buffer->record_disabled);
5079 	}
5080 
5081 	/* Make sure all commits have finished */
5082 	synchronize_rcu();
5083 
5084 	for_each_online_buffer_cpu(buffer, cpu) {
5085 		cpu_buffer = buffer->buffers[cpu];
5086 
5087 		reset_disabled_cpu_buffer(cpu_buffer);
5088 
5089 		atomic_dec(&cpu_buffer->record_disabled);
5090 		atomic_dec(&cpu_buffer->resize_disabled);
5091 	}
5092 
5093 	mutex_unlock(&buffer->mutex);
5094 }
5095 
5096 /**
5097  * ring_buffer_reset - reset a ring buffer
5098  * @buffer: The ring buffer to reset all cpu buffers
5099  */
ring_buffer_reset(struct trace_buffer * buffer)5100 void ring_buffer_reset(struct trace_buffer *buffer)
5101 {
5102 	struct ring_buffer_per_cpu *cpu_buffer;
5103 	int cpu;
5104 
5105 	/* prevent another thread from changing buffer sizes */
5106 	mutex_lock(&buffer->mutex);
5107 
5108 	for_each_buffer_cpu(buffer, cpu) {
5109 		cpu_buffer = buffer->buffers[cpu];
5110 
5111 		atomic_inc(&cpu_buffer->resize_disabled);
5112 		atomic_inc(&cpu_buffer->record_disabled);
5113 	}
5114 
5115 	/* Make sure all commits have finished */
5116 	synchronize_rcu();
5117 
5118 	for_each_buffer_cpu(buffer, cpu) {
5119 		cpu_buffer = buffer->buffers[cpu];
5120 
5121 		reset_disabled_cpu_buffer(cpu_buffer);
5122 
5123 		atomic_dec(&cpu_buffer->record_disabled);
5124 		atomic_dec(&cpu_buffer->resize_disabled);
5125 	}
5126 
5127 	mutex_unlock(&buffer->mutex);
5128 }
5129 EXPORT_SYMBOL_GPL(ring_buffer_reset);
5130 
5131 /**
5132  * rind_buffer_empty - is the ring buffer empty?
5133  * @buffer: The ring buffer to test
5134  */
ring_buffer_empty(struct trace_buffer * buffer)5135 bool ring_buffer_empty(struct trace_buffer *buffer)
5136 {
5137 	struct ring_buffer_per_cpu *cpu_buffer;
5138 	unsigned long flags;
5139 	bool dolock;
5140 	int cpu;
5141 	int ret;
5142 
5143 	/* yes this is racy, but if you don't like the race, lock the buffer */
5144 	for_each_buffer_cpu(buffer, cpu) {
5145 		cpu_buffer = buffer->buffers[cpu];
5146 		local_irq_save(flags);
5147 		dolock = rb_reader_lock(cpu_buffer);
5148 		ret = rb_per_cpu_empty(cpu_buffer);
5149 		rb_reader_unlock(cpu_buffer, dolock);
5150 		local_irq_restore(flags);
5151 
5152 		if (!ret)
5153 			return false;
5154 	}
5155 
5156 	return true;
5157 }
5158 EXPORT_SYMBOL_GPL(ring_buffer_empty);
5159 
5160 /**
5161  * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
5162  * @buffer: The ring buffer
5163  * @cpu: The CPU buffer to test
5164  */
ring_buffer_empty_cpu(struct trace_buffer * buffer,int cpu)5165 bool ring_buffer_empty_cpu(struct trace_buffer *buffer, int cpu)
5166 {
5167 	struct ring_buffer_per_cpu *cpu_buffer;
5168 	unsigned long flags;
5169 	bool dolock;
5170 	int ret;
5171 
5172 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
5173 		return true;
5174 
5175 	cpu_buffer = buffer->buffers[cpu];
5176 	local_irq_save(flags);
5177 	dolock = rb_reader_lock(cpu_buffer);
5178 	ret = rb_per_cpu_empty(cpu_buffer);
5179 	rb_reader_unlock(cpu_buffer, dolock);
5180 	local_irq_restore(flags);
5181 
5182 	return ret;
5183 }
5184 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
5185 
5186 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
5187 /**
5188  * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
5189  * @buffer_a: One buffer to swap with
5190  * @buffer_b: The other buffer to swap with
5191  * @cpu: the CPU of the buffers to swap
5192  *
5193  * This function is useful for tracers that want to take a "snapshot"
5194  * of a CPU buffer and has another back up buffer lying around.
5195  * it is expected that the tracer handles the cpu buffer not being
5196  * used at the moment.
5197  */
ring_buffer_swap_cpu(struct trace_buffer * buffer_a,struct trace_buffer * buffer_b,int cpu)5198 int ring_buffer_swap_cpu(struct trace_buffer *buffer_a,
5199 			 struct trace_buffer *buffer_b, int cpu)
5200 {
5201 	struct ring_buffer_per_cpu *cpu_buffer_a;
5202 	struct ring_buffer_per_cpu *cpu_buffer_b;
5203 	int ret = -EINVAL;
5204 
5205 	if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
5206 	    !cpumask_test_cpu(cpu, buffer_b->cpumask))
5207 		goto out;
5208 
5209 	cpu_buffer_a = buffer_a->buffers[cpu];
5210 	cpu_buffer_b = buffer_b->buffers[cpu];
5211 
5212 	/* At least make sure the two buffers are somewhat the same */
5213 	if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
5214 		goto out;
5215 
5216 	ret = -EAGAIN;
5217 
5218 	if (atomic_read(&buffer_a->record_disabled))
5219 		goto out;
5220 
5221 	if (atomic_read(&buffer_b->record_disabled))
5222 		goto out;
5223 
5224 	if (atomic_read(&cpu_buffer_a->record_disabled))
5225 		goto out;
5226 
5227 	if (atomic_read(&cpu_buffer_b->record_disabled))
5228 		goto out;
5229 
5230 	/*
5231 	 * We can't do a synchronize_rcu here because this
5232 	 * function can be called in atomic context.
5233 	 * Normally this will be called from the same CPU as cpu.
5234 	 * If not it's up to the caller to protect this.
5235 	 */
5236 	atomic_inc(&cpu_buffer_a->record_disabled);
5237 	atomic_inc(&cpu_buffer_b->record_disabled);
5238 
5239 	ret = -EBUSY;
5240 	if (local_read(&cpu_buffer_a->committing))
5241 		goto out_dec;
5242 	if (local_read(&cpu_buffer_b->committing))
5243 		goto out_dec;
5244 
5245 	buffer_a->buffers[cpu] = cpu_buffer_b;
5246 	buffer_b->buffers[cpu] = cpu_buffer_a;
5247 
5248 	cpu_buffer_b->buffer = buffer_a;
5249 	cpu_buffer_a->buffer = buffer_b;
5250 
5251 	ret = 0;
5252 
5253 out_dec:
5254 	atomic_dec(&cpu_buffer_a->record_disabled);
5255 	atomic_dec(&cpu_buffer_b->record_disabled);
5256 out:
5257 	return ret;
5258 }
5259 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
5260 #endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
5261 
5262 /**
5263  * ring_buffer_alloc_read_page - allocate a page to read from buffer
5264  * @buffer: the buffer to allocate for.
5265  * @cpu: the cpu buffer to allocate.
5266  *
5267  * This function is used in conjunction with ring_buffer_read_page.
5268  * When reading a full page from the ring buffer, these functions
5269  * can be used to speed up the process. The calling function should
5270  * allocate a few pages first with this function. Then when it
5271  * needs to get pages from the ring buffer, it passes the result
5272  * of this function into ring_buffer_read_page, which will swap
5273  * the page that was allocated, with the read page of the buffer.
5274  *
5275  * Returns:
5276  *  The page allocated, or ERR_PTR
5277  */
ring_buffer_alloc_read_page(struct trace_buffer * buffer,int cpu)5278 void *ring_buffer_alloc_read_page(struct trace_buffer *buffer, int cpu)
5279 {
5280 	struct ring_buffer_per_cpu *cpu_buffer;
5281 	struct buffer_data_page *bpage = NULL;
5282 	unsigned long flags;
5283 	struct page *page;
5284 
5285 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
5286 		return ERR_PTR(-ENODEV);
5287 
5288 	cpu_buffer = buffer->buffers[cpu];
5289 	local_irq_save(flags);
5290 	arch_spin_lock(&cpu_buffer->lock);
5291 
5292 	if (cpu_buffer->free_page) {
5293 		bpage = cpu_buffer->free_page;
5294 		cpu_buffer->free_page = NULL;
5295 	}
5296 
5297 	arch_spin_unlock(&cpu_buffer->lock);
5298 	local_irq_restore(flags);
5299 
5300 	if (bpage)
5301 		goto out;
5302 
5303 	page = alloc_pages_node(cpu_to_node(cpu),
5304 				GFP_KERNEL | __GFP_NORETRY, 0);
5305 	if (!page)
5306 		return ERR_PTR(-ENOMEM);
5307 
5308 	bpage = page_address(page);
5309 
5310  out:
5311 	rb_init_page(bpage);
5312 
5313 	return bpage;
5314 }
5315 EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
5316 
5317 /**
5318  * ring_buffer_free_read_page - free an allocated read page
5319  * @buffer: the buffer the page was allocate for
5320  * @cpu: the cpu buffer the page came from
5321  * @data: the page to free
5322  *
5323  * Free a page allocated from ring_buffer_alloc_read_page.
5324  */
ring_buffer_free_read_page(struct trace_buffer * buffer,int cpu,void * data)5325 void ring_buffer_free_read_page(struct trace_buffer *buffer, int cpu, void *data)
5326 {
5327 	struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
5328 	struct buffer_data_page *bpage = data;
5329 	struct page *page = virt_to_page(bpage);
5330 	unsigned long flags;
5331 
5332 	/* If the page is still in use someplace else, we can't reuse it */
5333 	if (page_ref_count(page) > 1)
5334 		goto out;
5335 
5336 	local_irq_save(flags);
5337 	arch_spin_lock(&cpu_buffer->lock);
5338 
5339 	if (!cpu_buffer->free_page) {
5340 		cpu_buffer->free_page = bpage;
5341 		bpage = NULL;
5342 	}
5343 
5344 	arch_spin_unlock(&cpu_buffer->lock);
5345 	local_irq_restore(flags);
5346 
5347  out:
5348 	free_page((unsigned long)bpage);
5349 }
5350 EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
5351 
5352 /**
5353  * ring_buffer_read_page - extract a page from the ring buffer
5354  * @buffer: buffer to extract from
5355  * @data_page: the page to use allocated from ring_buffer_alloc_read_page
5356  * @len: amount to extract
5357  * @cpu: the cpu of the buffer to extract
5358  * @full: should the extraction only happen when the page is full.
5359  *
5360  * This function will pull out a page from the ring buffer and consume it.
5361  * @data_page must be the address of the variable that was returned
5362  * from ring_buffer_alloc_read_page. This is because the page might be used
5363  * to swap with a page in the ring buffer.
5364  *
5365  * for example:
5366  *	rpage = ring_buffer_alloc_read_page(buffer, cpu);
5367  *	if (IS_ERR(rpage))
5368  *		return PTR_ERR(rpage);
5369  *	ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
5370  *	if (ret >= 0)
5371  *		process_page(rpage, ret);
5372  *
5373  * When @full is set, the function will not return true unless
5374  * the writer is off the reader page.
5375  *
5376  * Note: it is up to the calling functions to handle sleeps and wakeups.
5377  *  The ring buffer can be used anywhere in the kernel and can not
5378  *  blindly call wake_up. The layer that uses the ring buffer must be
5379  *  responsible for that.
5380  *
5381  * Returns:
5382  *  >=0 if data has been transferred, returns the offset of consumed data.
5383  *  <0 if no data has been transferred.
5384  */
ring_buffer_read_page(struct trace_buffer * buffer,void ** data_page,size_t len,int cpu,int full)5385 int ring_buffer_read_page(struct trace_buffer *buffer,
5386 			  void **data_page, size_t len, int cpu, int full)
5387 {
5388 	struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
5389 	struct ring_buffer_event *event;
5390 	struct buffer_data_page *bpage;
5391 	struct buffer_page *reader;
5392 	unsigned long missed_events;
5393 	unsigned long flags;
5394 	unsigned int commit;
5395 	unsigned int read;
5396 	u64 save_timestamp;
5397 	int ret = -1;
5398 
5399 	if (!cpumask_test_cpu(cpu, buffer->cpumask))
5400 		goto out;
5401 
5402 	/*
5403 	 * If len is not big enough to hold the page header, then
5404 	 * we can not copy anything.
5405 	 */
5406 	if (len <= BUF_PAGE_HDR_SIZE)
5407 		goto out;
5408 
5409 	len -= BUF_PAGE_HDR_SIZE;
5410 
5411 	if (!data_page)
5412 		goto out;
5413 
5414 	bpage = *data_page;
5415 	if (!bpage)
5416 		goto out;
5417 
5418 	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5419 
5420 	reader = rb_get_reader_page(cpu_buffer);
5421 	if (!reader)
5422 		goto out_unlock;
5423 
5424 	event = rb_reader_event(cpu_buffer);
5425 
5426 	read = reader->read;
5427 	commit = rb_page_commit(reader);
5428 
5429 	/* Check if any events were dropped */
5430 	missed_events = cpu_buffer->lost_events;
5431 
5432 	/*
5433 	 * If this page has been partially read or
5434 	 * if len is not big enough to read the rest of the page or
5435 	 * a writer is still on the page, then
5436 	 * we must copy the data from the page to the buffer.
5437 	 * Otherwise, we can simply swap the page with the one passed in.
5438 	 */
5439 	if (read || (len < (commit - read)) ||
5440 	    cpu_buffer->reader_page == cpu_buffer->commit_page) {
5441 		struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
5442 		unsigned int rpos = read;
5443 		unsigned int pos = 0;
5444 		unsigned int size;
5445 
5446 		/*
5447 		 * If a full page is expected, this can still be returned
5448 		 * if there's been a previous partial read and the
5449 		 * rest of the page can be read and the commit page is off
5450 		 * the reader page.
5451 		 */
5452 		if (full &&
5453 		    (!read || (len < (commit - read)) ||
5454 		     cpu_buffer->reader_page == cpu_buffer->commit_page))
5455 			goto out_unlock;
5456 
5457 		if (len > (commit - read))
5458 			len = (commit - read);
5459 
5460 		/* Always keep the time extend and data together */
5461 		size = rb_event_ts_length(event);
5462 
5463 		if (len < size)
5464 			goto out_unlock;
5465 
5466 		/* save the current timestamp, since the user will need it */
5467 		save_timestamp = cpu_buffer->read_stamp;
5468 
5469 		/* Need to copy one event at a time */
5470 		do {
5471 			/* We need the size of one event, because
5472 			 * rb_advance_reader only advances by one event,
5473 			 * whereas rb_event_ts_length may include the size of
5474 			 * one or two events.
5475 			 * We have already ensured there's enough space if this
5476 			 * is a time extend. */
5477 			size = rb_event_length(event);
5478 			memcpy(bpage->data + pos, rpage->data + rpos, size);
5479 
5480 			len -= size;
5481 
5482 			rb_advance_reader(cpu_buffer);
5483 			rpos = reader->read;
5484 			pos += size;
5485 
5486 			if (rpos >= commit)
5487 				break;
5488 
5489 			event = rb_reader_event(cpu_buffer);
5490 			/* Always keep the time extend and data together */
5491 			size = rb_event_ts_length(event);
5492 		} while (len >= size);
5493 
5494 		/* update bpage */
5495 		local_set(&bpage->commit, pos);
5496 		bpage->time_stamp = save_timestamp;
5497 
5498 		/* we copied everything to the beginning */
5499 		read = 0;
5500 	} else {
5501 		/* update the entry counter */
5502 		cpu_buffer->read += rb_page_entries(reader);
5503 		cpu_buffer->read_bytes += BUF_PAGE_SIZE;
5504 
5505 		/* swap the pages */
5506 		rb_init_page(bpage);
5507 		bpage = reader->page;
5508 		reader->page = *data_page;
5509 		local_set(&reader->write, 0);
5510 		local_set(&reader->entries, 0);
5511 		reader->read = 0;
5512 		*data_page = bpage;
5513 
5514 		/*
5515 		 * Use the real_end for the data size,
5516 		 * This gives us a chance to store the lost events
5517 		 * on the page.
5518 		 */
5519 		if (reader->real_end)
5520 			local_set(&bpage->commit, reader->real_end);
5521 	}
5522 	ret = read;
5523 
5524 	cpu_buffer->lost_events = 0;
5525 
5526 	commit = local_read(&bpage->commit);
5527 	/*
5528 	 * Set a flag in the commit field if we lost events
5529 	 */
5530 	if (missed_events) {
5531 		/* If there is room at the end of the page to save the
5532 		 * missed events, then record it there.
5533 		 */
5534 		if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
5535 			memcpy(&bpage->data[commit], &missed_events,
5536 			       sizeof(missed_events));
5537 			local_add(RB_MISSED_STORED, &bpage->commit);
5538 			commit += sizeof(missed_events);
5539 		}
5540 		local_add(RB_MISSED_EVENTS, &bpage->commit);
5541 	}
5542 
5543 	/*
5544 	 * This page may be off to user land. Zero it out here.
5545 	 */
5546 	if (commit < BUF_PAGE_SIZE)
5547 		memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
5548 
5549  out_unlock:
5550 	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5551 
5552  out:
5553 	return ret;
5554 }
5555 EXPORT_SYMBOL_GPL(ring_buffer_read_page);
5556 
5557 /*
5558  * We only allocate new buffers, never free them if the CPU goes down.
5559  * If we were to free the buffer, then the user would lose any trace that was in
5560  * the buffer.
5561  */
trace_rb_cpu_prepare(unsigned int cpu,struct hlist_node * node)5562 int trace_rb_cpu_prepare(unsigned int cpu, struct hlist_node *node)
5563 {
5564 	struct trace_buffer *buffer;
5565 	long nr_pages_same;
5566 	int cpu_i;
5567 	unsigned long nr_pages;
5568 
5569 	buffer = container_of(node, struct trace_buffer, node);
5570 	if (cpumask_test_cpu(cpu, buffer->cpumask))
5571 		return 0;
5572 
5573 	nr_pages = 0;
5574 	nr_pages_same = 1;
5575 	/* check if all cpu sizes are same */
5576 	for_each_buffer_cpu(buffer, cpu_i) {
5577 		/* fill in the size from first enabled cpu */
5578 		if (nr_pages == 0)
5579 			nr_pages = buffer->buffers[cpu_i]->nr_pages;
5580 		if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
5581 			nr_pages_same = 0;
5582 			break;
5583 		}
5584 	}
5585 	/* allocate minimum pages, user can later expand it */
5586 	if (!nr_pages_same)
5587 		nr_pages = 2;
5588 	buffer->buffers[cpu] =
5589 		rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
5590 	if (!buffer->buffers[cpu]) {
5591 		WARN(1, "failed to allocate ring buffer on CPU %u\n",
5592 		     cpu);
5593 		return -ENOMEM;
5594 	}
5595 	smp_wmb();
5596 	cpumask_set_cpu(cpu, buffer->cpumask);
5597 	return 0;
5598 }
5599 
5600 #ifdef CONFIG_RING_BUFFER_STARTUP_TEST
5601 /*
5602  * This is a basic integrity check of the ring buffer.
5603  * Late in the boot cycle this test will run when configured in.
5604  * It will kick off a thread per CPU that will go into a loop
5605  * writing to the per cpu ring buffer various sizes of data.
5606  * Some of the data will be large items, some small.
5607  *
5608  * Another thread is created that goes into a spin, sending out
5609  * IPIs to the other CPUs to also write into the ring buffer.
5610  * this is to test the nesting ability of the buffer.
5611  *
5612  * Basic stats are recorded and reported. If something in the
5613  * ring buffer should happen that's not expected, a big warning
5614  * is displayed and all ring buffers are disabled.
5615  */
5616 static struct task_struct *rb_threads[NR_CPUS] __initdata;
5617 
5618 struct rb_test_data {
5619 	struct trace_buffer *buffer;
5620 	unsigned long		events;
5621 	unsigned long		bytes_written;
5622 	unsigned long		bytes_alloc;
5623 	unsigned long		bytes_dropped;
5624 	unsigned long		events_nested;
5625 	unsigned long		bytes_written_nested;
5626 	unsigned long		bytes_alloc_nested;
5627 	unsigned long		bytes_dropped_nested;
5628 	int			min_size_nested;
5629 	int			max_size_nested;
5630 	int			max_size;
5631 	int			min_size;
5632 	int			cpu;
5633 	int			cnt;
5634 };
5635 
5636 static struct rb_test_data rb_data[NR_CPUS] __initdata;
5637 
5638 /* 1 meg per cpu */
5639 #define RB_TEST_BUFFER_SIZE	1048576
5640 
5641 static char rb_string[] __initdata =
5642 	"abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
5643 	"?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
5644 	"!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
5645 
5646 static bool rb_test_started __initdata;
5647 
5648 struct rb_item {
5649 	int size;
5650 	char str[];
5651 };
5652 
rb_write_something(struct rb_test_data * data,bool nested)5653 static __init int rb_write_something(struct rb_test_data *data, bool nested)
5654 {
5655 	struct ring_buffer_event *event;
5656 	struct rb_item *item;
5657 	bool started;
5658 	int event_len;
5659 	int size;
5660 	int len;
5661 	int cnt;
5662 
5663 	/* Have nested writes different that what is written */
5664 	cnt = data->cnt + (nested ? 27 : 0);
5665 
5666 	/* Multiply cnt by ~e, to make some unique increment */
5667 	size = (cnt * 68 / 25) % (sizeof(rb_string) - 1);
5668 
5669 	len = size + sizeof(struct rb_item);
5670 
5671 	started = rb_test_started;
5672 	/* read rb_test_started before checking buffer enabled */
5673 	smp_rmb();
5674 
5675 	event = ring_buffer_lock_reserve(data->buffer, len);
5676 	if (!event) {
5677 		/* Ignore dropped events before test starts. */
5678 		if (started) {
5679 			if (nested)
5680 				data->bytes_dropped += len;
5681 			else
5682 				data->bytes_dropped_nested += len;
5683 		}
5684 		return len;
5685 	}
5686 
5687 	event_len = ring_buffer_event_length(event);
5688 
5689 	if (RB_WARN_ON(data->buffer, event_len < len))
5690 		goto out;
5691 
5692 	item = ring_buffer_event_data(event);
5693 	item->size = size;
5694 	memcpy(item->str, rb_string, size);
5695 
5696 	if (nested) {
5697 		data->bytes_alloc_nested += event_len;
5698 		data->bytes_written_nested += len;
5699 		data->events_nested++;
5700 		if (!data->min_size_nested || len < data->min_size_nested)
5701 			data->min_size_nested = len;
5702 		if (len > data->max_size_nested)
5703 			data->max_size_nested = len;
5704 	} else {
5705 		data->bytes_alloc += event_len;
5706 		data->bytes_written += len;
5707 		data->events++;
5708 		if (!data->min_size || len < data->min_size)
5709 			data->max_size = len;
5710 		if (len > data->max_size)
5711 			data->max_size = len;
5712 	}
5713 
5714  out:
5715 	ring_buffer_unlock_commit(data->buffer, event);
5716 
5717 	return 0;
5718 }
5719 
rb_test(void * arg)5720 static __init int rb_test(void *arg)
5721 {
5722 	struct rb_test_data *data = arg;
5723 
5724 	while (!kthread_should_stop()) {
5725 		rb_write_something(data, false);
5726 		data->cnt++;
5727 
5728 		set_current_state(TASK_INTERRUPTIBLE);
5729 		/* Now sleep between a min of 100-300us and a max of 1ms */
5730 		usleep_range(((data->cnt % 3) + 1) * 100, 1000);
5731 	}
5732 
5733 	return 0;
5734 }
5735 
rb_ipi(void * ignore)5736 static __init void rb_ipi(void *ignore)
5737 {
5738 	struct rb_test_data *data;
5739 	int cpu = smp_processor_id();
5740 
5741 	data = &rb_data[cpu];
5742 	rb_write_something(data, true);
5743 }
5744 
rb_hammer_test(void * arg)5745 static __init int rb_hammer_test(void *arg)
5746 {
5747 	while (!kthread_should_stop()) {
5748 
5749 		/* Send an IPI to all cpus to write data! */
5750 		smp_call_function(rb_ipi, NULL, 1);
5751 		/* No sleep, but for non preempt, let others run */
5752 		schedule();
5753 	}
5754 
5755 	return 0;
5756 }
5757 
test_ringbuffer(void)5758 static __init int test_ringbuffer(void)
5759 {
5760 	struct task_struct *rb_hammer;
5761 	struct trace_buffer *buffer;
5762 	int cpu;
5763 	int ret = 0;
5764 
5765 	if (security_locked_down(LOCKDOWN_TRACEFS)) {
5766 		pr_warn("Lockdown is enabled, skipping ring buffer tests\n");
5767 		return 0;
5768 	}
5769 
5770 	pr_info("Running ring buffer tests...\n");
5771 
5772 	buffer = ring_buffer_alloc(RB_TEST_BUFFER_SIZE, RB_FL_OVERWRITE);
5773 	if (WARN_ON(!buffer))
5774 		return 0;
5775 
5776 	/* Disable buffer so that threads can't write to it yet */
5777 	ring_buffer_record_off(buffer);
5778 
5779 	for_each_online_cpu(cpu) {
5780 		rb_data[cpu].buffer = buffer;
5781 		rb_data[cpu].cpu = cpu;
5782 		rb_data[cpu].cnt = cpu;
5783 		rb_threads[cpu] = kthread_create(rb_test, &rb_data[cpu],
5784 						 "rbtester/%d", cpu);
5785 		if (WARN_ON(IS_ERR(rb_threads[cpu]))) {
5786 			pr_cont("FAILED\n");
5787 			ret = PTR_ERR(rb_threads[cpu]);
5788 			goto out_free;
5789 		}
5790 
5791 		kthread_bind(rb_threads[cpu], cpu);
5792  		wake_up_process(rb_threads[cpu]);
5793 	}
5794 
5795 	/* Now create the rb hammer! */
5796 	rb_hammer = kthread_run(rb_hammer_test, NULL, "rbhammer");
5797 	if (WARN_ON(IS_ERR(rb_hammer))) {
5798 		pr_cont("FAILED\n");
5799 		ret = PTR_ERR(rb_hammer);
5800 		goto out_free;
5801 	}
5802 
5803 	ring_buffer_record_on(buffer);
5804 	/*
5805 	 * Show buffer is enabled before setting rb_test_started.
5806 	 * Yes there's a small race window where events could be
5807 	 * dropped and the thread wont catch it. But when a ring
5808 	 * buffer gets enabled, there will always be some kind of
5809 	 * delay before other CPUs see it. Thus, we don't care about
5810 	 * those dropped events. We care about events dropped after
5811 	 * the threads see that the buffer is active.
5812 	 */
5813 	smp_wmb();
5814 	rb_test_started = true;
5815 
5816 	set_current_state(TASK_INTERRUPTIBLE);
5817 	/* Just run for 10 seconds */;
5818 	schedule_timeout(10 * HZ);
5819 
5820 	kthread_stop(rb_hammer);
5821 
5822  out_free:
5823 	for_each_online_cpu(cpu) {
5824 		if (!rb_threads[cpu])
5825 			break;
5826 		kthread_stop(rb_threads[cpu]);
5827 	}
5828 	if (ret) {
5829 		ring_buffer_free(buffer);
5830 		return ret;
5831 	}
5832 
5833 	/* Report! */
5834 	pr_info("finished\n");
5835 	for_each_online_cpu(cpu) {
5836 		struct ring_buffer_event *event;
5837 		struct rb_test_data *data = &rb_data[cpu];
5838 		struct rb_item *item;
5839 		unsigned long total_events;
5840 		unsigned long total_dropped;
5841 		unsigned long total_written;
5842 		unsigned long total_alloc;
5843 		unsigned long total_read = 0;
5844 		unsigned long total_size = 0;
5845 		unsigned long total_len = 0;
5846 		unsigned long total_lost = 0;
5847 		unsigned long lost;
5848 		int big_event_size;
5849 		int small_event_size;
5850 
5851 		ret = -1;
5852 
5853 		total_events = data->events + data->events_nested;
5854 		total_written = data->bytes_written + data->bytes_written_nested;
5855 		total_alloc = data->bytes_alloc + data->bytes_alloc_nested;
5856 		total_dropped = data->bytes_dropped + data->bytes_dropped_nested;
5857 
5858 		big_event_size = data->max_size + data->max_size_nested;
5859 		small_event_size = data->min_size + data->min_size_nested;
5860 
5861 		pr_info("CPU %d:\n", cpu);
5862 		pr_info("              events:    %ld\n", total_events);
5863 		pr_info("       dropped bytes:    %ld\n", total_dropped);
5864 		pr_info("       alloced bytes:    %ld\n", total_alloc);
5865 		pr_info("       written bytes:    %ld\n", total_written);
5866 		pr_info("       biggest event:    %d\n", big_event_size);
5867 		pr_info("      smallest event:    %d\n", small_event_size);
5868 
5869 		if (RB_WARN_ON(buffer, total_dropped))
5870 			break;
5871 
5872 		ret = 0;
5873 
5874 		while ((event = ring_buffer_consume(buffer, cpu, NULL, &lost))) {
5875 			total_lost += lost;
5876 			item = ring_buffer_event_data(event);
5877 			total_len += ring_buffer_event_length(event);
5878 			total_size += item->size + sizeof(struct rb_item);
5879 			if (memcmp(&item->str[0], rb_string, item->size) != 0) {
5880 				pr_info("FAILED!\n");
5881 				pr_info("buffer had: %.*s\n", item->size, item->str);
5882 				pr_info("expected:   %.*s\n", item->size, rb_string);
5883 				RB_WARN_ON(buffer, 1);
5884 				ret = -1;
5885 				break;
5886 			}
5887 			total_read++;
5888 		}
5889 		if (ret)
5890 			break;
5891 
5892 		ret = -1;
5893 
5894 		pr_info("         read events:   %ld\n", total_read);
5895 		pr_info("         lost events:   %ld\n", total_lost);
5896 		pr_info("        total events:   %ld\n", total_lost + total_read);
5897 		pr_info("  recorded len bytes:   %ld\n", total_len);
5898 		pr_info(" recorded size bytes:   %ld\n", total_size);
5899 		if (total_lost)
5900 			pr_info(" With dropped events, record len and size may not match\n"
5901 				" alloced and written from above\n");
5902 		if (!total_lost) {
5903 			if (RB_WARN_ON(buffer, total_len != total_alloc ||
5904 				       total_size != total_written))
5905 				break;
5906 		}
5907 		if (RB_WARN_ON(buffer, total_lost + total_read != total_events))
5908 			break;
5909 
5910 		ret = 0;
5911 	}
5912 	if (!ret)
5913 		pr_info("Ring buffer PASSED!\n");
5914 
5915 	ring_buffer_free(buffer);
5916 	return 0;
5917 }
5918 
5919 late_initcall(test_ringbuffer);
5920 #endif /* CONFIG_RING_BUFFER_STARTUP_TEST */
5921