1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * Generic ring buffer
4 *
5 * Copyright (C) 2008 Steven Rostedt <srostedt@redhat.com>
6 */
7 #include <linux/trace_recursion.h>
8 #include <linux/trace_events.h>
9 #include <linux/ring_buffer.h>
10 #include <linux/trace_clock.h>
11 #include <linux/sched/clock.h>
12 #include <linux/trace_seq.h>
13 #include <linux/spinlock.h>
14 #include <linux/irq_work.h>
15 #include <linux/security.h>
16 #include <linux/uaccess.h>
17 #include <linux/hardirq.h>
18 #include <linux/kthread.h> /* for self test */
19 #include <linux/module.h>
20 #include <linux/percpu.h>
21 #include <linux/mutex.h>
22 #include <linux/delay.h>
23 #include <linux/slab.h>
24 #include <linux/init.h>
25 #include <linux/hash.h>
26 #include <linux/list.h>
27 #include <linux/cpu.h>
28 #include <linux/oom.h>
29
30 #include <asm/local.h>
31
32 /*
33 * The "absolute" timestamp in the buffer is only 59 bits.
34 * If a clock has the 5 MSBs set, it needs to be saved and
35 * reinserted.
36 */
37 #define TS_MSB (0xf8ULL << 56)
38 #define ABS_TS_MASK (~TS_MSB)
39
40 static void update_pages_handler(struct work_struct *work);
41
42 /*
43 * The ring buffer header is special. We must manually up keep it.
44 */
ring_buffer_print_entry_header(struct trace_seq * s)45 int ring_buffer_print_entry_header(struct trace_seq *s)
46 {
47 trace_seq_puts(s, "# compressed entry header\n");
48 trace_seq_puts(s, "\ttype_len : 5 bits\n");
49 trace_seq_puts(s, "\ttime_delta : 27 bits\n");
50 trace_seq_puts(s, "\tarray : 32 bits\n");
51 trace_seq_putc(s, '\n');
52 trace_seq_printf(s, "\tpadding : type == %d\n",
53 RINGBUF_TYPE_PADDING);
54 trace_seq_printf(s, "\ttime_extend : type == %d\n",
55 RINGBUF_TYPE_TIME_EXTEND);
56 trace_seq_printf(s, "\ttime_stamp : type == %d\n",
57 RINGBUF_TYPE_TIME_STAMP);
58 trace_seq_printf(s, "\tdata max type_len == %d\n",
59 RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
60
61 return !trace_seq_has_overflowed(s);
62 }
63
64 /*
65 * The ring buffer is made up of a list of pages. A separate list of pages is
66 * allocated for each CPU. A writer may only write to a buffer that is
67 * associated with the CPU it is currently executing on. A reader may read
68 * from any per cpu buffer.
69 *
70 * The reader is special. For each per cpu buffer, the reader has its own
71 * reader page. When a reader has read the entire reader page, this reader
72 * page is swapped with another page in the ring buffer.
73 *
74 * Now, as long as the writer is off the reader page, the reader can do what
75 * ever it wants with that page. The writer will never write to that page
76 * again (as long as it is out of the ring buffer).
77 *
78 * Here's some silly ASCII art.
79 *
80 * +------+
81 * |reader| RING BUFFER
82 * |page |
83 * +------+ +---+ +---+ +---+
84 * | |-->| |-->| |
85 * +---+ +---+ +---+
86 * ^ |
87 * | |
88 * +---------------+
89 *
90 *
91 * +------+
92 * |reader| RING BUFFER
93 * |page |------------------v
94 * +------+ +---+ +---+ +---+
95 * | |-->| |-->| |
96 * +---+ +---+ +---+
97 * ^ |
98 * | |
99 * +---------------+
100 *
101 *
102 * +------+
103 * |reader| RING BUFFER
104 * |page |------------------v
105 * +------+ +---+ +---+ +---+
106 * ^ | |-->| |-->| |
107 * | +---+ +---+ +---+
108 * | |
109 * | |
110 * +------------------------------+
111 *
112 *
113 * +------+
114 * |buffer| RING BUFFER
115 * |page |------------------v
116 * +------+ +---+ +---+ +---+
117 * ^ | | | |-->| |
118 * | New +---+ +---+ +---+
119 * | Reader------^ |
120 * | page |
121 * +------------------------------+
122 *
123 *
124 * After we make this swap, the reader can hand this page off to the splice
125 * code and be done with it. It can even allocate a new page if it needs to
126 * and swap that into the ring buffer.
127 *
128 * We will be using cmpxchg soon to make all this lockless.
129 *
130 */
131
132 /* Used for individual buffers (after the counter) */
133 #define RB_BUFFER_OFF (1 << 20)
134
135 #define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
136
137 #define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
138 #define RB_ALIGNMENT 4U
139 #define RB_MAX_SMALL_DATA (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
140 #define RB_EVNT_MIN_SIZE 8U /* two 32bit words */
141
142 #ifndef CONFIG_HAVE_64BIT_ALIGNED_ACCESS
143 # define RB_FORCE_8BYTE_ALIGNMENT 0
144 # define RB_ARCH_ALIGNMENT RB_ALIGNMENT
145 #else
146 # define RB_FORCE_8BYTE_ALIGNMENT 1
147 # define RB_ARCH_ALIGNMENT 8U
148 #endif
149
150 #define RB_ALIGN_DATA __aligned(RB_ARCH_ALIGNMENT)
151
152 /* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
153 #define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
154
155 enum {
156 RB_LEN_TIME_EXTEND = 8,
157 RB_LEN_TIME_STAMP = 8,
158 };
159
160 #define skip_time_extend(event) \
161 ((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
162
163 #define extended_time(event) \
164 (event->type_len >= RINGBUF_TYPE_TIME_EXTEND)
165
rb_null_event(struct ring_buffer_event * event)166 static inline int rb_null_event(struct ring_buffer_event *event)
167 {
168 return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
169 }
170
rb_event_set_padding(struct ring_buffer_event * event)171 static void rb_event_set_padding(struct ring_buffer_event *event)
172 {
173 /* padding has a NULL time_delta */
174 event->type_len = RINGBUF_TYPE_PADDING;
175 event->time_delta = 0;
176 }
177
178 static unsigned
rb_event_data_length(struct ring_buffer_event * event)179 rb_event_data_length(struct ring_buffer_event *event)
180 {
181 unsigned length;
182
183 if (event->type_len)
184 length = event->type_len * RB_ALIGNMENT;
185 else
186 length = event->array[0];
187 return length + RB_EVNT_HDR_SIZE;
188 }
189
190 /*
191 * Return the length of the given event. Will return
192 * the length of the time extend if the event is a
193 * time extend.
194 */
195 static inline unsigned
rb_event_length(struct ring_buffer_event * event)196 rb_event_length(struct ring_buffer_event *event)
197 {
198 switch (event->type_len) {
199 case RINGBUF_TYPE_PADDING:
200 if (rb_null_event(event))
201 /* undefined */
202 return -1;
203 return event->array[0] + RB_EVNT_HDR_SIZE;
204
205 case RINGBUF_TYPE_TIME_EXTEND:
206 return RB_LEN_TIME_EXTEND;
207
208 case RINGBUF_TYPE_TIME_STAMP:
209 return RB_LEN_TIME_STAMP;
210
211 case RINGBUF_TYPE_DATA:
212 return rb_event_data_length(event);
213 default:
214 WARN_ON_ONCE(1);
215 }
216 /* not hit */
217 return 0;
218 }
219
220 /*
221 * Return total length of time extend and data,
222 * or just the event length for all other events.
223 */
224 static inline unsigned
rb_event_ts_length(struct ring_buffer_event * event)225 rb_event_ts_length(struct ring_buffer_event *event)
226 {
227 unsigned len = 0;
228
229 if (extended_time(event)) {
230 /* time extends include the data event after it */
231 len = RB_LEN_TIME_EXTEND;
232 event = skip_time_extend(event);
233 }
234 return len + rb_event_length(event);
235 }
236
237 /**
238 * ring_buffer_event_length - return the length of the event
239 * @event: the event to get the length of
240 *
241 * Returns the size of the data load of a data event.
242 * If the event is something other than a data event, it
243 * returns the size of the event itself. With the exception
244 * of a TIME EXTEND, where it still returns the size of the
245 * data load of the data event after it.
246 */
ring_buffer_event_length(struct ring_buffer_event * event)247 unsigned ring_buffer_event_length(struct ring_buffer_event *event)
248 {
249 unsigned length;
250
251 if (extended_time(event))
252 event = skip_time_extend(event);
253
254 length = rb_event_length(event);
255 if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
256 return length;
257 length -= RB_EVNT_HDR_SIZE;
258 if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
259 length -= sizeof(event->array[0]);
260 return length;
261 }
262 EXPORT_SYMBOL_GPL(ring_buffer_event_length);
263
264 /* inline for ring buffer fast paths */
265 static __always_inline void *
rb_event_data(struct ring_buffer_event * event)266 rb_event_data(struct ring_buffer_event *event)
267 {
268 if (extended_time(event))
269 event = skip_time_extend(event);
270 WARN_ON_ONCE(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
271 /* If length is in len field, then array[0] has the data */
272 if (event->type_len)
273 return (void *)&event->array[0];
274 /* Otherwise length is in array[0] and array[1] has the data */
275 return (void *)&event->array[1];
276 }
277
278 /**
279 * ring_buffer_event_data - return the data of the event
280 * @event: the event to get the data from
281 */
ring_buffer_event_data(struct ring_buffer_event * event)282 void *ring_buffer_event_data(struct ring_buffer_event *event)
283 {
284 return rb_event_data(event);
285 }
286 EXPORT_SYMBOL_GPL(ring_buffer_event_data);
287
288 #define for_each_buffer_cpu(buffer, cpu) \
289 for_each_cpu(cpu, buffer->cpumask)
290
291 #define for_each_online_buffer_cpu(buffer, cpu) \
292 for_each_cpu_and(cpu, buffer->cpumask, cpu_online_mask)
293
294 #define TS_SHIFT 27
295 #define TS_MASK ((1ULL << TS_SHIFT) - 1)
296 #define TS_DELTA_TEST (~TS_MASK)
297
rb_event_time_stamp(struct ring_buffer_event * event)298 static u64 rb_event_time_stamp(struct ring_buffer_event *event)
299 {
300 u64 ts;
301
302 ts = event->array[0];
303 ts <<= TS_SHIFT;
304 ts += event->time_delta;
305
306 return ts;
307 }
308
309 /* Flag when events were overwritten */
310 #define RB_MISSED_EVENTS (1 << 31)
311 /* Missed count stored at end */
312 #define RB_MISSED_STORED (1 << 30)
313
314 struct buffer_data_page {
315 u64 time_stamp; /* page time stamp */
316 local_t commit; /* write committed index */
317 unsigned char data[] RB_ALIGN_DATA; /* data of buffer page */
318 };
319
320 /*
321 * Note, the buffer_page list must be first. The buffer pages
322 * are allocated in cache lines, which means that each buffer
323 * page will be at the beginning of a cache line, and thus
324 * the least significant bits will be zero. We use this to
325 * add flags in the list struct pointers, to make the ring buffer
326 * lockless.
327 */
328 struct buffer_page {
329 struct list_head list; /* list of buffer pages */
330 local_t write; /* index for next write */
331 unsigned read; /* index for next read */
332 local_t entries; /* entries on this page */
333 unsigned long real_end; /* real end of data */
334 struct buffer_data_page *page; /* Actual data page */
335 };
336
337 /*
338 * The buffer page counters, write and entries, must be reset
339 * atomically when crossing page boundaries. To synchronize this
340 * update, two counters are inserted into the number. One is
341 * the actual counter for the write position or count on the page.
342 *
343 * The other is a counter of updaters. Before an update happens
344 * the update partition of the counter is incremented. This will
345 * allow the updater to update the counter atomically.
346 *
347 * The counter is 20 bits, and the state data is 12.
348 */
349 #define RB_WRITE_MASK 0xfffff
350 #define RB_WRITE_INTCNT (1 << 20)
351
rb_init_page(struct buffer_data_page * bpage)352 static void rb_init_page(struct buffer_data_page *bpage)
353 {
354 local_set(&bpage->commit, 0);
355 }
356
357 /*
358 * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
359 * this issue out.
360 */
free_buffer_page(struct buffer_page * bpage)361 static void free_buffer_page(struct buffer_page *bpage)
362 {
363 free_page((unsigned long)bpage->page);
364 kfree(bpage);
365 }
366
367 /*
368 * We need to fit the time_stamp delta into 27 bits.
369 */
test_time_stamp(u64 delta)370 static inline int test_time_stamp(u64 delta)
371 {
372 if (delta & TS_DELTA_TEST)
373 return 1;
374 return 0;
375 }
376
377 #define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
378
379 /* Max payload is BUF_PAGE_SIZE - header (8bytes) */
380 #define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
381
ring_buffer_print_page_header(struct trace_seq * s)382 int ring_buffer_print_page_header(struct trace_seq *s)
383 {
384 struct buffer_data_page field;
385
386 trace_seq_printf(s, "\tfield: u64 timestamp;\t"
387 "offset:0;\tsize:%u;\tsigned:%u;\n",
388 (unsigned int)sizeof(field.time_stamp),
389 (unsigned int)is_signed_type(u64));
390
391 trace_seq_printf(s, "\tfield: local_t commit;\t"
392 "offset:%u;\tsize:%u;\tsigned:%u;\n",
393 (unsigned int)offsetof(typeof(field), commit),
394 (unsigned int)sizeof(field.commit),
395 (unsigned int)is_signed_type(long));
396
397 trace_seq_printf(s, "\tfield: int overwrite;\t"
398 "offset:%u;\tsize:%u;\tsigned:%u;\n",
399 (unsigned int)offsetof(typeof(field), commit),
400 1,
401 (unsigned int)is_signed_type(long));
402
403 trace_seq_printf(s, "\tfield: char data;\t"
404 "offset:%u;\tsize:%u;\tsigned:%u;\n",
405 (unsigned int)offsetof(typeof(field), data),
406 (unsigned int)BUF_PAGE_SIZE,
407 (unsigned int)is_signed_type(char));
408
409 return !trace_seq_has_overflowed(s);
410 }
411
412 struct rb_irq_work {
413 struct irq_work work;
414 wait_queue_head_t waiters;
415 wait_queue_head_t full_waiters;
416 long wait_index;
417 bool waiters_pending;
418 bool full_waiters_pending;
419 bool wakeup_full;
420 };
421
422 /*
423 * Structure to hold event state and handle nested events.
424 */
425 struct rb_event_info {
426 u64 ts;
427 u64 delta;
428 u64 before;
429 u64 after;
430 unsigned long length;
431 struct buffer_page *tail_page;
432 int add_timestamp;
433 };
434
435 /*
436 * Used for the add_timestamp
437 * NONE
438 * EXTEND - wants a time extend
439 * ABSOLUTE - the buffer requests all events to have absolute time stamps
440 * FORCE - force a full time stamp.
441 */
442 enum {
443 RB_ADD_STAMP_NONE = 0,
444 RB_ADD_STAMP_EXTEND = BIT(1),
445 RB_ADD_STAMP_ABSOLUTE = BIT(2),
446 RB_ADD_STAMP_FORCE = BIT(3)
447 };
448 /*
449 * Used for which event context the event is in.
450 * TRANSITION = 0
451 * NMI = 1
452 * IRQ = 2
453 * SOFTIRQ = 3
454 * NORMAL = 4
455 *
456 * See trace_recursive_lock() comment below for more details.
457 */
458 enum {
459 RB_CTX_TRANSITION,
460 RB_CTX_NMI,
461 RB_CTX_IRQ,
462 RB_CTX_SOFTIRQ,
463 RB_CTX_NORMAL,
464 RB_CTX_MAX
465 };
466
467 #if BITS_PER_LONG == 32
468 #define RB_TIME_32
469 #endif
470
471 /* To test on 64 bit machines */
472 //#define RB_TIME_32
473
474 #ifdef RB_TIME_32
475
476 struct rb_time_struct {
477 local_t cnt;
478 local_t top;
479 local_t bottom;
480 local_t msb;
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 #define MAX_NEST 5
491
492 /*
493 * head_page == tail_page && head == tail then buffer is empty.
494 */
495 struct ring_buffer_per_cpu {
496 int cpu;
497 atomic_t record_disabled;
498 atomic_t resize_disabled;
499 struct trace_buffer *buffer;
500 raw_spinlock_t reader_lock; /* serialize readers */
501 arch_spinlock_t lock;
502 struct lock_class_key lock_key;
503 struct buffer_data_page *free_page;
504 unsigned long nr_pages;
505 unsigned int current_context;
506 struct list_head *pages;
507 struct buffer_page *head_page; /* read from head */
508 struct buffer_page *tail_page; /* write to tail */
509 struct buffer_page *commit_page; /* committed pages */
510 struct buffer_page *reader_page;
511 unsigned long lost_events;
512 unsigned long last_overrun;
513 unsigned long nest;
514 local_t entries_bytes;
515 local_t entries;
516 local_t overrun;
517 local_t commit_overrun;
518 local_t dropped_events;
519 local_t committing;
520 local_t commits;
521 local_t pages_touched;
522 local_t pages_lost;
523 local_t pages_read;
524 long last_pages_touch;
525 size_t shortest_full;
526 unsigned long read;
527 unsigned long read_bytes;
528 rb_time_t write_stamp;
529 rb_time_t before_stamp;
530 u64 event_stamp[MAX_NEST];
531 u64 read_stamp;
532 /* ring buffer pages to update, > 0 to add, < 0 to remove */
533 long nr_pages_to_update;
534 struct list_head new_pages; /* new pages to add */
535 struct work_struct update_pages_work;
536 struct completion update_done;
537
538 struct rb_irq_work irq_work;
539 };
540
541 struct trace_buffer {
542 unsigned flags;
543 int cpus;
544 atomic_t record_disabled;
545 cpumask_var_t cpumask;
546
547 struct lock_class_key *reader_lock_key;
548
549 struct mutex mutex;
550
551 struct ring_buffer_per_cpu **buffers;
552
553 struct hlist_node node;
554 u64 (*clock)(void);
555
556 struct rb_irq_work irq_work;
557 bool time_stamp_abs;
558 };
559
560 struct ring_buffer_iter {
561 struct ring_buffer_per_cpu *cpu_buffer;
562 unsigned long head;
563 unsigned long next_event;
564 struct buffer_page *head_page;
565 struct buffer_page *cache_reader_page;
566 unsigned long cache_read;
567 u64 read_stamp;
568 u64 page_stamp;
569 struct ring_buffer_event *event;
570 int missed_events;
571 };
572
573 #ifdef RB_TIME_32
574
575 /*
576 * On 32 bit machines, local64_t is very expensive. As the ring
577 * buffer doesn't need all the features of a true 64 bit atomic,
578 * on 32 bit, it uses these functions (64 still uses local64_t).
579 *
580 * For the ring buffer, 64 bit required operations for the time is
581 * the following:
582 *
583 * - Reads may fail if it interrupted a modification of the time stamp.
584 * It will succeed if it did not interrupt another write even if
585 * the read itself is interrupted by a write.
586 * It returns whether it was successful or not.
587 *
588 * - Writes always succeed and will overwrite other writes and writes
589 * that were done by events interrupting the current write.
590 *
591 * - A write followed by a read of the same time stamp will always succeed,
592 * but may not contain the same value.
593 *
594 * - A cmpxchg will fail if it interrupted another write or cmpxchg.
595 * Other than that, it acts like a normal cmpxchg.
596 *
597 * The 60 bit time stamp is broken up by 30 bits in a top and bottom half
598 * (bottom being the least significant 30 bits of the 60 bit time stamp).
599 *
600 * The two most significant bits of each half holds a 2 bit counter (0-3).
601 * Each update will increment this counter by one.
602 * When reading the top and bottom, if the two counter bits match then the
603 * top and bottom together make a valid 60 bit number.
604 */
605 #define RB_TIME_SHIFT 30
606 #define RB_TIME_VAL_MASK ((1 << RB_TIME_SHIFT) - 1)
607 #define RB_TIME_MSB_SHIFT 60
608
rb_time_cnt(unsigned long val)609 static inline int rb_time_cnt(unsigned long val)
610 {
611 return (val >> RB_TIME_SHIFT) & 3;
612 }
613
rb_time_val(unsigned long top,unsigned long bottom)614 static inline u64 rb_time_val(unsigned long top, unsigned long bottom)
615 {
616 u64 val;
617
618 val = top & RB_TIME_VAL_MASK;
619 val <<= RB_TIME_SHIFT;
620 val |= bottom & RB_TIME_VAL_MASK;
621
622 return val;
623 }
624
__rb_time_read(rb_time_t * t,u64 * ret,unsigned long * cnt)625 static inline bool __rb_time_read(rb_time_t *t, u64 *ret, unsigned long *cnt)
626 {
627 unsigned long top, bottom, msb;
628 unsigned long c;
629
630 /*
631 * If the read is interrupted by a write, then the cnt will
632 * be different. Loop until both top and bottom have been read
633 * without interruption.
634 */
635 do {
636 c = local_read(&t->cnt);
637 top = local_read(&t->top);
638 bottom = local_read(&t->bottom);
639 msb = local_read(&t->msb);
640 } while (c != local_read(&t->cnt));
641
642 *cnt = rb_time_cnt(top);
643
644 /* If top and bottom counts don't match, this interrupted a write */
645 if (*cnt != rb_time_cnt(bottom))
646 return false;
647
648 /* The shift to msb will lose its cnt bits */
649 *ret = rb_time_val(top, bottom) | ((u64)msb << RB_TIME_MSB_SHIFT);
650 return true;
651 }
652
rb_time_read(rb_time_t * t,u64 * ret)653 static bool rb_time_read(rb_time_t *t, u64 *ret)
654 {
655 unsigned long cnt;
656
657 return __rb_time_read(t, ret, &cnt);
658 }
659
rb_time_val_cnt(unsigned long val,unsigned long cnt)660 static inline unsigned long rb_time_val_cnt(unsigned long val, unsigned long cnt)
661 {
662 return (val & RB_TIME_VAL_MASK) | ((cnt & 3) << RB_TIME_SHIFT);
663 }
664
rb_time_split(u64 val,unsigned long * top,unsigned long * bottom,unsigned long * msb)665 static inline void rb_time_split(u64 val, unsigned long *top, unsigned long *bottom,
666 unsigned long *msb)
667 {
668 *top = (unsigned long)((val >> RB_TIME_SHIFT) & RB_TIME_VAL_MASK);
669 *bottom = (unsigned long)(val & RB_TIME_VAL_MASK);
670 *msb = (unsigned long)(val >> RB_TIME_MSB_SHIFT);
671 }
672
rb_time_val_set(local_t * t,unsigned long val,unsigned long cnt)673 static inline void rb_time_val_set(local_t *t, unsigned long val, unsigned long cnt)
674 {
675 val = rb_time_val_cnt(val, cnt);
676 local_set(t, val);
677 }
678
rb_time_set(rb_time_t * t,u64 val)679 static void rb_time_set(rb_time_t *t, u64 val)
680 {
681 unsigned long cnt, top, bottom, msb;
682
683 rb_time_split(val, &top, &bottom, &msb);
684
685 /* Writes always succeed with a valid number even if it gets interrupted. */
686 do {
687 cnt = local_inc_return(&t->cnt);
688 rb_time_val_set(&t->top, top, cnt);
689 rb_time_val_set(&t->bottom, bottom, cnt);
690 rb_time_val_set(&t->msb, val >> RB_TIME_MSB_SHIFT, cnt);
691 } while (cnt != local_read(&t->cnt));
692 }
693
694 static inline bool
rb_time_read_cmpxchg(local_t * l,unsigned long expect,unsigned long set)695 rb_time_read_cmpxchg(local_t *l, unsigned long expect, unsigned long set)
696 {
697 unsigned long ret;
698
699 ret = local_cmpxchg(l, expect, set);
700 return ret == expect;
701 }
702
rb_time_cmpxchg(rb_time_t * t,u64 expect,u64 set)703 static int rb_time_cmpxchg(rb_time_t *t, u64 expect, u64 set)
704 {
705 unsigned long cnt, top, bottom, msb;
706 unsigned long cnt2, top2, bottom2, msb2;
707 u64 val;
708
709 /* The cmpxchg always fails if it interrupted an update */
710 if (!__rb_time_read(t, &val, &cnt2))
711 return false;
712
713 if (val != expect)
714 return false;
715
716 cnt = local_read(&t->cnt);
717 if ((cnt & 3) != cnt2)
718 return false;
719
720 cnt2 = cnt + 1;
721
722 rb_time_split(val, &top, &bottom, &msb);
723 top = rb_time_val_cnt(top, cnt);
724 bottom = rb_time_val_cnt(bottom, cnt);
725
726 rb_time_split(set, &top2, &bottom2, &msb2);
727 top2 = rb_time_val_cnt(top2, cnt2);
728 bottom2 = rb_time_val_cnt(bottom2, cnt2);
729
730 if (!rb_time_read_cmpxchg(&t->cnt, cnt, cnt2))
731 return false;
732 if (!rb_time_read_cmpxchg(&t->msb, msb, msb2))
733 return false;
734 if (!rb_time_read_cmpxchg(&t->top, top, top2))
735 return false;
736 if (!rb_time_read_cmpxchg(&t->bottom, bottom, bottom2))
737 return false;
738 return true;
739 }
740
741 #else /* 64 bits */
742
743 /* local64_t always succeeds */
744
rb_time_read(rb_time_t * t,u64 * ret)745 static inline bool rb_time_read(rb_time_t *t, u64 *ret)
746 {
747 *ret = local64_read(&t->time);
748 return true;
749 }
rb_time_set(rb_time_t * t,u64 val)750 static void rb_time_set(rb_time_t *t, u64 val)
751 {
752 local64_set(&t->time, val);
753 }
754
rb_time_cmpxchg(rb_time_t * t,u64 expect,u64 set)755 static bool rb_time_cmpxchg(rb_time_t *t, u64 expect, u64 set)
756 {
757 u64 val;
758 val = local64_cmpxchg(&t->time, expect, set);
759 return val == expect;
760 }
761 #endif
762
763 /*
764 * Enable this to make sure that the event passed to
765 * ring_buffer_event_time_stamp() is not committed and also
766 * is on the buffer that it passed in.
767 */
768 //#define RB_VERIFY_EVENT
769 #ifdef RB_VERIFY_EVENT
770 static struct list_head *rb_list_head(struct list_head *list);
verify_event(struct ring_buffer_per_cpu * cpu_buffer,void * event)771 static void verify_event(struct ring_buffer_per_cpu *cpu_buffer,
772 void *event)
773 {
774 struct buffer_page *page = cpu_buffer->commit_page;
775 struct buffer_page *tail_page = READ_ONCE(cpu_buffer->tail_page);
776 struct list_head *next;
777 long commit, write;
778 unsigned long addr = (unsigned long)event;
779 bool done = false;
780 int stop = 0;
781
782 /* Make sure the event exists and is not committed yet */
783 do {
784 if (page == tail_page || WARN_ON_ONCE(stop++ > 100))
785 done = true;
786 commit = local_read(&page->page->commit);
787 write = local_read(&page->write);
788 if (addr >= (unsigned long)&page->page->data[commit] &&
789 addr < (unsigned long)&page->page->data[write])
790 return;
791
792 next = rb_list_head(page->list.next);
793 page = list_entry(next, struct buffer_page, list);
794 } while (!done);
795 WARN_ON_ONCE(1);
796 }
797 #else
verify_event(struct ring_buffer_per_cpu * cpu_buffer,void * event)798 static inline void verify_event(struct ring_buffer_per_cpu *cpu_buffer,
799 void *event)
800 {
801 }
802 #endif
803
804 /*
805 * The absolute time stamp drops the 5 MSBs and some clocks may
806 * require them. The rb_fix_abs_ts() will take a previous full
807 * time stamp, and add the 5 MSB of that time stamp on to the
808 * saved absolute time stamp. Then they are compared in case of
809 * the unlikely event that the latest time stamp incremented
810 * the 5 MSB.
811 */
rb_fix_abs_ts(u64 abs,u64 save_ts)812 static inline u64 rb_fix_abs_ts(u64 abs, u64 save_ts)
813 {
814 if (save_ts & TS_MSB) {
815 abs |= save_ts & TS_MSB;
816 /* Check for overflow */
817 if (unlikely(abs < save_ts))
818 abs += 1ULL << 59;
819 }
820 return abs;
821 }
822
823 static inline u64 rb_time_stamp(struct trace_buffer *buffer);
824
825 /**
826 * ring_buffer_event_time_stamp - return the event's current time stamp
827 * @buffer: The buffer that the event is on
828 * @event: the event to get the time stamp of
829 *
830 * Note, this must be called after @event is reserved, and before it is
831 * committed to the ring buffer. And must be called from the same
832 * context where the event was reserved (normal, softirq, irq, etc).
833 *
834 * Returns the time stamp associated with the current event.
835 * If the event has an extended time stamp, then that is used as
836 * the time stamp to return.
837 * In the highly unlikely case that the event was nested more than
838 * the max nesting, then the write_stamp of the buffer is returned,
839 * otherwise current time is returned, but that really neither of
840 * the last two cases should ever happen.
841 */
ring_buffer_event_time_stamp(struct trace_buffer * buffer,struct ring_buffer_event * event)842 u64 ring_buffer_event_time_stamp(struct trace_buffer *buffer,
843 struct ring_buffer_event *event)
844 {
845 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[smp_processor_id()];
846 unsigned int nest;
847 u64 ts;
848
849 /* If the event includes an absolute time, then just use that */
850 if (event->type_len == RINGBUF_TYPE_TIME_STAMP) {
851 ts = rb_event_time_stamp(event);
852 return rb_fix_abs_ts(ts, cpu_buffer->tail_page->page->time_stamp);
853 }
854
855 nest = local_read(&cpu_buffer->committing);
856 verify_event(cpu_buffer, event);
857 if (WARN_ON_ONCE(!nest))
858 goto fail;
859
860 /* Read the current saved nesting level time stamp */
861 if (likely(--nest < MAX_NEST))
862 return cpu_buffer->event_stamp[nest];
863
864 /* Shouldn't happen, warn if it does */
865 WARN_ONCE(1, "nest (%d) greater than max", nest);
866
867 fail:
868 /* Can only fail on 32 bit */
869 if (!rb_time_read(&cpu_buffer->write_stamp, &ts))
870 /* Screw it, just read the current time */
871 ts = rb_time_stamp(cpu_buffer->buffer);
872
873 return ts;
874 }
875
876 /**
877 * ring_buffer_nr_pages - get the number of buffer pages in the ring buffer
878 * @buffer: The ring_buffer to get the number of pages from
879 * @cpu: The cpu of the ring_buffer to get the number of pages from
880 *
881 * Returns the number of pages used by a per_cpu buffer of the ring buffer.
882 */
ring_buffer_nr_pages(struct trace_buffer * buffer,int cpu)883 size_t ring_buffer_nr_pages(struct trace_buffer *buffer, int cpu)
884 {
885 return buffer->buffers[cpu]->nr_pages;
886 }
887
888 /**
889 * ring_buffer_nr_dirty_pages - get the number of used pages in the ring buffer
890 * @buffer: The ring_buffer to get the number of pages from
891 * @cpu: The cpu of the ring_buffer to get the number of pages from
892 *
893 * Returns the number of pages that have content in the ring buffer.
894 */
ring_buffer_nr_dirty_pages(struct trace_buffer * buffer,int cpu)895 size_t ring_buffer_nr_dirty_pages(struct trace_buffer *buffer, int cpu)
896 {
897 size_t read;
898 size_t lost;
899 size_t cnt;
900
901 read = local_read(&buffer->buffers[cpu]->pages_read);
902 lost = local_read(&buffer->buffers[cpu]->pages_lost);
903 cnt = local_read(&buffer->buffers[cpu]->pages_touched);
904
905 if (WARN_ON_ONCE(cnt < lost))
906 return 0;
907
908 cnt -= lost;
909
910 /* The reader can read an empty page, but not more than that */
911 if (cnt < read) {
912 WARN_ON_ONCE(read > cnt + 1);
913 return 0;
914 }
915
916 return cnt - read;
917 }
918
full_hit(struct trace_buffer * buffer,int cpu,int full)919 static __always_inline bool full_hit(struct trace_buffer *buffer, int cpu, int full)
920 {
921 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
922 size_t nr_pages;
923 size_t dirty;
924
925 nr_pages = cpu_buffer->nr_pages;
926 if (!nr_pages || !full)
927 return true;
928
929 dirty = ring_buffer_nr_dirty_pages(buffer, cpu);
930
931 return (dirty * 100) > (full * nr_pages);
932 }
933
934 /*
935 * rb_wake_up_waiters - wake up tasks waiting for ring buffer input
936 *
937 * Schedules a delayed work to wake up any task that is blocked on the
938 * ring buffer waiters queue.
939 */
rb_wake_up_waiters(struct irq_work * work)940 static void rb_wake_up_waiters(struct irq_work *work)
941 {
942 struct rb_irq_work *rbwork = container_of(work, struct rb_irq_work, work);
943
944 wake_up_all(&rbwork->waiters);
945 if (rbwork->full_waiters_pending || rbwork->wakeup_full) {
946 rbwork->wakeup_full = false;
947 rbwork->full_waiters_pending = false;
948 wake_up_all(&rbwork->full_waiters);
949 }
950 }
951
952 /**
953 * ring_buffer_wake_waiters - wake up any waiters on this ring buffer
954 * @buffer: The ring buffer to wake waiters on
955 *
956 * In the case of a file that represents a ring buffer is closing,
957 * it is prudent to wake up any waiters that are on this.
958 */
ring_buffer_wake_waiters(struct trace_buffer * buffer,int cpu)959 void ring_buffer_wake_waiters(struct trace_buffer *buffer, int cpu)
960 {
961 struct ring_buffer_per_cpu *cpu_buffer;
962 struct rb_irq_work *rbwork;
963
964 if (!buffer)
965 return;
966
967 if (cpu == RING_BUFFER_ALL_CPUS) {
968
969 /* Wake up individual ones too. One level recursion */
970 for_each_buffer_cpu(buffer, cpu)
971 ring_buffer_wake_waiters(buffer, cpu);
972
973 rbwork = &buffer->irq_work;
974 } else {
975 if (WARN_ON_ONCE(!buffer->buffers))
976 return;
977 if (WARN_ON_ONCE(cpu >= nr_cpu_ids))
978 return;
979
980 cpu_buffer = buffer->buffers[cpu];
981 /* The CPU buffer may not have been initialized yet */
982 if (!cpu_buffer)
983 return;
984 rbwork = &cpu_buffer->irq_work;
985 }
986
987 rbwork->wait_index++;
988 /* make sure the waiters see the new index */
989 smp_wmb();
990
991 rb_wake_up_waiters(&rbwork->work);
992 }
993
994 /**
995 * ring_buffer_wait - wait for input to the ring buffer
996 * @buffer: buffer to wait on
997 * @cpu: the cpu buffer to wait on
998 * @full: wait until the percentage of pages are available, if @cpu != RING_BUFFER_ALL_CPUS
999 *
1000 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
1001 * as data is added to any of the @buffer's cpu buffers. Otherwise
1002 * it will wait for data to be added to a specific cpu buffer.
1003 */
ring_buffer_wait(struct trace_buffer * buffer,int cpu,int full)1004 int ring_buffer_wait(struct trace_buffer *buffer, int cpu, int full)
1005 {
1006 struct ring_buffer_per_cpu *cpu_buffer;
1007 DEFINE_WAIT(wait);
1008 struct rb_irq_work *work;
1009 long wait_index;
1010 int ret = 0;
1011
1012 /*
1013 * Depending on what the caller is waiting for, either any
1014 * data in any cpu buffer, or a specific buffer, put the
1015 * caller on the appropriate wait queue.
1016 */
1017 if (cpu == RING_BUFFER_ALL_CPUS) {
1018 work = &buffer->irq_work;
1019 /* Full only makes sense on per cpu reads */
1020 full = 0;
1021 } else {
1022 if (!cpumask_test_cpu(cpu, buffer->cpumask))
1023 return -ENODEV;
1024 cpu_buffer = buffer->buffers[cpu];
1025 work = &cpu_buffer->irq_work;
1026 }
1027
1028 wait_index = READ_ONCE(work->wait_index);
1029
1030 while (true) {
1031 if (full)
1032 prepare_to_wait(&work->full_waiters, &wait, TASK_INTERRUPTIBLE);
1033 else
1034 prepare_to_wait(&work->waiters, &wait, TASK_INTERRUPTIBLE);
1035
1036 /*
1037 * The events can happen in critical sections where
1038 * checking a work queue can cause deadlocks.
1039 * After adding a task to the queue, this flag is set
1040 * only to notify events to try to wake up the queue
1041 * using irq_work.
1042 *
1043 * We don't clear it even if the buffer is no longer
1044 * empty. The flag only causes the next event to run
1045 * irq_work to do the work queue wake up. The worse
1046 * that can happen if we race with !trace_empty() is that
1047 * an event will cause an irq_work to try to wake up
1048 * an empty queue.
1049 *
1050 * There's no reason to protect this flag either, as
1051 * the work queue and irq_work logic will do the necessary
1052 * synchronization for the wake ups. The only thing
1053 * that is necessary is that the wake up happens after
1054 * a task has been queued. It's OK for spurious wake ups.
1055 */
1056 if (full)
1057 work->full_waiters_pending = true;
1058 else
1059 work->waiters_pending = true;
1060
1061 if (signal_pending(current)) {
1062 ret = -EINTR;
1063 break;
1064 }
1065
1066 if (cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer))
1067 break;
1068
1069 if (cpu != RING_BUFFER_ALL_CPUS &&
1070 !ring_buffer_empty_cpu(buffer, cpu)) {
1071 unsigned long flags;
1072 bool pagebusy;
1073 bool done;
1074
1075 if (!full)
1076 break;
1077
1078 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
1079 pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
1080 done = !pagebusy && full_hit(buffer, cpu, full);
1081
1082 if (!cpu_buffer->shortest_full ||
1083 cpu_buffer->shortest_full > full)
1084 cpu_buffer->shortest_full = full;
1085 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
1086 if (done)
1087 break;
1088 }
1089
1090 schedule();
1091
1092 /* Make sure to see the new wait index */
1093 smp_rmb();
1094 if (wait_index != work->wait_index)
1095 break;
1096 }
1097
1098 if (full)
1099 finish_wait(&work->full_waiters, &wait);
1100 else
1101 finish_wait(&work->waiters, &wait);
1102
1103 return ret;
1104 }
1105
1106 /**
1107 * ring_buffer_poll_wait - poll on buffer input
1108 * @buffer: buffer to wait on
1109 * @cpu: the cpu buffer to wait on
1110 * @filp: the file descriptor
1111 * @poll_table: The poll descriptor
1112 * @full: wait until the percentage of pages are available, if @cpu != RING_BUFFER_ALL_CPUS
1113 *
1114 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
1115 * as data is added to any of the @buffer's cpu buffers. Otherwise
1116 * it will wait for data to be added to a specific cpu buffer.
1117 *
1118 * Returns EPOLLIN | EPOLLRDNORM if data exists in the buffers,
1119 * zero otherwise.
1120 */
ring_buffer_poll_wait(struct trace_buffer * buffer,int cpu,struct file * filp,poll_table * poll_table,int full)1121 __poll_t ring_buffer_poll_wait(struct trace_buffer *buffer, int cpu,
1122 struct file *filp, poll_table *poll_table, int full)
1123 {
1124 struct ring_buffer_per_cpu *cpu_buffer;
1125 struct rb_irq_work *work;
1126
1127 if (cpu == RING_BUFFER_ALL_CPUS) {
1128 work = &buffer->irq_work;
1129 full = 0;
1130 } else {
1131 if (!cpumask_test_cpu(cpu, buffer->cpumask))
1132 return -EINVAL;
1133
1134 cpu_buffer = buffer->buffers[cpu];
1135 work = &cpu_buffer->irq_work;
1136 }
1137
1138 if (full) {
1139 poll_wait(filp, &work->full_waiters, poll_table);
1140 work->full_waiters_pending = true;
1141 } else {
1142 poll_wait(filp, &work->waiters, poll_table);
1143 work->waiters_pending = true;
1144 }
1145
1146 /*
1147 * There's a tight race between setting the waiters_pending and
1148 * checking if the ring buffer is empty. Once the waiters_pending bit
1149 * is set, the next event will wake the task up, but we can get stuck
1150 * if there's only a single event in.
1151 *
1152 * FIXME: Ideally, we need a memory barrier on the writer side as well,
1153 * but adding a memory barrier to all events will cause too much of a
1154 * performance hit in the fast path. We only need a memory barrier when
1155 * the buffer goes from empty to having content. But as this race is
1156 * extremely small, and it's not a problem if another event comes in, we
1157 * will fix it later.
1158 */
1159 smp_mb();
1160
1161 if (full)
1162 return full_hit(buffer, cpu, full) ? EPOLLIN | EPOLLRDNORM : 0;
1163
1164 if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
1165 (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
1166 return EPOLLIN | EPOLLRDNORM;
1167 return 0;
1168 }
1169
1170 /* buffer may be either ring_buffer or ring_buffer_per_cpu */
1171 #define RB_WARN_ON(b, cond) \
1172 ({ \
1173 int _____ret = unlikely(cond); \
1174 if (_____ret) { \
1175 if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
1176 struct ring_buffer_per_cpu *__b = \
1177 (void *)b; \
1178 atomic_inc(&__b->buffer->record_disabled); \
1179 } else \
1180 atomic_inc(&b->record_disabled); \
1181 WARN_ON(1); \
1182 } \
1183 _____ret; \
1184 })
1185
1186 /* Up this if you want to test the TIME_EXTENTS and normalization */
1187 #define DEBUG_SHIFT 0
1188
rb_time_stamp(struct trace_buffer * buffer)1189 static inline u64 rb_time_stamp(struct trace_buffer *buffer)
1190 {
1191 u64 ts;
1192
1193 /* Skip retpolines :-( */
1194 if (IS_ENABLED(CONFIG_RETPOLINE) && likely(buffer->clock == trace_clock_local))
1195 ts = trace_clock_local();
1196 else
1197 ts = buffer->clock();
1198
1199 /* shift to debug/test normalization and TIME_EXTENTS */
1200 return ts << DEBUG_SHIFT;
1201 }
1202
ring_buffer_time_stamp(struct trace_buffer * buffer)1203 u64 ring_buffer_time_stamp(struct trace_buffer *buffer)
1204 {
1205 u64 time;
1206
1207 preempt_disable_notrace();
1208 time = rb_time_stamp(buffer);
1209 preempt_enable_notrace();
1210
1211 return time;
1212 }
1213 EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
1214
ring_buffer_normalize_time_stamp(struct trace_buffer * buffer,int cpu,u64 * ts)1215 void ring_buffer_normalize_time_stamp(struct trace_buffer *buffer,
1216 int cpu, u64 *ts)
1217 {
1218 /* Just stupid testing the normalize function and deltas */
1219 *ts >>= DEBUG_SHIFT;
1220 }
1221 EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
1222
1223 /*
1224 * Making the ring buffer lockless makes things tricky.
1225 * Although writes only happen on the CPU that they are on,
1226 * and they only need to worry about interrupts. Reads can
1227 * happen on any CPU.
1228 *
1229 * The reader page is always off the ring buffer, but when the
1230 * reader finishes with a page, it needs to swap its page with
1231 * a new one from the buffer. The reader needs to take from
1232 * the head (writes go to the tail). But if a writer is in overwrite
1233 * mode and wraps, it must push the head page forward.
1234 *
1235 * Here lies the problem.
1236 *
1237 * The reader must be careful to replace only the head page, and
1238 * not another one. As described at the top of the file in the
1239 * ASCII art, the reader sets its old page to point to the next
1240 * page after head. It then sets the page after head to point to
1241 * the old reader page. But if the writer moves the head page
1242 * during this operation, the reader could end up with the tail.
1243 *
1244 * We use cmpxchg to help prevent this race. We also do something
1245 * special with the page before head. We set the LSB to 1.
1246 *
1247 * When the writer must push the page forward, it will clear the
1248 * bit that points to the head page, move the head, and then set
1249 * the bit that points to the new head page.
1250 *
1251 * We also don't want an interrupt coming in and moving the head
1252 * page on another writer. Thus we use the second LSB to catch
1253 * that too. Thus:
1254 *
1255 * head->list->prev->next bit 1 bit 0
1256 * ------- -------
1257 * Normal page 0 0
1258 * Points to head page 0 1
1259 * New head page 1 0
1260 *
1261 * Note we can not trust the prev pointer of the head page, because:
1262 *
1263 * +----+ +-----+ +-----+
1264 * | |------>| T |---X--->| N |
1265 * | |<------| | | |
1266 * +----+ +-----+ +-----+
1267 * ^ ^ |
1268 * | +-----+ | |
1269 * +----------| R |----------+ |
1270 * | |<-----------+
1271 * +-----+
1272 *
1273 * Key: ---X--> HEAD flag set in pointer
1274 * T Tail page
1275 * R Reader page
1276 * N Next page
1277 *
1278 * (see __rb_reserve_next() to see where this happens)
1279 *
1280 * What the above shows is that the reader just swapped out
1281 * the reader page with a page in the buffer, but before it
1282 * could make the new header point back to the new page added
1283 * it was preempted by a writer. The writer moved forward onto
1284 * the new page added by the reader and is about to move forward
1285 * again.
1286 *
1287 * You can see, it is legitimate for the previous pointer of
1288 * the head (or any page) not to point back to itself. But only
1289 * temporarily.
1290 */
1291
1292 #define RB_PAGE_NORMAL 0UL
1293 #define RB_PAGE_HEAD 1UL
1294 #define RB_PAGE_UPDATE 2UL
1295
1296
1297 #define RB_FLAG_MASK 3UL
1298
1299 /* PAGE_MOVED is not part of the mask */
1300 #define RB_PAGE_MOVED 4UL
1301
1302 /*
1303 * rb_list_head - remove any bit
1304 */
rb_list_head(struct list_head * list)1305 static struct list_head *rb_list_head(struct list_head *list)
1306 {
1307 unsigned long val = (unsigned long)list;
1308
1309 return (struct list_head *)(val & ~RB_FLAG_MASK);
1310 }
1311
1312 /*
1313 * rb_is_head_page - test if the given page is the head page
1314 *
1315 * Because the reader may move the head_page pointer, we can
1316 * not trust what the head page is (it may be pointing to
1317 * the reader page). But if the next page is a header page,
1318 * its flags will be non zero.
1319 */
1320 static inline int
rb_is_head_page(struct buffer_page * page,struct list_head * list)1321 rb_is_head_page(struct buffer_page *page, struct list_head *list)
1322 {
1323 unsigned long val;
1324
1325 val = (unsigned long)list->next;
1326
1327 if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
1328 return RB_PAGE_MOVED;
1329
1330 return val & RB_FLAG_MASK;
1331 }
1332
1333 /*
1334 * rb_is_reader_page
1335 *
1336 * The unique thing about the reader page, is that, if the
1337 * writer is ever on it, the previous pointer never points
1338 * back to the reader page.
1339 */
rb_is_reader_page(struct buffer_page * page)1340 static bool rb_is_reader_page(struct buffer_page *page)
1341 {
1342 struct list_head *list = page->list.prev;
1343
1344 return rb_list_head(list->next) != &page->list;
1345 }
1346
1347 /*
1348 * rb_set_list_to_head - set a list_head to be pointing to head.
1349 */
rb_set_list_to_head(struct list_head * list)1350 static void rb_set_list_to_head(struct list_head *list)
1351 {
1352 unsigned long *ptr;
1353
1354 ptr = (unsigned long *)&list->next;
1355 *ptr |= RB_PAGE_HEAD;
1356 *ptr &= ~RB_PAGE_UPDATE;
1357 }
1358
1359 /*
1360 * rb_head_page_activate - sets up head page
1361 */
rb_head_page_activate(struct ring_buffer_per_cpu * cpu_buffer)1362 static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
1363 {
1364 struct buffer_page *head;
1365
1366 head = cpu_buffer->head_page;
1367 if (!head)
1368 return;
1369
1370 /*
1371 * Set the previous list pointer to have the HEAD flag.
1372 */
1373 rb_set_list_to_head(head->list.prev);
1374 }
1375
rb_list_head_clear(struct list_head * list)1376 static void rb_list_head_clear(struct list_head *list)
1377 {
1378 unsigned long *ptr = (unsigned long *)&list->next;
1379
1380 *ptr &= ~RB_FLAG_MASK;
1381 }
1382
1383 /*
1384 * rb_head_page_deactivate - clears head page ptr (for free list)
1385 */
1386 static void
rb_head_page_deactivate(struct ring_buffer_per_cpu * cpu_buffer)1387 rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
1388 {
1389 struct list_head *hd;
1390
1391 /* Go through the whole list and clear any pointers found. */
1392 rb_list_head_clear(cpu_buffer->pages);
1393
1394 list_for_each(hd, cpu_buffer->pages)
1395 rb_list_head_clear(hd);
1396 }
1397
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)1398 static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
1399 struct buffer_page *head,
1400 struct buffer_page *prev,
1401 int old_flag, int new_flag)
1402 {
1403 struct list_head *list;
1404 unsigned long val = (unsigned long)&head->list;
1405 unsigned long ret;
1406
1407 list = &prev->list;
1408
1409 val &= ~RB_FLAG_MASK;
1410
1411 ret = cmpxchg((unsigned long *)&list->next,
1412 val | old_flag, val | new_flag);
1413
1414 /* check if the reader took the page */
1415 if ((ret & ~RB_FLAG_MASK) != val)
1416 return RB_PAGE_MOVED;
1417
1418 return ret & RB_FLAG_MASK;
1419 }
1420
rb_head_page_set_update(struct ring_buffer_per_cpu * cpu_buffer,struct buffer_page * head,struct buffer_page * prev,int old_flag)1421 static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
1422 struct buffer_page *head,
1423 struct buffer_page *prev,
1424 int old_flag)
1425 {
1426 return rb_head_page_set(cpu_buffer, head, prev,
1427 old_flag, RB_PAGE_UPDATE);
1428 }
1429
rb_head_page_set_head(struct ring_buffer_per_cpu * cpu_buffer,struct buffer_page * head,struct buffer_page * prev,int old_flag)1430 static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
1431 struct buffer_page *head,
1432 struct buffer_page *prev,
1433 int old_flag)
1434 {
1435 return rb_head_page_set(cpu_buffer, head, prev,
1436 old_flag, RB_PAGE_HEAD);
1437 }
1438
rb_head_page_set_normal(struct ring_buffer_per_cpu * cpu_buffer,struct buffer_page * head,struct buffer_page * prev,int old_flag)1439 static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
1440 struct buffer_page *head,
1441 struct buffer_page *prev,
1442 int old_flag)
1443 {
1444 return rb_head_page_set(cpu_buffer, head, prev,
1445 old_flag, RB_PAGE_NORMAL);
1446 }
1447
rb_inc_page(struct buffer_page ** bpage)1448 static inline void rb_inc_page(struct buffer_page **bpage)
1449 {
1450 struct list_head *p = rb_list_head((*bpage)->list.next);
1451
1452 *bpage = list_entry(p, struct buffer_page, list);
1453 }
1454
1455 static struct buffer_page *
rb_set_head_page(struct ring_buffer_per_cpu * cpu_buffer)1456 rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
1457 {
1458 struct buffer_page *head;
1459 struct buffer_page *page;
1460 struct list_head *list;
1461 int i;
1462
1463 if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
1464 return NULL;
1465
1466 /* sanity check */
1467 list = cpu_buffer->pages;
1468 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
1469 return NULL;
1470
1471 page = head = cpu_buffer->head_page;
1472 /*
1473 * It is possible that the writer moves the header behind
1474 * where we started, and we miss in one loop.
1475 * A second loop should grab the header, but we'll do
1476 * three loops just because I'm paranoid.
1477 */
1478 for (i = 0; i < 3; i++) {
1479 do {
1480 if (rb_is_head_page(page, page->list.prev)) {
1481 cpu_buffer->head_page = page;
1482 return page;
1483 }
1484 rb_inc_page(&page);
1485 } while (page != head);
1486 }
1487
1488 RB_WARN_ON(cpu_buffer, 1);
1489
1490 return NULL;
1491 }
1492
rb_head_page_replace(struct buffer_page * old,struct buffer_page * new)1493 static int rb_head_page_replace(struct buffer_page *old,
1494 struct buffer_page *new)
1495 {
1496 unsigned long *ptr = (unsigned long *)&old->list.prev->next;
1497 unsigned long val;
1498 unsigned long ret;
1499
1500 val = *ptr & ~RB_FLAG_MASK;
1501 val |= RB_PAGE_HEAD;
1502
1503 ret = cmpxchg(ptr, val, (unsigned long)&new->list);
1504
1505 return ret == val;
1506 }
1507
1508 /*
1509 * rb_tail_page_update - move the tail page forward
1510 */
rb_tail_page_update(struct ring_buffer_per_cpu * cpu_buffer,struct buffer_page * tail_page,struct buffer_page * next_page)1511 static void rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
1512 struct buffer_page *tail_page,
1513 struct buffer_page *next_page)
1514 {
1515 unsigned long old_entries;
1516 unsigned long old_write;
1517
1518 /*
1519 * The tail page now needs to be moved forward.
1520 *
1521 * We need to reset the tail page, but without messing
1522 * with possible erasing of data brought in by interrupts
1523 * that have moved the tail page and are currently on it.
1524 *
1525 * We add a counter to the write field to denote this.
1526 */
1527 old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
1528 old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
1529
1530 local_inc(&cpu_buffer->pages_touched);
1531 /*
1532 * Just make sure we have seen our old_write and synchronize
1533 * with any interrupts that come in.
1534 */
1535 barrier();
1536
1537 /*
1538 * If the tail page is still the same as what we think
1539 * it is, then it is up to us to update the tail
1540 * pointer.
1541 */
1542 if (tail_page == READ_ONCE(cpu_buffer->tail_page)) {
1543 /* Zero the write counter */
1544 unsigned long val = old_write & ~RB_WRITE_MASK;
1545 unsigned long eval = old_entries & ~RB_WRITE_MASK;
1546
1547 /*
1548 * This will only succeed if an interrupt did
1549 * not come in and change it. In which case, we
1550 * do not want to modify it.
1551 *
1552 * We add (void) to let the compiler know that we do not care
1553 * about the return value of these functions. We use the
1554 * cmpxchg to only update if an interrupt did not already
1555 * do it for us. If the cmpxchg fails, we don't care.
1556 */
1557 (void)local_cmpxchg(&next_page->write, old_write, val);
1558 (void)local_cmpxchg(&next_page->entries, old_entries, eval);
1559
1560 /*
1561 * No need to worry about races with clearing out the commit.
1562 * it only can increment when a commit takes place. But that
1563 * only happens in the outer most nested commit.
1564 */
1565 local_set(&next_page->page->commit, 0);
1566
1567 /* Again, either we update tail_page or an interrupt does */
1568 (void)cmpxchg(&cpu_buffer->tail_page, tail_page, next_page);
1569 }
1570 }
1571
rb_check_bpage(struct ring_buffer_per_cpu * cpu_buffer,struct buffer_page * bpage)1572 static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
1573 struct buffer_page *bpage)
1574 {
1575 unsigned long val = (unsigned long)bpage;
1576
1577 if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK))
1578 return 1;
1579
1580 return 0;
1581 }
1582
1583 /**
1584 * rb_check_list - make sure a pointer to a list has the last bits zero
1585 */
rb_check_list(struct ring_buffer_per_cpu * cpu_buffer,struct list_head * list)1586 static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer,
1587 struct list_head *list)
1588 {
1589 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev))
1590 return 1;
1591 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next))
1592 return 1;
1593 return 0;
1594 }
1595
1596 /**
1597 * rb_check_pages - integrity check of buffer pages
1598 * @cpu_buffer: CPU buffer with pages to test
1599 *
1600 * As a safety measure we check to make sure the data pages have not
1601 * been corrupted.
1602 */
rb_check_pages(struct ring_buffer_per_cpu * cpu_buffer)1603 static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
1604 {
1605 struct list_head *head = cpu_buffer->pages;
1606 struct buffer_page *bpage, *tmp;
1607
1608 /* Reset the head page if it exists */
1609 if (cpu_buffer->head_page)
1610 rb_set_head_page(cpu_buffer);
1611
1612 rb_head_page_deactivate(cpu_buffer);
1613
1614 if (RB_WARN_ON(cpu_buffer, head->next->prev != head))
1615 return -1;
1616 if (RB_WARN_ON(cpu_buffer, head->prev->next != head))
1617 return -1;
1618
1619 if (rb_check_list(cpu_buffer, head))
1620 return -1;
1621
1622 list_for_each_entry_safe(bpage, tmp, head, list) {
1623 if (RB_WARN_ON(cpu_buffer,
1624 bpage->list.next->prev != &bpage->list))
1625 return -1;
1626 if (RB_WARN_ON(cpu_buffer,
1627 bpage->list.prev->next != &bpage->list))
1628 return -1;
1629 if (rb_check_list(cpu_buffer, &bpage->list))
1630 return -1;
1631 }
1632
1633 rb_head_page_activate(cpu_buffer);
1634
1635 return 0;
1636 }
1637
__rb_allocate_pages(struct ring_buffer_per_cpu * cpu_buffer,long nr_pages,struct list_head * pages)1638 static int __rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1639 long nr_pages, struct list_head *pages)
1640 {
1641 struct buffer_page *bpage, *tmp;
1642 bool user_thread = current->mm != NULL;
1643 gfp_t mflags;
1644 long i;
1645
1646 /*
1647 * Check if the available memory is there first.
1648 * Note, si_mem_available() only gives us a rough estimate of available
1649 * memory. It may not be accurate. But we don't care, we just want
1650 * to prevent doing any allocation when it is obvious that it is
1651 * not going to succeed.
1652 */
1653 i = si_mem_available();
1654 if (i < nr_pages)
1655 return -ENOMEM;
1656
1657 /*
1658 * __GFP_RETRY_MAYFAIL flag makes sure that the allocation fails
1659 * gracefully without invoking oom-killer and the system is not
1660 * destabilized.
1661 */
1662 mflags = GFP_KERNEL | __GFP_RETRY_MAYFAIL;
1663
1664 /*
1665 * If a user thread allocates too much, and si_mem_available()
1666 * reports there's enough memory, even though there is not.
1667 * Make sure the OOM killer kills this thread. This can happen
1668 * even with RETRY_MAYFAIL because another task may be doing
1669 * an allocation after this task has taken all memory.
1670 * This is the task the OOM killer needs to take out during this
1671 * loop, even if it was triggered by an allocation somewhere else.
1672 */
1673 if (user_thread)
1674 set_current_oom_origin();
1675 for (i = 0; i < nr_pages; i++) {
1676 struct page *page;
1677
1678 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1679 mflags, cpu_to_node(cpu_buffer->cpu));
1680 if (!bpage)
1681 goto free_pages;
1682
1683 rb_check_bpage(cpu_buffer, bpage);
1684
1685 list_add(&bpage->list, pages);
1686
1687 page = alloc_pages_node(cpu_to_node(cpu_buffer->cpu), mflags, 0);
1688 if (!page)
1689 goto free_pages;
1690 bpage->page = page_address(page);
1691 rb_init_page(bpage->page);
1692
1693 if (user_thread && fatal_signal_pending(current))
1694 goto free_pages;
1695 }
1696 if (user_thread)
1697 clear_current_oom_origin();
1698
1699 return 0;
1700
1701 free_pages:
1702 list_for_each_entry_safe(bpage, tmp, pages, list) {
1703 list_del_init(&bpage->list);
1704 free_buffer_page(bpage);
1705 }
1706 if (user_thread)
1707 clear_current_oom_origin();
1708
1709 return -ENOMEM;
1710 }
1711
rb_allocate_pages(struct ring_buffer_per_cpu * cpu_buffer,unsigned long nr_pages)1712 static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1713 unsigned long nr_pages)
1714 {
1715 LIST_HEAD(pages);
1716
1717 WARN_ON(!nr_pages);
1718
1719 if (__rb_allocate_pages(cpu_buffer, nr_pages, &pages))
1720 return -ENOMEM;
1721
1722 /*
1723 * The ring buffer page list is a circular list that does not
1724 * start and end with a list head. All page list items point to
1725 * other pages.
1726 */
1727 cpu_buffer->pages = pages.next;
1728 list_del(&pages);
1729
1730 cpu_buffer->nr_pages = nr_pages;
1731
1732 rb_check_pages(cpu_buffer);
1733
1734 return 0;
1735 }
1736
1737 static struct ring_buffer_per_cpu *
rb_allocate_cpu_buffer(struct trace_buffer * buffer,long nr_pages,int cpu)1738 rb_allocate_cpu_buffer(struct trace_buffer *buffer, long nr_pages, int cpu)
1739 {
1740 struct ring_buffer_per_cpu *cpu_buffer;
1741 struct buffer_page *bpage;
1742 struct page *page;
1743 int ret;
1744
1745 cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1746 GFP_KERNEL, cpu_to_node(cpu));
1747 if (!cpu_buffer)
1748 return NULL;
1749
1750 cpu_buffer->cpu = cpu;
1751 cpu_buffer->buffer = buffer;
1752 raw_spin_lock_init(&cpu_buffer->reader_lock);
1753 lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1754 cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1755 INIT_WORK(&cpu_buffer->update_pages_work, update_pages_handler);
1756 init_completion(&cpu_buffer->update_done);
1757 init_irq_work(&cpu_buffer->irq_work.work, rb_wake_up_waiters);
1758 init_waitqueue_head(&cpu_buffer->irq_work.waiters);
1759 init_waitqueue_head(&cpu_buffer->irq_work.full_waiters);
1760
1761 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1762 GFP_KERNEL, cpu_to_node(cpu));
1763 if (!bpage)
1764 goto fail_free_buffer;
1765
1766 rb_check_bpage(cpu_buffer, bpage);
1767
1768 cpu_buffer->reader_page = bpage;
1769 page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0);
1770 if (!page)
1771 goto fail_free_reader;
1772 bpage->page = page_address(page);
1773 rb_init_page(bpage->page);
1774
1775 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1776 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1777
1778 ret = rb_allocate_pages(cpu_buffer, nr_pages);
1779 if (ret < 0)
1780 goto fail_free_reader;
1781
1782 cpu_buffer->head_page
1783 = list_entry(cpu_buffer->pages, struct buffer_page, list);
1784 cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1785
1786 rb_head_page_activate(cpu_buffer);
1787
1788 return cpu_buffer;
1789
1790 fail_free_reader:
1791 free_buffer_page(cpu_buffer->reader_page);
1792
1793 fail_free_buffer:
1794 kfree(cpu_buffer);
1795 return NULL;
1796 }
1797
rb_free_cpu_buffer(struct ring_buffer_per_cpu * cpu_buffer)1798 static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1799 {
1800 struct list_head *head = cpu_buffer->pages;
1801 struct buffer_page *bpage, *tmp;
1802
1803 free_buffer_page(cpu_buffer->reader_page);
1804
1805 if (head) {
1806 rb_head_page_deactivate(cpu_buffer);
1807
1808 list_for_each_entry_safe(bpage, tmp, head, list) {
1809 list_del_init(&bpage->list);
1810 free_buffer_page(bpage);
1811 }
1812 bpage = list_entry(head, struct buffer_page, list);
1813 free_buffer_page(bpage);
1814 }
1815
1816 kfree(cpu_buffer);
1817 }
1818
1819 /**
1820 * __ring_buffer_alloc - allocate a new ring_buffer
1821 * @size: the size in bytes per cpu that is needed.
1822 * @flags: attributes to set for the ring buffer.
1823 * @key: ring buffer reader_lock_key.
1824 *
1825 * Currently the only flag that is available is the RB_FL_OVERWRITE
1826 * flag. This flag means that the buffer will overwrite old data
1827 * when the buffer wraps. If this flag is not set, the buffer will
1828 * drop data when the tail hits the head.
1829 */
__ring_buffer_alloc(unsigned long size,unsigned flags,struct lock_class_key * key)1830 struct trace_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1831 struct lock_class_key *key)
1832 {
1833 struct trace_buffer *buffer;
1834 long nr_pages;
1835 int bsize;
1836 int cpu;
1837 int ret;
1838
1839 /* keep it in its own cache line */
1840 buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1841 GFP_KERNEL);
1842 if (!buffer)
1843 return NULL;
1844
1845 if (!zalloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1846 goto fail_free_buffer;
1847
1848 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1849 buffer->flags = flags;
1850 buffer->clock = trace_clock_local;
1851 buffer->reader_lock_key = key;
1852
1853 init_irq_work(&buffer->irq_work.work, rb_wake_up_waiters);
1854 init_waitqueue_head(&buffer->irq_work.waiters);
1855
1856 /* need at least two pages */
1857 if (nr_pages < 2)
1858 nr_pages = 2;
1859
1860 buffer->cpus = nr_cpu_ids;
1861
1862 bsize = sizeof(void *) * nr_cpu_ids;
1863 buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1864 GFP_KERNEL);
1865 if (!buffer->buffers)
1866 goto fail_free_cpumask;
1867
1868 cpu = raw_smp_processor_id();
1869 cpumask_set_cpu(cpu, buffer->cpumask);
1870 buffer->buffers[cpu] = rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
1871 if (!buffer->buffers[cpu])
1872 goto fail_free_buffers;
1873
1874 ret = cpuhp_state_add_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1875 if (ret < 0)
1876 goto fail_free_buffers;
1877
1878 mutex_init(&buffer->mutex);
1879
1880 return buffer;
1881
1882 fail_free_buffers:
1883 for_each_buffer_cpu(buffer, cpu) {
1884 if (buffer->buffers[cpu])
1885 rb_free_cpu_buffer(buffer->buffers[cpu]);
1886 }
1887 kfree(buffer->buffers);
1888
1889 fail_free_cpumask:
1890 free_cpumask_var(buffer->cpumask);
1891
1892 fail_free_buffer:
1893 kfree(buffer);
1894 return NULL;
1895 }
1896 EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1897
1898 /**
1899 * ring_buffer_free - free a ring buffer.
1900 * @buffer: the buffer to free.
1901 */
1902 void
ring_buffer_free(struct trace_buffer * buffer)1903 ring_buffer_free(struct trace_buffer *buffer)
1904 {
1905 int cpu;
1906
1907 cpuhp_state_remove_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1908
1909 for_each_buffer_cpu(buffer, cpu)
1910 rb_free_cpu_buffer(buffer->buffers[cpu]);
1911
1912 kfree(buffer->buffers);
1913 free_cpumask_var(buffer->cpumask);
1914
1915 kfree(buffer);
1916 }
1917 EXPORT_SYMBOL_GPL(ring_buffer_free);
1918
ring_buffer_set_clock(struct trace_buffer * buffer,u64 (* clock)(void))1919 void ring_buffer_set_clock(struct trace_buffer *buffer,
1920 u64 (*clock)(void))
1921 {
1922 buffer->clock = clock;
1923 }
1924
ring_buffer_set_time_stamp_abs(struct trace_buffer * buffer,bool abs)1925 void ring_buffer_set_time_stamp_abs(struct trace_buffer *buffer, bool abs)
1926 {
1927 buffer->time_stamp_abs = abs;
1928 }
1929
ring_buffer_time_stamp_abs(struct trace_buffer * buffer)1930 bool ring_buffer_time_stamp_abs(struct trace_buffer *buffer)
1931 {
1932 return buffer->time_stamp_abs;
1933 }
1934
1935 static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1936
rb_page_entries(struct buffer_page * bpage)1937 static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1938 {
1939 return local_read(&bpage->entries) & RB_WRITE_MASK;
1940 }
1941
rb_page_write(struct buffer_page * bpage)1942 static inline unsigned long rb_page_write(struct buffer_page *bpage)
1943 {
1944 return local_read(&bpage->write) & RB_WRITE_MASK;
1945 }
1946
1947 static int
rb_remove_pages(struct ring_buffer_per_cpu * cpu_buffer,unsigned long nr_pages)1948 rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned long nr_pages)
1949 {
1950 struct list_head *tail_page, *to_remove, *next_page;
1951 struct buffer_page *to_remove_page, *tmp_iter_page;
1952 struct buffer_page *last_page, *first_page;
1953 unsigned long nr_removed;
1954 unsigned long head_bit;
1955 int page_entries;
1956
1957 head_bit = 0;
1958
1959 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1960 atomic_inc(&cpu_buffer->record_disabled);
1961 /*
1962 * We don't race with the readers since we have acquired the reader
1963 * lock. We also don't race with writers after disabling recording.
1964 * This makes it easy to figure out the first and the last page to be
1965 * removed from the list. We unlink all the pages in between including
1966 * the first and last pages. This is done in a busy loop so that we
1967 * lose the least number of traces.
1968 * The pages are freed after we restart recording and unlock readers.
1969 */
1970 tail_page = &cpu_buffer->tail_page->list;
1971
1972 /*
1973 * tail page might be on reader page, we remove the next page
1974 * from the ring buffer
1975 */
1976 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
1977 tail_page = rb_list_head(tail_page->next);
1978 to_remove = tail_page;
1979
1980 /* start of pages to remove */
1981 first_page = list_entry(rb_list_head(to_remove->next),
1982 struct buffer_page, list);
1983
1984 for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
1985 to_remove = rb_list_head(to_remove)->next;
1986 head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
1987 }
1988
1989 next_page = rb_list_head(to_remove)->next;
1990
1991 /*
1992 * Now we remove all pages between tail_page and next_page.
1993 * Make sure that we have head_bit value preserved for the
1994 * next page
1995 */
1996 tail_page->next = (struct list_head *)((unsigned long)next_page |
1997 head_bit);
1998 next_page = rb_list_head(next_page);
1999 next_page->prev = tail_page;
2000
2001 /* make sure pages points to a valid page in the ring buffer */
2002 cpu_buffer->pages = next_page;
2003
2004 /* update head page */
2005 if (head_bit)
2006 cpu_buffer->head_page = list_entry(next_page,
2007 struct buffer_page, list);
2008
2009 /*
2010 * change read pointer to make sure any read iterators reset
2011 * themselves
2012 */
2013 cpu_buffer->read = 0;
2014
2015 /* pages are removed, resume tracing and then free the pages */
2016 atomic_dec(&cpu_buffer->record_disabled);
2017 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
2018
2019 RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
2020
2021 /* last buffer page to remove */
2022 last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
2023 list);
2024 tmp_iter_page = first_page;
2025
2026 do {
2027 cond_resched();
2028
2029 to_remove_page = tmp_iter_page;
2030 rb_inc_page(&tmp_iter_page);
2031
2032 /* update the counters */
2033 page_entries = rb_page_entries(to_remove_page);
2034 if (page_entries) {
2035 /*
2036 * If something was added to this page, it was full
2037 * since it is not the tail page. So we deduct the
2038 * bytes consumed in ring buffer from here.
2039 * Increment overrun to account for the lost events.
2040 */
2041 local_add(page_entries, &cpu_buffer->overrun);
2042 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
2043 local_inc(&cpu_buffer->pages_lost);
2044 }
2045
2046 /*
2047 * We have already removed references to this list item, just
2048 * free up the buffer_page and its page
2049 */
2050 free_buffer_page(to_remove_page);
2051 nr_removed--;
2052
2053 } while (to_remove_page != last_page);
2054
2055 RB_WARN_ON(cpu_buffer, nr_removed);
2056
2057 return nr_removed == 0;
2058 }
2059
2060 static int
rb_insert_pages(struct ring_buffer_per_cpu * cpu_buffer)2061 rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
2062 {
2063 struct list_head *pages = &cpu_buffer->new_pages;
2064 int retries, success;
2065
2066 raw_spin_lock_irq(&cpu_buffer->reader_lock);
2067 /*
2068 * We are holding the reader lock, so the reader page won't be swapped
2069 * in the ring buffer. Now we are racing with the writer trying to
2070 * move head page and the tail page.
2071 * We are going to adapt the reader page update process where:
2072 * 1. We first splice the start and end of list of new pages between
2073 * the head page and its previous page.
2074 * 2. We cmpxchg the prev_page->next to point from head page to the
2075 * start of new pages list.
2076 * 3. Finally, we update the head->prev to the end of new list.
2077 *
2078 * We will try this process 10 times, to make sure that we don't keep
2079 * spinning.
2080 */
2081 retries = 10;
2082 success = 0;
2083 while (retries--) {
2084 struct list_head *head_page, *prev_page, *r;
2085 struct list_head *last_page, *first_page;
2086 struct list_head *head_page_with_bit;
2087
2088 head_page = &rb_set_head_page(cpu_buffer)->list;
2089 if (!head_page)
2090 break;
2091 prev_page = head_page->prev;
2092
2093 first_page = pages->next;
2094 last_page = pages->prev;
2095
2096 head_page_with_bit = (struct list_head *)
2097 ((unsigned long)head_page | RB_PAGE_HEAD);
2098
2099 last_page->next = head_page_with_bit;
2100 first_page->prev = prev_page;
2101
2102 r = cmpxchg(&prev_page->next, head_page_with_bit, first_page);
2103
2104 if (r == head_page_with_bit) {
2105 /*
2106 * yay, we replaced the page pointer to our new list,
2107 * now, we just have to update to head page's prev
2108 * pointer to point to end of list
2109 */
2110 head_page->prev = last_page;
2111 success = 1;
2112 break;
2113 }
2114 }
2115
2116 if (success)
2117 INIT_LIST_HEAD(pages);
2118 /*
2119 * If we weren't successful in adding in new pages, warn and stop
2120 * tracing
2121 */
2122 RB_WARN_ON(cpu_buffer, !success);
2123 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
2124
2125 /* free pages if they weren't inserted */
2126 if (!success) {
2127 struct buffer_page *bpage, *tmp;
2128 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
2129 list) {
2130 list_del_init(&bpage->list);
2131 free_buffer_page(bpage);
2132 }
2133 }
2134 return success;
2135 }
2136
rb_update_pages(struct ring_buffer_per_cpu * cpu_buffer)2137 static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
2138 {
2139 int success;
2140
2141 if (cpu_buffer->nr_pages_to_update > 0)
2142 success = rb_insert_pages(cpu_buffer);
2143 else
2144 success = rb_remove_pages(cpu_buffer,
2145 -cpu_buffer->nr_pages_to_update);
2146
2147 if (success)
2148 cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
2149 }
2150
update_pages_handler(struct work_struct * work)2151 static void update_pages_handler(struct work_struct *work)
2152 {
2153 struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
2154 struct ring_buffer_per_cpu, update_pages_work);
2155 rb_update_pages(cpu_buffer);
2156 complete(&cpu_buffer->update_done);
2157 }
2158
2159 /**
2160 * ring_buffer_resize - resize the ring buffer
2161 * @buffer: the buffer to resize.
2162 * @size: the new size.
2163 * @cpu_id: the cpu buffer to resize
2164 *
2165 * Minimum size is 2 * BUF_PAGE_SIZE.
2166 *
2167 * Returns 0 on success and < 0 on failure.
2168 */
ring_buffer_resize(struct trace_buffer * buffer,unsigned long size,int cpu_id)2169 int ring_buffer_resize(struct trace_buffer *buffer, unsigned long size,
2170 int cpu_id)
2171 {
2172 struct ring_buffer_per_cpu *cpu_buffer;
2173 unsigned long nr_pages;
2174 int cpu, err;
2175
2176 /*
2177 * Always succeed at resizing a non-existent buffer:
2178 */
2179 if (!buffer)
2180 return 0;
2181
2182 /* Make sure the requested buffer exists */
2183 if (cpu_id != RING_BUFFER_ALL_CPUS &&
2184 !cpumask_test_cpu(cpu_id, buffer->cpumask))
2185 return 0;
2186
2187 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
2188
2189 /* we need a minimum of two pages */
2190 if (nr_pages < 2)
2191 nr_pages = 2;
2192
2193 /* prevent another thread from changing buffer sizes */
2194 mutex_lock(&buffer->mutex);
2195
2196
2197 if (cpu_id == RING_BUFFER_ALL_CPUS) {
2198 /*
2199 * Don't succeed if resizing is disabled, as a reader might be
2200 * manipulating the ring buffer and is expecting a sane state while
2201 * this is true.
2202 */
2203 for_each_buffer_cpu(buffer, cpu) {
2204 cpu_buffer = buffer->buffers[cpu];
2205 if (atomic_read(&cpu_buffer->resize_disabled)) {
2206 err = -EBUSY;
2207 goto out_err_unlock;
2208 }
2209 }
2210
2211 /* calculate the pages to update */
2212 for_each_buffer_cpu(buffer, cpu) {
2213 cpu_buffer = buffer->buffers[cpu];
2214
2215 cpu_buffer->nr_pages_to_update = nr_pages -
2216 cpu_buffer->nr_pages;
2217 /*
2218 * nothing more to do for removing pages or no update
2219 */
2220 if (cpu_buffer->nr_pages_to_update <= 0)
2221 continue;
2222 /*
2223 * to add pages, make sure all new pages can be
2224 * allocated without receiving ENOMEM
2225 */
2226 INIT_LIST_HEAD(&cpu_buffer->new_pages);
2227 if (__rb_allocate_pages(cpu_buffer, cpu_buffer->nr_pages_to_update,
2228 &cpu_buffer->new_pages)) {
2229 /* not enough memory for new pages */
2230 err = -ENOMEM;
2231 goto out_err;
2232 }
2233 }
2234
2235 cpus_read_lock();
2236 /*
2237 * Fire off all the required work handlers
2238 * We can't schedule on offline CPUs, but it's not necessary
2239 * since we can change their buffer sizes without any race.
2240 */
2241 for_each_buffer_cpu(buffer, cpu) {
2242 cpu_buffer = buffer->buffers[cpu];
2243 if (!cpu_buffer->nr_pages_to_update)
2244 continue;
2245
2246 /* Can't run something on an offline CPU. */
2247 if (!cpu_online(cpu)) {
2248 rb_update_pages(cpu_buffer);
2249 cpu_buffer->nr_pages_to_update = 0;
2250 } else {
2251 schedule_work_on(cpu,
2252 &cpu_buffer->update_pages_work);
2253 }
2254 }
2255
2256 /* wait for all the updates to complete */
2257 for_each_buffer_cpu(buffer, cpu) {
2258 cpu_buffer = buffer->buffers[cpu];
2259 if (!cpu_buffer->nr_pages_to_update)
2260 continue;
2261
2262 if (cpu_online(cpu))
2263 wait_for_completion(&cpu_buffer->update_done);
2264 cpu_buffer->nr_pages_to_update = 0;
2265 }
2266
2267 cpus_read_unlock();
2268 } else {
2269 cpu_buffer = buffer->buffers[cpu_id];
2270
2271 if (nr_pages == cpu_buffer->nr_pages)
2272 goto out;
2273
2274 /*
2275 * Don't succeed if resizing is disabled, as a reader might be
2276 * manipulating the ring buffer and is expecting a sane state while
2277 * this is true.
2278 */
2279 if (atomic_read(&cpu_buffer->resize_disabled)) {
2280 err = -EBUSY;
2281 goto out_err_unlock;
2282 }
2283
2284 cpu_buffer->nr_pages_to_update = nr_pages -
2285 cpu_buffer->nr_pages;
2286
2287 INIT_LIST_HEAD(&cpu_buffer->new_pages);
2288 if (cpu_buffer->nr_pages_to_update > 0 &&
2289 __rb_allocate_pages(cpu_buffer, cpu_buffer->nr_pages_to_update,
2290 &cpu_buffer->new_pages)) {
2291 err = -ENOMEM;
2292 goto out_err;
2293 }
2294
2295 cpus_read_lock();
2296
2297 /* Can't run something on an offline CPU. */
2298 if (!cpu_online(cpu_id))
2299 rb_update_pages(cpu_buffer);
2300 else {
2301 schedule_work_on(cpu_id,
2302 &cpu_buffer->update_pages_work);
2303 wait_for_completion(&cpu_buffer->update_done);
2304 }
2305
2306 cpu_buffer->nr_pages_to_update = 0;
2307 cpus_read_unlock();
2308 }
2309
2310 out:
2311 /*
2312 * The ring buffer resize can happen with the ring buffer
2313 * enabled, so that the update disturbs the tracing as little
2314 * as possible. But if the buffer is disabled, we do not need
2315 * to worry about that, and we can take the time to verify
2316 * that the buffer is not corrupt.
2317 */
2318 if (atomic_read(&buffer->record_disabled)) {
2319 atomic_inc(&buffer->record_disabled);
2320 /*
2321 * Even though the buffer was disabled, we must make sure
2322 * that it is truly disabled before calling rb_check_pages.
2323 * There could have been a race between checking
2324 * record_disable and incrementing it.
2325 */
2326 synchronize_rcu();
2327 for_each_buffer_cpu(buffer, cpu) {
2328 cpu_buffer = buffer->buffers[cpu];
2329 rb_check_pages(cpu_buffer);
2330 }
2331 atomic_dec(&buffer->record_disabled);
2332 }
2333
2334 mutex_unlock(&buffer->mutex);
2335 return 0;
2336
2337 out_err:
2338 for_each_buffer_cpu(buffer, cpu) {
2339 struct buffer_page *bpage, *tmp;
2340
2341 cpu_buffer = buffer->buffers[cpu];
2342 cpu_buffer->nr_pages_to_update = 0;
2343
2344 if (list_empty(&cpu_buffer->new_pages))
2345 continue;
2346
2347 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
2348 list) {
2349 list_del_init(&bpage->list);
2350 free_buffer_page(bpage);
2351 }
2352 }
2353 out_err_unlock:
2354 mutex_unlock(&buffer->mutex);
2355 return err;
2356 }
2357 EXPORT_SYMBOL_GPL(ring_buffer_resize);
2358
ring_buffer_change_overwrite(struct trace_buffer * buffer,int val)2359 void ring_buffer_change_overwrite(struct trace_buffer *buffer, int val)
2360 {
2361 mutex_lock(&buffer->mutex);
2362 if (val)
2363 buffer->flags |= RB_FL_OVERWRITE;
2364 else
2365 buffer->flags &= ~RB_FL_OVERWRITE;
2366 mutex_unlock(&buffer->mutex);
2367 }
2368 EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
2369
__rb_page_index(struct buffer_page * bpage,unsigned index)2370 static __always_inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
2371 {
2372 return bpage->page->data + index;
2373 }
2374
2375 static __always_inline struct ring_buffer_event *
rb_reader_event(struct ring_buffer_per_cpu * cpu_buffer)2376 rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
2377 {
2378 return __rb_page_index(cpu_buffer->reader_page,
2379 cpu_buffer->reader_page->read);
2380 }
2381
rb_page_commit(struct buffer_page * bpage)2382 static __always_inline unsigned rb_page_commit(struct buffer_page *bpage)
2383 {
2384 return local_read(&bpage->page->commit);
2385 }
2386
2387 static struct ring_buffer_event *
rb_iter_head_event(struct ring_buffer_iter * iter)2388 rb_iter_head_event(struct ring_buffer_iter *iter)
2389 {
2390 struct ring_buffer_event *event;
2391 struct buffer_page *iter_head_page = iter->head_page;
2392 unsigned long commit;
2393 unsigned length;
2394
2395 if (iter->head != iter->next_event)
2396 return iter->event;
2397
2398 /*
2399 * When the writer goes across pages, it issues a cmpxchg which
2400 * is a mb(), which will synchronize with the rmb here.
2401 * (see rb_tail_page_update() and __rb_reserve_next())
2402 */
2403 commit = rb_page_commit(iter_head_page);
2404 smp_rmb();
2405 event = __rb_page_index(iter_head_page, iter->head);
2406 length = rb_event_length(event);
2407
2408 /*
2409 * READ_ONCE() doesn't work on functions and we don't want the
2410 * compiler doing any crazy optimizations with length.
2411 */
2412 barrier();
2413
2414 if ((iter->head + length) > commit || length > BUF_MAX_DATA_SIZE)
2415 /* Writer corrupted the read? */
2416 goto reset;
2417
2418 memcpy(iter->event, event, length);
2419 /*
2420 * If the page stamp is still the same after this rmb() then the
2421 * event was safely copied without the writer entering the page.
2422 */
2423 smp_rmb();
2424
2425 /* Make sure the page didn't change since we read this */
2426 if (iter->page_stamp != iter_head_page->page->time_stamp ||
2427 commit > rb_page_commit(iter_head_page))
2428 goto reset;
2429
2430 iter->next_event = iter->head + length;
2431 return iter->event;
2432 reset:
2433 /* Reset to the beginning */
2434 iter->page_stamp = iter->read_stamp = iter->head_page->page->time_stamp;
2435 iter->head = 0;
2436 iter->next_event = 0;
2437 iter->missed_events = 1;
2438 return NULL;
2439 }
2440
2441 /* Size is determined by what has been committed */
rb_page_size(struct buffer_page * bpage)2442 static __always_inline unsigned rb_page_size(struct buffer_page *bpage)
2443 {
2444 return rb_page_commit(bpage);
2445 }
2446
2447 static __always_inline unsigned
rb_commit_index(struct ring_buffer_per_cpu * cpu_buffer)2448 rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
2449 {
2450 return rb_page_commit(cpu_buffer->commit_page);
2451 }
2452
2453 static __always_inline unsigned
rb_event_index(struct ring_buffer_event * event)2454 rb_event_index(struct ring_buffer_event *event)
2455 {
2456 unsigned long addr = (unsigned long)event;
2457
2458 return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
2459 }
2460
rb_inc_iter(struct ring_buffer_iter * iter)2461 static void rb_inc_iter(struct ring_buffer_iter *iter)
2462 {
2463 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
2464
2465 /*
2466 * The iterator could be on the reader page (it starts there).
2467 * But the head could have moved, since the reader was
2468 * found. Check for this case and assign the iterator
2469 * to the head page instead of next.
2470 */
2471 if (iter->head_page == cpu_buffer->reader_page)
2472 iter->head_page = rb_set_head_page(cpu_buffer);
2473 else
2474 rb_inc_page(&iter->head_page);
2475
2476 iter->page_stamp = iter->read_stamp = iter->head_page->page->time_stamp;
2477 iter->head = 0;
2478 iter->next_event = 0;
2479 }
2480
2481 /*
2482 * rb_handle_head_page - writer hit the head page
2483 *
2484 * Returns: +1 to retry page
2485 * 0 to continue
2486 * -1 on error
2487 */
2488 static int
rb_handle_head_page(struct ring_buffer_per_cpu * cpu_buffer,struct buffer_page * tail_page,struct buffer_page * next_page)2489 rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
2490 struct buffer_page *tail_page,
2491 struct buffer_page *next_page)
2492 {
2493 struct buffer_page *new_head;
2494 int entries;
2495 int type;
2496 int ret;
2497
2498 entries = rb_page_entries(next_page);
2499
2500 /*
2501 * The hard part is here. We need to move the head
2502 * forward, and protect against both readers on
2503 * other CPUs and writers coming in via interrupts.
2504 */
2505 type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
2506 RB_PAGE_HEAD);
2507
2508 /*
2509 * type can be one of four:
2510 * NORMAL - an interrupt already moved it for us
2511 * HEAD - we are the first to get here.
2512 * UPDATE - we are the interrupt interrupting
2513 * a current move.
2514 * MOVED - a reader on another CPU moved the next
2515 * pointer to its reader page. Give up
2516 * and try again.
2517 */
2518
2519 switch (type) {
2520 case RB_PAGE_HEAD:
2521 /*
2522 * We changed the head to UPDATE, thus
2523 * it is our responsibility to update
2524 * the counters.
2525 */
2526 local_add(entries, &cpu_buffer->overrun);
2527 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
2528 local_inc(&cpu_buffer->pages_lost);
2529
2530 /*
2531 * The entries will be zeroed out when we move the
2532 * tail page.
2533 */
2534
2535 /* still more to do */
2536 break;
2537
2538 case RB_PAGE_UPDATE:
2539 /*
2540 * This is an interrupt that interrupt the
2541 * previous update. Still more to do.
2542 */
2543 break;
2544 case RB_PAGE_NORMAL:
2545 /*
2546 * An interrupt came in before the update
2547 * and processed this for us.
2548 * Nothing left to do.
2549 */
2550 return 1;
2551 case RB_PAGE_MOVED:
2552 /*
2553 * The reader is on another CPU and just did
2554 * a swap with our next_page.
2555 * Try again.
2556 */
2557 return 1;
2558 default:
2559 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
2560 return -1;
2561 }
2562
2563 /*
2564 * Now that we are here, the old head pointer is
2565 * set to UPDATE. This will keep the reader from
2566 * swapping the head page with the reader page.
2567 * The reader (on another CPU) will spin till
2568 * we are finished.
2569 *
2570 * We just need to protect against interrupts
2571 * doing the job. We will set the next pointer
2572 * to HEAD. After that, we set the old pointer
2573 * to NORMAL, but only if it was HEAD before.
2574 * otherwise we are an interrupt, and only
2575 * want the outer most commit to reset it.
2576 */
2577 new_head = next_page;
2578 rb_inc_page(&new_head);
2579
2580 ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
2581 RB_PAGE_NORMAL);
2582
2583 /*
2584 * Valid returns are:
2585 * HEAD - an interrupt came in and already set it.
2586 * NORMAL - One of two things:
2587 * 1) We really set it.
2588 * 2) A bunch of interrupts came in and moved
2589 * the page forward again.
2590 */
2591 switch (ret) {
2592 case RB_PAGE_HEAD:
2593 case RB_PAGE_NORMAL:
2594 /* OK */
2595 break;
2596 default:
2597 RB_WARN_ON(cpu_buffer, 1);
2598 return -1;
2599 }
2600
2601 /*
2602 * It is possible that an interrupt came in,
2603 * set the head up, then more interrupts came in
2604 * and moved it again. When we get back here,
2605 * the page would have been set to NORMAL but we
2606 * just set it back to HEAD.
2607 *
2608 * How do you detect this? Well, if that happened
2609 * the tail page would have moved.
2610 */
2611 if (ret == RB_PAGE_NORMAL) {
2612 struct buffer_page *buffer_tail_page;
2613
2614 buffer_tail_page = READ_ONCE(cpu_buffer->tail_page);
2615 /*
2616 * If the tail had moved passed next, then we need
2617 * to reset the pointer.
2618 */
2619 if (buffer_tail_page != tail_page &&
2620 buffer_tail_page != next_page)
2621 rb_head_page_set_normal(cpu_buffer, new_head,
2622 next_page,
2623 RB_PAGE_HEAD);
2624 }
2625
2626 /*
2627 * If this was the outer most commit (the one that
2628 * changed the original pointer from HEAD to UPDATE),
2629 * then it is up to us to reset it to NORMAL.
2630 */
2631 if (type == RB_PAGE_HEAD) {
2632 ret = rb_head_page_set_normal(cpu_buffer, next_page,
2633 tail_page,
2634 RB_PAGE_UPDATE);
2635 if (RB_WARN_ON(cpu_buffer,
2636 ret != RB_PAGE_UPDATE))
2637 return -1;
2638 }
2639
2640 return 0;
2641 }
2642
2643 static inline void
rb_reset_tail(struct ring_buffer_per_cpu * cpu_buffer,unsigned long tail,struct rb_event_info * info)2644 rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2645 unsigned long tail, struct rb_event_info *info)
2646 {
2647 struct buffer_page *tail_page = info->tail_page;
2648 struct ring_buffer_event *event;
2649 unsigned long length = info->length;
2650
2651 /*
2652 * Only the event that crossed the page boundary
2653 * must fill the old tail_page with padding.
2654 */
2655 if (tail >= BUF_PAGE_SIZE) {
2656 /*
2657 * If the page was filled, then we still need
2658 * to update the real_end. Reset it to zero
2659 * and the reader will ignore it.
2660 */
2661 if (tail == BUF_PAGE_SIZE)
2662 tail_page->real_end = 0;
2663
2664 local_sub(length, &tail_page->write);
2665 return;
2666 }
2667
2668 event = __rb_page_index(tail_page, tail);
2669
2670 /* account for padding bytes */
2671 local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes);
2672
2673 /*
2674 * Save the original length to the meta data.
2675 * This will be used by the reader to add lost event
2676 * counter.
2677 */
2678 tail_page->real_end = tail;
2679
2680 /*
2681 * If this event is bigger than the minimum size, then
2682 * we need to be careful that we don't subtract the
2683 * write counter enough to allow another writer to slip
2684 * in on this page.
2685 * We put in a discarded commit instead, to make sure
2686 * that this space is not used again.
2687 *
2688 * If we are less than the minimum size, we don't need to
2689 * worry about it.
2690 */
2691 if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
2692 /* No room for any events */
2693
2694 /* Mark the rest of the page with padding */
2695 rb_event_set_padding(event);
2696
2697 /* Make sure the padding is visible before the write update */
2698 smp_wmb();
2699
2700 /* Set the write back to the previous setting */
2701 local_sub(length, &tail_page->write);
2702 return;
2703 }
2704
2705 /* Put in a discarded event */
2706 event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
2707 event->type_len = RINGBUF_TYPE_PADDING;
2708 /* time delta must be non zero */
2709 event->time_delta = 1;
2710
2711 /* Make sure the padding is visible before the tail_page->write update */
2712 smp_wmb();
2713
2714 /* Set write to end of buffer */
2715 length = (tail + length) - BUF_PAGE_SIZE;
2716 local_sub(length, &tail_page->write);
2717 }
2718
2719 static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer);
2720
2721 /*
2722 * This is the slow path, force gcc not to inline it.
2723 */
2724 static noinline struct ring_buffer_event *
rb_move_tail(struct ring_buffer_per_cpu * cpu_buffer,unsigned long tail,struct rb_event_info * info)2725 rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2726 unsigned long tail, struct rb_event_info *info)
2727 {
2728 struct buffer_page *tail_page = info->tail_page;
2729 struct buffer_page *commit_page = cpu_buffer->commit_page;
2730 struct trace_buffer *buffer = cpu_buffer->buffer;
2731 struct buffer_page *next_page;
2732 int ret;
2733
2734 next_page = tail_page;
2735
2736 rb_inc_page(&next_page);
2737
2738 /*
2739 * If for some reason, we had an interrupt storm that made
2740 * it all the way around the buffer, bail, and warn
2741 * about it.
2742 */
2743 if (unlikely(next_page == commit_page)) {
2744 local_inc(&cpu_buffer->commit_overrun);
2745 goto out_reset;
2746 }
2747
2748 /*
2749 * This is where the fun begins!
2750 *
2751 * We are fighting against races between a reader that
2752 * could be on another CPU trying to swap its reader
2753 * page with the buffer head.
2754 *
2755 * We are also fighting against interrupts coming in and
2756 * moving the head or tail on us as well.
2757 *
2758 * If the next page is the head page then we have filled
2759 * the buffer, unless the commit page is still on the
2760 * reader page.
2761 */
2762 if (rb_is_head_page(next_page, &tail_page->list)) {
2763
2764 /*
2765 * If the commit is not on the reader page, then
2766 * move the header page.
2767 */
2768 if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2769 /*
2770 * If we are not in overwrite mode,
2771 * this is easy, just stop here.
2772 */
2773 if (!(buffer->flags & RB_FL_OVERWRITE)) {
2774 local_inc(&cpu_buffer->dropped_events);
2775 goto out_reset;
2776 }
2777
2778 ret = rb_handle_head_page(cpu_buffer,
2779 tail_page,
2780 next_page);
2781 if (ret < 0)
2782 goto out_reset;
2783 if (ret)
2784 goto out_again;
2785 } else {
2786 /*
2787 * We need to be careful here too. The
2788 * commit page could still be on the reader
2789 * page. We could have a small buffer, and
2790 * have filled up the buffer with events
2791 * from interrupts and such, and wrapped.
2792 *
2793 * Note, if the tail page is also on the
2794 * reader_page, we let it move out.
2795 */
2796 if (unlikely((cpu_buffer->commit_page !=
2797 cpu_buffer->tail_page) &&
2798 (cpu_buffer->commit_page ==
2799 cpu_buffer->reader_page))) {
2800 local_inc(&cpu_buffer->commit_overrun);
2801 goto out_reset;
2802 }
2803 }
2804 }
2805
2806 rb_tail_page_update(cpu_buffer, tail_page, next_page);
2807
2808 out_again:
2809
2810 rb_reset_tail(cpu_buffer, tail, info);
2811
2812 /* Commit what we have for now. */
2813 rb_end_commit(cpu_buffer);
2814 /* rb_end_commit() decs committing */
2815 local_inc(&cpu_buffer->committing);
2816
2817 /* fail and let the caller try again */
2818 return ERR_PTR(-EAGAIN);
2819
2820 out_reset:
2821 /* reset write */
2822 rb_reset_tail(cpu_buffer, tail, info);
2823
2824 return NULL;
2825 }
2826
2827 /* Slow path */
2828 static struct ring_buffer_event *
rb_add_time_stamp(struct ring_buffer_event * event,u64 delta,bool abs)2829 rb_add_time_stamp(struct ring_buffer_event *event, u64 delta, bool abs)
2830 {
2831 if (abs)
2832 event->type_len = RINGBUF_TYPE_TIME_STAMP;
2833 else
2834 event->type_len = RINGBUF_TYPE_TIME_EXTEND;
2835
2836 /* Not the first event on the page, or not delta? */
2837 if (abs || rb_event_index(event)) {
2838 event->time_delta = delta & TS_MASK;
2839 event->array[0] = delta >> TS_SHIFT;
2840 } else {
2841 /* nope, just zero it */
2842 event->time_delta = 0;
2843 event->array[0] = 0;
2844 }
2845
2846 return skip_time_extend(event);
2847 }
2848
2849 #ifndef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
sched_clock_stable(void)2850 static inline bool sched_clock_stable(void)
2851 {
2852 return true;
2853 }
2854 #endif
2855
2856 static void
rb_check_timestamp(struct ring_buffer_per_cpu * cpu_buffer,struct rb_event_info * info)2857 rb_check_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2858 struct rb_event_info *info)
2859 {
2860 u64 write_stamp;
2861
2862 WARN_ONCE(1, "Delta way too big! %llu ts=%llu before=%llu after=%llu write stamp=%llu\n%s",
2863 (unsigned long long)info->delta,
2864 (unsigned long long)info->ts,
2865 (unsigned long long)info->before,
2866 (unsigned long long)info->after,
2867 (unsigned long long)(rb_time_read(&cpu_buffer->write_stamp, &write_stamp) ? write_stamp : 0),
2868 sched_clock_stable() ? "" :
2869 "If you just came from a suspend/resume,\n"
2870 "please switch to the trace global clock:\n"
2871 " echo global > /sys/kernel/debug/tracing/trace_clock\n"
2872 "or add trace_clock=global to the kernel command line\n");
2873 }
2874
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)2875 static void rb_add_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2876 struct ring_buffer_event **event,
2877 struct rb_event_info *info,
2878 u64 *delta,
2879 unsigned int *length)
2880 {
2881 bool abs = info->add_timestamp &
2882 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE);
2883
2884 if (unlikely(info->delta > (1ULL << 59))) {
2885 /*
2886 * Some timers can use more than 59 bits, and when a timestamp
2887 * is added to the buffer, it will lose those bits.
2888 */
2889 if (abs && (info->ts & TS_MSB)) {
2890 info->delta &= ABS_TS_MASK;
2891
2892 /* did the clock go backwards */
2893 } else if (info->before == info->after && info->before > info->ts) {
2894 /* not interrupted */
2895 static int once;
2896
2897 /*
2898 * This is possible with a recalibrating of the TSC.
2899 * Do not produce a call stack, but just report it.
2900 */
2901 if (!once) {
2902 once++;
2903 pr_warn("Ring buffer clock went backwards: %llu -> %llu\n",
2904 info->before, info->ts);
2905 }
2906 } else
2907 rb_check_timestamp(cpu_buffer, info);
2908 if (!abs)
2909 info->delta = 0;
2910 }
2911 *event = rb_add_time_stamp(*event, info->delta, abs);
2912 *length -= RB_LEN_TIME_EXTEND;
2913 *delta = 0;
2914 }
2915
2916 /**
2917 * rb_update_event - update event type and data
2918 * @cpu_buffer: The per cpu buffer of the @event
2919 * @event: the event to update
2920 * @info: The info to update the @event with (contains length and delta)
2921 *
2922 * Update the type and data fields of the @event. The length
2923 * is the actual size that is written to the ring buffer,
2924 * and with this, we can determine what to place into the
2925 * data field.
2926 */
2927 static void
rb_update_event(struct ring_buffer_per_cpu * cpu_buffer,struct ring_buffer_event * event,struct rb_event_info * info)2928 rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
2929 struct ring_buffer_event *event,
2930 struct rb_event_info *info)
2931 {
2932 unsigned length = info->length;
2933 u64 delta = info->delta;
2934 unsigned int nest = local_read(&cpu_buffer->committing) - 1;
2935
2936 if (!WARN_ON_ONCE(nest >= MAX_NEST))
2937 cpu_buffer->event_stamp[nest] = info->ts;
2938
2939 /*
2940 * If we need to add a timestamp, then we
2941 * add it to the start of the reserved space.
2942 */
2943 if (unlikely(info->add_timestamp))
2944 rb_add_timestamp(cpu_buffer, &event, info, &delta, &length);
2945
2946 event->time_delta = delta;
2947 length -= RB_EVNT_HDR_SIZE;
2948 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
2949 event->type_len = 0;
2950 event->array[0] = length;
2951 } else
2952 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
2953 }
2954
rb_calculate_event_length(unsigned length)2955 static unsigned rb_calculate_event_length(unsigned length)
2956 {
2957 struct ring_buffer_event event; /* Used only for sizeof array */
2958
2959 /* zero length can cause confusions */
2960 if (!length)
2961 length++;
2962
2963 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
2964 length += sizeof(event.array[0]);
2965
2966 length += RB_EVNT_HDR_SIZE;
2967 length = ALIGN(length, RB_ARCH_ALIGNMENT);
2968
2969 /*
2970 * In case the time delta is larger than the 27 bits for it
2971 * in the header, we need to add a timestamp. If another
2972 * event comes in when trying to discard this one to increase
2973 * the length, then the timestamp will be added in the allocated
2974 * space of this event. If length is bigger than the size needed
2975 * for the TIME_EXTEND, then padding has to be used. The events
2976 * length must be either RB_LEN_TIME_EXTEND, or greater than or equal
2977 * to RB_LEN_TIME_EXTEND + 8, as 8 is the minimum size for padding.
2978 * As length is a multiple of 4, we only need to worry if it
2979 * is 12 (RB_LEN_TIME_EXTEND + 4).
2980 */
2981 if (length == RB_LEN_TIME_EXTEND + RB_ALIGNMENT)
2982 length += RB_ALIGNMENT;
2983
2984 return length;
2985 }
2986
rb_time_delta(struct ring_buffer_event * event)2987 static u64 rb_time_delta(struct ring_buffer_event *event)
2988 {
2989 switch (event->type_len) {
2990 case RINGBUF_TYPE_PADDING:
2991 return 0;
2992
2993 case RINGBUF_TYPE_TIME_EXTEND:
2994 return rb_event_time_stamp(event);
2995
2996 case RINGBUF_TYPE_TIME_STAMP:
2997 return 0;
2998
2999 case RINGBUF_TYPE_DATA:
3000 return event->time_delta;
3001 default:
3002 return 0;
3003 }
3004 }
3005
3006 static inline int
rb_try_to_discard(struct ring_buffer_per_cpu * cpu_buffer,struct ring_buffer_event * event)3007 rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
3008 struct ring_buffer_event *event)
3009 {
3010 unsigned long new_index, old_index;
3011 struct buffer_page *bpage;
3012 unsigned long index;
3013 unsigned long addr;
3014 u64 write_stamp;
3015 u64 delta;
3016
3017 new_index = rb_event_index(event);
3018 old_index = new_index + rb_event_ts_length(event);
3019 addr = (unsigned long)event;
3020 addr &= PAGE_MASK;
3021
3022 bpage = READ_ONCE(cpu_buffer->tail_page);
3023
3024 delta = rb_time_delta(event);
3025
3026 if (!rb_time_read(&cpu_buffer->write_stamp, &write_stamp))
3027 return 0;
3028
3029 /* Make sure the write stamp is read before testing the location */
3030 barrier();
3031
3032 if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
3033 unsigned long write_mask =
3034 local_read(&bpage->write) & ~RB_WRITE_MASK;
3035 unsigned long event_length = rb_event_length(event);
3036
3037 /* Something came in, can't discard */
3038 if (!rb_time_cmpxchg(&cpu_buffer->write_stamp,
3039 write_stamp, write_stamp - delta))
3040 return 0;
3041
3042 /*
3043 * It's possible that the event time delta is zero
3044 * (has the same time stamp as the previous event)
3045 * in which case write_stamp and before_stamp could
3046 * be the same. In such a case, force before_stamp
3047 * to be different than write_stamp. It doesn't
3048 * matter what it is, as long as its different.
3049 */
3050 if (!delta)
3051 rb_time_set(&cpu_buffer->before_stamp, 0);
3052
3053 /*
3054 * If an event were to come in now, it would see that the
3055 * write_stamp and the before_stamp are different, and assume
3056 * that this event just added itself before updating
3057 * the write stamp. The interrupting event will fix the
3058 * write stamp for us, and use the before stamp as its delta.
3059 */
3060
3061 /*
3062 * This is on the tail page. It is possible that
3063 * a write could come in and move the tail page
3064 * and write to the next page. That is fine
3065 * because we just shorten what is on this page.
3066 */
3067 old_index += write_mask;
3068 new_index += write_mask;
3069 index = local_cmpxchg(&bpage->write, old_index, new_index);
3070 if (index == old_index) {
3071 /* update counters */
3072 local_sub(event_length, &cpu_buffer->entries_bytes);
3073 return 1;
3074 }
3075 }
3076
3077 /* could not discard */
3078 return 0;
3079 }
3080
rb_start_commit(struct ring_buffer_per_cpu * cpu_buffer)3081 static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
3082 {
3083 local_inc(&cpu_buffer->committing);
3084 local_inc(&cpu_buffer->commits);
3085 }
3086
3087 static __always_inline void
rb_set_commit_to_write(struct ring_buffer_per_cpu * cpu_buffer)3088 rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
3089 {
3090 unsigned long max_count;
3091
3092 /*
3093 * We only race with interrupts and NMIs on this CPU.
3094 * If we own the commit event, then we can commit
3095 * all others that interrupted us, since the interruptions
3096 * are in stack format (they finish before they come
3097 * back to us). This allows us to do a simple loop to
3098 * assign the commit to the tail.
3099 */
3100 again:
3101 max_count = cpu_buffer->nr_pages * 100;
3102
3103 while (cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)) {
3104 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
3105 return;
3106 if (RB_WARN_ON(cpu_buffer,
3107 rb_is_reader_page(cpu_buffer->tail_page)))
3108 return;
3109 local_set(&cpu_buffer->commit_page->page->commit,
3110 rb_page_write(cpu_buffer->commit_page));
3111 rb_inc_page(&cpu_buffer->commit_page);
3112 /* add barrier to keep gcc from optimizing too much */
3113 barrier();
3114 }
3115 while (rb_commit_index(cpu_buffer) !=
3116 rb_page_write(cpu_buffer->commit_page)) {
3117
3118 local_set(&cpu_buffer->commit_page->page->commit,
3119 rb_page_write(cpu_buffer->commit_page));
3120 RB_WARN_ON(cpu_buffer,
3121 local_read(&cpu_buffer->commit_page->page->commit) &
3122 ~RB_WRITE_MASK);
3123 barrier();
3124 }
3125
3126 /* again, keep gcc from optimizing */
3127 barrier();
3128
3129 /*
3130 * If an interrupt came in just after the first while loop
3131 * and pushed the tail page forward, we will be left with
3132 * a dangling commit that will never go forward.
3133 */
3134 if (unlikely(cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)))
3135 goto again;
3136 }
3137
rb_end_commit(struct ring_buffer_per_cpu * cpu_buffer)3138 static __always_inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
3139 {
3140 unsigned long commits;
3141
3142 if (RB_WARN_ON(cpu_buffer,
3143 !local_read(&cpu_buffer->committing)))
3144 return;
3145
3146 again:
3147 commits = local_read(&cpu_buffer->commits);
3148 /* synchronize with interrupts */
3149 barrier();
3150 if (local_read(&cpu_buffer->committing) == 1)
3151 rb_set_commit_to_write(cpu_buffer);
3152
3153 local_dec(&cpu_buffer->committing);
3154
3155 /* synchronize with interrupts */
3156 barrier();
3157
3158 /*
3159 * Need to account for interrupts coming in between the
3160 * updating of the commit page and the clearing of the
3161 * committing counter.
3162 */
3163 if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
3164 !local_read(&cpu_buffer->committing)) {
3165 local_inc(&cpu_buffer->committing);
3166 goto again;
3167 }
3168 }
3169
rb_event_discard(struct ring_buffer_event * event)3170 static inline void rb_event_discard(struct ring_buffer_event *event)
3171 {
3172 if (extended_time(event))
3173 event = skip_time_extend(event);
3174
3175 /* array[0] holds the actual length for the discarded event */
3176 event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
3177 event->type_len = RINGBUF_TYPE_PADDING;
3178 /* time delta must be non zero */
3179 if (!event->time_delta)
3180 event->time_delta = 1;
3181 }
3182
rb_commit(struct ring_buffer_per_cpu * cpu_buffer,struct ring_buffer_event * event)3183 static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
3184 struct ring_buffer_event *event)
3185 {
3186 local_inc(&cpu_buffer->entries);
3187 rb_end_commit(cpu_buffer);
3188 }
3189
3190 static __always_inline void
rb_wakeups(struct trace_buffer * buffer,struct ring_buffer_per_cpu * cpu_buffer)3191 rb_wakeups(struct trace_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer)
3192 {
3193 if (buffer->irq_work.waiters_pending) {
3194 buffer->irq_work.waiters_pending = false;
3195 /* irq_work_queue() supplies it's own memory barriers */
3196 irq_work_queue(&buffer->irq_work.work);
3197 }
3198
3199 if (cpu_buffer->irq_work.waiters_pending) {
3200 cpu_buffer->irq_work.waiters_pending = false;
3201 /* irq_work_queue() supplies it's own memory barriers */
3202 irq_work_queue(&cpu_buffer->irq_work.work);
3203 }
3204
3205 if (cpu_buffer->last_pages_touch == local_read(&cpu_buffer->pages_touched))
3206 return;
3207
3208 if (cpu_buffer->reader_page == cpu_buffer->commit_page)
3209 return;
3210
3211 if (!cpu_buffer->irq_work.full_waiters_pending)
3212 return;
3213
3214 cpu_buffer->last_pages_touch = local_read(&cpu_buffer->pages_touched);
3215
3216 if (!full_hit(buffer, cpu_buffer->cpu, cpu_buffer->shortest_full))
3217 return;
3218
3219 cpu_buffer->irq_work.wakeup_full = true;
3220 cpu_buffer->irq_work.full_waiters_pending = false;
3221 /* irq_work_queue() supplies it's own memory barriers */
3222 irq_work_queue(&cpu_buffer->irq_work.work);
3223 }
3224
3225 #ifdef CONFIG_RING_BUFFER_RECORD_RECURSION
3226 # define do_ring_buffer_record_recursion() \
3227 do_ftrace_record_recursion(_THIS_IP_, _RET_IP_)
3228 #else
3229 # define do_ring_buffer_record_recursion() do { } while (0)
3230 #endif
3231
3232 /*
3233 * The lock and unlock are done within a preempt disable section.
3234 * The current_context per_cpu variable can only be modified
3235 * by the current task between lock and unlock. But it can
3236 * be modified more than once via an interrupt. To pass this
3237 * information from the lock to the unlock without having to
3238 * access the 'in_interrupt()' functions again (which do show
3239 * a bit of overhead in something as critical as function tracing,
3240 * we use a bitmask trick.
3241 *
3242 * bit 1 = NMI context
3243 * bit 2 = IRQ context
3244 * bit 3 = SoftIRQ context
3245 * bit 4 = normal context.
3246 *
3247 * This works because this is the order of contexts that can
3248 * preempt other contexts. A SoftIRQ never preempts an IRQ
3249 * context.
3250 *
3251 * When the context is determined, the corresponding bit is
3252 * checked and set (if it was set, then a recursion of that context
3253 * happened).
3254 *
3255 * On unlock, we need to clear this bit. To do so, just subtract
3256 * 1 from the current_context and AND it to itself.
3257 *
3258 * (binary)
3259 * 101 - 1 = 100
3260 * 101 & 100 = 100 (clearing bit zero)
3261 *
3262 * 1010 - 1 = 1001
3263 * 1010 & 1001 = 1000 (clearing bit 1)
3264 *
3265 * The least significant bit can be cleared this way, and it
3266 * just so happens that it is the same bit corresponding to
3267 * the current context.
3268 *
3269 * Now the TRANSITION bit breaks the above slightly. The TRANSITION bit
3270 * is set when a recursion is detected at the current context, and if
3271 * the TRANSITION bit is already set, it will fail the recursion.
3272 * This is needed because there's a lag between the changing of
3273 * interrupt context and updating the preempt count. In this case,
3274 * a false positive will be found. To handle this, one extra recursion
3275 * is allowed, and this is done by the TRANSITION bit. If the TRANSITION
3276 * bit is already set, then it is considered a recursion and the function
3277 * ends. Otherwise, the TRANSITION bit is set, and that bit is returned.
3278 *
3279 * On the trace_recursive_unlock(), the TRANSITION bit will be the first
3280 * to be cleared. Even if it wasn't the context that set it. That is,
3281 * if an interrupt comes in while NORMAL bit is set and the ring buffer
3282 * is called before preempt_count() is updated, since the check will
3283 * be on the NORMAL bit, the TRANSITION bit will then be set. If an
3284 * NMI then comes in, it will set the NMI bit, but when the NMI code
3285 * does the trace_recursive_unlock() it will clear the TRANSITION bit
3286 * and leave the NMI bit set. But this is fine, because the interrupt
3287 * code that set the TRANSITION bit will then clear the NMI bit when it
3288 * calls trace_recursive_unlock(). If another NMI comes in, it will
3289 * set the TRANSITION bit and continue.
3290 *
3291 * Note: The TRANSITION bit only handles a single transition between context.
3292 */
3293
3294 static __always_inline int
trace_recursive_lock(struct ring_buffer_per_cpu * cpu_buffer)3295 trace_recursive_lock(struct ring_buffer_per_cpu *cpu_buffer)
3296 {
3297 unsigned int val = cpu_buffer->current_context;
3298 int bit = interrupt_context_level();
3299
3300 bit = RB_CTX_NORMAL - bit;
3301
3302 if (unlikely(val & (1 << (bit + cpu_buffer->nest)))) {
3303 /*
3304 * It is possible that this was called by transitioning
3305 * between interrupt context, and preempt_count() has not
3306 * been updated yet. In this case, use the TRANSITION bit.
3307 */
3308 bit = RB_CTX_TRANSITION;
3309 if (val & (1 << (bit + cpu_buffer->nest))) {
3310 do_ring_buffer_record_recursion();
3311 return 1;
3312 }
3313 }
3314
3315 val |= (1 << (bit + cpu_buffer->nest));
3316 cpu_buffer->current_context = val;
3317
3318 return 0;
3319 }
3320
3321 static __always_inline void
trace_recursive_unlock(struct ring_buffer_per_cpu * cpu_buffer)3322 trace_recursive_unlock(struct ring_buffer_per_cpu *cpu_buffer)
3323 {
3324 cpu_buffer->current_context &=
3325 cpu_buffer->current_context - (1 << cpu_buffer->nest);
3326 }
3327
3328 /* The recursive locking above uses 5 bits */
3329 #define NESTED_BITS 5
3330
3331 /**
3332 * ring_buffer_nest_start - Allow to trace while nested
3333 * @buffer: The ring buffer to modify
3334 *
3335 * The ring buffer has a safety mechanism to prevent recursion.
3336 * But there may be a case where a trace needs to be done while
3337 * tracing something else. In this case, calling this function
3338 * will allow this function to nest within a currently active
3339 * ring_buffer_lock_reserve().
3340 *
3341 * Call this function before calling another ring_buffer_lock_reserve() and
3342 * call ring_buffer_nest_end() after the nested ring_buffer_unlock_commit().
3343 */
ring_buffer_nest_start(struct trace_buffer * buffer)3344 void ring_buffer_nest_start(struct trace_buffer *buffer)
3345 {
3346 struct ring_buffer_per_cpu *cpu_buffer;
3347 int cpu;
3348
3349 /* Enabled by ring_buffer_nest_end() */
3350 preempt_disable_notrace();
3351 cpu = raw_smp_processor_id();
3352 cpu_buffer = buffer->buffers[cpu];
3353 /* This is the shift value for the above recursive locking */
3354 cpu_buffer->nest += NESTED_BITS;
3355 }
3356
3357 /**
3358 * ring_buffer_nest_end - Allow to trace while nested
3359 * @buffer: The ring buffer to modify
3360 *
3361 * Must be called after ring_buffer_nest_start() and after the
3362 * ring_buffer_unlock_commit().
3363 */
ring_buffer_nest_end(struct trace_buffer * buffer)3364 void ring_buffer_nest_end(struct trace_buffer *buffer)
3365 {
3366 struct ring_buffer_per_cpu *cpu_buffer;
3367 int cpu;
3368
3369 /* disabled by ring_buffer_nest_start() */
3370 cpu = raw_smp_processor_id();
3371 cpu_buffer = buffer->buffers[cpu];
3372 /* This is the shift value for the above recursive locking */
3373 cpu_buffer->nest -= NESTED_BITS;
3374 preempt_enable_notrace();
3375 }
3376
3377 /**
3378 * ring_buffer_unlock_commit - commit a reserved
3379 * @buffer: The buffer to commit to
3380 * @event: The event pointer to commit.
3381 *
3382 * This commits the data to the ring buffer, and releases any locks held.
3383 *
3384 * Must be paired with ring_buffer_lock_reserve.
3385 */
ring_buffer_unlock_commit(struct trace_buffer * buffer,struct ring_buffer_event * event)3386 int ring_buffer_unlock_commit(struct trace_buffer *buffer,
3387 struct ring_buffer_event *event)
3388 {
3389 struct ring_buffer_per_cpu *cpu_buffer;
3390 int cpu = raw_smp_processor_id();
3391
3392 cpu_buffer = buffer->buffers[cpu];
3393
3394 rb_commit(cpu_buffer, event);
3395
3396 rb_wakeups(buffer, cpu_buffer);
3397
3398 trace_recursive_unlock(cpu_buffer);
3399
3400 preempt_enable_notrace();
3401
3402 return 0;
3403 }
3404 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
3405
3406 /* Special value to validate all deltas on a page. */
3407 #define CHECK_FULL_PAGE 1L
3408
3409 #ifdef CONFIG_RING_BUFFER_VALIDATE_TIME_DELTAS
dump_buffer_page(struct buffer_data_page * bpage,struct rb_event_info * info,unsigned long tail)3410 static void dump_buffer_page(struct buffer_data_page *bpage,
3411 struct rb_event_info *info,
3412 unsigned long tail)
3413 {
3414 struct ring_buffer_event *event;
3415 u64 ts, delta;
3416 int e;
3417
3418 ts = bpage->time_stamp;
3419 pr_warn(" [%lld] PAGE TIME STAMP\n", ts);
3420
3421 for (e = 0; e < tail; e += rb_event_length(event)) {
3422
3423 event = (struct ring_buffer_event *)(bpage->data + e);
3424
3425 switch (event->type_len) {
3426
3427 case RINGBUF_TYPE_TIME_EXTEND:
3428 delta = rb_event_time_stamp(event);
3429 ts += delta;
3430 pr_warn(" [%lld] delta:%lld TIME EXTEND\n", ts, delta);
3431 break;
3432
3433 case RINGBUF_TYPE_TIME_STAMP:
3434 delta = rb_event_time_stamp(event);
3435 ts = rb_fix_abs_ts(delta, ts);
3436 pr_warn(" [%lld] absolute:%lld TIME STAMP\n", ts, delta);
3437 break;
3438
3439 case RINGBUF_TYPE_PADDING:
3440 ts += event->time_delta;
3441 pr_warn(" [%lld] delta:%d PADDING\n", ts, event->time_delta);
3442 break;
3443
3444 case RINGBUF_TYPE_DATA:
3445 ts += event->time_delta;
3446 pr_warn(" [%lld] delta:%d\n", ts, event->time_delta);
3447 break;
3448
3449 default:
3450 break;
3451 }
3452 }
3453 }
3454
3455 static DEFINE_PER_CPU(atomic_t, checking);
3456 static atomic_t ts_dump;
3457
3458 /*
3459 * Check if the current event time stamp matches the deltas on
3460 * the buffer page.
3461 */
check_buffer(struct ring_buffer_per_cpu * cpu_buffer,struct rb_event_info * info,unsigned long tail)3462 static void check_buffer(struct ring_buffer_per_cpu *cpu_buffer,
3463 struct rb_event_info *info,
3464 unsigned long tail)
3465 {
3466 struct ring_buffer_event *event;
3467 struct buffer_data_page *bpage;
3468 u64 ts, delta;
3469 bool full = false;
3470 int e;
3471
3472 bpage = info->tail_page->page;
3473
3474 if (tail == CHECK_FULL_PAGE) {
3475 full = true;
3476 tail = local_read(&bpage->commit);
3477 } else if (info->add_timestamp &
3478 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE)) {
3479 /* Ignore events with absolute time stamps */
3480 return;
3481 }
3482
3483 /*
3484 * Do not check the first event (skip possible extends too).
3485 * Also do not check if previous events have not been committed.
3486 */
3487 if (tail <= 8 || tail > local_read(&bpage->commit))
3488 return;
3489
3490 /*
3491 * If this interrupted another event,
3492 */
3493 if (atomic_inc_return(this_cpu_ptr(&checking)) != 1)
3494 goto out;
3495
3496 ts = bpage->time_stamp;
3497
3498 for (e = 0; e < tail; e += rb_event_length(event)) {
3499
3500 event = (struct ring_buffer_event *)(bpage->data + e);
3501
3502 switch (event->type_len) {
3503
3504 case RINGBUF_TYPE_TIME_EXTEND:
3505 delta = rb_event_time_stamp(event);
3506 ts += delta;
3507 break;
3508
3509 case RINGBUF_TYPE_TIME_STAMP:
3510 delta = rb_event_time_stamp(event);
3511 ts = rb_fix_abs_ts(delta, ts);
3512 break;
3513
3514 case RINGBUF_TYPE_PADDING:
3515 if (event->time_delta == 1)
3516 break;
3517 fallthrough;
3518 case RINGBUF_TYPE_DATA:
3519 ts += event->time_delta;
3520 break;
3521
3522 default:
3523 RB_WARN_ON(cpu_buffer, 1);
3524 }
3525 }
3526 if ((full && ts > info->ts) ||
3527 (!full && ts + info->delta != info->ts)) {
3528 /* If another report is happening, ignore this one */
3529 if (atomic_inc_return(&ts_dump) != 1) {
3530 atomic_dec(&ts_dump);
3531 goto out;
3532 }
3533 atomic_inc(&cpu_buffer->record_disabled);
3534 /* There's some cases in boot up that this can happen */
3535 WARN_ON_ONCE(system_state != SYSTEM_BOOTING);
3536 pr_warn("[CPU: %d]TIME DOES NOT MATCH expected:%lld actual:%lld delta:%lld before:%lld after:%lld%s\n",
3537 cpu_buffer->cpu,
3538 ts + info->delta, info->ts, info->delta,
3539 info->before, info->after,
3540 full ? " (full)" : "");
3541 dump_buffer_page(bpage, info, tail);
3542 atomic_dec(&ts_dump);
3543 /* Do not re-enable checking */
3544 return;
3545 }
3546 out:
3547 atomic_dec(this_cpu_ptr(&checking));
3548 }
3549 #else
check_buffer(struct ring_buffer_per_cpu * cpu_buffer,struct rb_event_info * info,unsigned long tail)3550 static inline void check_buffer(struct ring_buffer_per_cpu *cpu_buffer,
3551 struct rb_event_info *info,
3552 unsigned long tail)
3553 {
3554 }
3555 #endif /* CONFIG_RING_BUFFER_VALIDATE_TIME_DELTAS */
3556
3557 static struct ring_buffer_event *
__rb_reserve_next(struct ring_buffer_per_cpu * cpu_buffer,struct rb_event_info * info)3558 __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
3559 struct rb_event_info *info)
3560 {
3561 struct ring_buffer_event *event;
3562 struct buffer_page *tail_page;
3563 unsigned long tail, write, w;
3564 bool a_ok;
3565 bool b_ok;
3566
3567 /* Don't let the compiler play games with cpu_buffer->tail_page */
3568 tail_page = info->tail_page = READ_ONCE(cpu_buffer->tail_page);
3569
3570 /*A*/ w = local_read(&tail_page->write) & RB_WRITE_MASK;
3571 barrier();
3572 b_ok = rb_time_read(&cpu_buffer->before_stamp, &info->before);
3573 a_ok = rb_time_read(&cpu_buffer->write_stamp, &info->after);
3574 barrier();
3575 info->ts = rb_time_stamp(cpu_buffer->buffer);
3576
3577 if ((info->add_timestamp & RB_ADD_STAMP_ABSOLUTE)) {
3578 info->delta = info->ts;
3579 } else {
3580 /*
3581 * If interrupting an event time update, we may need an
3582 * absolute timestamp.
3583 * Don't bother if this is the start of a new page (w == 0).
3584 */
3585 if (unlikely(!a_ok || !b_ok || (info->before != info->after && w))) {
3586 info->add_timestamp |= RB_ADD_STAMP_FORCE | RB_ADD_STAMP_EXTEND;
3587 info->length += RB_LEN_TIME_EXTEND;
3588 } else {
3589 info->delta = info->ts - info->after;
3590 if (unlikely(test_time_stamp(info->delta))) {
3591 info->add_timestamp |= RB_ADD_STAMP_EXTEND;
3592 info->length += RB_LEN_TIME_EXTEND;
3593 }
3594 }
3595 }
3596
3597 /*B*/ rb_time_set(&cpu_buffer->before_stamp, info->ts);
3598
3599 /*C*/ write = local_add_return(info->length, &tail_page->write);
3600
3601 /* set write to only the index of the write */
3602 write &= RB_WRITE_MASK;
3603
3604 tail = write - info->length;
3605
3606 /* See if we shot pass the end of this buffer page */
3607 if (unlikely(write > BUF_PAGE_SIZE)) {
3608 /* before and after may now different, fix it up*/
3609 b_ok = rb_time_read(&cpu_buffer->before_stamp, &info->before);
3610 a_ok = rb_time_read(&cpu_buffer->write_stamp, &info->after);
3611 if (a_ok && b_ok && info->before != info->after)
3612 (void)rb_time_cmpxchg(&cpu_buffer->before_stamp,
3613 info->before, info->after);
3614 if (a_ok && b_ok)
3615 check_buffer(cpu_buffer, info, CHECK_FULL_PAGE);
3616 return rb_move_tail(cpu_buffer, tail, info);
3617 }
3618
3619 if (likely(tail == w)) {
3620 u64 save_before;
3621 bool s_ok;
3622
3623 /* Nothing interrupted us between A and C */
3624 /*D*/ rb_time_set(&cpu_buffer->write_stamp, info->ts);
3625 barrier();
3626 /*E*/ s_ok = rb_time_read(&cpu_buffer->before_stamp, &save_before);
3627 RB_WARN_ON(cpu_buffer, !s_ok);
3628 if (likely(!(info->add_timestamp &
3629 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE))))
3630 /* This did not interrupt any time update */
3631 info->delta = info->ts - info->after;
3632 else
3633 /* Just use full timestamp for interrupting event */
3634 info->delta = info->ts;
3635 barrier();
3636 check_buffer(cpu_buffer, info, tail);
3637 if (unlikely(info->ts != save_before)) {
3638 /* SLOW PATH - Interrupted between C and E */
3639
3640 a_ok = rb_time_read(&cpu_buffer->write_stamp, &info->after);
3641 RB_WARN_ON(cpu_buffer, !a_ok);
3642
3643 /* Write stamp must only go forward */
3644 if (save_before > info->after) {
3645 /*
3646 * We do not care about the result, only that
3647 * it gets updated atomically.
3648 */
3649 (void)rb_time_cmpxchg(&cpu_buffer->write_stamp,
3650 info->after, save_before);
3651 }
3652 }
3653 } else {
3654 u64 ts;
3655 /* SLOW PATH - Interrupted between A and C */
3656 a_ok = rb_time_read(&cpu_buffer->write_stamp, &info->after);
3657 /* Was interrupted before here, write_stamp must be valid */
3658 RB_WARN_ON(cpu_buffer, !a_ok);
3659 ts = rb_time_stamp(cpu_buffer->buffer);
3660 barrier();
3661 /*E*/ if (write == (local_read(&tail_page->write) & RB_WRITE_MASK) &&
3662 info->after < ts &&
3663 rb_time_cmpxchg(&cpu_buffer->write_stamp,
3664 info->after, ts)) {
3665 /* Nothing came after this event between C and E */
3666 info->delta = ts - info->after;
3667 } else {
3668 /*
3669 * Interrupted between C and E:
3670 * Lost the previous events time stamp. Just set the
3671 * delta to zero, and this will be the same time as
3672 * the event this event interrupted. And the events that
3673 * came after this will still be correct (as they would
3674 * have built their delta on the previous event.
3675 */
3676 info->delta = 0;
3677 }
3678 info->ts = ts;
3679 info->add_timestamp &= ~RB_ADD_STAMP_FORCE;
3680 }
3681
3682 /*
3683 * If this is the first commit on the page, then it has the same
3684 * timestamp as the page itself.
3685 */
3686 if (unlikely(!tail && !(info->add_timestamp &
3687 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE))))
3688 info->delta = 0;
3689
3690 /* We reserved something on the buffer */
3691
3692 event = __rb_page_index(tail_page, tail);
3693 rb_update_event(cpu_buffer, event, info);
3694
3695 local_inc(&tail_page->entries);
3696
3697 /*
3698 * If this is the first commit on the page, then update
3699 * its timestamp.
3700 */
3701 if (unlikely(!tail))
3702 tail_page->page->time_stamp = info->ts;
3703
3704 /* account for these added bytes */
3705 local_add(info->length, &cpu_buffer->entries_bytes);
3706
3707 return event;
3708 }
3709
3710 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)3711 rb_reserve_next_event(struct trace_buffer *buffer,
3712 struct ring_buffer_per_cpu *cpu_buffer,
3713 unsigned long length)
3714 {
3715 struct ring_buffer_event *event;
3716 struct rb_event_info info;
3717 int nr_loops = 0;
3718 int add_ts_default;
3719
3720 rb_start_commit(cpu_buffer);
3721 /* The commit page can not change after this */
3722
3723 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
3724 /*
3725 * Due to the ability to swap a cpu buffer from a buffer
3726 * it is possible it was swapped before we committed.
3727 * (committing stops a swap). We check for it here and
3728 * if it happened, we have to fail the write.
3729 */
3730 barrier();
3731 if (unlikely(READ_ONCE(cpu_buffer->buffer) != buffer)) {
3732 local_dec(&cpu_buffer->committing);
3733 local_dec(&cpu_buffer->commits);
3734 return NULL;
3735 }
3736 #endif
3737
3738 info.length = rb_calculate_event_length(length);
3739
3740 if (ring_buffer_time_stamp_abs(cpu_buffer->buffer)) {
3741 add_ts_default = RB_ADD_STAMP_ABSOLUTE;
3742 info.length += RB_LEN_TIME_EXTEND;
3743 } else {
3744 add_ts_default = RB_ADD_STAMP_NONE;
3745 }
3746
3747 again:
3748 info.add_timestamp = add_ts_default;
3749 info.delta = 0;
3750
3751 /*
3752 * We allow for interrupts to reenter here and do a trace.
3753 * If one does, it will cause this original code to loop
3754 * back here. Even with heavy interrupts happening, this
3755 * should only happen a few times in a row. If this happens
3756 * 1000 times in a row, there must be either an interrupt
3757 * storm or we have something buggy.
3758 * Bail!
3759 */
3760 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
3761 goto out_fail;
3762
3763 event = __rb_reserve_next(cpu_buffer, &info);
3764
3765 if (unlikely(PTR_ERR(event) == -EAGAIN)) {
3766 if (info.add_timestamp & (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_EXTEND))
3767 info.length -= RB_LEN_TIME_EXTEND;
3768 goto again;
3769 }
3770
3771 if (likely(event))
3772 return event;
3773 out_fail:
3774 rb_end_commit(cpu_buffer);
3775 return NULL;
3776 }
3777
3778 /**
3779 * ring_buffer_lock_reserve - reserve a part of the buffer
3780 * @buffer: the ring buffer to reserve from
3781 * @length: the length of the data to reserve (excluding event header)
3782 *
3783 * Returns a reserved event on the ring buffer to copy directly to.
3784 * The user of this interface will need to get the body to write into
3785 * and can use the ring_buffer_event_data() interface.
3786 *
3787 * The length is the length of the data needed, not the event length
3788 * which also includes the event header.
3789 *
3790 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
3791 * If NULL is returned, then nothing has been allocated or locked.
3792 */
3793 struct ring_buffer_event *
ring_buffer_lock_reserve(struct trace_buffer * buffer,unsigned long length)3794 ring_buffer_lock_reserve(struct trace_buffer *buffer, unsigned long length)
3795 {
3796 struct ring_buffer_per_cpu *cpu_buffer;
3797 struct ring_buffer_event *event;
3798 int cpu;
3799
3800 /* If we are tracing schedule, we don't want to recurse */
3801 preempt_disable_notrace();
3802
3803 if (unlikely(atomic_read(&buffer->record_disabled)))
3804 goto out;
3805
3806 cpu = raw_smp_processor_id();
3807
3808 if (unlikely(!cpumask_test_cpu(cpu, buffer->cpumask)))
3809 goto out;
3810
3811 cpu_buffer = buffer->buffers[cpu];
3812
3813 if (unlikely(atomic_read(&cpu_buffer->record_disabled)))
3814 goto out;
3815
3816 if (unlikely(length > BUF_MAX_DATA_SIZE))
3817 goto out;
3818
3819 if (unlikely(trace_recursive_lock(cpu_buffer)))
3820 goto out;
3821
3822 event = rb_reserve_next_event(buffer, cpu_buffer, length);
3823 if (!event)
3824 goto out_unlock;
3825
3826 return event;
3827
3828 out_unlock:
3829 trace_recursive_unlock(cpu_buffer);
3830 out:
3831 preempt_enable_notrace();
3832 return NULL;
3833 }
3834 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
3835
3836 /*
3837 * Decrement the entries to the page that an event is on.
3838 * The event does not even need to exist, only the pointer
3839 * to the page it is on. This may only be called before the commit
3840 * takes place.
3841 */
3842 static inline void
rb_decrement_entry(struct ring_buffer_per_cpu * cpu_buffer,struct ring_buffer_event * event)3843 rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
3844 struct ring_buffer_event *event)
3845 {
3846 unsigned long addr = (unsigned long)event;
3847 struct buffer_page *bpage = cpu_buffer->commit_page;
3848 struct buffer_page *start;
3849
3850 addr &= PAGE_MASK;
3851
3852 /* Do the likely case first */
3853 if (likely(bpage->page == (void *)addr)) {
3854 local_dec(&bpage->entries);
3855 return;
3856 }
3857
3858 /*
3859 * Because the commit page may be on the reader page we
3860 * start with the next page and check the end loop there.
3861 */
3862 rb_inc_page(&bpage);
3863 start = bpage;
3864 do {
3865 if (bpage->page == (void *)addr) {
3866 local_dec(&bpage->entries);
3867 return;
3868 }
3869 rb_inc_page(&bpage);
3870 } while (bpage != start);
3871
3872 /* commit not part of this buffer?? */
3873 RB_WARN_ON(cpu_buffer, 1);
3874 }
3875
3876 /**
3877 * ring_buffer_discard_commit - discard an event that has not been committed
3878 * @buffer: the ring buffer
3879 * @event: non committed event to discard
3880 *
3881 * Sometimes an event that is in the ring buffer needs to be ignored.
3882 * This function lets the user discard an event in the ring buffer
3883 * and then that event will not be read later.
3884 *
3885 * This function only works if it is called before the item has been
3886 * committed. It will try to free the event from the ring buffer
3887 * if another event has not been added behind it.
3888 *
3889 * If another event has been added behind it, it will set the event
3890 * up as discarded, and perform the commit.
3891 *
3892 * If this function is called, do not call ring_buffer_unlock_commit on
3893 * the event.
3894 */
ring_buffer_discard_commit(struct trace_buffer * buffer,struct ring_buffer_event * event)3895 void ring_buffer_discard_commit(struct trace_buffer *buffer,
3896 struct ring_buffer_event *event)
3897 {
3898 struct ring_buffer_per_cpu *cpu_buffer;
3899 int cpu;
3900
3901 /* The event is discarded regardless */
3902 rb_event_discard(event);
3903
3904 cpu = smp_processor_id();
3905 cpu_buffer = buffer->buffers[cpu];
3906
3907 /*
3908 * This must only be called if the event has not been
3909 * committed yet. Thus we can assume that preemption
3910 * is still disabled.
3911 */
3912 RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
3913
3914 rb_decrement_entry(cpu_buffer, event);
3915 if (rb_try_to_discard(cpu_buffer, event))
3916 goto out;
3917
3918 out:
3919 rb_end_commit(cpu_buffer);
3920
3921 trace_recursive_unlock(cpu_buffer);
3922
3923 preempt_enable_notrace();
3924
3925 }
3926 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
3927
3928 /**
3929 * ring_buffer_write - write data to the buffer without reserving
3930 * @buffer: The ring buffer to write to.
3931 * @length: The length of the data being written (excluding the event header)
3932 * @data: The data to write to the buffer.
3933 *
3934 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
3935 * one function. If you already have the data to write to the buffer, it
3936 * may be easier to simply call this function.
3937 *
3938 * Note, like ring_buffer_lock_reserve, the length is the length of the data
3939 * and not the length of the event which would hold the header.
3940 */
ring_buffer_write(struct trace_buffer * buffer,unsigned long length,void * data)3941 int ring_buffer_write(struct trace_buffer *buffer,
3942 unsigned long length,
3943 void *data)
3944 {
3945 struct ring_buffer_per_cpu *cpu_buffer;
3946 struct ring_buffer_event *event;
3947 void *body;
3948 int ret = -EBUSY;
3949 int cpu;
3950
3951 preempt_disable_notrace();
3952
3953 if (atomic_read(&buffer->record_disabled))
3954 goto out;
3955
3956 cpu = raw_smp_processor_id();
3957
3958 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3959 goto out;
3960
3961 cpu_buffer = buffer->buffers[cpu];
3962
3963 if (atomic_read(&cpu_buffer->record_disabled))
3964 goto out;
3965
3966 if (length > BUF_MAX_DATA_SIZE)
3967 goto out;
3968
3969 if (unlikely(trace_recursive_lock(cpu_buffer)))
3970 goto out;
3971
3972 event = rb_reserve_next_event(buffer, cpu_buffer, length);
3973 if (!event)
3974 goto out_unlock;
3975
3976 body = rb_event_data(event);
3977
3978 memcpy(body, data, length);
3979
3980 rb_commit(cpu_buffer, event);
3981
3982 rb_wakeups(buffer, cpu_buffer);
3983
3984 ret = 0;
3985
3986 out_unlock:
3987 trace_recursive_unlock(cpu_buffer);
3988
3989 out:
3990 preempt_enable_notrace();
3991
3992 return ret;
3993 }
3994 EXPORT_SYMBOL_GPL(ring_buffer_write);
3995
rb_per_cpu_empty(struct ring_buffer_per_cpu * cpu_buffer)3996 static bool rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
3997 {
3998 struct buffer_page *reader = cpu_buffer->reader_page;
3999 struct buffer_page *head = rb_set_head_page(cpu_buffer);
4000 struct buffer_page *commit = cpu_buffer->commit_page;
4001
4002 /* In case of error, head will be NULL */
4003 if (unlikely(!head))
4004 return true;
4005
4006 /* Reader should exhaust content in reader page */
4007 if (reader->read != rb_page_commit(reader))
4008 return false;
4009
4010 /*
4011 * If writers are committing on the reader page, knowing all
4012 * committed content has been read, the ring buffer is empty.
4013 */
4014 if (commit == reader)
4015 return true;
4016
4017 /*
4018 * If writers are committing on a page other than reader page
4019 * and head page, there should always be content to read.
4020 */
4021 if (commit != head)
4022 return false;
4023
4024 /*
4025 * Writers are committing on the head page, we just need
4026 * to care about there're committed data, and the reader will
4027 * swap reader page with head page when it is to read data.
4028 */
4029 return rb_page_commit(commit) == 0;
4030 }
4031
4032 /**
4033 * ring_buffer_record_disable - stop all writes into the buffer
4034 * @buffer: The ring buffer to stop writes to.
4035 *
4036 * This prevents all writes to the buffer. Any attempt to write
4037 * to the buffer after this will fail and return NULL.
4038 *
4039 * The caller should call synchronize_rcu() after this.
4040 */
ring_buffer_record_disable(struct trace_buffer * buffer)4041 void ring_buffer_record_disable(struct trace_buffer *buffer)
4042 {
4043 atomic_inc(&buffer->record_disabled);
4044 }
4045 EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
4046
4047 /**
4048 * ring_buffer_record_enable - enable writes to the buffer
4049 * @buffer: The ring buffer to enable writes
4050 *
4051 * Note, multiple disables will need the same number of enables
4052 * to truly enable the writing (much like preempt_disable).
4053 */
ring_buffer_record_enable(struct trace_buffer * buffer)4054 void ring_buffer_record_enable(struct trace_buffer *buffer)
4055 {
4056 atomic_dec(&buffer->record_disabled);
4057 }
4058 EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
4059
4060 /**
4061 * ring_buffer_record_off - stop all writes into the buffer
4062 * @buffer: The ring buffer to stop writes to.
4063 *
4064 * This prevents all writes to the buffer. Any attempt to write
4065 * to the buffer after this will fail and return NULL.
4066 *
4067 * This is different than ring_buffer_record_disable() as
4068 * it works like an on/off switch, where as the disable() version
4069 * must be paired with a enable().
4070 */
ring_buffer_record_off(struct trace_buffer * buffer)4071 void ring_buffer_record_off(struct trace_buffer *buffer)
4072 {
4073 unsigned int rd;
4074 unsigned int new_rd;
4075
4076 do {
4077 rd = atomic_read(&buffer->record_disabled);
4078 new_rd = rd | RB_BUFFER_OFF;
4079 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
4080 }
4081 EXPORT_SYMBOL_GPL(ring_buffer_record_off);
4082
4083 /**
4084 * ring_buffer_record_on - restart writes into the buffer
4085 * @buffer: The ring buffer to start writes to.
4086 *
4087 * This enables all writes to the buffer that was disabled by
4088 * ring_buffer_record_off().
4089 *
4090 * This is different than ring_buffer_record_enable() as
4091 * it works like an on/off switch, where as the enable() version
4092 * must be paired with a disable().
4093 */
ring_buffer_record_on(struct trace_buffer * buffer)4094 void ring_buffer_record_on(struct trace_buffer *buffer)
4095 {
4096 unsigned int rd;
4097 unsigned int new_rd;
4098
4099 do {
4100 rd = atomic_read(&buffer->record_disabled);
4101 new_rd = rd & ~RB_BUFFER_OFF;
4102 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
4103 }
4104 EXPORT_SYMBOL_GPL(ring_buffer_record_on);
4105
4106 /**
4107 * ring_buffer_record_is_on - return true if the ring buffer can write
4108 * @buffer: The ring buffer to see if write is enabled
4109 *
4110 * Returns true if the ring buffer is in a state that it accepts writes.
4111 */
ring_buffer_record_is_on(struct trace_buffer * buffer)4112 bool ring_buffer_record_is_on(struct trace_buffer *buffer)
4113 {
4114 return !atomic_read(&buffer->record_disabled);
4115 }
4116
4117 /**
4118 * ring_buffer_record_is_set_on - return true if the ring buffer is set writable
4119 * @buffer: The ring buffer to see if write is set enabled
4120 *
4121 * Returns true if the ring buffer is set writable by ring_buffer_record_on().
4122 * Note that this does NOT mean it is in a writable state.
4123 *
4124 * It may return true when the ring buffer has been disabled by
4125 * ring_buffer_record_disable(), as that is a temporary disabling of
4126 * the ring buffer.
4127 */
ring_buffer_record_is_set_on(struct trace_buffer * buffer)4128 bool ring_buffer_record_is_set_on(struct trace_buffer *buffer)
4129 {
4130 return !(atomic_read(&buffer->record_disabled) & RB_BUFFER_OFF);
4131 }
4132
4133 /**
4134 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
4135 * @buffer: The ring buffer to stop writes to.
4136 * @cpu: The CPU buffer to stop
4137 *
4138 * This prevents all writes to the buffer. Any attempt to write
4139 * to the buffer after this will fail and return NULL.
4140 *
4141 * The caller should call synchronize_rcu() after this.
4142 */
ring_buffer_record_disable_cpu(struct trace_buffer * buffer,int cpu)4143 void ring_buffer_record_disable_cpu(struct trace_buffer *buffer, int cpu)
4144 {
4145 struct ring_buffer_per_cpu *cpu_buffer;
4146
4147 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4148 return;
4149
4150 cpu_buffer = buffer->buffers[cpu];
4151 atomic_inc(&cpu_buffer->record_disabled);
4152 }
4153 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
4154
4155 /**
4156 * ring_buffer_record_enable_cpu - enable writes to the buffer
4157 * @buffer: The ring buffer to enable writes
4158 * @cpu: The CPU to enable.
4159 *
4160 * Note, multiple disables will need the same number of enables
4161 * to truly enable the writing (much like preempt_disable).
4162 */
ring_buffer_record_enable_cpu(struct trace_buffer * buffer,int cpu)4163 void ring_buffer_record_enable_cpu(struct trace_buffer *buffer, int cpu)
4164 {
4165 struct ring_buffer_per_cpu *cpu_buffer;
4166
4167 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4168 return;
4169
4170 cpu_buffer = buffer->buffers[cpu];
4171 atomic_dec(&cpu_buffer->record_disabled);
4172 }
4173 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
4174
4175 /*
4176 * The total entries in the ring buffer is the running counter
4177 * of entries entered into the ring buffer, minus the sum of
4178 * the entries read from the ring buffer and the number of
4179 * entries that were overwritten.
4180 */
4181 static inline unsigned long
rb_num_of_entries(struct ring_buffer_per_cpu * cpu_buffer)4182 rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
4183 {
4184 return local_read(&cpu_buffer->entries) -
4185 (local_read(&cpu_buffer->overrun) + cpu_buffer->read);
4186 }
4187
4188 /**
4189 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
4190 * @buffer: The ring buffer
4191 * @cpu: The per CPU buffer to read from.
4192 */
ring_buffer_oldest_event_ts(struct trace_buffer * buffer,int cpu)4193 u64 ring_buffer_oldest_event_ts(struct trace_buffer *buffer, int cpu)
4194 {
4195 unsigned long flags;
4196 struct ring_buffer_per_cpu *cpu_buffer;
4197 struct buffer_page *bpage;
4198 u64 ret = 0;
4199
4200 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4201 return 0;
4202
4203 cpu_buffer = buffer->buffers[cpu];
4204 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4205 /*
4206 * if the tail is on reader_page, oldest time stamp is on the reader
4207 * page
4208 */
4209 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
4210 bpage = cpu_buffer->reader_page;
4211 else
4212 bpage = rb_set_head_page(cpu_buffer);
4213 if (bpage)
4214 ret = bpage->page->time_stamp;
4215 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4216
4217 return ret;
4218 }
4219 EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
4220
4221 /**
4222 * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
4223 * @buffer: The ring buffer
4224 * @cpu: The per CPU buffer to read from.
4225 */
ring_buffer_bytes_cpu(struct trace_buffer * buffer,int cpu)4226 unsigned long ring_buffer_bytes_cpu(struct trace_buffer *buffer, int cpu)
4227 {
4228 struct ring_buffer_per_cpu *cpu_buffer;
4229 unsigned long ret;
4230
4231 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4232 return 0;
4233
4234 cpu_buffer = buffer->buffers[cpu];
4235 ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
4236
4237 return ret;
4238 }
4239 EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
4240
4241 /**
4242 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
4243 * @buffer: The ring buffer
4244 * @cpu: The per CPU buffer to get the entries from.
4245 */
ring_buffer_entries_cpu(struct trace_buffer * buffer,int cpu)4246 unsigned long ring_buffer_entries_cpu(struct trace_buffer *buffer, int cpu)
4247 {
4248 struct ring_buffer_per_cpu *cpu_buffer;
4249
4250 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4251 return 0;
4252
4253 cpu_buffer = buffer->buffers[cpu];
4254
4255 return rb_num_of_entries(cpu_buffer);
4256 }
4257 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
4258
4259 /**
4260 * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
4261 * buffer wrapping around (only if RB_FL_OVERWRITE is on).
4262 * @buffer: The ring buffer
4263 * @cpu: The per CPU buffer to get the number of overruns from
4264 */
ring_buffer_overrun_cpu(struct trace_buffer * buffer,int cpu)4265 unsigned long ring_buffer_overrun_cpu(struct trace_buffer *buffer, int cpu)
4266 {
4267 struct ring_buffer_per_cpu *cpu_buffer;
4268 unsigned long ret;
4269
4270 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4271 return 0;
4272
4273 cpu_buffer = buffer->buffers[cpu];
4274 ret = local_read(&cpu_buffer->overrun);
4275
4276 return ret;
4277 }
4278 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
4279
4280 /**
4281 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
4282 * commits failing due to the buffer wrapping around while there are uncommitted
4283 * events, such as during an interrupt storm.
4284 * @buffer: The ring buffer
4285 * @cpu: The per CPU buffer to get the number of overruns from
4286 */
4287 unsigned long
ring_buffer_commit_overrun_cpu(struct trace_buffer * buffer,int cpu)4288 ring_buffer_commit_overrun_cpu(struct trace_buffer *buffer, int cpu)
4289 {
4290 struct ring_buffer_per_cpu *cpu_buffer;
4291 unsigned long ret;
4292
4293 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4294 return 0;
4295
4296 cpu_buffer = buffer->buffers[cpu];
4297 ret = local_read(&cpu_buffer->commit_overrun);
4298
4299 return ret;
4300 }
4301 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
4302
4303 /**
4304 * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
4305 * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
4306 * @buffer: The ring buffer
4307 * @cpu: The per CPU buffer to get the number of overruns from
4308 */
4309 unsigned long
ring_buffer_dropped_events_cpu(struct trace_buffer * buffer,int cpu)4310 ring_buffer_dropped_events_cpu(struct trace_buffer *buffer, int cpu)
4311 {
4312 struct ring_buffer_per_cpu *cpu_buffer;
4313 unsigned long ret;
4314
4315 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4316 return 0;
4317
4318 cpu_buffer = buffer->buffers[cpu];
4319 ret = local_read(&cpu_buffer->dropped_events);
4320
4321 return ret;
4322 }
4323 EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
4324
4325 /**
4326 * ring_buffer_read_events_cpu - get the number of events successfully read
4327 * @buffer: The ring buffer
4328 * @cpu: The per CPU buffer to get the number of events read
4329 */
4330 unsigned long
ring_buffer_read_events_cpu(struct trace_buffer * buffer,int cpu)4331 ring_buffer_read_events_cpu(struct trace_buffer *buffer, int cpu)
4332 {
4333 struct ring_buffer_per_cpu *cpu_buffer;
4334
4335 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4336 return 0;
4337
4338 cpu_buffer = buffer->buffers[cpu];
4339 return cpu_buffer->read;
4340 }
4341 EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu);
4342
4343 /**
4344 * ring_buffer_entries - get the number of entries in a buffer
4345 * @buffer: The ring buffer
4346 *
4347 * Returns the total number of entries in the ring buffer
4348 * (all CPU entries)
4349 */
ring_buffer_entries(struct trace_buffer * buffer)4350 unsigned long ring_buffer_entries(struct trace_buffer *buffer)
4351 {
4352 struct ring_buffer_per_cpu *cpu_buffer;
4353 unsigned long entries = 0;
4354 int cpu;
4355
4356 /* if you care about this being correct, lock the buffer */
4357 for_each_buffer_cpu(buffer, cpu) {
4358 cpu_buffer = buffer->buffers[cpu];
4359 entries += rb_num_of_entries(cpu_buffer);
4360 }
4361
4362 return entries;
4363 }
4364 EXPORT_SYMBOL_GPL(ring_buffer_entries);
4365
4366 /**
4367 * ring_buffer_overruns - get the number of overruns in buffer
4368 * @buffer: The ring buffer
4369 *
4370 * Returns the total number of overruns in the ring buffer
4371 * (all CPU entries)
4372 */
ring_buffer_overruns(struct trace_buffer * buffer)4373 unsigned long ring_buffer_overruns(struct trace_buffer *buffer)
4374 {
4375 struct ring_buffer_per_cpu *cpu_buffer;
4376 unsigned long overruns = 0;
4377 int cpu;
4378
4379 /* if you care about this being correct, lock the buffer */
4380 for_each_buffer_cpu(buffer, cpu) {
4381 cpu_buffer = buffer->buffers[cpu];
4382 overruns += local_read(&cpu_buffer->overrun);
4383 }
4384
4385 return overruns;
4386 }
4387 EXPORT_SYMBOL_GPL(ring_buffer_overruns);
4388
rb_iter_reset(struct ring_buffer_iter * iter)4389 static void rb_iter_reset(struct ring_buffer_iter *iter)
4390 {
4391 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4392
4393 /* Iterator usage is expected to have record disabled */
4394 iter->head_page = cpu_buffer->reader_page;
4395 iter->head = cpu_buffer->reader_page->read;
4396 iter->next_event = iter->head;
4397
4398 iter->cache_reader_page = iter->head_page;
4399 iter->cache_read = cpu_buffer->read;
4400
4401 if (iter->head) {
4402 iter->read_stamp = cpu_buffer->read_stamp;
4403 iter->page_stamp = cpu_buffer->reader_page->page->time_stamp;
4404 } else {
4405 iter->read_stamp = iter->head_page->page->time_stamp;
4406 iter->page_stamp = iter->read_stamp;
4407 }
4408 }
4409
4410 /**
4411 * ring_buffer_iter_reset - reset an iterator
4412 * @iter: The iterator to reset
4413 *
4414 * Resets the iterator, so that it will start from the beginning
4415 * again.
4416 */
ring_buffer_iter_reset(struct ring_buffer_iter * iter)4417 void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
4418 {
4419 struct ring_buffer_per_cpu *cpu_buffer;
4420 unsigned long flags;
4421
4422 if (!iter)
4423 return;
4424
4425 cpu_buffer = iter->cpu_buffer;
4426
4427 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4428 rb_iter_reset(iter);
4429 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4430 }
4431 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
4432
4433 /**
4434 * ring_buffer_iter_empty - check if an iterator has no more to read
4435 * @iter: The iterator to check
4436 */
ring_buffer_iter_empty(struct ring_buffer_iter * iter)4437 int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
4438 {
4439 struct ring_buffer_per_cpu *cpu_buffer;
4440 struct buffer_page *reader;
4441 struct buffer_page *head_page;
4442 struct buffer_page *commit_page;
4443 struct buffer_page *curr_commit_page;
4444 unsigned commit;
4445 u64 curr_commit_ts;
4446 u64 commit_ts;
4447
4448 cpu_buffer = iter->cpu_buffer;
4449 reader = cpu_buffer->reader_page;
4450 head_page = cpu_buffer->head_page;
4451 commit_page = cpu_buffer->commit_page;
4452 commit_ts = commit_page->page->time_stamp;
4453
4454 /*
4455 * When the writer goes across pages, it issues a cmpxchg which
4456 * is a mb(), which will synchronize with the rmb here.
4457 * (see rb_tail_page_update())
4458 */
4459 smp_rmb();
4460 commit = rb_page_commit(commit_page);
4461 /* We want to make sure that the commit page doesn't change */
4462 smp_rmb();
4463
4464 /* Make sure commit page didn't change */
4465 curr_commit_page = READ_ONCE(cpu_buffer->commit_page);
4466 curr_commit_ts = READ_ONCE(curr_commit_page->page->time_stamp);
4467
4468 /* If the commit page changed, then there's more data */
4469 if (curr_commit_page != commit_page ||
4470 curr_commit_ts != commit_ts)
4471 return 0;
4472
4473 /* Still racy, as it may return a false positive, but that's OK */
4474 return ((iter->head_page == commit_page && iter->head >= commit) ||
4475 (iter->head_page == reader && commit_page == head_page &&
4476 head_page->read == commit &&
4477 iter->head == rb_page_commit(cpu_buffer->reader_page)));
4478 }
4479 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
4480
4481 static void
rb_update_read_stamp(struct ring_buffer_per_cpu * cpu_buffer,struct ring_buffer_event * event)4482 rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
4483 struct ring_buffer_event *event)
4484 {
4485 u64 delta;
4486
4487 switch (event->type_len) {
4488 case RINGBUF_TYPE_PADDING:
4489 return;
4490
4491 case RINGBUF_TYPE_TIME_EXTEND:
4492 delta = rb_event_time_stamp(event);
4493 cpu_buffer->read_stamp += delta;
4494 return;
4495
4496 case RINGBUF_TYPE_TIME_STAMP:
4497 delta = rb_event_time_stamp(event);
4498 delta = rb_fix_abs_ts(delta, cpu_buffer->read_stamp);
4499 cpu_buffer->read_stamp = delta;
4500 return;
4501
4502 case RINGBUF_TYPE_DATA:
4503 cpu_buffer->read_stamp += event->time_delta;
4504 return;
4505
4506 default:
4507 RB_WARN_ON(cpu_buffer, 1);
4508 }
4509 return;
4510 }
4511
4512 static void
rb_update_iter_read_stamp(struct ring_buffer_iter * iter,struct ring_buffer_event * event)4513 rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
4514 struct ring_buffer_event *event)
4515 {
4516 u64 delta;
4517
4518 switch (event->type_len) {
4519 case RINGBUF_TYPE_PADDING:
4520 return;
4521
4522 case RINGBUF_TYPE_TIME_EXTEND:
4523 delta = rb_event_time_stamp(event);
4524 iter->read_stamp += delta;
4525 return;
4526
4527 case RINGBUF_TYPE_TIME_STAMP:
4528 delta = rb_event_time_stamp(event);
4529 delta = rb_fix_abs_ts(delta, iter->read_stamp);
4530 iter->read_stamp = delta;
4531 return;
4532
4533 case RINGBUF_TYPE_DATA:
4534 iter->read_stamp += event->time_delta;
4535 return;
4536
4537 default:
4538 RB_WARN_ON(iter->cpu_buffer, 1);
4539 }
4540 return;
4541 }
4542
4543 static struct buffer_page *
rb_get_reader_page(struct ring_buffer_per_cpu * cpu_buffer)4544 rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
4545 {
4546 struct buffer_page *reader = NULL;
4547 unsigned long overwrite;
4548 unsigned long flags;
4549 int nr_loops = 0;
4550 int ret;
4551
4552 local_irq_save(flags);
4553 arch_spin_lock(&cpu_buffer->lock);
4554
4555 again:
4556 /*
4557 * This should normally only loop twice. But because the
4558 * start of the reader inserts an empty page, it causes
4559 * a case where we will loop three times. There should be no
4560 * reason to loop four times (that I know of).
4561 */
4562 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
4563 reader = NULL;
4564 goto out;
4565 }
4566
4567 reader = cpu_buffer->reader_page;
4568
4569 /* If there's more to read, return this page */
4570 if (cpu_buffer->reader_page->read < rb_page_size(reader))
4571 goto out;
4572
4573 /* Never should we have an index greater than the size */
4574 if (RB_WARN_ON(cpu_buffer,
4575 cpu_buffer->reader_page->read > rb_page_size(reader)))
4576 goto out;
4577
4578 /* check if we caught up to the tail */
4579 reader = NULL;
4580 if (cpu_buffer->commit_page == cpu_buffer->reader_page)
4581 goto out;
4582
4583 /* Don't bother swapping if the ring buffer is empty */
4584 if (rb_num_of_entries(cpu_buffer) == 0)
4585 goto out;
4586
4587 /*
4588 * Reset the reader page to size zero.
4589 */
4590 local_set(&cpu_buffer->reader_page->write, 0);
4591 local_set(&cpu_buffer->reader_page->entries, 0);
4592 local_set(&cpu_buffer->reader_page->page->commit, 0);
4593 cpu_buffer->reader_page->real_end = 0;
4594
4595 spin:
4596 /*
4597 * Splice the empty reader page into the list around the head.
4598 */
4599 reader = rb_set_head_page(cpu_buffer);
4600 if (!reader)
4601 goto out;
4602 cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
4603 cpu_buffer->reader_page->list.prev = reader->list.prev;
4604
4605 /*
4606 * cpu_buffer->pages just needs to point to the buffer, it
4607 * has no specific buffer page to point to. Lets move it out
4608 * of our way so we don't accidentally swap it.
4609 */
4610 cpu_buffer->pages = reader->list.prev;
4611
4612 /* The reader page will be pointing to the new head */
4613 rb_set_list_to_head(&cpu_buffer->reader_page->list);
4614
4615 /*
4616 * We want to make sure we read the overruns after we set up our
4617 * pointers to the next object. The writer side does a
4618 * cmpxchg to cross pages which acts as the mb on the writer
4619 * side. Note, the reader will constantly fail the swap
4620 * while the writer is updating the pointers, so this
4621 * guarantees that the overwrite recorded here is the one we
4622 * want to compare with the last_overrun.
4623 */
4624 smp_mb();
4625 overwrite = local_read(&(cpu_buffer->overrun));
4626
4627 /*
4628 * Here's the tricky part.
4629 *
4630 * We need to move the pointer past the header page.
4631 * But we can only do that if a writer is not currently
4632 * moving it. The page before the header page has the
4633 * flag bit '1' set if it is pointing to the page we want.
4634 * but if the writer is in the process of moving it
4635 * than it will be '2' or already moved '0'.
4636 */
4637
4638 ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
4639
4640 /*
4641 * If we did not convert it, then we must try again.
4642 */
4643 if (!ret)
4644 goto spin;
4645
4646 /*
4647 * Yay! We succeeded in replacing the page.
4648 *
4649 * Now make the new head point back to the reader page.
4650 */
4651 rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
4652 rb_inc_page(&cpu_buffer->head_page);
4653
4654 local_inc(&cpu_buffer->pages_read);
4655
4656 /* Finally update the reader page to the new head */
4657 cpu_buffer->reader_page = reader;
4658 cpu_buffer->reader_page->read = 0;
4659
4660 if (overwrite != cpu_buffer->last_overrun) {
4661 cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
4662 cpu_buffer->last_overrun = overwrite;
4663 }
4664
4665 goto again;
4666
4667 out:
4668 /* Update the read_stamp on the first event */
4669 if (reader && reader->read == 0)
4670 cpu_buffer->read_stamp = reader->page->time_stamp;
4671
4672 arch_spin_unlock(&cpu_buffer->lock);
4673 local_irq_restore(flags);
4674
4675 /*
4676 * The writer has preempt disable, wait for it. But not forever
4677 * Although, 1 second is pretty much "forever"
4678 */
4679 #define USECS_WAIT 1000000
4680 for (nr_loops = 0; nr_loops < USECS_WAIT; nr_loops++) {
4681 /* If the write is past the end of page, a writer is still updating it */
4682 if (likely(!reader || rb_page_write(reader) <= BUF_PAGE_SIZE))
4683 break;
4684
4685 udelay(1);
4686
4687 /* Get the latest version of the reader write value */
4688 smp_rmb();
4689 }
4690
4691 /* The writer is not moving forward? Something is wrong */
4692 if (RB_WARN_ON(cpu_buffer, nr_loops == USECS_WAIT))
4693 reader = NULL;
4694
4695 /*
4696 * Make sure we see any padding after the write update
4697 * (see rb_reset_tail())
4698 */
4699 smp_rmb();
4700
4701
4702 return reader;
4703 }
4704
rb_advance_reader(struct ring_buffer_per_cpu * cpu_buffer)4705 static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
4706 {
4707 struct ring_buffer_event *event;
4708 struct buffer_page *reader;
4709 unsigned length;
4710
4711 reader = rb_get_reader_page(cpu_buffer);
4712
4713 /* This function should not be called when buffer is empty */
4714 if (RB_WARN_ON(cpu_buffer, !reader))
4715 return;
4716
4717 event = rb_reader_event(cpu_buffer);
4718
4719 if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
4720 cpu_buffer->read++;
4721
4722 rb_update_read_stamp(cpu_buffer, event);
4723
4724 length = rb_event_length(event);
4725 cpu_buffer->reader_page->read += length;
4726 }
4727
rb_advance_iter(struct ring_buffer_iter * iter)4728 static void rb_advance_iter(struct ring_buffer_iter *iter)
4729 {
4730 struct ring_buffer_per_cpu *cpu_buffer;
4731
4732 cpu_buffer = iter->cpu_buffer;
4733
4734 /* If head == next_event then we need to jump to the next event */
4735 if (iter->head == iter->next_event) {
4736 /* If the event gets overwritten again, there's nothing to do */
4737 if (rb_iter_head_event(iter) == NULL)
4738 return;
4739 }
4740
4741 iter->head = iter->next_event;
4742
4743 /*
4744 * Check if we are at the end of the buffer.
4745 */
4746 if (iter->next_event >= rb_page_size(iter->head_page)) {
4747 /* discarded commits can make the page empty */
4748 if (iter->head_page == cpu_buffer->commit_page)
4749 return;
4750 rb_inc_iter(iter);
4751 return;
4752 }
4753
4754 rb_update_iter_read_stamp(iter, iter->event);
4755 }
4756
rb_lost_events(struct ring_buffer_per_cpu * cpu_buffer)4757 static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
4758 {
4759 return cpu_buffer->lost_events;
4760 }
4761
4762 static struct ring_buffer_event *
rb_buffer_peek(struct ring_buffer_per_cpu * cpu_buffer,u64 * ts,unsigned long * lost_events)4763 rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
4764 unsigned long *lost_events)
4765 {
4766 struct ring_buffer_event *event;
4767 struct buffer_page *reader;
4768 int nr_loops = 0;
4769
4770 if (ts)
4771 *ts = 0;
4772 again:
4773 /*
4774 * We repeat when a time extend is encountered.
4775 * Since the time extend is always attached to a data event,
4776 * we should never loop more than once.
4777 * (We never hit the following condition more than twice).
4778 */
4779 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
4780 return NULL;
4781
4782 reader = rb_get_reader_page(cpu_buffer);
4783 if (!reader)
4784 return NULL;
4785
4786 event = rb_reader_event(cpu_buffer);
4787
4788 switch (event->type_len) {
4789 case RINGBUF_TYPE_PADDING:
4790 if (rb_null_event(event))
4791 RB_WARN_ON(cpu_buffer, 1);
4792 /*
4793 * Because the writer could be discarding every
4794 * event it creates (which would probably be bad)
4795 * if we were to go back to "again" then we may never
4796 * catch up, and will trigger the warn on, or lock
4797 * the box. Return the padding, and we will release
4798 * the current locks, and try again.
4799 */
4800 return event;
4801
4802 case RINGBUF_TYPE_TIME_EXTEND:
4803 /* Internal data, OK to advance */
4804 rb_advance_reader(cpu_buffer);
4805 goto again;
4806
4807 case RINGBUF_TYPE_TIME_STAMP:
4808 if (ts) {
4809 *ts = rb_event_time_stamp(event);
4810 *ts = rb_fix_abs_ts(*ts, reader->page->time_stamp);
4811 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
4812 cpu_buffer->cpu, ts);
4813 }
4814 /* Internal data, OK to advance */
4815 rb_advance_reader(cpu_buffer);
4816 goto again;
4817
4818 case RINGBUF_TYPE_DATA:
4819 if (ts && !(*ts)) {
4820 *ts = cpu_buffer->read_stamp + event->time_delta;
4821 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
4822 cpu_buffer->cpu, ts);
4823 }
4824 if (lost_events)
4825 *lost_events = rb_lost_events(cpu_buffer);
4826 return event;
4827
4828 default:
4829 RB_WARN_ON(cpu_buffer, 1);
4830 }
4831
4832 return NULL;
4833 }
4834 EXPORT_SYMBOL_GPL(ring_buffer_peek);
4835
4836 static struct ring_buffer_event *
rb_iter_peek(struct ring_buffer_iter * iter,u64 * ts)4837 rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
4838 {
4839 struct trace_buffer *buffer;
4840 struct ring_buffer_per_cpu *cpu_buffer;
4841 struct ring_buffer_event *event;
4842 int nr_loops = 0;
4843
4844 if (ts)
4845 *ts = 0;
4846
4847 cpu_buffer = iter->cpu_buffer;
4848 buffer = cpu_buffer->buffer;
4849
4850 /*
4851 * Check if someone performed a consuming read to
4852 * the buffer. A consuming read invalidates the iterator
4853 * and we need to reset the iterator in this case.
4854 */
4855 if (unlikely(iter->cache_read != cpu_buffer->read ||
4856 iter->cache_reader_page != cpu_buffer->reader_page))
4857 rb_iter_reset(iter);
4858
4859 again:
4860 if (ring_buffer_iter_empty(iter))
4861 return NULL;
4862
4863 /*
4864 * As the writer can mess with what the iterator is trying
4865 * to read, just give up if we fail to get an event after
4866 * three tries. The iterator is not as reliable when reading
4867 * the ring buffer with an active write as the consumer is.
4868 * Do not warn if the three failures is reached.
4869 */
4870 if (++nr_loops > 3)
4871 return NULL;
4872
4873 if (rb_per_cpu_empty(cpu_buffer))
4874 return NULL;
4875
4876 if (iter->head >= rb_page_size(iter->head_page)) {
4877 rb_inc_iter(iter);
4878 goto again;
4879 }
4880
4881 event = rb_iter_head_event(iter);
4882 if (!event)
4883 goto again;
4884
4885 switch (event->type_len) {
4886 case RINGBUF_TYPE_PADDING:
4887 if (rb_null_event(event)) {
4888 rb_inc_iter(iter);
4889 goto again;
4890 }
4891 rb_advance_iter(iter);
4892 return event;
4893
4894 case RINGBUF_TYPE_TIME_EXTEND:
4895 /* Internal data, OK to advance */
4896 rb_advance_iter(iter);
4897 goto again;
4898
4899 case RINGBUF_TYPE_TIME_STAMP:
4900 if (ts) {
4901 *ts = rb_event_time_stamp(event);
4902 *ts = rb_fix_abs_ts(*ts, iter->head_page->page->time_stamp);
4903 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
4904 cpu_buffer->cpu, ts);
4905 }
4906 /* Internal data, OK to advance */
4907 rb_advance_iter(iter);
4908 goto again;
4909
4910 case RINGBUF_TYPE_DATA:
4911 if (ts && !(*ts)) {
4912 *ts = iter->read_stamp + event->time_delta;
4913 ring_buffer_normalize_time_stamp(buffer,
4914 cpu_buffer->cpu, ts);
4915 }
4916 return event;
4917
4918 default:
4919 RB_WARN_ON(cpu_buffer, 1);
4920 }
4921
4922 return NULL;
4923 }
4924 EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
4925
rb_reader_lock(struct ring_buffer_per_cpu * cpu_buffer)4926 static inline bool rb_reader_lock(struct ring_buffer_per_cpu *cpu_buffer)
4927 {
4928 if (likely(!in_nmi())) {
4929 raw_spin_lock(&cpu_buffer->reader_lock);
4930 return true;
4931 }
4932
4933 /*
4934 * If an NMI die dumps out the content of the ring buffer
4935 * trylock must be used to prevent a deadlock if the NMI
4936 * preempted a task that holds the ring buffer locks. If
4937 * we get the lock then all is fine, if not, then continue
4938 * to do the read, but this can corrupt the ring buffer,
4939 * so it must be permanently disabled from future writes.
4940 * Reading from NMI is a oneshot deal.
4941 */
4942 if (raw_spin_trylock(&cpu_buffer->reader_lock))
4943 return true;
4944
4945 /* Continue without locking, but disable the ring buffer */
4946 atomic_inc(&cpu_buffer->record_disabled);
4947 return false;
4948 }
4949
4950 static inline void
rb_reader_unlock(struct ring_buffer_per_cpu * cpu_buffer,bool locked)4951 rb_reader_unlock(struct ring_buffer_per_cpu *cpu_buffer, bool locked)
4952 {
4953 if (likely(locked))
4954 raw_spin_unlock(&cpu_buffer->reader_lock);
4955 return;
4956 }
4957
4958 /**
4959 * ring_buffer_peek - peek at the next event to be read
4960 * @buffer: The ring buffer to read
4961 * @cpu: The cpu to peak at
4962 * @ts: The timestamp counter of this event.
4963 * @lost_events: a variable to store if events were lost (may be NULL)
4964 *
4965 * This will return the event that will be read next, but does
4966 * not consume the data.
4967 */
4968 struct ring_buffer_event *
ring_buffer_peek(struct trace_buffer * buffer,int cpu,u64 * ts,unsigned long * lost_events)4969 ring_buffer_peek(struct trace_buffer *buffer, int cpu, u64 *ts,
4970 unsigned long *lost_events)
4971 {
4972 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4973 struct ring_buffer_event *event;
4974 unsigned long flags;
4975 bool dolock;
4976
4977 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4978 return NULL;
4979
4980 again:
4981 local_irq_save(flags);
4982 dolock = rb_reader_lock(cpu_buffer);
4983 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
4984 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4985 rb_advance_reader(cpu_buffer);
4986 rb_reader_unlock(cpu_buffer, dolock);
4987 local_irq_restore(flags);
4988
4989 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4990 goto again;
4991
4992 return event;
4993 }
4994
4995 /** ring_buffer_iter_dropped - report if there are dropped events
4996 * @iter: The ring buffer iterator
4997 *
4998 * Returns true if there was dropped events since the last peek.
4999 */
ring_buffer_iter_dropped(struct ring_buffer_iter * iter)5000 bool ring_buffer_iter_dropped(struct ring_buffer_iter *iter)
5001 {
5002 bool ret = iter->missed_events != 0;
5003
5004 iter->missed_events = 0;
5005 return ret;
5006 }
5007 EXPORT_SYMBOL_GPL(ring_buffer_iter_dropped);
5008
5009 /**
5010 * ring_buffer_iter_peek - peek at the next event to be read
5011 * @iter: The ring buffer iterator
5012 * @ts: The timestamp counter of this event.
5013 *
5014 * This will return the event that will be read next, but does
5015 * not increment the iterator.
5016 */
5017 struct ring_buffer_event *
ring_buffer_iter_peek(struct ring_buffer_iter * iter,u64 * ts)5018 ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
5019 {
5020 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
5021 struct ring_buffer_event *event;
5022 unsigned long flags;
5023
5024 again:
5025 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5026 event = rb_iter_peek(iter, ts);
5027 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5028
5029 if (event && event->type_len == RINGBUF_TYPE_PADDING)
5030 goto again;
5031
5032 return event;
5033 }
5034
5035 /**
5036 * ring_buffer_consume - return an event and consume it
5037 * @buffer: The ring buffer to get the next event from
5038 * @cpu: the cpu to read the buffer from
5039 * @ts: a variable to store the timestamp (may be NULL)
5040 * @lost_events: a variable to store if events were lost (may be NULL)
5041 *
5042 * Returns the next event in the ring buffer, and that event is consumed.
5043 * Meaning, that sequential reads will keep returning a different event,
5044 * and eventually empty the ring buffer if the producer is slower.
5045 */
5046 struct ring_buffer_event *
ring_buffer_consume(struct trace_buffer * buffer,int cpu,u64 * ts,unsigned long * lost_events)5047 ring_buffer_consume(struct trace_buffer *buffer, int cpu, u64 *ts,
5048 unsigned long *lost_events)
5049 {
5050 struct ring_buffer_per_cpu *cpu_buffer;
5051 struct ring_buffer_event *event = NULL;
5052 unsigned long flags;
5053 bool dolock;
5054
5055 again:
5056 /* might be called in atomic */
5057 preempt_disable();
5058
5059 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5060 goto out;
5061
5062 cpu_buffer = buffer->buffers[cpu];
5063 local_irq_save(flags);
5064 dolock = rb_reader_lock(cpu_buffer);
5065
5066 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
5067 if (event) {
5068 cpu_buffer->lost_events = 0;
5069 rb_advance_reader(cpu_buffer);
5070 }
5071
5072 rb_reader_unlock(cpu_buffer, dolock);
5073 local_irq_restore(flags);
5074
5075 out:
5076 preempt_enable();
5077
5078 if (event && event->type_len == RINGBUF_TYPE_PADDING)
5079 goto again;
5080
5081 return event;
5082 }
5083 EXPORT_SYMBOL_GPL(ring_buffer_consume);
5084
5085 /**
5086 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
5087 * @buffer: The ring buffer to read from
5088 * @cpu: The cpu buffer to iterate over
5089 * @flags: gfp flags to use for memory allocation
5090 *
5091 * This performs the initial preparations necessary to iterate
5092 * through the buffer. Memory is allocated, buffer recording
5093 * is disabled, and the iterator pointer is returned to the caller.
5094 *
5095 * Disabling buffer recording prevents the reading from being
5096 * corrupted. This is not a consuming read, so a producer is not
5097 * expected.
5098 *
5099 * After a sequence of ring_buffer_read_prepare calls, the user is
5100 * expected to make at least one call to ring_buffer_read_prepare_sync.
5101 * Afterwards, ring_buffer_read_start is invoked to get things going
5102 * for real.
5103 *
5104 * This overall must be paired with ring_buffer_read_finish.
5105 */
5106 struct ring_buffer_iter *
ring_buffer_read_prepare(struct trace_buffer * buffer,int cpu,gfp_t flags)5107 ring_buffer_read_prepare(struct trace_buffer *buffer, int cpu, gfp_t flags)
5108 {
5109 struct ring_buffer_per_cpu *cpu_buffer;
5110 struct ring_buffer_iter *iter;
5111
5112 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5113 return NULL;
5114
5115 iter = kzalloc(sizeof(*iter), flags);
5116 if (!iter)
5117 return NULL;
5118
5119 iter->event = kmalloc(BUF_MAX_DATA_SIZE, flags);
5120 if (!iter->event) {
5121 kfree(iter);
5122 return NULL;
5123 }
5124
5125 cpu_buffer = buffer->buffers[cpu];
5126
5127 iter->cpu_buffer = cpu_buffer;
5128
5129 atomic_inc(&cpu_buffer->resize_disabled);
5130
5131 return iter;
5132 }
5133 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
5134
5135 /**
5136 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
5137 *
5138 * All previously invoked ring_buffer_read_prepare calls to prepare
5139 * iterators will be synchronized. Afterwards, read_buffer_read_start
5140 * calls on those iterators are allowed.
5141 */
5142 void
ring_buffer_read_prepare_sync(void)5143 ring_buffer_read_prepare_sync(void)
5144 {
5145 synchronize_rcu();
5146 }
5147 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
5148
5149 /**
5150 * ring_buffer_read_start - start a non consuming read of the buffer
5151 * @iter: The iterator returned by ring_buffer_read_prepare
5152 *
5153 * This finalizes the startup of an iteration through the buffer.
5154 * The iterator comes from a call to ring_buffer_read_prepare and
5155 * an intervening ring_buffer_read_prepare_sync must have been
5156 * performed.
5157 *
5158 * Must be paired with ring_buffer_read_finish.
5159 */
5160 void
ring_buffer_read_start(struct ring_buffer_iter * iter)5161 ring_buffer_read_start(struct ring_buffer_iter *iter)
5162 {
5163 struct ring_buffer_per_cpu *cpu_buffer;
5164 unsigned long flags;
5165
5166 if (!iter)
5167 return;
5168
5169 cpu_buffer = iter->cpu_buffer;
5170
5171 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5172 arch_spin_lock(&cpu_buffer->lock);
5173 rb_iter_reset(iter);
5174 arch_spin_unlock(&cpu_buffer->lock);
5175 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5176 }
5177 EXPORT_SYMBOL_GPL(ring_buffer_read_start);
5178
5179 /**
5180 * ring_buffer_read_finish - finish reading the iterator of the buffer
5181 * @iter: The iterator retrieved by ring_buffer_start
5182 *
5183 * This re-enables the recording to the buffer, and frees the
5184 * iterator.
5185 */
5186 void
ring_buffer_read_finish(struct ring_buffer_iter * iter)5187 ring_buffer_read_finish(struct ring_buffer_iter *iter)
5188 {
5189 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
5190 unsigned long flags;
5191
5192 /*
5193 * Ring buffer is disabled from recording, here's a good place
5194 * to check the integrity of the ring buffer.
5195 * Must prevent readers from trying to read, as the check
5196 * clears the HEAD page and readers require it.
5197 */
5198 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5199 rb_check_pages(cpu_buffer);
5200 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5201
5202 atomic_dec(&cpu_buffer->resize_disabled);
5203 kfree(iter->event);
5204 kfree(iter);
5205 }
5206 EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
5207
5208 /**
5209 * ring_buffer_iter_advance - advance the iterator to the next location
5210 * @iter: The ring buffer iterator
5211 *
5212 * Move the location of the iterator such that the next read will
5213 * be the next location of the iterator.
5214 */
ring_buffer_iter_advance(struct ring_buffer_iter * iter)5215 void ring_buffer_iter_advance(struct ring_buffer_iter *iter)
5216 {
5217 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
5218 unsigned long flags;
5219
5220 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5221
5222 rb_advance_iter(iter);
5223
5224 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5225 }
5226 EXPORT_SYMBOL_GPL(ring_buffer_iter_advance);
5227
5228 /**
5229 * ring_buffer_size - return the size of the ring buffer (in bytes)
5230 * @buffer: The ring buffer.
5231 * @cpu: The CPU to get ring buffer size from.
5232 */
ring_buffer_size(struct trace_buffer * buffer,int cpu)5233 unsigned long ring_buffer_size(struct trace_buffer *buffer, int cpu)
5234 {
5235 /*
5236 * Earlier, this method returned
5237 * BUF_PAGE_SIZE * buffer->nr_pages
5238 * Since the nr_pages field is now removed, we have converted this to
5239 * return the per cpu buffer value.
5240 */
5241 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5242 return 0;
5243
5244 return BUF_PAGE_SIZE * buffer->buffers[cpu]->nr_pages;
5245 }
5246 EXPORT_SYMBOL_GPL(ring_buffer_size);
5247
5248 static void
rb_reset_cpu(struct ring_buffer_per_cpu * cpu_buffer)5249 rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
5250 {
5251 rb_head_page_deactivate(cpu_buffer);
5252
5253 cpu_buffer->head_page
5254 = list_entry(cpu_buffer->pages, struct buffer_page, list);
5255 local_set(&cpu_buffer->head_page->write, 0);
5256 local_set(&cpu_buffer->head_page->entries, 0);
5257 local_set(&cpu_buffer->head_page->page->commit, 0);
5258
5259 cpu_buffer->head_page->read = 0;
5260
5261 cpu_buffer->tail_page = cpu_buffer->head_page;
5262 cpu_buffer->commit_page = cpu_buffer->head_page;
5263
5264 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
5265 INIT_LIST_HEAD(&cpu_buffer->new_pages);
5266 local_set(&cpu_buffer->reader_page->write, 0);
5267 local_set(&cpu_buffer->reader_page->entries, 0);
5268 local_set(&cpu_buffer->reader_page->page->commit, 0);
5269 cpu_buffer->reader_page->read = 0;
5270
5271 local_set(&cpu_buffer->entries_bytes, 0);
5272 local_set(&cpu_buffer->overrun, 0);
5273 local_set(&cpu_buffer->commit_overrun, 0);
5274 local_set(&cpu_buffer->dropped_events, 0);
5275 local_set(&cpu_buffer->entries, 0);
5276 local_set(&cpu_buffer->committing, 0);
5277 local_set(&cpu_buffer->commits, 0);
5278 local_set(&cpu_buffer->pages_touched, 0);
5279 local_set(&cpu_buffer->pages_lost, 0);
5280 local_set(&cpu_buffer->pages_read, 0);
5281 cpu_buffer->last_pages_touch = 0;
5282 cpu_buffer->shortest_full = 0;
5283 cpu_buffer->read = 0;
5284 cpu_buffer->read_bytes = 0;
5285
5286 rb_time_set(&cpu_buffer->write_stamp, 0);
5287 rb_time_set(&cpu_buffer->before_stamp, 0);
5288
5289 memset(cpu_buffer->event_stamp, 0, sizeof(cpu_buffer->event_stamp));
5290
5291 cpu_buffer->lost_events = 0;
5292 cpu_buffer->last_overrun = 0;
5293
5294 rb_head_page_activate(cpu_buffer);
5295 }
5296
5297 /* Must have disabled the cpu buffer then done a synchronize_rcu */
reset_disabled_cpu_buffer(struct ring_buffer_per_cpu * cpu_buffer)5298 static void reset_disabled_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
5299 {
5300 unsigned long flags;
5301
5302 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5303
5304 if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
5305 goto out;
5306
5307 arch_spin_lock(&cpu_buffer->lock);
5308
5309 rb_reset_cpu(cpu_buffer);
5310
5311 arch_spin_unlock(&cpu_buffer->lock);
5312
5313 out:
5314 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5315 }
5316
5317 /**
5318 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
5319 * @buffer: The ring buffer to reset a per cpu buffer of
5320 * @cpu: The CPU buffer to be reset
5321 */
ring_buffer_reset_cpu(struct trace_buffer * buffer,int cpu)5322 void ring_buffer_reset_cpu(struct trace_buffer *buffer, int cpu)
5323 {
5324 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
5325
5326 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5327 return;
5328
5329 /* prevent another thread from changing buffer sizes */
5330 mutex_lock(&buffer->mutex);
5331
5332 atomic_inc(&cpu_buffer->resize_disabled);
5333 atomic_inc(&cpu_buffer->record_disabled);
5334
5335 /* Make sure all commits have finished */
5336 synchronize_rcu();
5337
5338 reset_disabled_cpu_buffer(cpu_buffer);
5339
5340 atomic_dec(&cpu_buffer->record_disabled);
5341 atomic_dec(&cpu_buffer->resize_disabled);
5342
5343 mutex_unlock(&buffer->mutex);
5344 }
5345 EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
5346
5347 /**
5348 * ring_buffer_reset_online_cpus - reset a ring buffer per CPU buffer
5349 * @buffer: The ring buffer to reset a per cpu buffer of
5350 * @cpu: The CPU buffer to be reset
5351 */
ring_buffer_reset_online_cpus(struct trace_buffer * buffer)5352 void ring_buffer_reset_online_cpus(struct trace_buffer *buffer)
5353 {
5354 struct ring_buffer_per_cpu *cpu_buffer;
5355 int cpu;
5356
5357 /* prevent another thread from changing buffer sizes */
5358 mutex_lock(&buffer->mutex);
5359
5360 for_each_online_buffer_cpu(buffer, cpu) {
5361 cpu_buffer = buffer->buffers[cpu];
5362
5363 atomic_inc(&cpu_buffer->resize_disabled);
5364 atomic_inc(&cpu_buffer->record_disabled);
5365 }
5366
5367 /* Make sure all commits have finished */
5368 synchronize_rcu();
5369
5370 for_each_online_buffer_cpu(buffer, cpu) {
5371 cpu_buffer = buffer->buffers[cpu];
5372
5373 reset_disabled_cpu_buffer(cpu_buffer);
5374
5375 atomic_dec(&cpu_buffer->record_disabled);
5376 atomic_dec(&cpu_buffer->resize_disabled);
5377 }
5378
5379 mutex_unlock(&buffer->mutex);
5380 }
5381
5382 /**
5383 * ring_buffer_reset - reset a ring buffer
5384 * @buffer: The ring buffer to reset all cpu buffers
5385 */
ring_buffer_reset(struct trace_buffer * buffer)5386 void ring_buffer_reset(struct trace_buffer *buffer)
5387 {
5388 struct ring_buffer_per_cpu *cpu_buffer;
5389 int cpu;
5390
5391 /* prevent another thread from changing buffer sizes */
5392 mutex_lock(&buffer->mutex);
5393
5394 for_each_buffer_cpu(buffer, cpu) {
5395 cpu_buffer = buffer->buffers[cpu];
5396
5397 atomic_inc(&cpu_buffer->resize_disabled);
5398 atomic_inc(&cpu_buffer->record_disabled);
5399 }
5400
5401 /* Make sure all commits have finished */
5402 synchronize_rcu();
5403
5404 for_each_buffer_cpu(buffer, cpu) {
5405 cpu_buffer = buffer->buffers[cpu];
5406
5407 reset_disabled_cpu_buffer(cpu_buffer);
5408
5409 atomic_dec(&cpu_buffer->record_disabled);
5410 atomic_dec(&cpu_buffer->resize_disabled);
5411 }
5412
5413 mutex_unlock(&buffer->mutex);
5414 }
5415 EXPORT_SYMBOL_GPL(ring_buffer_reset);
5416
5417 /**
5418 * ring_buffer_empty - is the ring buffer empty?
5419 * @buffer: The ring buffer to test
5420 */
ring_buffer_empty(struct trace_buffer * buffer)5421 bool ring_buffer_empty(struct trace_buffer *buffer)
5422 {
5423 struct ring_buffer_per_cpu *cpu_buffer;
5424 unsigned long flags;
5425 bool dolock;
5426 int cpu;
5427 int ret;
5428
5429 /* yes this is racy, but if you don't like the race, lock the buffer */
5430 for_each_buffer_cpu(buffer, cpu) {
5431 cpu_buffer = buffer->buffers[cpu];
5432 local_irq_save(flags);
5433 dolock = rb_reader_lock(cpu_buffer);
5434 ret = rb_per_cpu_empty(cpu_buffer);
5435 rb_reader_unlock(cpu_buffer, dolock);
5436 local_irq_restore(flags);
5437
5438 if (!ret)
5439 return false;
5440 }
5441
5442 return true;
5443 }
5444 EXPORT_SYMBOL_GPL(ring_buffer_empty);
5445
5446 /**
5447 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
5448 * @buffer: The ring buffer
5449 * @cpu: The CPU buffer to test
5450 */
ring_buffer_empty_cpu(struct trace_buffer * buffer,int cpu)5451 bool ring_buffer_empty_cpu(struct trace_buffer *buffer, int cpu)
5452 {
5453 struct ring_buffer_per_cpu *cpu_buffer;
5454 unsigned long flags;
5455 bool dolock;
5456 int ret;
5457
5458 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5459 return true;
5460
5461 cpu_buffer = buffer->buffers[cpu];
5462 local_irq_save(flags);
5463 dolock = rb_reader_lock(cpu_buffer);
5464 ret = rb_per_cpu_empty(cpu_buffer);
5465 rb_reader_unlock(cpu_buffer, dolock);
5466 local_irq_restore(flags);
5467
5468 return ret;
5469 }
5470 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
5471
5472 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
5473 /**
5474 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
5475 * @buffer_a: One buffer to swap with
5476 * @buffer_b: The other buffer to swap with
5477 * @cpu: the CPU of the buffers to swap
5478 *
5479 * This function is useful for tracers that want to take a "snapshot"
5480 * of a CPU buffer and has another back up buffer lying around.
5481 * it is expected that the tracer handles the cpu buffer not being
5482 * used at the moment.
5483 */
ring_buffer_swap_cpu(struct trace_buffer * buffer_a,struct trace_buffer * buffer_b,int cpu)5484 int ring_buffer_swap_cpu(struct trace_buffer *buffer_a,
5485 struct trace_buffer *buffer_b, int cpu)
5486 {
5487 struct ring_buffer_per_cpu *cpu_buffer_a;
5488 struct ring_buffer_per_cpu *cpu_buffer_b;
5489 int ret = -EINVAL;
5490
5491 if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
5492 !cpumask_test_cpu(cpu, buffer_b->cpumask))
5493 goto out;
5494
5495 cpu_buffer_a = buffer_a->buffers[cpu];
5496 cpu_buffer_b = buffer_b->buffers[cpu];
5497
5498 /* At least make sure the two buffers are somewhat the same */
5499 if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
5500 goto out;
5501
5502 ret = -EAGAIN;
5503
5504 if (atomic_read(&buffer_a->record_disabled))
5505 goto out;
5506
5507 if (atomic_read(&buffer_b->record_disabled))
5508 goto out;
5509
5510 if (atomic_read(&cpu_buffer_a->record_disabled))
5511 goto out;
5512
5513 if (atomic_read(&cpu_buffer_b->record_disabled))
5514 goto out;
5515
5516 /*
5517 * We can't do a synchronize_rcu here because this
5518 * function can be called in atomic context.
5519 * Normally this will be called from the same CPU as cpu.
5520 * If not it's up to the caller to protect this.
5521 */
5522 atomic_inc(&cpu_buffer_a->record_disabled);
5523 atomic_inc(&cpu_buffer_b->record_disabled);
5524
5525 ret = -EBUSY;
5526 if (local_read(&cpu_buffer_a->committing))
5527 goto out_dec;
5528 if (local_read(&cpu_buffer_b->committing))
5529 goto out_dec;
5530
5531 buffer_a->buffers[cpu] = cpu_buffer_b;
5532 buffer_b->buffers[cpu] = cpu_buffer_a;
5533
5534 cpu_buffer_b->buffer = buffer_a;
5535 cpu_buffer_a->buffer = buffer_b;
5536
5537 ret = 0;
5538
5539 out_dec:
5540 atomic_dec(&cpu_buffer_a->record_disabled);
5541 atomic_dec(&cpu_buffer_b->record_disabled);
5542 out:
5543 return ret;
5544 }
5545 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
5546 #endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
5547
5548 /**
5549 * ring_buffer_alloc_read_page - allocate a page to read from buffer
5550 * @buffer: the buffer to allocate for.
5551 * @cpu: the cpu buffer to allocate.
5552 *
5553 * This function is used in conjunction with ring_buffer_read_page.
5554 * When reading a full page from the ring buffer, these functions
5555 * can be used to speed up the process. The calling function should
5556 * allocate a few pages first with this function. Then when it
5557 * needs to get pages from the ring buffer, it passes the result
5558 * of this function into ring_buffer_read_page, which will swap
5559 * the page that was allocated, with the read page of the buffer.
5560 *
5561 * Returns:
5562 * The page allocated, or ERR_PTR
5563 */
ring_buffer_alloc_read_page(struct trace_buffer * buffer,int cpu)5564 void *ring_buffer_alloc_read_page(struct trace_buffer *buffer, int cpu)
5565 {
5566 struct ring_buffer_per_cpu *cpu_buffer;
5567 struct buffer_data_page *bpage = NULL;
5568 unsigned long flags;
5569 struct page *page;
5570
5571 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5572 return ERR_PTR(-ENODEV);
5573
5574 cpu_buffer = buffer->buffers[cpu];
5575 local_irq_save(flags);
5576 arch_spin_lock(&cpu_buffer->lock);
5577
5578 if (cpu_buffer->free_page) {
5579 bpage = cpu_buffer->free_page;
5580 cpu_buffer->free_page = NULL;
5581 }
5582
5583 arch_spin_unlock(&cpu_buffer->lock);
5584 local_irq_restore(flags);
5585
5586 if (bpage)
5587 goto out;
5588
5589 page = alloc_pages_node(cpu_to_node(cpu),
5590 GFP_KERNEL | __GFP_NORETRY, 0);
5591 if (!page)
5592 return ERR_PTR(-ENOMEM);
5593
5594 bpage = page_address(page);
5595
5596 out:
5597 rb_init_page(bpage);
5598
5599 return bpage;
5600 }
5601 EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
5602
5603 /**
5604 * ring_buffer_free_read_page - free an allocated read page
5605 * @buffer: the buffer the page was allocate for
5606 * @cpu: the cpu buffer the page came from
5607 * @data: the page to free
5608 *
5609 * Free a page allocated from ring_buffer_alloc_read_page.
5610 */
ring_buffer_free_read_page(struct trace_buffer * buffer,int cpu,void * data)5611 void ring_buffer_free_read_page(struct trace_buffer *buffer, int cpu, void *data)
5612 {
5613 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
5614 struct buffer_data_page *bpage = data;
5615 struct page *page = virt_to_page(bpage);
5616 unsigned long flags;
5617
5618 /* If the page is still in use someplace else, we can't reuse it */
5619 if (page_ref_count(page) > 1)
5620 goto out;
5621
5622 local_irq_save(flags);
5623 arch_spin_lock(&cpu_buffer->lock);
5624
5625 if (!cpu_buffer->free_page) {
5626 cpu_buffer->free_page = bpage;
5627 bpage = NULL;
5628 }
5629
5630 arch_spin_unlock(&cpu_buffer->lock);
5631 local_irq_restore(flags);
5632
5633 out:
5634 free_page((unsigned long)bpage);
5635 }
5636 EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
5637
5638 /**
5639 * ring_buffer_read_page - extract a page from the ring buffer
5640 * @buffer: buffer to extract from
5641 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
5642 * @len: amount to extract
5643 * @cpu: the cpu of the buffer to extract
5644 * @full: should the extraction only happen when the page is full.
5645 *
5646 * This function will pull out a page from the ring buffer and consume it.
5647 * @data_page must be the address of the variable that was returned
5648 * from ring_buffer_alloc_read_page. This is because the page might be used
5649 * to swap with a page in the ring buffer.
5650 *
5651 * for example:
5652 * rpage = ring_buffer_alloc_read_page(buffer, cpu);
5653 * if (IS_ERR(rpage))
5654 * return PTR_ERR(rpage);
5655 * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
5656 * if (ret >= 0)
5657 * process_page(rpage, ret);
5658 *
5659 * When @full is set, the function will not return true unless
5660 * the writer is off the reader page.
5661 *
5662 * Note: it is up to the calling functions to handle sleeps and wakeups.
5663 * The ring buffer can be used anywhere in the kernel and can not
5664 * blindly call wake_up. The layer that uses the ring buffer must be
5665 * responsible for that.
5666 *
5667 * Returns:
5668 * >=0 if data has been transferred, returns the offset of consumed data.
5669 * <0 if no data has been transferred.
5670 */
ring_buffer_read_page(struct trace_buffer * buffer,void ** data_page,size_t len,int cpu,int full)5671 int ring_buffer_read_page(struct trace_buffer *buffer,
5672 void **data_page, size_t len, int cpu, int full)
5673 {
5674 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
5675 struct ring_buffer_event *event;
5676 struct buffer_data_page *bpage;
5677 struct buffer_page *reader;
5678 unsigned long missed_events;
5679 unsigned long flags;
5680 unsigned int commit;
5681 unsigned int read;
5682 u64 save_timestamp;
5683 int ret = -1;
5684
5685 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5686 goto out;
5687
5688 /*
5689 * If len is not big enough to hold the page header, then
5690 * we can not copy anything.
5691 */
5692 if (len <= BUF_PAGE_HDR_SIZE)
5693 goto out;
5694
5695 len -= BUF_PAGE_HDR_SIZE;
5696
5697 if (!data_page)
5698 goto out;
5699
5700 bpage = *data_page;
5701 if (!bpage)
5702 goto out;
5703
5704 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5705
5706 reader = rb_get_reader_page(cpu_buffer);
5707 if (!reader)
5708 goto out_unlock;
5709
5710 event = rb_reader_event(cpu_buffer);
5711
5712 read = reader->read;
5713 commit = rb_page_commit(reader);
5714
5715 /* Check if any events were dropped */
5716 missed_events = cpu_buffer->lost_events;
5717
5718 /*
5719 * If this page has been partially read or
5720 * if len is not big enough to read the rest of the page or
5721 * a writer is still on the page, then
5722 * we must copy the data from the page to the buffer.
5723 * Otherwise, we can simply swap the page with the one passed in.
5724 */
5725 if (read || (len < (commit - read)) ||
5726 cpu_buffer->reader_page == cpu_buffer->commit_page) {
5727 struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
5728 unsigned int rpos = read;
5729 unsigned int pos = 0;
5730 unsigned int size;
5731
5732 /*
5733 * If a full page is expected, this can still be returned
5734 * if there's been a previous partial read and the
5735 * rest of the page can be read and the commit page is off
5736 * the reader page.
5737 */
5738 if (full &&
5739 (!read || (len < (commit - read)) ||
5740 cpu_buffer->reader_page == cpu_buffer->commit_page))
5741 goto out_unlock;
5742
5743 if (len > (commit - read))
5744 len = (commit - read);
5745
5746 /* Always keep the time extend and data together */
5747 size = rb_event_ts_length(event);
5748
5749 if (len < size)
5750 goto out_unlock;
5751
5752 /* save the current timestamp, since the user will need it */
5753 save_timestamp = cpu_buffer->read_stamp;
5754
5755 /* Need to copy one event at a time */
5756 do {
5757 /* We need the size of one event, because
5758 * rb_advance_reader only advances by one event,
5759 * whereas rb_event_ts_length may include the size of
5760 * one or two events.
5761 * We have already ensured there's enough space if this
5762 * is a time extend. */
5763 size = rb_event_length(event);
5764 memcpy(bpage->data + pos, rpage->data + rpos, size);
5765
5766 len -= size;
5767
5768 rb_advance_reader(cpu_buffer);
5769 rpos = reader->read;
5770 pos += size;
5771
5772 if (rpos >= commit)
5773 break;
5774
5775 event = rb_reader_event(cpu_buffer);
5776 /* Always keep the time extend and data together */
5777 size = rb_event_ts_length(event);
5778 } while (len >= size);
5779
5780 /* update bpage */
5781 local_set(&bpage->commit, pos);
5782 bpage->time_stamp = save_timestamp;
5783
5784 /* we copied everything to the beginning */
5785 read = 0;
5786 } else {
5787 /* update the entry counter */
5788 cpu_buffer->read += rb_page_entries(reader);
5789 cpu_buffer->read_bytes += BUF_PAGE_SIZE;
5790
5791 /* swap the pages */
5792 rb_init_page(bpage);
5793 bpage = reader->page;
5794 reader->page = *data_page;
5795 local_set(&reader->write, 0);
5796 local_set(&reader->entries, 0);
5797 reader->read = 0;
5798 *data_page = bpage;
5799
5800 /*
5801 * Use the real_end for the data size,
5802 * This gives us a chance to store the lost events
5803 * on the page.
5804 */
5805 if (reader->real_end)
5806 local_set(&bpage->commit, reader->real_end);
5807 }
5808 ret = read;
5809
5810 cpu_buffer->lost_events = 0;
5811
5812 commit = local_read(&bpage->commit);
5813 /*
5814 * Set a flag in the commit field if we lost events
5815 */
5816 if (missed_events) {
5817 /* If there is room at the end of the page to save the
5818 * missed events, then record it there.
5819 */
5820 if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
5821 memcpy(&bpage->data[commit], &missed_events,
5822 sizeof(missed_events));
5823 local_add(RB_MISSED_STORED, &bpage->commit);
5824 commit += sizeof(missed_events);
5825 }
5826 local_add(RB_MISSED_EVENTS, &bpage->commit);
5827 }
5828
5829 /*
5830 * This page may be off to user land. Zero it out here.
5831 */
5832 if (commit < BUF_PAGE_SIZE)
5833 memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
5834
5835 out_unlock:
5836 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5837
5838 out:
5839 return ret;
5840 }
5841 EXPORT_SYMBOL_GPL(ring_buffer_read_page);
5842
5843 /*
5844 * We only allocate new buffers, never free them if the CPU goes down.
5845 * If we were to free the buffer, then the user would lose any trace that was in
5846 * the buffer.
5847 */
trace_rb_cpu_prepare(unsigned int cpu,struct hlist_node * node)5848 int trace_rb_cpu_prepare(unsigned int cpu, struct hlist_node *node)
5849 {
5850 struct trace_buffer *buffer;
5851 long nr_pages_same;
5852 int cpu_i;
5853 unsigned long nr_pages;
5854
5855 buffer = container_of(node, struct trace_buffer, node);
5856 if (cpumask_test_cpu(cpu, buffer->cpumask))
5857 return 0;
5858
5859 nr_pages = 0;
5860 nr_pages_same = 1;
5861 /* check if all cpu sizes are same */
5862 for_each_buffer_cpu(buffer, cpu_i) {
5863 /* fill in the size from first enabled cpu */
5864 if (nr_pages == 0)
5865 nr_pages = buffer->buffers[cpu_i]->nr_pages;
5866 if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
5867 nr_pages_same = 0;
5868 break;
5869 }
5870 }
5871 /* allocate minimum pages, user can later expand it */
5872 if (!nr_pages_same)
5873 nr_pages = 2;
5874 buffer->buffers[cpu] =
5875 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
5876 if (!buffer->buffers[cpu]) {
5877 WARN(1, "failed to allocate ring buffer on CPU %u\n",
5878 cpu);
5879 return -ENOMEM;
5880 }
5881 smp_wmb();
5882 cpumask_set_cpu(cpu, buffer->cpumask);
5883 return 0;
5884 }
5885
5886 #ifdef CONFIG_RING_BUFFER_STARTUP_TEST
5887 /*
5888 * This is a basic integrity check of the ring buffer.
5889 * Late in the boot cycle this test will run when configured in.
5890 * It will kick off a thread per CPU that will go into a loop
5891 * writing to the per cpu ring buffer various sizes of data.
5892 * Some of the data will be large items, some small.
5893 *
5894 * Another thread is created that goes into a spin, sending out
5895 * IPIs to the other CPUs to also write into the ring buffer.
5896 * this is to test the nesting ability of the buffer.
5897 *
5898 * Basic stats are recorded and reported. If something in the
5899 * ring buffer should happen that's not expected, a big warning
5900 * is displayed and all ring buffers are disabled.
5901 */
5902 static struct task_struct *rb_threads[NR_CPUS] __initdata;
5903
5904 struct rb_test_data {
5905 struct trace_buffer *buffer;
5906 unsigned long events;
5907 unsigned long bytes_written;
5908 unsigned long bytes_alloc;
5909 unsigned long bytes_dropped;
5910 unsigned long events_nested;
5911 unsigned long bytes_written_nested;
5912 unsigned long bytes_alloc_nested;
5913 unsigned long bytes_dropped_nested;
5914 int min_size_nested;
5915 int max_size_nested;
5916 int max_size;
5917 int min_size;
5918 int cpu;
5919 int cnt;
5920 };
5921
5922 static struct rb_test_data rb_data[NR_CPUS] __initdata;
5923
5924 /* 1 meg per cpu */
5925 #define RB_TEST_BUFFER_SIZE 1048576
5926
5927 static char rb_string[] __initdata =
5928 "abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
5929 "?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
5930 "!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
5931
5932 static bool rb_test_started __initdata;
5933
5934 struct rb_item {
5935 int size;
5936 char str[];
5937 };
5938
rb_write_something(struct rb_test_data * data,bool nested)5939 static __init int rb_write_something(struct rb_test_data *data, bool nested)
5940 {
5941 struct ring_buffer_event *event;
5942 struct rb_item *item;
5943 bool started;
5944 int event_len;
5945 int size;
5946 int len;
5947 int cnt;
5948
5949 /* Have nested writes different that what is written */
5950 cnt = data->cnt + (nested ? 27 : 0);
5951
5952 /* Multiply cnt by ~e, to make some unique increment */
5953 size = (cnt * 68 / 25) % (sizeof(rb_string) - 1);
5954
5955 len = size + sizeof(struct rb_item);
5956
5957 started = rb_test_started;
5958 /* read rb_test_started before checking buffer enabled */
5959 smp_rmb();
5960
5961 event = ring_buffer_lock_reserve(data->buffer, len);
5962 if (!event) {
5963 /* Ignore dropped events before test starts. */
5964 if (started) {
5965 if (nested)
5966 data->bytes_dropped += len;
5967 else
5968 data->bytes_dropped_nested += len;
5969 }
5970 return len;
5971 }
5972
5973 event_len = ring_buffer_event_length(event);
5974
5975 if (RB_WARN_ON(data->buffer, event_len < len))
5976 goto out;
5977
5978 item = ring_buffer_event_data(event);
5979 item->size = size;
5980 memcpy(item->str, rb_string, size);
5981
5982 if (nested) {
5983 data->bytes_alloc_nested += event_len;
5984 data->bytes_written_nested += len;
5985 data->events_nested++;
5986 if (!data->min_size_nested || len < data->min_size_nested)
5987 data->min_size_nested = len;
5988 if (len > data->max_size_nested)
5989 data->max_size_nested = len;
5990 } else {
5991 data->bytes_alloc += event_len;
5992 data->bytes_written += len;
5993 data->events++;
5994 if (!data->min_size || len < data->min_size)
5995 data->max_size = len;
5996 if (len > data->max_size)
5997 data->max_size = len;
5998 }
5999
6000 out:
6001 ring_buffer_unlock_commit(data->buffer, event);
6002
6003 return 0;
6004 }
6005
rb_test(void * arg)6006 static __init int rb_test(void *arg)
6007 {
6008 struct rb_test_data *data = arg;
6009
6010 while (!kthread_should_stop()) {
6011 rb_write_something(data, false);
6012 data->cnt++;
6013
6014 set_current_state(TASK_INTERRUPTIBLE);
6015 /* Now sleep between a min of 100-300us and a max of 1ms */
6016 usleep_range(((data->cnt % 3) + 1) * 100, 1000);
6017 }
6018
6019 return 0;
6020 }
6021
rb_ipi(void * ignore)6022 static __init void rb_ipi(void *ignore)
6023 {
6024 struct rb_test_data *data;
6025 int cpu = smp_processor_id();
6026
6027 data = &rb_data[cpu];
6028 rb_write_something(data, true);
6029 }
6030
rb_hammer_test(void * arg)6031 static __init int rb_hammer_test(void *arg)
6032 {
6033 while (!kthread_should_stop()) {
6034
6035 /* Send an IPI to all cpus to write data! */
6036 smp_call_function(rb_ipi, NULL, 1);
6037 /* No sleep, but for non preempt, let others run */
6038 schedule();
6039 }
6040
6041 return 0;
6042 }
6043
test_ringbuffer(void)6044 static __init int test_ringbuffer(void)
6045 {
6046 struct task_struct *rb_hammer;
6047 struct trace_buffer *buffer;
6048 int cpu;
6049 int ret = 0;
6050
6051 if (security_locked_down(LOCKDOWN_TRACEFS)) {
6052 pr_warn("Lockdown is enabled, skipping ring buffer tests\n");
6053 return 0;
6054 }
6055
6056 pr_info("Running ring buffer tests...\n");
6057
6058 buffer = ring_buffer_alloc(RB_TEST_BUFFER_SIZE, RB_FL_OVERWRITE);
6059 if (WARN_ON(!buffer))
6060 return 0;
6061
6062 /* Disable buffer so that threads can't write to it yet */
6063 ring_buffer_record_off(buffer);
6064
6065 for_each_online_cpu(cpu) {
6066 rb_data[cpu].buffer = buffer;
6067 rb_data[cpu].cpu = cpu;
6068 rb_data[cpu].cnt = cpu;
6069 rb_threads[cpu] = kthread_run_on_cpu(rb_test, &rb_data[cpu],
6070 cpu, "rbtester/%u");
6071 if (WARN_ON(IS_ERR(rb_threads[cpu]))) {
6072 pr_cont("FAILED\n");
6073 ret = PTR_ERR(rb_threads[cpu]);
6074 goto out_free;
6075 }
6076 }
6077
6078 /* Now create the rb hammer! */
6079 rb_hammer = kthread_run(rb_hammer_test, NULL, "rbhammer");
6080 if (WARN_ON(IS_ERR(rb_hammer))) {
6081 pr_cont("FAILED\n");
6082 ret = PTR_ERR(rb_hammer);
6083 goto out_free;
6084 }
6085
6086 ring_buffer_record_on(buffer);
6087 /*
6088 * Show buffer is enabled before setting rb_test_started.
6089 * Yes there's a small race window where events could be
6090 * dropped and the thread wont catch it. But when a ring
6091 * buffer gets enabled, there will always be some kind of
6092 * delay before other CPUs see it. Thus, we don't care about
6093 * those dropped events. We care about events dropped after
6094 * the threads see that the buffer is active.
6095 */
6096 smp_wmb();
6097 rb_test_started = true;
6098
6099 set_current_state(TASK_INTERRUPTIBLE);
6100 /* Just run for 10 seconds */;
6101 schedule_timeout(10 * HZ);
6102
6103 kthread_stop(rb_hammer);
6104
6105 out_free:
6106 for_each_online_cpu(cpu) {
6107 if (!rb_threads[cpu])
6108 break;
6109 kthread_stop(rb_threads[cpu]);
6110 }
6111 if (ret) {
6112 ring_buffer_free(buffer);
6113 return ret;
6114 }
6115
6116 /* Report! */
6117 pr_info("finished\n");
6118 for_each_online_cpu(cpu) {
6119 struct ring_buffer_event *event;
6120 struct rb_test_data *data = &rb_data[cpu];
6121 struct rb_item *item;
6122 unsigned long total_events;
6123 unsigned long total_dropped;
6124 unsigned long total_written;
6125 unsigned long total_alloc;
6126 unsigned long total_read = 0;
6127 unsigned long total_size = 0;
6128 unsigned long total_len = 0;
6129 unsigned long total_lost = 0;
6130 unsigned long lost;
6131 int big_event_size;
6132 int small_event_size;
6133
6134 ret = -1;
6135
6136 total_events = data->events + data->events_nested;
6137 total_written = data->bytes_written + data->bytes_written_nested;
6138 total_alloc = data->bytes_alloc + data->bytes_alloc_nested;
6139 total_dropped = data->bytes_dropped + data->bytes_dropped_nested;
6140
6141 big_event_size = data->max_size + data->max_size_nested;
6142 small_event_size = data->min_size + data->min_size_nested;
6143
6144 pr_info("CPU %d:\n", cpu);
6145 pr_info(" events: %ld\n", total_events);
6146 pr_info(" dropped bytes: %ld\n", total_dropped);
6147 pr_info(" alloced bytes: %ld\n", total_alloc);
6148 pr_info(" written bytes: %ld\n", total_written);
6149 pr_info(" biggest event: %d\n", big_event_size);
6150 pr_info(" smallest event: %d\n", small_event_size);
6151
6152 if (RB_WARN_ON(buffer, total_dropped))
6153 break;
6154
6155 ret = 0;
6156
6157 while ((event = ring_buffer_consume(buffer, cpu, NULL, &lost))) {
6158 total_lost += lost;
6159 item = ring_buffer_event_data(event);
6160 total_len += ring_buffer_event_length(event);
6161 total_size += item->size + sizeof(struct rb_item);
6162 if (memcmp(&item->str[0], rb_string, item->size) != 0) {
6163 pr_info("FAILED!\n");
6164 pr_info("buffer had: %.*s\n", item->size, item->str);
6165 pr_info("expected: %.*s\n", item->size, rb_string);
6166 RB_WARN_ON(buffer, 1);
6167 ret = -1;
6168 break;
6169 }
6170 total_read++;
6171 }
6172 if (ret)
6173 break;
6174
6175 ret = -1;
6176
6177 pr_info(" read events: %ld\n", total_read);
6178 pr_info(" lost events: %ld\n", total_lost);
6179 pr_info(" total events: %ld\n", total_lost + total_read);
6180 pr_info(" recorded len bytes: %ld\n", total_len);
6181 pr_info(" recorded size bytes: %ld\n", total_size);
6182 if (total_lost) {
6183 pr_info(" With dropped events, record len and size may not match\n"
6184 " alloced and written from above\n");
6185 } else {
6186 if (RB_WARN_ON(buffer, total_len != total_alloc ||
6187 total_size != total_written))
6188 break;
6189 }
6190 if (RB_WARN_ON(buffer, total_lost + total_read != total_events))
6191 break;
6192
6193 ret = 0;
6194 }
6195 if (!ret)
6196 pr_info("Ring buffer PASSED!\n");
6197
6198 ring_buffer_free(buffer);
6199 return 0;
6200 }
6201
6202 late_initcall(test_ringbuffer);
6203 #endif /* CONFIG_RING_BUFFER_STARTUP_TEST */
6204