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