Loading...
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
32static void update_pages_handler(struct work_struct *work);
33
34/*
35 * The ring buffer header is special. We must manually up keep it.
36 */
37int ring_buffer_print_entry_header(struct trace_seq *s)
38{
39 trace_seq_puts(s, "# compressed entry header\n");
40 trace_seq_puts(s, "\ttype_len : 5 bits\n");
41 trace_seq_puts(s, "\ttime_delta : 27 bits\n");
42 trace_seq_puts(s, "\tarray : 32 bits\n");
43 trace_seq_putc(s, '\n');
44 trace_seq_printf(s, "\tpadding : type == %d\n",
45 RINGBUF_TYPE_PADDING);
46 trace_seq_printf(s, "\ttime_extend : type == %d\n",
47 RINGBUF_TYPE_TIME_EXTEND);
48 trace_seq_printf(s, "\ttime_stamp : type == %d\n",
49 RINGBUF_TYPE_TIME_STAMP);
50 trace_seq_printf(s, "\tdata max type_len == %d\n",
51 RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
52
53 return !trace_seq_has_overflowed(s);
54}
55
56/*
57 * The ring buffer is made up of a list of pages. A separate list of pages is
58 * allocated for each CPU. A writer may only write to a buffer that is
59 * associated with the CPU it is currently executing on. A reader may read
60 * from any per cpu buffer.
61 *
62 * The reader is special. For each per cpu buffer, the reader has its own
63 * reader page. When a reader has read the entire reader page, this reader
64 * page is swapped with another page in the ring buffer.
65 *
66 * Now, as long as the writer is off the reader page, the reader can do what
67 * ever it wants with that page. The writer will never write to that page
68 * again (as long as it is out of the ring buffer).
69 *
70 * Here's some silly ASCII art.
71 *
72 * +------+
73 * |reader| RING BUFFER
74 * |page |
75 * +------+ +---+ +---+ +---+
76 * | |-->| |-->| |
77 * +---+ +---+ +---+
78 * ^ |
79 * | |
80 * +---------------+
81 *
82 *
83 * +------+
84 * |reader| RING BUFFER
85 * |page |------------------v
86 * +------+ +---+ +---+ +---+
87 * | |-->| |-->| |
88 * +---+ +---+ +---+
89 * ^ |
90 * | |
91 * +---------------+
92 *
93 *
94 * +------+
95 * |reader| RING BUFFER
96 * |page |------------------v
97 * +------+ +---+ +---+ +---+
98 * ^ | |-->| |-->| |
99 * | +---+ +---+ +---+
100 * | |
101 * | |
102 * +------------------------------+
103 *
104 *
105 * +------+
106 * |buffer| RING BUFFER
107 * |page |------------------v
108 * +------+ +---+ +---+ +---+
109 * ^ | | | |-->| |
110 * | New +---+ +---+ +---+
111 * | Reader------^ |
112 * | page |
113 * +------------------------------+
114 *
115 *
116 * After we make this swap, the reader can hand this page off to the splice
117 * code and be done with it. It can even allocate a new page if it needs to
118 * and swap that into the ring buffer.
119 *
120 * We will be using cmpxchg soon to make all this lockless.
121 *
122 */
123
124/* Used for individual buffers (after the counter) */
125#define RB_BUFFER_OFF (1 << 20)
126
127#define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
128
129#define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
130#define RB_ALIGNMENT 4U
131#define RB_MAX_SMALL_DATA (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
132#define RB_EVNT_MIN_SIZE 8U /* two 32bit words */
133
134#ifndef CONFIG_HAVE_64BIT_ALIGNED_ACCESS
135# define RB_FORCE_8BYTE_ALIGNMENT 0
136# define RB_ARCH_ALIGNMENT RB_ALIGNMENT
137#else
138# define RB_FORCE_8BYTE_ALIGNMENT 1
139# define RB_ARCH_ALIGNMENT 8U
140#endif
141
142#define RB_ALIGN_DATA __aligned(RB_ARCH_ALIGNMENT)
143
144/* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
145#define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
146
147enum {
148 RB_LEN_TIME_EXTEND = 8,
149 RB_LEN_TIME_STAMP = 8,
150};
151
152#define skip_time_extend(event) \
153 ((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
154
155#define extended_time(event) \
156 (event->type_len >= RINGBUF_TYPE_TIME_EXTEND)
157
158static inline int rb_null_event(struct ring_buffer_event *event)
159{
160 return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
161}
162
163static void rb_event_set_padding(struct ring_buffer_event *event)
164{
165 /* padding has a NULL time_delta */
166 event->type_len = RINGBUF_TYPE_PADDING;
167 event->time_delta = 0;
168}
169
170static unsigned
171rb_event_data_length(struct ring_buffer_event *event)
172{
173 unsigned length;
174
175 if (event->type_len)
176 length = event->type_len * RB_ALIGNMENT;
177 else
178 length = event->array[0];
179 return length + RB_EVNT_HDR_SIZE;
180}
181
182/*
183 * Return the length of the given event. Will return
184 * the length of the time extend if the event is a
185 * time extend.
186 */
187static inline unsigned
188rb_event_length(struct ring_buffer_event *event)
189{
190 switch (event->type_len) {
191 case RINGBUF_TYPE_PADDING:
192 if (rb_null_event(event))
193 /* undefined */
194 return -1;
195 return event->array[0] + RB_EVNT_HDR_SIZE;
196
197 case RINGBUF_TYPE_TIME_EXTEND:
198 return RB_LEN_TIME_EXTEND;
199
200 case RINGBUF_TYPE_TIME_STAMP:
201 return RB_LEN_TIME_STAMP;
202
203 case RINGBUF_TYPE_DATA:
204 return rb_event_data_length(event);
205 default:
206 WARN_ON_ONCE(1);
207 }
208 /* not hit */
209 return 0;
210}
211
212/*
213 * Return total length of time extend and data,
214 * or just the event length for all other events.
215 */
216static inline unsigned
217rb_event_ts_length(struct ring_buffer_event *event)
218{
219 unsigned len = 0;
220
221 if (extended_time(event)) {
222 /* time extends include the data event after it */
223 len = RB_LEN_TIME_EXTEND;
224 event = skip_time_extend(event);
225 }
226 return len + rb_event_length(event);
227}
228
229/**
230 * ring_buffer_event_length - return the length of the event
231 * @event: the event to get the length of
232 *
233 * Returns the size of the data load of a data event.
234 * If the event is something other than a data event, it
235 * returns the size of the event itself. With the exception
236 * of a TIME EXTEND, where it still returns the size of the
237 * data load of the data event after it.
238 */
239unsigned ring_buffer_event_length(struct ring_buffer_event *event)
240{
241 unsigned length;
242
243 if (extended_time(event))
244 event = skip_time_extend(event);
245
246 length = rb_event_length(event);
247 if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
248 return length;
249 length -= RB_EVNT_HDR_SIZE;
250 if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
251 length -= sizeof(event->array[0]);
252 return length;
253}
254EXPORT_SYMBOL_GPL(ring_buffer_event_length);
255
256/* inline for ring buffer fast paths */
257static __always_inline void *
258rb_event_data(struct ring_buffer_event *event)
259{
260 if (extended_time(event))
261 event = skip_time_extend(event);
262 WARN_ON_ONCE(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
263 /* If length is in len field, then array[0] has the data */
264 if (event->type_len)
265 return (void *)&event->array[0];
266 /* Otherwise length is in array[0] and array[1] has the data */
267 return (void *)&event->array[1];
268}
269
270/**
271 * ring_buffer_event_data - return the data of the event
272 * @event: the event to get the data from
273 */
274void *ring_buffer_event_data(struct ring_buffer_event *event)
275{
276 return rb_event_data(event);
277}
278EXPORT_SYMBOL_GPL(ring_buffer_event_data);
279
280#define for_each_buffer_cpu(buffer, cpu) \
281 for_each_cpu(cpu, buffer->cpumask)
282
283#define for_each_online_buffer_cpu(buffer, cpu) \
284 for_each_cpu_and(cpu, buffer->cpumask, cpu_online_mask)
285
286#define TS_SHIFT 27
287#define TS_MASK ((1ULL << TS_SHIFT) - 1)
288#define TS_DELTA_TEST (~TS_MASK)
289
290static u64 rb_event_time_stamp(struct ring_buffer_event *event)
291{
292 u64 ts;
293
294 ts = event->array[0];
295 ts <<= TS_SHIFT;
296 ts += event->time_delta;
297
298 return ts;
299}
300
301/* Flag when events were overwritten */
302#define RB_MISSED_EVENTS (1 << 31)
303/* Missed count stored at end */
304#define RB_MISSED_STORED (1 << 30)
305
306struct buffer_data_page {
307 u64 time_stamp; /* page time stamp */
308 local_t commit; /* write committed index */
309 unsigned char data[] RB_ALIGN_DATA; /* data of buffer page */
310};
311
312/*
313 * Note, the buffer_page list must be first. The buffer pages
314 * are allocated in cache lines, which means that each buffer
315 * page will be at the beginning of a cache line, and thus
316 * the least significant bits will be zero. We use this to
317 * add flags in the list struct pointers, to make the ring buffer
318 * lockless.
319 */
320struct buffer_page {
321 struct list_head list; /* list of buffer pages */
322 local_t write; /* index for next write */
323 unsigned read; /* index for next read */
324 local_t entries; /* entries on this page */
325 unsigned long real_end; /* real end of data */
326 struct buffer_data_page *page; /* Actual data page */
327};
328
329/*
330 * The buffer page counters, write and entries, must be reset
331 * atomically when crossing page boundaries. To synchronize this
332 * update, two counters are inserted into the number. One is
333 * the actual counter for the write position or count on the page.
334 *
335 * The other is a counter of updaters. Before an update happens
336 * the update partition of the counter is incremented. This will
337 * allow the updater to update the counter atomically.
338 *
339 * The counter is 20 bits, and the state data is 12.
340 */
341#define RB_WRITE_MASK 0xfffff
342#define RB_WRITE_INTCNT (1 << 20)
343
344static void rb_init_page(struct buffer_data_page *bpage)
345{
346 local_set(&bpage->commit, 0);
347}
348
349/*
350 * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
351 * this issue out.
352 */
353static void free_buffer_page(struct buffer_page *bpage)
354{
355 free_page((unsigned long)bpage->page);
356 kfree(bpage);
357}
358
359/*
360 * We need to fit the time_stamp delta into 27 bits.
361 */
362static inline int test_time_stamp(u64 delta)
363{
364 if (delta & TS_DELTA_TEST)
365 return 1;
366 return 0;
367}
368
369#define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
370
371/* Max payload is BUF_PAGE_SIZE - header (8bytes) */
372#define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
373
374int ring_buffer_print_page_header(struct trace_seq *s)
375{
376 struct buffer_data_page field;
377
378 trace_seq_printf(s, "\tfield: u64 timestamp;\t"
379 "offset:0;\tsize:%u;\tsigned:%u;\n",
380 (unsigned int)sizeof(field.time_stamp),
381 (unsigned int)is_signed_type(u64));
382
383 trace_seq_printf(s, "\tfield: local_t commit;\t"
384 "offset:%u;\tsize:%u;\tsigned:%u;\n",
385 (unsigned int)offsetof(typeof(field), commit),
386 (unsigned int)sizeof(field.commit),
387 (unsigned int)is_signed_type(long));
388
389 trace_seq_printf(s, "\tfield: int overwrite;\t"
390 "offset:%u;\tsize:%u;\tsigned:%u;\n",
391 (unsigned int)offsetof(typeof(field), commit),
392 1,
393 (unsigned int)is_signed_type(long));
394
395 trace_seq_printf(s, "\tfield: char data;\t"
396 "offset:%u;\tsize:%u;\tsigned:%u;\n",
397 (unsigned int)offsetof(typeof(field), data),
398 (unsigned int)BUF_PAGE_SIZE,
399 (unsigned int)is_signed_type(char));
400
401 return !trace_seq_has_overflowed(s);
402}
403
404struct rb_irq_work {
405 struct irq_work work;
406 wait_queue_head_t waiters;
407 wait_queue_head_t full_waiters;
408 bool waiters_pending;
409 bool full_waiters_pending;
410 bool wakeup_full;
411};
412
413/*
414 * Structure to hold event state and handle nested events.
415 */
416struct rb_event_info {
417 u64 ts;
418 u64 delta;
419 u64 before;
420 u64 after;
421 unsigned long length;
422 struct buffer_page *tail_page;
423 int add_timestamp;
424};
425
426/*
427 * Used for the add_timestamp
428 * NONE
429 * EXTEND - wants a time extend
430 * ABSOLUTE - the buffer requests all events to have absolute time stamps
431 * FORCE - force a full time stamp.
432 */
433enum {
434 RB_ADD_STAMP_NONE = 0,
435 RB_ADD_STAMP_EXTEND = BIT(1),
436 RB_ADD_STAMP_ABSOLUTE = BIT(2),
437 RB_ADD_STAMP_FORCE = BIT(3)
438};
439/*
440 * Used for which event context the event is in.
441 * TRANSITION = 0
442 * NMI = 1
443 * IRQ = 2
444 * SOFTIRQ = 3
445 * NORMAL = 4
446 *
447 * See trace_recursive_lock() comment below for more details.
448 */
449enum {
450 RB_CTX_TRANSITION,
451 RB_CTX_NMI,
452 RB_CTX_IRQ,
453 RB_CTX_SOFTIRQ,
454 RB_CTX_NORMAL,
455 RB_CTX_MAX
456};
457
458#if BITS_PER_LONG == 32
459#define RB_TIME_32
460#endif
461
462/* To test on 64 bit machines */
463//#define RB_TIME_32
464
465#ifdef RB_TIME_32
466
467struct rb_time_struct {
468 local_t cnt;
469 local_t top;
470 local_t bottom;
471};
472#else
473#include <asm/local64.h>
474struct rb_time_struct {
475 local64_t time;
476};
477#endif
478typedef struct rb_time_struct rb_time_t;
479
480#define MAX_NEST 5
481
482/*
483 * head_page == tail_page && head == tail then buffer is empty.
484 */
485struct ring_buffer_per_cpu {
486 int cpu;
487 atomic_t record_disabled;
488 atomic_t resize_disabled;
489 struct trace_buffer *buffer;
490 raw_spinlock_t reader_lock; /* serialize readers */
491 arch_spinlock_t lock;
492 struct lock_class_key lock_key;
493 struct buffer_data_page *free_page;
494 unsigned long nr_pages;
495 unsigned int current_context;
496 struct list_head *pages;
497 struct buffer_page *head_page; /* read from head */
498 struct buffer_page *tail_page; /* write to tail */
499 struct buffer_page *commit_page; /* committed pages */
500 struct buffer_page *reader_page;
501 unsigned long lost_events;
502 unsigned long last_overrun;
503 unsigned long nest;
504 local_t entries_bytes;
505 local_t entries;
506 local_t overrun;
507 local_t commit_overrun;
508 local_t dropped_events;
509 local_t committing;
510 local_t commits;
511 local_t pages_touched;
512 local_t pages_read;
513 long last_pages_touch;
514 size_t shortest_full;
515 unsigned long read;
516 unsigned long read_bytes;
517 rb_time_t write_stamp;
518 rb_time_t before_stamp;
519 u64 event_stamp[MAX_NEST];
520 u64 read_stamp;
521 /* ring buffer pages to update, > 0 to add, < 0 to remove */
522 long nr_pages_to_update;
523 struct list_head new_pages; /* new pages to add */
524 struct work_struct update_pages_work;
525 struct completion update_done;
526
527 struct rb_irq_work irq_work;
528};
529
530struct trace_buffer {
531 unsigned flags;
532 int cpus;
533 atomic_t record_disabled;
534 cpumask_var_t cpumask;
535
536 struct lock_class_key *reader_lock_key;
537
538 struct mutex mutex;
539
540 struct ring_buffer_per_cpu **buffers;
541
542 struct hlist_node node;
543 u64 (*clock)(void);
544
545 struct rb_irq_work irq_work;
546 bool time_stamp_abs;
547};
548
549struct ring_buffer_iter {
550 struct ring_buffer_per_cpu *cpu_buffer;
551 unsigned long head;
552 unsigned long next_event;
553 struct buffer_page *head_page;
554 struct buffer_page *cache_reader_page;
555 unsigned long cache_read;
556 u64 read_stamp;
557 u64 page_stamp;
558 struct ring_buffer_event *event;
559 int missed_events;
560};
561
562#ifdef RB_TIME_32
563
564/*
565 * On 32 bit machines, local64_t is very expensive. As the ring
566 * buffer doesn't need all the features of a true 64 bit atomic,
567 * on 32 bit, it uses these functions (64 still uses local64_t).
568 *
569 * For the ring buffer, 64 bit required operations for the time is
570 * the following:
571 *
572 * - Only need 59 bits (uses 60 to make it even).
573 * - Reads may fail if it interrupted a modification of the time stamp.
574 * It will succeed if it did not interrupt another write even if
575 * the read itself is interrupted by a write.
576 * It returns whether it was successful or not.
577 *
578 * - Writes always succeed and will overwrite other writes and writes
579 * that were done by events interrupting the current write.
580 *
581 * - A write followed by a read of the same time stamp will always succeed,
582 * but may not contain the same value.
583 *
584 * - A cmpxchg will fail if it interrupted another write or cmpxchg.
585 * Other than that, it acts like a normal cmpxchg.
586 *
587 * The 60 bit time stamp is broken up by 30 bits in a top and bottom half
588 * (bottom being the least significant 30 bits of the 60 bit time stamp).
589 *
590 * The two most significant bits of each half holds a 2 bit counter (0-3).
591 * Each update will increment this counter by one.
592 * When reading the top and bottom, if the two counter bits match then the
593 * top and bottom together make a valid 60 bit number.
594 */
595#define RB_TIME_SHIFT 30
596#define RB_TIME_VAL_MASK ((1 << RB_TIME_SHIFT) - 1)
597
598static inline int rb_time_cnt(unsigned long val)
599{
600 return (val >> RB_TIME_SHIFT) & 3;
601}
602
603static inline u64 rb_time_val(unsigned long top, unsigned long bottom)
604{
605 u64 val;
606
607 val = top & RB_TIME_VAL_MASK;
608 val <<= RB_TIME_SHIFT;
609 val |= bottom & RB_TIME_VAL_MASK;
610
611 return val;
612}
613
614static inline bool __rb_time_read(rb_time_t *t, u64 *ret, unsigned long *cnt)
615{
616 unsigned long top, bottom;
617 unsigned long c;
618
619 /*
620 * If the read is interrupted by a write, then the cnt will
621 * be different. Loop until both top and bottom have been read
622 * without interruption.
623 */
624 do {
625 c = local_read(&t->cnt);
626 top = local_read(&t->top);
627 bottom = local_read(&t->bottom);
628 } while (c != local_read(&t->cnt));
629
630 *cnt = rb_time_cnt(top);
631
632 /* If top and bottom counts don't match, this interrupted a write */
633 if (*cnt != rb_time_cnt(bottom))
634 return false;
635
636 *ret = rb_time_val(top, bottom);
637 return true;
638}
639
640static bool rb_time_read(rb_time_t *t, u64 *ret)
641{
642 unsigned long cnt;
643
644 return __rb_time_read(t, ret, &cnt);
645}
646
647static inline unsigned long rb_time_val_cnt(unsigned long val, unsigned long cnt)
648{
649 return (val & RB_TIME_VAL_MASK) | ((cnt & 3) << RB_TIME_SHIFT);
650}
651
652static inline void rb_time_split(u64 val, unsigned long *top, unsigned long *bottom)
653{
654 *top = (unsigned long)((val >> RB_TIME_SHIFT) & RB_TIME_VAL_MASK);
655 *bottom = (unsigned long)(val & RB_TIME_VAL_MASK);
656}
657
658static inline void rb_time_val_set(local_t *t, unsigned long val, unsigned long cnt)
659{
660 val = rb_time_val_cnt(val, cnt);
661 local_set(t, val);
662}
663
664static void rb_time_set(rb_time_t *t, u64 val)
665{
666 unsigned long cnt, top, bottom;
667
668 rb_time_split(val, &top, &bottom);
669
670 /* Writes always succeed with a valid number even if it gets interrupted. */
671 do {
672 cnt = local_inc_return(&t->cnt);
673 rb_time_val_set(&t->top, top, cnt);
674 rb_time_val_set(&t->bottom, bottom, cnt);
675 } while (cnt != local_read(&t->cnt));
676}
677
678static inline bool
679rb_time_read_cmpxchg(local_t *l, unsigned long expect, unsigned long set)
680{
681 unsigned long ret;
682
683 ret = local_cmpxchg(l, expect, set);
684 return ret == expect;
685}
686
687static int rb_time_cmpxchg(rb_time_t *t, u64 expect, u64 set)
688{
689 unsigned long cnt, top, bottom;
690 unsigned long cnt2, top2, bottom2;
691 u64 val;
692
693 /* The cmpxchg always fails if it interrupted an update */
694 if (!__rb_time_read(t, &val, &cnt2))
695 return false;
696
697 if (val != expect)
698 return false;
699
700 cnt = local_read(&t->cnt);
701 if ((cnt & 3) != cnt2)
702 return false;
703
704 cnt2 = cnt + 1;
705
706 rb_time_split(val, &top, &bottom);
707 top = rb_time_val_cnt(top, cnt);
708 bottom = rb_time_val_cnt(bottom, cnt);
709
710 rb_time_split(set, &top2, &bottom2);
711 top2 = rb_time_val_cnt(top2, cnt2);
712 bottom2 = rb_time_val_cnt(bottom2, cnt2);
713
714 if (!rb_time_read_cmpxchg(&t->cnt, cnt, cnt2))
715 return false;
716 if (!rb_time_read_cmpxchg(&t->top, top, top2))
717 return false;
718 if (!rb_time_read_cmpxchg(&t->bottom, bottom, bottom2))
719 return false;
720 return true;
721}
722
723#else /* 64 bits */
724
725/* local64_t always succeeds */
726
727static inline bool rb_time_read(rb_time_t *t, u64 *ret)
728{
729 *ret = local64_read(&t->time);
730 return true;
731}
732static void rb_time_set(rb_time_t *t, u64 val)
733{
734 local64_set(&t->time, val);
735}
736
737static bool rb_time_cmpxchg(rb_time_t *t, u64 expect, u64 set)
738{
739 u64 val;
740 val = local64_cmpxchg(&t->time, expect, set);
741 return val == expect;
742}
743#endif
744
745/*
746 * Enable this to make sure that the event passed to
747 * ring_buffer_event_time_stamp() is not committed and also
748 * is on the buffer that it passed in.
749 */
750//#define RB_VERIFY_EVENT
751#ifdef RB_VERIFY_EVENT
752static struct list_head *rb_list_head(struct list_head *list);
753static void verify_event(struct ring_buffer_per_cpu *cpu_buffer,
754 void *event)
755{
756 struct buffer_page *page = cpu_buffer->commit_page;
757 struct buffer_page *tail_page = READ_ONCE(cpu_buffer->tail_page);
758 struct list_head *next;
759 long commit, write;
760 unsigned long addr = (unsigned long)event;
761 bool done = false;
762 int stop = 0;
763
764 /* Make sure the event exists and is not committed yet */
765 do {
766 if (page == tail_page || WARN_ON_ONCE(stop++ > 100))
767 done = true;
768 commit = local_read(&page->page->commit);
769 write = local_read(&page->write);
770 if (addr >= (unsigned long)&page->page->data[commit] &&
771 addr < (unsigned long)&page->page->data[write])
772 return;
773
774 next = rb_list_head(page->list.next);
775 page = list_entry(next, struct buffer_page, list);
776 } while (!done);
777 WARN_ON_ONCE(1);
778}
779#else
780static inline void verify_event(struct ring_buffer_per_cpu *cpu_buffer,
781 void *event)
782{
783}
784#endif
785
786
787static inline u64 rb_time_stamp(struct trace_buffer *buffer);
788
789/**
790 * ring_buffer_event_time_stamp - return the event's current time stamp
791 * @buffer: The buffer that the event is on
792 * @event: the event to get the time stamp of
793 *
794 * Note, this must be called after @event is reserved, and before it is
795 * committed to the ring buffer. And must be called from the same
796 * context where the event was reserved (normal, softirq, irq, etc).
797 *
798 * Returns the time stamp associated with the current event.
799 * If the event has an extended time stamp, then that is used as
800 * the time stamp to return.
801 * In the highly unlikely case that the event was nested more than
802 * the max nesting, then the write_stamp of the buffer is returned,
803 * otherwise current time is returned, but that really neither of
804 * the last two cases should ever happen.
805 */
806u64 ring_buffer_event_time_stamp(struct trace_buffer *buffer,
807 struct ring_buffer_event *event)
808{
809 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[smp_processor_id()];
810 unsigned int nest;
811 u64 ts;
812
813 /* If the event includes an absolute time, then just use that */
814 if (event->type_len == RINGBUF_TYPE_TIME_STAMP)
815 return rb_event_time_stamp(event);
816
817 nest = local_read(&cpu_buffer->committing);
818 verify_event(cpu_buffer, event);
819 if (WARN_ON_ONCE(!nest))
820 goto fail;
821
822 /* Read the current saved nesting level time stamp */
823 if (likely(--nest < MAX_NEST))
824 return cpu_buffer->event_stamp[nest];
825
826 /* Shouldn't happen, warn if it does */
827 WARN_ONCE(1, "nest (%d) greater than max", nest);
828
829 fail:
830 /* Can only fail on 32 bit */
831 if (!rb_time_read(&cpu_buffer->write_stamp, &ts))
832 /* Screw it, just read the current time */
833 ts = rb_time_stamp(cpu_buffer->buffer);
834
835 return ts;
836}
837
838/**
839 * ring_buffer_nr_pages - get the number of buffer pages in the ring buffer
840 * @buffer: The ring_buffer to get the number of pages from
841 * @cpu: The cpu of the ring_buffer to get the number of pages from
842 *
843 * Returns the number of pages used by a per_cpu buffer of the ring buffer.
844 */
845size_t ring_buffer_nr_pages(struct trace_buffer *buffer, int cpu)
846{
847 return buffer->buffers[cpu]->nr_pages;
848}
849
850/**
851 * ring_buffer_nr_pages_dirty - get the number of used pages in the ring buffer
852 * @buffer: The ring_buffer to get the number of pages from
853 * @cpu: The cpu of the ring_buffer to get the number of pages from
854 *
855 * Returns the number of pages that have content in the ring buffer.
856 */
857size_t ring_buffer_nr_dirty_pages(struct trace_buffer *buffer, int cpu)
858{
859 size_t read;
860 size_t cnt;
861
862 read = local_read(&buffer->buffers[cpu]->pages_read);
863 cnt = local_read(&buffer->buffers[cpu]->pages_touched);
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
873/*
874 * rb_wake_up_waiters - wake up tasks waiting for ring buffer input
875 *
876 * Schedules a delayed work to wake up any task that is blocked on the
877 * ring buffer waiters queue.
878 */
879static void rb_wake_up_waiters(struct irq_work *work)
880{
881 struct rb_irq_work *rbwork = container_of(work, struct rb_irq_work, work);
882
883 wake_up_all(&rbwork->waiters);
884 if (rbwork->wakeup_full) {
885 rbwork->wakeup_full = false;
886 wake_up_all(&rbwork->full_waiters);
887 }
888}
889
890/**
891 * ring_buffer_wait - wait for input to the ring buffer
892 * @buffer: buffer to wait on
893 * @cpu: the cpu buffer to wait on
894 * @full: wait until the percentage of pages are available, if @cpu != RING_BUFFER_ALL_CPUS
895 *
896 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
897 * as data is added to any of the @buffer's cpu buffers. Otherwise
898 * it will wait for data to be added to a specific cpu buffer.
899 */
900int ring_buffer_wait(struct trace_buffer *buffer, int cpu, int full)
901{
902 struct ring_buffer_per_cpu *cpu_buffer;
903 DEFINE_WAIT(wait);
904 struct rb_irq_work *work;
905 int ret = 0;
906
907 /*
908 * Depending on what the caller is waiting for, either any
909 * data in any cpu buffer, or a specific buffer, put the
910 * caller on the appropriate wait queue.
911 */
912 if (cpu == RING_BUFFER_ALL_CPUS) {
913 work = &buffer->irq_work;
914 /* Full only makes sense on per cpu reads */
915 full = 0;
916 } else {
917 if (!cpumask_test_cpu(cpu, buffer->cpumask))
918 return -ENODEV;
919 cpu_buffer = buffer->buffers[cpu];
920 work = &cpu_buffer->irq_work;
921 }
922
923
924 while (true) {
925 if (full)
926 prepare_to_wait(&work->full_waiters, &wait, TASK_INTERRUPTIBLE);
927 else
928 prepare_to_wait(&work->waiters, &wait, TASK_INTERRUPTIBLE);
929
930 /*
931 * The events can happen in critical sections where
932 * checking a work queue can cause deadlocks.
933 * After adding a task to the queue, this flag is set
934 * only to notify events to try to wake up the queue
935 * using irq_work.
936 *
937 * We don't clear it even if the buffer is no longer
938 * empty. The flag only causes the next event to run
939 * irq_work to do the work queue wake up. The worse
940 * that can happen if we race with !trace_empty() is that
941 * an event will cause an irq_work to try to wake up
942 * an empty queue.
943 *
944 * There's no reason to protect this flag either, as
945 * the work queue and irq_work logic will do the necessary
946 * synchronization for the wake ups. The only thing
947 * that is necessary is that the wake up happens after
948 * a task has been queued. It's OK for spurious wake ups.
949 */
950 if (full)
951 work->full_waiters_pending = true;
952 else
953 work->waiters_pending = true;
954
955 if (signal_pending(current)) {
956 ret = -EINTR;
957 break;
958 }
959
960 if (cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer))
961 break;
962
963 if (cpu != RING_BUFFER_ALL_CPUS &&
964 !ring_buffer_empty_cpu(buffer, cpu)) {
965 unsigned long flags;
966 bool pagebusy;
967 size_t nr_pages;
968 size_t dirty;
969
970 if (!full)
971 break;
972
973 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
974 pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
975 nr_pages = cpu_buffer->nr_pages;
976 dirty = ring_buffer_nr_dirty_pages(buffer, cpu);
977 if (!cpu_buffer->shortest_full ||
978 cpu_buffer->shortest_full < full)
979 cpu_buffer->shortest_full = full;
980 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
981 if (!pagebusy &&
982 (!nr_pages || (dirty * 100) > full * nr_pages))
983 break;
984 }
985
986 schedule();
987 }
988
989 if (full)
990 finish_wait(&work->full_waiters, &wait);
991 else
992 finish_wait(&work->waiters, &wait);
993
994 return ret;
995}
996
997/**
998 * ring_buffer_poll_wait - poll on buffer input
999 * @buffer: buffer to wait on
1000 * @cpu: the cpu buffer to wait on
1001 * @filp: the file descriptor
1002 * @poll_table: The poll descriptor
1003 *
1004 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
1005 * as data is added to any of the @buffer's cpu buffers. Otherwise
1006 * it will wait for data to be added to a specific cpu buffer.
1007 *
1008 * Returns EPOLLIN | EPOLLRDNORM if data exists in the buffers,
1009 * zero otherwise.
1010 */
1011__poll_t ring_buffer_poll_wait(struct trace_buffer *buffer, int cpu,
1012 struct file *filp, poll_table *poll_table)
1013{
1014 struct ring_buffer_per_cpu *cpu_buffer;
1015 struct rb_irq_work *work;
1016
1017 if (cpu == RING_BUFFER_ALL_CPUS)
1018 work = &buffer->irq_work;
1019 else {
1020 if (!cpumask_test_cpu(cpu, buffer->cpumask))
1021 return -EINVAL;
1022
1023 cpu_buffer = buffer->buffers[cpu];
1024 work = &cpu_buffer->irq_work;
1025 }
1026
1027 poll_wait(filp, &work->waiters, poll_table);
1028 work->waiters_pending = true;
1029 /*
1030 * There's a tight race between setting the waiters_pending and
1031 * checking if the ring buffer is empty. Once the waiters_pending bit
1032 * is set, the next event will wake the task up, but we can get stuck
1033 * if there's only a single event in.
1034 *
1035 * FIXME: Ideally, we need a memory barrier on the writer side as well,
1036 * but adding a memory barrier to all events will cause too much of a
1037 * performance hit in the fast path. We only need a memory barrier when
1038 * the buffer goes from empty to having content. But as this race is
1039 * extremely small, and it's not a problem if another event comes in, we
1040 * will fix it later.
1041 */
1042 smp_mb();
1043
1044 if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
1045 (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
1046 return EPOLLIN | EPOLLRDNORM;
1047 return 0;
1048}
1049
1050/* buffer may be either ring_buffer or ring_buffer_per_cpu */
1051#define RB_WARN_ON(b, cond) \
1052 ({ \
1053 int _____ret = unlikely(cond); \
1054 if (_____ret) { \
1055 if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
1056 struct ring_buffer_per_cpu *__b = \
1057 (void *)b; \
1058 atomic_inc(&__b->buffer->record_disabled); \
1059 } else \
1060 atomic_inc(&b->record_disabled); \
1061 WARN_ON(1); \
1062 } \
1063 _____ret; \
1064 })
1065
1066/* Up this if you want to test the TIME_EXTENTS and normalization */
1067#define DEBUG_SHIFT 0
1068
1069static inline u64 rb_time_stamp(struct trace_buffer *buffer)
1070{
1071 u64 ts;
1072
1073 /* Skip retpolines :-( */
1074 if (IS_ENABLED(CONFIG_RETPOLINE) && likely(buffer->clock == trace_clock_local))
1075 ts = trace_clock_local();
1076 else
1077 ts = buffer->clock();
1078
1079 /* shift to debug/test normalization and TIME_EXTENTS */
1080 return ts << DEBUG_SHIFT;
1081}
1082
1083u64 ring_buffer_time_stamp(struct trace_buffer *buffer)
1084{
1085 u64 time;
1086
1087 preempt_disable_notrace();
1088 time = rb_time_stamp(buffer);
1089 preempt_enable_notrace();
1090
1091 return time;
1092}
1093EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
1094
1095void ring_buffer_normalize_time_stamp(struct trace_buffer *buffer,
1096 int cpu, u64 *ts)
1097{
1098 /* Just stupid testing the normalize function and deltas */
1099 *ts >>= DEBUG_SHIFT;
1100}
1101EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
1102
1103/*
1104 * Making the ring buffer lockless makes things tricky.
1105 * Although writes only happen on the CPU that they are on,
1106 * and they only need to worry about interrupts. Reads can
1107 * happen on any CPU.
1108 *
1109 * The reader page is always off the ring buffer, but when the
1110 * reader finishes with a page, it needs to swap its page with
1111 * a new one from the buffer. The reader needs to take from
1112 * the head (writes go to the tail). But if a writer is in overwrite
1113 * mode and wraps, it must push the head page forward.
1114 *
1115 * Here lies the problem.
1116 *
1117 * The reader must be careful to replace only the head page, and
1118 * not another one. As described at the top of the file in the
1119 * ASCII art, the reader sets its old page to point to the next
1120 * page after head. It then sets the page after head to point to
1121 * the old reader page. But if the writer moves the head page
1122 * during this operation, the reader could end up with the tail.
1123 *
1124 * We use cmpxchg to help prevent this race. We also do something
1125 * special with the page before head. We set the LSB to 1.
1126 *
1127 * When the writer must push the page forward, it will clear the
1128 * bit that points to the head page, move the head, and then set
1129 * the bit that points to the new head page.
1130 *
1131 * We also don't want an interrupt coming in and moving the head
1132 * page on another writer. Thus we use the second LSB to catch
1133 * that too. Thus:
1134 *
1135 * head->list->prev->next bit 1 bit 0
1136 * ------- -------
1137 * Normal page 0 0
1138 * Points to head page 0 1
1139 * New head page 1 0
1140 *
1141 * Note we can not trust the prev pointer of the head page, because:
1142 *
1143 * +----+ +-----+ +-----+
1144 * | |------>| T |---X--->| N |
1145 * | |<------| | | |
1146 * +----+ +-----+ +-----+
1147 * ^ ^ |
1148 * | +-----+ | |
1149 * +----------| R |----------+ |
1150 * | |<-----------+
1151 * +-----+
1152 *
1153 * Key: ---X--> HEAD flag set in pointer
1154 * T Tail page
1155 * R Reader page
1156 * N Next page
1157 *
1158 * (see __rb_reserve_next() to see where this happens)
1159 *
1160 * What the above shows is that the reader just swapped out
1161 * the reader page with a page in the buffer, but before it
1162 * could make the new header point back to the new page added
1163 * it was preempted by a writer. The writer moved forward onto
1164 * the new page added by the reader and is about to move forward
1165 * again.
1166 *
1167 * You can see, it is legitimate for the previous pointer of
1168 * the head (or any page) not to point back to itself. But only
1169 * temporarily.
1170 */
1171
1172#define RB_PAGE_NORMAL 0UL
1173#define RB_PAGE_HEAD 1UL
1174#define RB_PAGE_UPDATE 2UL
1175
1176
1177#define RB_FLAG_MASK 3UL
1178
1179/* PAGE_MOVED is not part of the mask */
1180#define RB_PAGE_MOVED 4UL
1181
1182/*
1183 * rb_list_head - remove any bit
1184 */
1185static struct list_head *rb_list_head(struct list_head *list)
1186{
1187 unsigned long val = (unsigned long)list;
1188
1189 return (struct list_head *)(val & ~RB_FLAG_MASK);
1190}
1191
1192/*
1193 * rb_is_head_page - test if the given page is the head page
1194 *
1195 * Because the reader may move the head_page pointer, we can
1196 * not trust what the head page is (it may be pointing to
1197 * the reader page). But if the next page is a header page,
1198 * its flags will be non zero.
1199 */
1200static inline int
1201rb_is_head_page(struct buffer_page *page, struct list_head *list)
1202{
1203 unsigned long val;
1204
1205 val = (unsigned long)list->next;
1206
1207 if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
1208 return RB_PAGE_MOVED;
1209
1210 return val & RB_FLAG_MASK;
1211}
1212
1213/*
1214 * rb_is_reader_page
1215 *
1216 * The unique thing about the reader page, is that, if the
1217 * writer is ever on it, the previous pointer never points
1218 * back to the reader page.
1219 */
1220static bool rb_is_reader_page(struct buffer_page *page)
1221{
1222 struct list_head *list = page->list.prev;
1223
1224 return rb_list_head(list->next) != &page->list;
1225}
1226
1227/*
1228 * rb_set_list_to_head - set a list_head to be pointing to head.
1229 */
1230static void rb_set_list_to_head(struct list_head *list)
1231{
1232 unsigned long *ptr;
1233
1234 ptr = (unsigned long *)&list->next;
1235 *ptr |= RB_PAGE_HEAD;
1236 *ptr &= ~RB_PAGE_UPDATE;
1237}
1238
1239/*
1240 * rb_head_page_activate - sets up head page
1241 */
1242static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
1243{
1244 struct buffer_page *head;
1245
1246 head = cpu_buffer->head_page;
1247 if (!head)
1248 return;
1249
1250 /*
1251 * Set the previous list pointer to have the HEAD flag.
1252 */
1253 rb_set_list_to_head(head->list.prev);
1254}
1255
1256static void rb_list_head_clear(struct list_head *list)
1257{
1258 unsigned long *ptr = (unsigned long *)&list->next;
1259
1260 *ptr &= ~RB_FLAG_MASK;
1261}
1262
1263/*
1264 * rb_head_page_deactivate - clears head page ptr (for free list)
1265 */
1266static void
1267rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
1268{
1269 struct list_head *hd;
1270
1271 /* Go through the whole list and clear any pointers found. */
1272 rb_list_head_clear(cpu_buffer->pages);
1273
1274 list_for_each(hd, cpu_buffer->pages)
1275 rb_list_head_clear(hd);
1276}
1277
1278static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
1279 struct buffer_page *head,
1280 struct buffer_page *prev,
1281 int old_flag, int new_flag)
1282{
1283 struct list_head *list;
1284 unsigned long val = (unsigned long)&head->list;
1285 unsigned long ret;
1286
1287 list = &prev->list;
1288
1289 val &= ~RB_FLAG_MASK;
1290
1291 ret = cmpxchg((unsigned long *)&list->next,
1292 val | old_flag, val | new_flag);
1293
1294 /* check if the reader took the page */
1295 if ((ret & ~RB_FLAG_MASK) != val)
1296 return RB_PAGE_MOVED;
1297
1298 return ret & RB_FLAG_MASK;
1299}
1300
1301static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
1302 struct buffer_page *head,
1303 struct buffer_page *prev,
1304 int old_flag)
1305{
1306 return rb_head_page_set(cpu_buffer, head, prev,
1307 old_flag, RB_PAGE_UPDATE);
1308}
1309
1310static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
1311 struct buffer_page *head,
1312 struct buffer_page *prev,
1313 int old_flag)
1314{
1315 return rb_head_page_set(cpu_buffer, head, prev,
1316 old_flag, RB_PAGE_HEAD);
1317}
1318
1319static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
1320 struct buffer_page *head,
1321 struct buffer_page *prev,
1322 int old_flag)
1323{
1324 return rb_head_page_set(cpu_buffer, head, prev,
1325 old_flag, RB_PAGE_NORMAL);
1326}
1327
1328static inline void rb_inc_page(struct buffer_page **bpage)
1329{
1330 struct list_head *p = rb_list_head((*bpage)->list.next);
1331
1332 *bpage = list_entry(p, struct buffer_page, list);
1333}
1334
1335static struct buffer_page *
1336rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
1337{
1338 struct buffer_page *head;
1339 struct buffer_page *page;
1340 struct list_head *list;
1341 int i;
1342
1343 if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
1344 return NULL;
1345
1346 /* sanity check */
1347 list = cpu_buffer->pages;
1348 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
1349 return NULL;
1350
1351 page = head = cpu_buffer->head_page;
1352 /*
1353 * It is possible that the writer moves the header behind
1354 * where we started, and we miss in one loop.
1355 * A second loop should grab the header, but we'll do
1356 * three loops just because I'm paranoid.
1357 */
1358 for (i = 0; i < 3; i++) {
1359 do {
1360 if (rb_is_head_page(page, page->list.prev)) {
1361 cpu_buffer->head_page = page;
1362 return page;
1363 }
1364 rb_inc_page(&page);
1365 } while (page != head);
1366 }
1367
1368 RB_WARN_ON(cpu_buffer, 1);
1369
1370 return NULL;
1371}
1372
1373static int rb_head_page_replace(struct buffer_page *old,
1374 struct buffer_page *new)
1375{
1376 unsigned long *ptr = (unsigned long *)&old->list.prev->next;
1377 unsigned long val;
1378 unsigned long ret;
1379
1380 val = *ptr & ~RB_FLAG_MASK;
1381 val |= RB_PAGE_HEAD;
1382
1383 ret = cmpxchg(ptr, val, (unsigned long)&new->list);
1384
1385 return ret == val;
1386}
1387
1388/*
1389 * rb_tail_page_update - move the tail page forward
1390 */
1391static void rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
1392 struct buffer_page *tail_page,
1393 struct buffer_page *next_page)
1394{
1395 unsigned long old_entries;
1396 unsigned long old_write;
1397
1398 /*
1399 * The tail page now needs to be moved forward.
1400 *
1401 * We need to reset the tail page, but without messing
1402 * with possible erasing of data brought in by interrupts
1403 * that have moved the tail page and are currently on it.
1404 *
1405 * We add a counter to the write field to denote this.
1406 */
1407 old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
1408 old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
1409
1410 local_inc(&cpu_buffer->pages_touched);
1411 /*
1412 * Just make sure we have seen our old_write and synchronize
1413 * with any interrupts that come in.
1414 */
1415 barrier();
1416
1417 /*
1418 * If the tail page is still the same as what we think
1419 * it is, then it is up to us to update the tail
1420 * pointer.
1421 */
1422 if (tail_page == READ_ONCE(cpu_buffer->tail_page)) {
1423 /* Zero the write counter */
1424 unsigned long val = old_write & ~RB_WRITE_MASK;
1425 unsigned long eval = old_entries & ~RB_WRITE_MASK;
1426
1427 /*
1428 * This will only succeed if an interrupt did
1429 * not come in and change it. In which case, we
1430 * do not want to modify it.
1431 *
1432 * We add (void) to let the compiler know that we do not care
1433 * about the return value of these functions. We use the
1434 * cmpxchg to only update if an interrupt did not already
1435 * do it for us. If the cmpxchg fails, we don't care.
1436 */
1437 (void)local_cmpxchg(&next_page->write, old_write, val);
1438 (void)local_cmpxchg(&next_page->entries, old_entries, eval);
1439
1440 /*
1441 * No need to worry about races with clearing out the commit.
1442 * it only can increment when a commit takes place. But that
1443 * only happens in the outer most nested commit.
1444 */
1445 local_set(&next_page->page->commit, 0);
1446
1447 /* Again, either we update tail_page or an interrupt does */
1448 (void)cmpxchg(&cpu_buffer->tail_page, tail_page, next_page);
1449 }
1450}
1451
1452static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
1453 struct buffer_page *bpage)
1454{
1455 unsigned long val = (unsigned long)bpage;
1456
1457 if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK))
1458 return 1;
1459
1460 return 0;
1461}
1462
1463/**
1464 * rb_check_list - make sure a pointer to a list has the last bits zero
1465 */
1466static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer,
1467 struct list_head *list)
1468{
1469 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev))
1470 return 1;
1471 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next))
1472 return 1;
1473 return 0;
1474}
1475
1476/**
1477 * rb_check_pages - integrity check of buffer pages
1478 * @cpu_buffer: CPU buffer with pages to test
1479 *
1480 * As a safety measure we check to make sure the data pages have not
1481 * been corrupted.
1482 */
1483static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
1484{
1485 struct list_head *head = cpu_buffer->pages;
1486 struct buffer_page *bpage, *tmp;
1487
1488 /* Reset the head page if it exists */
1489 if (cpu_buffer->head_page)
1490 rb_set_head_page(cpu_buffer);
1491
1492 rb_head_page_deactivate(cpu_buffer);
1493
1494 if (RB_WARN_ON(cpu_buffer, head->next->prev != head))
1495 return -1;
1496 if (RB_WARN_ON(cpu_buffer, head->prev->next != head))
1497 return -1;
1498
1499 if (rb_check_list(cpu_buffer, head))
1500 return -1;
1501
1502 list_for_each_entry_safe(bpage, tmp, head, list) {
1503 if (RB_WARN_ON(cpu_buffer,
1504 bpage->list.next->prev != &bpage->list))
1505 return -1;
1506 if (RB_WARN_ON(cpu_buffer,
1507 bpage->list.prev->next != &bpage->list))
1508 return -1;
1509 if (rb_check_list(cpu_buffer, &bpage->list))
1510 return -1;
1511 }
1512
1513 rb_head_page_activate(cpu_buffer);
1514
1515 return 0;
1516}
1517
1518static int __rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1519 long nr_pages, struct list_head *pages)
1520{
1521 struct buffer_page *bpage, *tmp;
1522 bool user_thread = current->mm != NULL;
1523 gfp_t mflags;
1524 long i;
1525
1526 /*
1527 * Check if the available memory is there first.
1528 * Note, si_mem_available() only gives us a rough estimate of available
1529 * memory. It may not be accurate. But we don't care, we just want
1530 * to prevent doing any allocation when it is obvious that it is
1531 * not going to succeed.
1532 */
1533 i = si_mem_available();
1534 if (i < nr_pages)
1535 return -ENOMEM;
1536
1537 /*
1538 * __GFP_RETRY_MAYFAIL flag makes sure that the allocation fails
1539 * gracefully without invoking oom-killer and the system is not
1540 * destabilized.
1541 */
1542 mflags = GFP_KERNEL | __GFP_RETRY_MAYFAIL;
1543
1544 /*
1545 * If a user thread allocates too much, and si_mem_available()
1546 * reports there's enough memory, even though there is not.
1547 * Make sure the OOM killer kills this thread. This can happen
1548 * even with RETRY_MAYFAIL because another task may be doing
1549 * an allocation after this task has taken all memory.
1550 * This is the task the OOM killer needs to take out during this
1551 * loop, even if it was triggered by an allocation somewhere else.
1552 */
1553 if (user_thread)
1554 set_current_oom_origin();
1555 for (i = 0; i < nr_pages; i++) {
1556 struct page *page;
1557
1558 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1559 mflags, cpu_to_node(cpu_buffer->cpu));
1560 if (!bpage)
1561 goto free_pages;
1562
1563 rb_check_bpage(cpu_buffer, bpage);
1564
1565 list_add(&bpage->list, pages);
1566
1567 page = alloc_pages_node(cpu_to_node(cpu_buffer->cpu), mflags, 0);
1568 if (!page)
1569 goto free_pages;
1570 bpage->page = page_address(page);
1571 rb_init_page(bpage->page);
1572
1573 if (user_thread && fatal_signal_pending(current))
1574 goto free_pages;
1575 }
1576 if (user_thread)
1577 clear_current_oom_origin();
1578
1579 return 0;
1580
1581free_pages:
1582 list_for_each_entry_safe(bpage, tmp, pages, list) {
1583 list_del_init(&bpage->list);
1584 free_buffer_page(bpage);
1585 }
1586 if (user_thread)
1587 clear_current_oom_origin();
1588
1589 return -ENOMEM;
1590}
1591
1592static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1593 unsigned long nr_pages)
1594{
1595 LIST_HEAD(pages);
1596
1597 WARN_ON(!nr_pages);
1598
1599 if (__rb_allocate_pages(cpu_buffer, nr_pages, &pages))
1600 return -ENOMEM;
1601
1602 /*
1603 * The ring buffer page list is a circular list that does not
1604 * start and end with a list head. All page list items point to
1605 * other pages.
1606 */
1607 cpu_buffer->pages = pages.next;
1608 list_del(&pages);
1609
1610 cpu_buffer->nr_pages = nr_pages;
1611
1612 rb_check_pages(cpu_buffer);
1613
1614 return 0;
1615}
1616
1617static struct ring_buffer_per_cpu *
1618rb_allocate_cpu_buffer(struct trace_buffer *buffer, long nr_pages, int cpu)
1619{
1620 struct ring_buffer_per_cpu *cpu_buffer;
1621 struct buffer_page *bpage;
1622 struct page *page;
1623 int ret;
1624
1625 cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1626 GFP_KERNEL, cpu_to_node(cpu));
1627 if (!cpu_buffer)
1628 return NULL;
1629
1630 cpu_buffer->cpu = cpu;
1631 cpu_buffer->buffer = buffer;
1632 raw_spin_lock_init(&cpu_buffer->reader_lock);
1633 lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1634 cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1635 INIT_WORK(&cpu_buffer->update_pages_work, update_pages_handler);
1636 init_completion(&cpu_buffer->update_done);
1637 init_irq_work(&cpu_buffer->irq_work.work, rb_wake_up_waiters);
1638 init_waitqueue_head(&cpu_buffer->irq_work.waiters);
1639 init_waitqueue_head(&cpu_buffer->irq_work.full_waiters);
1640
1641 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1642 GFP_KERNEL, cpu_to_node(cpu));
1643 if (!bpage)
1644 goto fail_free_buffer;
1645
1646 rb_check_bpage(cpu_buffer, bpage);
1647
1648 cpu_buffer->reader_page = bpage;
1649 page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0);
1650 if (!page)
1651 goto fail_free_reader;
1652 bpage->page = page_address(page);
1653 rb_init_page(bpage->page);
1654
1655 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1656 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1657
1658 ret = rb_allocate_pages(cpu_buffer, nr_pages);
1659 if (ret < 0)
1660 goto fail_free_reader;
1661
1662 cpu_buffer->head_page
1663 = list_entry(cpu_buffer->pages, struct buffer_page, list);
1664 cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1665
1666 rb_head_page_activate(cpu_buffer);
1667
1668 return cpu_buffer;
1669
1670 fail_free_reader:
1671 free_buffer_page(cpu_buffer->reader_page);
1672
1673 fail_free_buffer:
1674 kfree(cpu_buffer);
1675 return NULL;
1676}
1677
1678static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1679{
1680 struct list_head *head = cpu_buffer->pages;
1681 struct buffer_page *bpage, *tmp;
1682
1683 free_buffer_page(cpu_buffer->reader_page);
1684
1685 rb_head_page_deactivate(cpu_buffer);
1686
1687 if (head) {
1688 list_for_each_entry_safe(bpage, tmp, head, list) {
1689 list_del_init(&bpage->list);
1690 free_buffer_page(bpage);
1691 }
1692 bpage = list_entry(head, struct buffer_page, list);
1693 free_buffer_page(bpage);
1694 }
1695
1696 kfree(cpu_buffer);
1697}
1698
1699/**
1700 * __ring_buffer_alloc - allocate a new ring_buffer
1701 * @size: the size in bytes per cpu that is needed.
1702 * @flags: attributes to set for the ring buffer.
1703 * @key: ring buffer reader_lock_key.
1704 *
1705 * Currently the only flag that is available is the RB_FL_OVERWRITE
1706 * flag. This flag means that the buffer will overwrite old data
1707 * when the buffer wraps. If this flag is not set, the buffer will
1708 * drop data when the tail hits the head.
1709 */
1710struct trace_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1711 struct lock_class_key *key)
1712{
1713 struct trace_buffer *buffer;
1714 long nr_pages;
1715 int bsize;
1716 int cpu;
1717 int ret;
1718
1719 /* keep it in its own cache line */
1720 buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1721 GFP_KERNEL);
1722 if (!buffer)
1723 return NULL;
1724
1725 if (!zalloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1726 goto fail_free_buffer;
1727
1728 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1729 buffer->flags = flags;
1730 buffer->clock = trace_clock_local;
1731 buffer->reader_lock_key = key;
1732
1733 init_irq_work(&buffer->irq_work.work, rb_wake_up_waiters);
1734 init_waitqueue_head(&buffer->irq_work.waiters);
1735
1736 /* need at least two pages */
1737 if (nr_pages < 2)
1738 nr_pages = 2;
1739
1740 buffer->cpus = nr_cpu_ids;
1741
1742 bsize = sizeof(void *) * nr_cpu_ids;
1743 buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1744 GFP_KERNEL);
1745 if (!buffer->buffers)
1746 goto fail_free_cpumask;
1747
1748 cpu = raw_smp_processor_id();
1749 cpumask_set_cpu(cpu, buffer->cpumask);
1750 buffer->buffers[cpu] = rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
1751 if (!buffer->buffers[cpu])
1752 goto fail_free_buffers;
1753
1754 ret = cpuhp_state_add_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1755 if (ret < 0)
1756 goto fail_free_buffers;
1757
1758 mutex_init(&buffer->mutex);
1759
1760 return buffer;
1761
1762 fail_free_buffers:
1763 for_each_buffer_cpu(buffer, cpu) {
1764 if (buffer->buffers[cpu])
1765 rb_free_cpu_buffer(buffer->buffers[cpu]);
1766 }
1767 kfree(buffer->buffers);
1768
1769 fail_free_cpumask:
1770 free_cpumask_var(buffer->cpumask);
1771
1772 fail_free_buffer:
1773 kfree(buffer);
1774 return NULL;
1775}
1776EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1777
1778/**
1779 * ring_buffer_free - free a ring buffer.
1780 * @buffer: the buffer to free.
1781 */
1782void
1783ring_buffer_free(struct trace_buffer *buffer)
1784{
1785 int cpu;
1786
1787 cpuhp_state_remove_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1788
1789 for_each_buffer_cpu(buffer, cpu)
1790 rb_free_cpu_buffer(buffer->buffers[cpu]);
1791
1792 kfree(buffer->buffers);
1793 free_cpumask_var(buffer->cpumask);
1794
1795 kfree(buffer);
1796}
1797EXPORT_SYMBOL_GPL(ring_buffer_free);
1798
1799void ring_buffer_set_clock(struct trace_buffer *buffer,
1800 u64 (*clock)(void))
1801{
1802 buffer->clock = clock;
1803}
1804
1805void ring_buffer_set_time_stamp_abs(struct trace_buffer *buffer, bool abs)
1806{
1807 buffer->time_stamp_abs = abs;
1808}
1809
1810bool ring_buffer_time_stamp_abs(struct trace_buffer *buffer)
1811{
1812 return buffer->time_stamp_abs;
1813}
1814
1815static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1816
1817static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1818{
1819 return local_read(&bpage->entries) & RB_WRITE_MASK;
1820}
1821
1822static inline unsigned long rb_page_write(struct buffer_page *bpage)
1823{
1824 return local_read(&bpage->write) & RB_WRITE_MASK;
1825}
1826
1827static int
1828rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned long nr_pages)
1829{
1830 struct list_head *tail_page, *to_remove, *next_page;
1831 struct buffer_page *to_remove_page, *tmp_iter_page;
1832 struct buffer_page *last_page, *first_page;
1833 unsigned long nr_removed;
1834 unsigned long head_bit;
1835 int page_entries;
1836
1837 head_bit = 0;
1838
1839 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1840 atomic_inc(&cpu_buffer->record_disabled);
1841 /*
1842 * We don't race with the readers since we have acquired the reader
1843 * lock. We also don't race with writers after disabling recording.
1844 * This makes it easy to figure out the first and the last page to be
1845 * removed from the list. We unlink all the pages in between including
1846 * the first and last pages. This is done in a busy loop so that we
1847 * lose the least number of traces.
1848 * The pages are freed after we restart recording and unlock readers.
1849 */
1850 tail_page = &cpu_buffer->tail_page->list;
1851
1852 /*
1853 * tail page might be on reader page, we remove the next page
1854 * from the ring buffer
1855 */
1856 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
1857 tail_page = rb_list_head(tail_page->next);
1858 to_remove = tail_page;
1859
1860 /* start of pages to remove */
1861 first_page = list_entry(rb_list_head(to_remove->next),
1862 struct buffer_page, list);
1863
1864 for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
1865 to_remove = rb_list_head(to_remove)->next;
1866 head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
1867 }
1868
1869 next_page = rb_list_head(to_remove)->next;
1870
1871 /*
1872 * Now we remove all pages between tail_page and next_page.
1873 * Make sure that we have head_bit value preserved for the
1874 * next page
1875 */
1876 tail_page->next = (struct list_head *)((unsigned long)next_page |
1877 head_bit);
1878 next_page = rb_list_head(next_page);
1879 next_page->prev = tail_page;
1880
1881 /* make sure pages points to a valid page in the ring buffer */
1882 cpu_buffer->pages = next_page;
1883
1884 /* update head page */
1885 if (head_bit)
1886 cpu_buffer->head_page = list_entry(next_page,
1887 struct buffer_page, list);
1888
1889 /*
1890 * change read pointer to make sure any read iterators reset
1891 * themselves
1892 */
1893 cpu_buffer->read = 0;
1894
1895 /* pages are removed, resume tracing and then free the pages */
1896 atomic_dec(&cpu_buffer->record_disabled);
1897 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1898
1899 RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
1900
1901 /* last buffer page to remove */
1902 last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
1903 list);
1904 tmp_iter_page = first_page;
1905
1906 do {
1907 cond_resched();
1908
1909 to_remove_page = tmp_iter_page;
1910 rb_inc_page(&tmp_iter_page);
1911
1912 /* update the counters */
1913 page_entries = rb_page_entries(to_remove_page);
1914 if (page_entries) {
1915 /*
1916 * If something was added to this page, it was full
1917 * since it is not the tail page. So we deduct the
1918 * bytes consumed in ring buffer from here.
1919 * Increment overrun to account for the lost events.
1920 */
1921 local_add(page_entries, &cpu_buffer->overrun);
1922 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1923 }
1924
1925 /*
1926 * We have already removed references to this list item, just
1927 * free up the buffer_page and its page
1928 */
1929 free_buffer_page(to_remove_page);
1930 nr_removed--;
1931
1932 } while (to_remove_page != last_page);
1933
1934 RB_WARN_ON(cpu_buffer, nr_removed);
1935
1936 return nr_removed == 0;
1937}
1938
1939static int
1940rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
1941{
1942 struct list_head *pages = &cpu_buffer->new_pages;
1943 int retries, success;
1944
1945 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1946 /*
1947 * We are holding the reader lock, so the reader page won't be swapped
1948 * in the ring buffer. Now we are racing with the writer trying to
1949 * move head page and the tail page.
1950 * We are going to adapt the reader page update process where:
1951 * 1. We first splice the start and end of list of new pages between
1952 * the head page and its previous page.
1953 * 2. We cmpxchg the prev_page->next to point from head page to the
1954 * start of new pages list.
1955 * 3. Finally, we update the head->prev to the end of new list.
1956 *
1957 * We will try this process 10 times, to make sure that we don't keep
1958 * spinning.
1959 */
1960 retries = 10;
1961 success = 0;
1962 while (retries--) {
1963 struct list_head *head_page, *prev_page, *r;
1964 struct list_head *last_page, *first_page;
1965 struct list_head *head_page_with_bit;
1966
1967 head_page = &rb_set_head_page(cpu_buffer)->list;
1968 if (!head_page)
1969 break;
1970 prev_page = head_page->prev;
1971
1972 first_page = pages->next;
1973 last_page = pages->prev;
1974
1975 head_page_with_bit = (struct list_head *)
1976 ((unsigned long)head_page | RB_PAGE_HEAD);
1977
1978 last_page->next = head_page_with_bit;
1979 first_page->prev = prev_page;
1980
1981 r = cmpxchg(&prev_page->next, head_page_with_bit, first_page);
1982
1983 if (r == head_page_with_bit) {
1984 /*
1985 * yay, we replaced the page pointer to our new list,
1986 * now, we just have to update to head page's prev
1987 * pointer to point to end of list
1988 */
1989 head_page->prev = last_page;
1990 success = 1;
1991 break;
1992 }
1993 }
1994
1995 if (success)
1996 INIT_LIST_HEAD(pages);
1997 /*
1998 * If we weren't successful in adding in new pages, warn and stop
1999 * tracing
2000 */
2001 RB_WARN_ON(cpu_buffer, !success);
2002 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
2003
2004 /* free pages if they weren't inserted */
2005 if (!success) {
2006 struct buffer_page *bpage, *tmp;
2007 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
2008 list) {
2009 list_del_init(&bpage->list);
2010 free_buffer_page(bpage);
2011 }
2012 }
2013 return success;
2014}
2015
2016static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
2017{
2018 int success;
2019
2020 if (cpu_buffer->nr_pages_to_update > 0)
2021 success = rb_insert_pages(cpu_buffer);
2022 else
2023 success = rb_remove_pages(cpu_buffer,
2024 -cpu_buffer->nr_pages_to_update);
2025
2026 if (success)
2027 cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
2028}
2029
2030static void update_pages_handler(struct work_struct *work)
2031{
2032 struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
2033 struct ring_buffer_per_cpu, update_pages_work);
2034 rb_update_pages(cpu_buffer);
2035 complete(&cpu_buffer->update_done);
2036}
2037
2038/**
2039 * ring_buffer_resize - resize the ring buffer
2040 * @buffer: the buffer to resize.
2041 * @size: the new size.
2042 * @cpu_id: the cpu buffer to resize
2043 *
2044 * Minimum size is 2 * BUF_PAGE_SIZE.
2045 *
2046 * Returns 0 on success and < 0 on failure.
2047 */
2048int ring_buffer_resize(struct trace_buffer *buffer, unsigned long size,
2049 int cpu_id)
2050{
2051 struct ring_buffer_per_cpu *cpu_buffer;
2052 unsigned long nr_pages;
2053 int cpu, err;
2054
2055 /*
2056 * Always succeed at resizing a non-existent buffer:
2057 */
2058 if (!buffer)
2059 return 0;
2060
2061 /* Make sure the requested buffer exists */
2062 if (cpu_id != RING_BUFFER_ALL_CPUS &&
2063 !cpumask_test_cpu(cpu_id, buffer->cpumask))
2064 return 0;
2065
2066 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
2067
2068 /* we need a minimum of two pages */
2069 if (nr_pages < 2)
2070 nr_pages = 2;
2071
2072 /* prevent another thread from changing buffer sizes */
2073 mutex_lock(&buffer->mutex);
2074
2075
2076 if (cpu_id == RING_BUFFER_ALL_CPUS) {
2077 /*
2078 * Don't succeed if resizing is disabled, as a reader might be
2079 * manipulating the ring buffer and is expecting a sane state while
2080 * this is true.
2081 */
2082 for_each_buffer_cpu(buffer, cpu) {
2083 cpu_buffer = buffer->buffers[cpu];
2084 if (atomic_read(&cpu_buffer->resize_disabled)) {
2085 err = -EBUSY;
2086 goto out_err_unlock;
2087 }
2088 }
2089
2090 /* calculate the pages to update */
2091 for_each_buffer_cpu(buffer, cpu) {
2092 cpu_buffer = buffer->buffers[cpu];
2093
2094 cpu_buffer->nr_pages_to_update = nr_pages -
2095 cpu_buffer->nr_pages;
2096 /*
2097 * nothing more to do for removing pages or no update
2098 */
2099 if (cpu_buffer->nr_pages_to_update <= 0)
2100 continue;
2101 /*
2102 * to add pages, make sure all new pages can be
2103 * allocated without receiving ENOMEM
2104 */
2105 INIT_LIST_HEAD(&cpu_buffer->new_pages);
2106 if (__rb_allocate_pages(cpu_buffer, cpu_buffer->nr_pages_to_update,
2107 &cpu_buffer->new_pages)) {
2108 /* not enough memory for new pages */
2109 err = -ENOMEM;
2110 goto out_err;
2111 }
2112 }
2113
2114 get_online_cpus();
2115 /*
2116 * Fire off all the required work handlers
2117 * We can't schedule on offline CPUs, but it's not necessary
2118 * since we can change their buffer sizes without any race.
2119 */
2120 for_each_buffer_cpu(buffer, cpu) {
2121 cpu_buffer = buffer->buffers[cpu];
2122 if (!cpu_buffer->nr_pages_to_update)
2123 continue;
2124
2125 /* Can't run something on an offline CPU. */
2126 if (!cpu_online(cpu)) {
2127 rb_update_pages(cpu_buffer);
2128 cpu_buffer->nr_pages_to_update = 0;
2129 } else {
2130 schedule_work_on(cpu,
2131 &cpu_buffer->update_pages_work);
2132 }
2133 }
2134
2135 /* wait for all the updates to complete */
2136 for_each_buffer_cpu(buffer, cpu) {
2137 cpu_buffer = buffer->buffers[cpu];
2138 if (!cpu_buffer->nr_pages_to_update)
2139 continue;
2140
2141 if (cpu_online(cpu))
2142 wait_for_completion(&cpu_buffer->update_done);
2143 cpu_buffer->nr_pages_to_update = 0;
2144 }
2145
2146 put_online_cpus();
2147 } else {
2148 cpu_buffer = buffer->buffers[cpu_id];
2149
2150 if (nr_pages == cpu_buffer->nr_pages)
2151 goto out;
2152
2153 /*
2154 * Don't succeed if resizing is disabled, as a reader might be
2155 * manipulating the ring buffer and is expecting a sane state while
2156 * this is true.
2157 */
2158 if (atomic_read(&cpu_buffer->resize_disabled)) {
2159 err = -EBUSY;
2160 goto out_err_unlock;
2161 }
2162
2163 cpu_buffer->nr_pages_to_update = nr_pages -
2164 cpu_buffer->nr_pages;
2165
2166 INIT_LIST_HEAD(&cpu_buffer->new_pages);
2167 if (cpu_buffer->nr_pages_to_update > 0 &&
2168 __rb_allocate_pages(cpu_buffer, cpu_buffer->nr_pages_to_update,
2169 &cpu_buffer->new_pages)) {
2170 err = -ENOMEM;
2171 goto out_err;
2172 }
2173
2174 get_online_cpus();
2175
2176 /* Can't run something on an offline CPU. */
2177 if (!cpu_online(cpu_id))
2178 rb_update_pages(cpu_buffer);
2179 else {
2180 schedule_work_on(cpu_id,
2181 &cpu_buffer->update_pages_work);
2182 wait_for_completion(&cpu_buffer->update_done);
2183 }
2184
2185 cpu_buffer->nr_pages_to_update = 0;
2186 put_online_cpus();
2187 }
2188
2189 out:
2190 /*
2191 * The ring buffer resize can happen with the ring buffer
2192 * enabled, so that the update disturbs the tracing as little
2193 * as possible. But if the buffer is disabled, we do not need
2194 * to worry about that, and we can take the time to verify
2195 * that the buffer is not corrupt.
2196 */
2197 if (atomic_read(&buffer->record_disabled)) {
2198 atomic_inc(&buffer->record_disabled);
2199 /*
2200 * Even though the buffer was disabled, we must make sure
2201 * that it is truly disabled before calling rb_check_pages.
2202 * There could have been a race between checking
2203 * record_disable and incrementing it.
2204 */
2205 synchronize_rcu();
2206 for_each_buffer_cpu(buffer, cpu) {
2207 cpu_buffer = buffer->buffers[cpu];
2208 rb_check_pages(cpu_buffer);
2209 }
2210 atomic_dec(&buffer->record_disabled);
2211 }
2212
2213 mutex_unlock(&buffer->mutex);
2214 return 0;
2215
2216 out_err:
2217 for_each_buffer_cpu(buffer, cpu) {
2218 struct buffer_page *bpage, *tmp;
2219
2220 cpu_buffer = buffer->buffers[cpu];
2221 cpu_buffer->nr_pages_to_update = 0;
2222
2223 if (list_empty(&cpu_buffer->new_pages))
2224 continue;
2225
2226 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
2227 list) {
2228 list_del_init(&bpage->list);
2229 free_buffer_page(bpage);
2230 }
2231 }
2232 out_err_unlock:
2233 mutex_unlock(&buffer->mutex);
2234 return err;
2235}
2236EXPORT_SYMBOL_GPL(ring_buffer_resize);
2237
2238void ring_buffer_change_overwrite(struct trace_buffer *buffer, int val)
2239{
2240 mutex_lock(&buffer->mutex);
2241 if (val)
2242 buffer->flags |= RB_FL_OVERWRITE;
2243 else
2244 buffer->flags &= ~RB_FL_OVERWRITE;
2245 mutex_unlock(&buffer->mutex);
2246}
2247EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
2248
2249static __always_inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
2250{
2251 return bpage->page->data + index;
2252}
2253
2254static __always_inline struct ring_buffer_event *
2255rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
2256{
2257 return __rb_page_index(cpu_buffer->reader_page,
2258 cpu_buffer->reader_page->read);
2259}
2260
2261static __always_inline unsigned rb_page_commit(struct buffer_page *bpage)
2262{
2263 return local_read(&bpage->page->commit);
2264}
2265
2266static struct ring_buffer_event *
2267rb_iter_head_event(struct ring_buffer_iter *iter)
2268{
2269 struct ring_buffer_event *event;
2270 struct buffer_page *iter_head_page = iter->head_page;
2271 unsigned long commit;
2272 unsigned length;
2273
2274 if (iter->head != iter->next_event)
2275 return iter->event;
2276
2277 /*
2278 * When the writer goes across pages, it issues a cmpxchg which
2279 * is a mb(), which will synchronize with the rmb here.
2280 * (see rb_tail_page_update() and __rb_reserve_next())
2281 */
2282 commit = rb_page_commit(iter_head_page);
2283 smp_rmb();
2284 event = __rb_page_index(iter_head_page, iter->head);
2285 length = rb_event_length(event);
2286
2287 /*
2288 * READ_ONCE() doesn't work on functions and we don't want the
2289 * compiler doing any crazy optimizations with length.
2290 */
2291 barrier();
2292
2293 if ((iter->head + length) > commit || length > BUF_MAX_DATA_SIZE)
2294 /* Writer corrupted the read? */
2295 goto reset;
2296
2297 memcpy(iter->event, event, length);
2298 /*
2299 * If the page stamp is still the same after this rmb() then the
2300 * event was safely copied without the writer entering the page.
2301 */
2302 smp_rmb();
2303
2304 /* Make sure the page didn't change since we read this */
2305 if (iter->page_stamp != iter_head_page->page->time_stamp ||
2306 commit > rb_page_commit(iter_head_page))
2307 goto reset;
2308
2309 iter->next_event = iter->head + length;
2310 return iter->event;
2311 reset:
2312 /* Reset to the beginning */
2313 iter->page_stamp = iter->read_stamp = iter->head_page->page->time_stamp;
2314 iter->head = 0;
2315 iter->next_event = 0;
2316 iter->missed_events = 1;
2317 return NULL;
2318}
2319
2320/* Size is determined by what has been committed */
2321static __always_inline unsigned rb_page_size(struct buffer_page *bpage)
2322{
2323 return rb_page_commit(bpage);
2324}
2325
2326static __always_inline unsigned
2327rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
2328{
2329 return rb_page_commit(cpu_buffer->commit_page);
2330}
2331
2332static __always_inline unsigned
2333rb_event_index(struct ring_buffer_event *event)
2334{
2335 unsigned long addr = (unsigned long)event;
2336
2337 return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
2338}
2339
2340static void rb_inc_iter(struct ring_buffer_iter *iter)
2341{
2342 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
2343
2344 /*
2345 * The iterator could be on the reader page (it starts there).
2346 * But the head could have moved, since the reader was
2347 * found. Check for this case and assign the iterator
2348 * to the head page instead of next.
2349 */
2350 if (iter->head_page == cpu_buffer->reader_page)
2351 iter->head_page = rb_set_head_page(cpu_buffer);
2352 else
2353 rb_inc_page(&iter->head_page);
2354
2355 iter->page_stamp = iter->read_stamp = iter->head_page->page->time_stamp;
2356 iter->head = 0;
2357 iter->next_event = 0;
2358}
2359
2360/*
2361 * rb_handle_head_page - writer hit the head page
2362 *
2363 * Returns: +1 to retry page
2364 * 0 to continue
2365 * -1 on error
2366 */
2367static int
2368rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
2369 struct buffer_page *tail_page,
2370 struct buffer_page *next_page)
2371{
2372 struct buffer_page *new_head;
2373 int entries;
2374 int type;
2375 int ret;
2376
2377 entries = rb_page_entries(next_page);
2378
2379 /*
2380 * The hard part is here. We need to move the head
2381 * forward, and protect against both readers on
2382 * other CPUs and writers coming in via interrupts.
2383 */
2384 type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
2385 RB_PAGE_HEAD);
2386
2387 /*
2388 * type can be one of four:
2389 * NORMAL - an interrupt already moved it for us
2390 * HEAD - we are the first to get here.
2391 * UPDATE - we are the interrupt interrupting
2392 * a current move.
2393 * MOVED - a reader on another CPU moved the next
2394 * pointer to its reader page. Give up
2395 * and try again.
2396 */
2397
2398 switch (type) {
2399 case RB_PAGE_HEAD:
2400 /*
2401 * We changed the head to UPDATE, thus
2402 * it is our responsibility to update
2403 * the counters.
2404 */
2405 local_add(entries, &cpu_buffer->overrun);
2406 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
2407
2408 /*
2409 * The entries will be zeroed out when we move the
2410 * tail page.
2411 */
2412
2413 /* still more to do */
2414 break;
2415
2416 case RB_PAGE_UPDATE:
2417 /*
2418 * This is an interrupt that interrupt the
2419 * previous update. Still more to do.
2420 */
2421 break;
2422 case RB_PAGE_NORMAL:
2423 /*
2424 * An interrupt came in before the update
2425 * and processed this for us.
2426 * Nothing left to do.
2427 */
2428 return 1;
2429 case RB_PAGE_MOVED:
2430 /*
2431 * The reader is on another CPU and just did
2432 * a swap with our next_page.
2433 * Try again.
2434 */
2435 return 1;
2436 default:
2437 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
2438 return -1;
2439 }
2440
2441 /*
2442 * Now that we are here, the old head pointer is
2443 * set to UPDATE. This will keep the reader from
2444 * swapping the head page with the reader page.
2445 * The reader (on another CPU) will spin till
2446 * we are finished.
2447 *
2448 * We just need to protect against interrupts
2449 * doing the job. We will set the next pointer
2450 * to HEAD. After that, we set the old pointer
2451 * to NORMAL, but only if it was HEAD before.
2452 * otherwise we are an interrupt, and only
2453 * want the outer most commit to reset it.
2454 */
2455 new_head = next_page;
2456 rb_inc_page(&new_head);
2457
2458 ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
2459 RB_PAGE_NORMAL);
2460
2461 /*
2462 * Valid returns are:
2463 * HEAD - an interrupt came in and already set it.
2464 * NORMAL - One of two things:
2465 * 1) We really set it.
2466 * 2) A bunch of interrupts came in and moved
2467 * the page forward again.
2468 */
2469 switch (ret) {
2470 case RB_PAGE_HEAD:
2471 case RB_PAGE_NORMAL:
2472 /* OK */
2473 break;
2474 default:
2475 RB_WARN_ON(cpu_buffer, 1);
2476 return -1;
2477 }
2478
2479 /*
2480 * It is possible that an interrupt came in,
2481 * set the head up, then more interrupts came in
2482 * and moved it again. When we get back here,
2483 * the page would have been set to NORMAL but we
2484 * just set it back to HEAD.
2485 *
2486 * How do you detect this? Well, if that happened
2487 * the tail page would have moved.
2488 */
2489 if (ret == RB_PAGE_NORMAL) {
2490 struct buffer_page *buffer_tail_page;
2491
2492 buffer_tail_page = READ_ONCE(cpu_buffer->tail_page);
2493 /*
2494 * If the tail had moved passed next, then we need
2495 * to reset the pointer.
2496 */
2497 if (buffer_tail_page != tail_page &&
2498 buffer_tail_page != next_page)
2499 rb_head_page_set_normal(cpu_buffer, new_head,
2500 next_page,
2501 RB_PAGE_HEAD);
2502 }
2503
2504 /*
2505 * If this was the outer most commit (the one that
2506 * changed the original pointer from HEAD to UPDATE),
2507 * then it is up to us to reset it to NORMAL.
2508 */
2509 if (type == RB_PAGE_HEAD) {
2510 ret = rb_head_page_set_normal(cpu_buffer, next_page,
2511 tail_page,
2512 RB_PAGE_UPDATE);
2513 if (RB_WARN_ON(cpu_buffer,
2514 ret != RB_PAGE_UPDATE))
2515 return -1;
2516 }
2517
2518 return 0;
2519}
2520
2521static inline void
2522rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2523 unsigned long tail, struct rb_event_info *info)
2524{
2525 struct buffer_page *tail_page = info->tail_page;
2526 struct ring_buffer_event *event;
2527 unsigned long length = info->length;
2528
2529 /*
2530 * Only the event that crossed the page boundary
2531 * must fill the old tail_page with padding.
2532 */
2533 if (tail >= BUF_PAGE_SIZE) {
2534 /*
2535 * If the page was filled, then we still need
2536 * to update the real_end. Reset it to zero
2537 * and the reader will ignore it.
2538 */
2539 if (tail == BUF_PAGE_SIZE)
2540 tail_page->real_end = 0;
2541
2542 local_sub(length, &tail_page->write);
2543 return;
2544 }
2545
2546 event = __rb_page_index(tail_page, tail);
2547
2548 /* account for padding bytes */
2549 local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes);
2550
2551 /*
2552 * Save the original length to the meta data.
2553 * This will be used by the reader to add lost event
2554 * counter.
2555 */
2556 tail_page->real_end = tail;
2557
2558 /*
2559 * If this event is bigger than the minimum size, then
2560 * we need to be careful that we don't subtract the
2561 * write counter enough to allow another writer to slip
2562 * in on this page.
2563 * We put in a discarded commit instead, to make sure
2564 * that this space is not used again.
2565 *
2566 * If we are less than the minimum size, we don't need to
2567 * worry about it.
2568 */
2569 if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
2570 /* No room for any events */
2571
2572 /* Mark the rest of the page with padding */
2573 rb_event_set_padding(event);
2574
2575 /* Set the write back to the previous setting */
2576 local_sub(length, &tail_page->write);
2577 return;
2578 }
2579
2580 /* Put in a discarded event */
2581 event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
2582 event->type_len = RINGBUF_TYPE_PADDING;
2583 /* time delta must be non zero */
2584 event->time_delta = 1;
2585
2586 /* Set write to end of buffer */
2587 length = (tail + length) - BUF_PAGE_SIZE;
2588 local_sub(length, &tail_page->write);
2589}
2590
2591static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer);
2592
2593/*
2594 * This is the slow path, force gcc not to inline it.
2595 */
2596static noinline struct ring_buffer_event *
2597rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2598 unsigned long tail, struct rb_event_info *info)
2599{
2600 struct buffer_page *tail_page = info->tail_page;
2601 struct buffer_page *commit_page = cpu_buffer->commit_page;
2602 struct trace_buffer *buffer = cpu_buffer->buffer;
2603 struct buffer_page *next_page;
2604 int ret;
2605
2606 next_page = tail_page;
2607
2608 rb_inc_page(&next_page);
2609
2610 /*
2611 * If for some reason, we had an interrupt storm that made
2612 * it all the way around the buffer, bail, and warn
2613 * about it.
2614 */
2615 if (unlikely(next_page == commit_page)) {
2616 local_inc(&cpu_buffer->commit_overrun);
2617 goto out_reset;
2618 }
2619
2620 /*
2621 * This is where the fun begins!
2622 *
2623 * We are fighting against races between a reader that
2624 * could be on another CPU trying to swap its reader
2625 * page with the buffer head.
2626 *
2627 * We are also fighting against interrupts coming in and
2628 * moving the head or tail on us as well.
2629 *
2630 * If the next page is the head page then we have filled
2631 * the buffer, unless the commit page is still on the
2632 * reader page.
2633 */
2634 if (rb_is_head_page(next_page, &tail_page->list)) {
2635
2636 /*
2637 * If the commit is not on the reader page, then
2638 * move the header page.
2639 */
2640 if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2641 /*
2642 * If we are not in overwrite mode,
2643 * this is easy, just stop here.
2644 */
2645 if (!(buffer->flags & RB_FL_OVERWRITE)) {
2646 local_inc(&cpu_buffer->dropped_events);
2647 goto out_reset;
2648 }
2649
2650 ret = rb_handle_head_page(cpu_buffer,
2651 tail_page,
2652 next_page);
2653 if (ret < 0)
2654 goto out_reset;
2655 if (ret)
2656 goto out_again;
2657 } else {
2658 /*
2659 * We need to be careful here too. The
2660 * commit page could still be on the reader
2661 * page. We could have a small buffer, and
2662 * have filled up the buffer with events
2663 * from interrupts and such, and wrapped.
2664 *
2665 * Note, if the tail page is also on the
2666 * reader_page, we let it move out.
2667 */
2668 if (unlikely((cpu_buffer->commit_page !=
2669 cpu_buffer->tail_page) &&
2670 (cpu_buffer->commit_page ==
2671 cpu_buffer->reader_page))) {
2672 local_inc(&cpu_buffer->commit_overrun);
2673 goto out_reset;
2674 }
2675 }
2676 }
2677
2678 rb_tail_page_update(cpu_buffer, tail_page, next_page);
2679
2680 out_again:
2681
2682 rb_reset_tail(cpu_buffer, tail, info);
2683
2684 /* Commit what we have for now. */
2685 rb_end_commit(cpu_buffer);
2686 /* rb_end_commit() decs committing */
2687 local_inc(&cpu_buffer->committing);
2688
2689 /* fail and let the caller try again */
2690 return ERR_PTR(-EAGAIN);
2691
2692 out_reset:
2693 /* reset write */
2694 rb_reset_tail(cpu_buffer, tail, info);
2695
2696 return NULL;
2697}
2698
2699/* Slow path */
2700static struct ring_buffer_event *
2701rb_add_time_stamp(struct ring_buffer_event *event, u64 delta, bool abs)
2702{
2703 if (abs)
2704 event->type_len = RINGBUF_TYPE_TIME_STAMP;
2705 else
2706 event->type_len = RINGBUF_TYPE_TIME_EXTEND;
2707
2708 /* Not the first event on the page, or not delta? */
2709 if (abs || rb_event_index(event)) {
2710 event->time_delta = delta & TS_MASK;
2711 event->array[0] = delta >> TS_SHIFT;
2712 } else {
2713 /* nope, just zero it */
2714 event->time_delta = 0;
2715 event->array[0] = 0;
2716 }
2717
2718 return skip_time_extend(event);
2719}
2720
2721#ifndef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2722static inline bool sched_clock_stable(void)
2723{
2724 return true;
2725}
2726#endif
2727
2728static void
2729rb_check_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2730 struct rb_event_info *info)
2731{
2732 u64 write_stamp;
2733
2734 WARN_ONCE(1, "Delta way too big! %llu ts=%llu before=%llu after=%llu write stamp=%llu\n%s",
2735 (unsigned long long)info->delta,
2736 (unsigned long long)info->ts,
2737 (unsigned long long)info->before,
2738 (unsigned long long)info->after,
2739 (unsigned long long)(rb_time_read(&cpu_buffer->write_stamp, &write_stamp) ? write_stamp : 0),
2740 sched_clock_stable() ? "" :
2741 "If you just came from a suspend/resume,\n"
2742 "please switch to the trace global clock:\n"
2743 " echo global > /sys/kernel/debug/tracing/trace_clock\n"
2744 "or add trace_clock=global to the kernel command line\n");
2745}
2746
2747static void rb_add_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2748 struct ring_buffer_event **event,
2749 struct rb_event_info *info,
2750 u64 *delta,
2751 unsigned int *length)
2752{
2753 bool abs = info->add_timestamp &
2754 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE);
2755
2756 if (unlikely(info->delta > (1ULL << 59))) {
2757 /* did the clock go backwards */
2758 if (info->before == info->after && info->before > info->ts) {
2759 /* not interrupted */
2760 static int once;
2761
2762 /*
2763 * This is possible with a recalibrating of the TSC.
2764 * Do not produce a call stack, but just report it.
2765 */
2766 if (!once) {
2767 once++;
2768 pr_warn("Ring buffer clock went backwards: %llu -> %llu\n",
2769 info->before, info->ts);
2770 }
2771 } else
2772 rb_check_timestamp(cpu_buffer, info);
2773 if (!abs)
2774 info->delta = 0;
2775 }
2776 *event = rb_add_time_stamp(*event, info->delta, abs);
2777 *length -= RB_LEN_TIME_EXTEND;
2778 *delta = 0;
2779}
2780
2781/**
2782 * rb_update_event - update event type and data
2783 * @cpu_buffer: The per cpu buffer of the @event
2784 * @event: the event to update
2785 * @info: The info to update the @event with (contains length and delta)
2786 *
2787 * Update the type and data fields of the @event. The length
2788 * is the actual size that is written to the ring buffer,
2789 * and with this, we can determine what to place into the
2790 * data field.
2791 */
2792static void
2793rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
2794 struct ring_buffer_event *event,
2795 struct rb_event_info *info)
2796{
2797 unsigned length = info->length;
2798 u64 delta = info->delta;
2799 unsigned int nest = local_read(&cpu_buffer->committing) - 1;
2800
2801 if (!WARN_ON_ONCE(nest >= MAX_NEST))
2802 cpu_buffer->event_stamp[nest] = info->ts;
2803
2804 /*
2805 * If we need to add a timestamp, then we
2806 * add it to the start of the reserved space.
2807 */
2808 if (unlikely(info->add_timestamp))
2809 rb_add_timestamp(cpu_buffer, &event, info, &delta, &length);
2810
2811 event->time_delta = delta;
2812 length -= RB_EVNT_HDR_SIZE;
2813 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
2814 event->type_len = 0;
2815 event->array[0] = length;
2816 } else
2817 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
2818}
2819
2820static unsigned rb_calculate_event_length(unsigned length)
2821{
2822 struct ring_buffer_event event; /* Used only for sizeof array */
2823
2824 /* zero length can cause confusions */
2825 if (!length)
2826 length++;
2827
2828 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
2829 length += sizeof(event.array[0]);
2830
2831 length += RB_EVNT_HDR_SIZE;
2832 length = ALIGN(length, RB_ARCH_ALIGNMENT);
2833
2834 /*
2835 * In case the time delta is larger than the 27 bits for it
2836 * in the header, we need to add a timestamp. If another
2837 * event comes in when trying to discard this one to increase
2838 * the length, then the timestamp will be added in the allocated
2839 * space of this event. If length is bigger than the size needed
2840 * for the TIME_EXTEND, then padding has to be used. The events
2841 * length must be either RB_LEN_TIME_EXTEND, or greater than or equal
2842 * to RB_LEN_TIME_EXTEND + 8, as 8 is the minimum size for padding.
2843 * As length is a multiple of 4, we only need to worry if it
2844 * is 12 (RB_LEN_TIME_EXTEND + 4).
2845 */
2846 if (length == RB_LEN_TIME_EXTEND + RB_ALIGNMENT)
2847 length += RB_ALIGNMENT;
2848
2849 return length;
2850}
2851
2852static u64 rb_time_delta(struct ring_buffer_event *event)
2853{
2854 switch (event->type_len) {
2855 case RINGBUF_TYPE_PADDING:
2856 return 0;
2857
2858 case RINGBUF_TYPE_TIME_EXTEND:
2859 return rb_event_time_stamp(event);
2860
2861 case RINGBUF_TYPE_TIME_STAMP:
2862 return 0;
2863
2864 case RINGBUF_TYPE_DATA:
2865 return event->time_delta;
2866 default:
2867 return 0;
2868 }
2869}
2870
2871static inline int
2872rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2873 struct ring_buffer_event *event)
2874{
2875 unsigned long new_index, old_index;
2876 struct buffer_page *bpage;
2877 unsigned long index;
2878 unsigned long addr;
2879 u64 write_stamp;
2880 u64 delta;
2881
2882 new_index = rb_event_index(event);
2883 old_index = new_index + rb_event_ts_length(event);
2884 addr = (unsigned long)event;
2885 addr &= PAGE_MASK;
2886
2887 bpage = READ_ONCE(cpu_buffer->tail_page);
2888
2889 delta = rb_time_delta(event);
2890
2891 if (!rb_time_read(&cpu_buffer->write_stamp, &write_stamp))
2892 return 0;
2893
2894 /* Make sure the write stamp is read before testing the location */
2895 barrier();
2896
2897 if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2898 unsigned long write_mask =
2899 local_read(&bpage->write) & ~RB_WRITE_MASK;
2900 unsigned long event_length = rb_event_length(event);
2901
2902 /* Something came in, can't discard */
2903 if (!rb_time_cmpxchg(&cpu_buffer->write_stamp,
2904 write_stamp, write_stamp - delta))
2905 return 0;
2906
2907 /*
2908 * It's possible that the event time delta is zero
2909 * (has the same time stamp as the previous event)
2910 * in which case write_stamp and before_stamp could
2911 * be the same. In such a case, force before_stamp
2912 * to be different than write_stamp. It doesn't
2913 * matter what it is, as long as its different.
2914 */
2915 if (!delta)
2916 rb_time_set(&cpu_buffer->before_stamp, 0);
2917
2918 /*
2919 * If an event were to come in now, it would see that the
2920 * write_stamp and the before_stamp are different, and assume
2921 * that this event just added itself before updating
2922 * the write stamp. The interrupting event will fix the
2923 * write stamp for us, and use the before stamp as its delta.
2924 */
2925
2926 /*
2927 * This is on the tail page. It is possible that
2928 * a write could come in and move the tail page
2929 * and write to the next page. That is fine
2930 * because we just shorten what is on this page.
2931 */
2932 old_index += write_mask;
2933 new_index += write_mask;
2934 index = local_cmpxchg(&bpage->write, old_index, new_index);
2935 if (index == old_index) {
2936 /* update counters */
2937 local_sub(event_length, &cpu_buffer->entries_bytes);
2938 return 1;
2939 }
2940 }
2941
2942 /* could not discard */
2943 return 0;
2944}
2945
2946static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2947{
2948 local_inc(&cpu_buffer->committing);
2949 local_inc(&cpu_buffer->commits);
2950}
2951
2952static __always_inline void
2953rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
2954{
2955 unsigned long max_count;
2956
2957 /*
2958 * We only race with interrupts and NMIs on this CPU.
2959 * If we own the commit event, then we can commit
2960 * all others that interrupted us, since the interruptions
2961 * are in stack format (they finish before they come
2962 * back to us). This allows us to do a simple loop to
2963 * assign the commit to the tail.
2964 */
2965 again:
2966 max_count = cpu_buffer->nr_pages * 100;
2967
2968 while (cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)) {
2969 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
2970 return;
2971 if (RB_WARN_ON(cpu_buffer,
2972 rb_is_reader_page(cpu_buffer->tail_page)))
2973 return;
2974 local_set(&cpu_buffer->commit_page->page->commit,
2975 rb_page_write(cpu_buffer->commit_page));
2976 rb_inc_page(&cpu_buffer->commit_page);
2977 /* add barrier to keep gcc from optimizing too much */
2978 barrier();
2979 }
2980 while (rb_commit_index(cpu_buffer) !=
2981 rb_page_write(cpu_buffer->commit_page)) {
2982
2983 local_set(&cpu_buffer->commit_page->page->commit,
2984 rb_page_write(cpu_buffer->commit_page));
2985 RB_WARN_ON(cpu_buffer,
2986 local_read(&cpu_buffer->commit_page->page->commit) &
2987 ~RB_WRITE_MASK);
2988 barrier();
2989 }
2990
2991 /* again, keep gcc from optimizing */
2992 barrier();
2993
2994 /*
2995 * If an interrupt came in just after the first while loop
2996 * and pushed the tail page forward, we will be left with
2997 * a dangling commit that will never go forward.
2998 */
2999 if (unlikely(cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)))
3000 goto again;
3001}
3002
3003static __always_inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
3004{
3005 unsigned long commits;
3006
3007 if (RB_WARN_ON(cpu_buffer,
3008 !local_read(&cpu_buffer->committing)))
3009 return;
3010
3011 again:
3012 commits = local_read(&cpu_buffer->commits);
3013 /* synchronize with interrupts */
3014 barrier();
3015 if (local_read(&cpu_buffer->committing) == 1)
3016 rb_set_commit_to_write(cpu_buffer);
3017
3018 local_dec(&cpu_buffer->committing);
3019
3020 /* synchronize with interrupts */
3021 barrier();
3022
3023 /*
3024 * Need to account for interrupts coming in between the
3025 * updating of the commit page and the clearing of the
3026 * committing counter.
3027 */
3028 if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
3029 !local_read(&cpu_buffer->committing)) {
3030 local_inc(&cpu_buffer->committing);
3031 goto again;
3032 }
3033}
3034
3035static inline void rb_event_discard(struct ring_buffer_event *event)
3036{
3037 if (extended_time(event))
3038 event = skip_time_extend(event);
3039
3040 /* array[0] holds the actual length for the discarded event */
3041 event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
3042 event->type_len = RINGBUF_TYPE_PADDING;
3043 /* time delta must be non zero */
3044 if (!event->time_delta)
3045 event->time_delta = 1;
3046}
3047
3048static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
3049 struct ring_buffer_event *event)
3050{
3051 local_inc(&cpu_buffer->entries);
3052 rb_end_commit(cpu_buffer);
3053}
3054
3055static __always_inline void
3056rb_wakeups(struct trace_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer)
3057{
3058 size_t nr_pages;
3059 size_t dirty;
3060 size_t full;
3061
3062 if (buffer->irq_work.waiters_pending) {
3063 buffer->irq_work.waiters_pending = false;
3064 /* irq_work_queue() supplies it's own memory barriers */
3065 irq_work_queue(&buffer->irq_work.work);
3066 }
3067
3068 if (cpu_buffer->irq_work.waiters_pending) {
3069 cpu_buffer->irq_work.waiters_pending = false;
3070 /* irq_work_queue() supplies it's own memory barriers */
3071 irq_work_queue(&cpu_buffer->irq_work.work);
3072 }
3073
3074 if (cpu_buffer->last_pages_touch == local_read(&cpu_buffer->pages_touched))
3075 return;
3076
3077 if (cpu_buffer->reader_page == cpu_buffer->commit_page)
3078 return;
3079
3080 if (!cpu_buffer->irq_work.full_waiters_pending)
3081 return;
3082
3083 cpu_buffer->last_pages_touch = local_read(&cpu_buffer->pages_touched);
3084
3085 full = cpu_buffer->shortest_full;
3086 nr_pages = cpu_buffer->nr_pages;
3087 dirty = ring_buffer_nr_dirty_pages(buffer, cpu_buffer->cpu);
3088 if (full && nr_pages && (dirty * 100) <= full * nr_pages)
3089 return;
3090
3091 cpu_buffer->irq_work.wakeup_full = true;
3092 cpu_buffer->irq_work.full_waiters_pending = false;
3093 /* irq_work_queue() supplies it's own memory barriers */
3094 irq_work_queue(&cpu_buffer->irq_work.work);
3095}
3096
3097#ifdef CONFIG_RING_BUFFER_RECORD_RECURSION
3098# define do_ring_buffer_record_recursion() \
3099 do_ftrace_record_recursion(_THIS_IP_, _RET_IP_)
3100#else
3101# define do_ring_buffer_record_recursion() do { } while (0)
3102#endif
3103
3104/*
3105 * The lock and unlock are done within a preempt disable section.
3106 * The current_context per_cpu variable can only be modified
3107 * by the current task between lock and unlock. But it can
3108 * be modified more than once via an interrupt. To pass this
3109 * information from the lock to the unlock without having to
3110 * access the 'in_interrupt()' functions again (which do show
3111 * a bit of overhead in something as critical as function tracing,
3112 * we use a bitmask trick.
3113 *
3114 * bit 1 = NMI context
3115 * bit 2 = IRQ context
3116 * bit 3 = SoftIRQ context
3117 * bit 4 = normal context.
3118 *
3119 * This works because this is the order of contexts that can
3120 * preempt other contexts. A SoftIRQ never preempts an IRQ
3121 * context.
3122 *
3123 * When the context is determined, the corresponding bit is
3124 * checked and set (if it was set, then a recursion of that context
3125 * happened).
3126 *
3127 * On unlock, we need to clear this bit. To do so, just subtract
3128 * 1 from the current_context and AND it to itself.
3129 *
3130 * (binary)
3131 * 101 - 1 = 100
3132 * 101 & 100 = 100 (clearing bit zero)
3133 *
3134 * 1010 - 1 = 1001
3135 * 1010 & 1001 = 1000 (clearing bit 1)
3136 *
3137 * The least significant bit can be cleared this way, and it
3138 * just so happens that it is the same bit corresponding to
3139 * the current context.
3140 *
3141 * Now the TRANSITION bit breaks the above slightly. The TRANSITION bit
3142 * is set when a recursion is detected at the current context, and if
3143 * the TRANSITION bit is already set, it will fail the recursion.
3144 * This is needed because there's a lag between the changing of
3145 * interrupt context and updating the preempt count. In this case,
3146 * a false positive will be found. To handle this, one extra recursion
3147 * is allowed, and this is done by the TRANSITION bit. If the TRANSITION
3148 * bit is already set, then it is considered a recursion and the function
3149 * ends. Otherwise, the TRANSITION bit is set, and that bit is returned.
3150 *
3151 * On the trace_recursive_unlock(), the TRANSITION bit will be the first
3152 * to be cleared. Even if it wasn't the context that set it. That is,
3153 * if an interrupt comes in while NORMAL bit is set and the ring buffer
3154 * is called before preempt_count() is updated, since the check will
3155 * be on the NORMAL bit, the TRANSITION bit will then be set. If an
3156 * NMI then comes in, it will set the NMI bit, but when the NMI code
3157 * does the trace_recursive_unlock() it will clear the TRANSITION bit
3158 * and leave the NMI bit set. But this is fine, because the interrupt
3159 * code that set the TRANSITION bit will then clear the NMI bit when it
3160 * calls trace_recursive_unlock(). If another NMI comes in, it will
3161 * set the TRANSITION bit and continue.
3162 *
3163 * Note: The TRANSITION bit only handles a single transition between context.
3164 */
3165
3166static __always_inline int
3167trace_recursive_lock(struct ring_buffer_per_cpu *cpu_buffer)
3168{
3169 unsigned int val = cpu_buffer->current_context;
3170 unsigned long pc = preempt_count();
3171 int bit;
3172
3173 if (!(pc & (NMI_MASK | HARDIRQ_MASK | SOFTIRQ_OFFSET)))
3174 bit = RB_CTX_NORMAL;
3175 else
3176 bit = pc & NMI_MASK ? RB_CTX_NMI :
3177 pc & HARDIRQ_MASK ? RB_CTX_IRQ : RB_CTX_SOFTIRQ;
3178
3179 if (unlikely(val & (1 << (bit + cpu_buffer->nest)))) {
3180 /*
3181 * It is possible that this was called by transitioning
3182 * between interrupt context, and preempt_count() has not
3183 * been updated yet. In this case, use the TRANSITION bit.
3184 */
3185 bit = RB_CTX_TRANSITION;
3186 if (val & (1 << (bit + cpu_buffer->nest))) {
3187 do_ring_buffer_record_recursion();
3188 return 1;
3189 }
3190 }
3191
3192 val |= (1 << (bit + cpu_buffer->nest));
3193 cpu_buffer->current_context = val;
3194
3195 return 0;
3196}
3197
3198static __always_inline void
3199trace_recursive_unlock(struct ring_buffer_per_cpu *cpu_buffer)
3200{
3201 cpu_buffer->current_context &=
3202 cpu_buffer->current_context - (1 << cpu_buffer->nest);
3203}
3204
3205/* The recursive locking above uses 5 bits */
3206#define NESTED_BITS 5
3207
3208/**
3209 * ring_buffer_nest_start - Allow to trace while nested
3210 * @buffer: The ring buffer to modify
3211 *
3212 * The ring buffer has a safety mechanism to prevent recursion.
3213 * But there may be a case where a trace needs to be done while
3214 * tracing something else. In this case, calling this function
3215 * will allow this function to nest within a currently active
3216 * ring_buffer_lock_reserve().
3217 *
3218 * Call this function before calling another ring_buffer_lock_reserve() and
3219 * call ring_buffer_nest_end() after the nested ring_buffer_unlock_commit().
3220 */
3221void ring_buffer_nest_start(struct trace_buffer *buffer)
3222{
3223 struct ring_buffer_per_cpu *cpu_buffer;
3224 int cpu;
3225
3226 /* Enabled by ring_buffer_nest_end() */
3227 preempt_disable_notrace();
3228 cpu = raw_smp_processor_id();
3229 cpu_buffer = buffer->buffers[cpu];
3230 /* This is the shift value for the above recursive locking */
3231 cpu_buffer->nest += NESTED_BITS;
3232}
3233
3234/**
3235 * ring_buffer_nest_end - Allow to trace while nested
3236 * @buffer: The ring buffer to modify
3237 *
3238 * Must be called after ring_buffer_nest_start() and after the
3239 * ring_buffer_unlock_commit().
3240 */
3241void ring_buffer_nest_end(struct trace_buffer *buffer)
3242{
3243 struct ring_buffer_per_cpu *cpu_buffer;
3244 int cpu;
3245
3246 /* disabled by ring_buffer_nest_start() */
3247 cpu = raw_smp_processor_id();
3248 cpu_buffer = buffer->buffers[cpu];
3249 /* This is the shift value for the above recursive locking */
3250 cpu_buffer->nest -= NESTED_BITS;
3251 preempt_enable_notrace();
3252}
3253
3254/**
3255 * ring_buffer_unlock_commit - commit a reserved
3256 * @buffer: The buffer to commit to
3257 * @event: The event pointer to commit.
3258 *
3259 * This commits the data to the ring buffer, and releases any locks held.
3260 *
3261 * Must be paired with ring_buffer_lock_reserve.
3262 */
3263int ring_buffer_unlock_commit(struct trace_buffer *buffer,
3264 struct ring_buffer_event *event)
3265{
3266 struct ring_buffer_per_cpu *cpu_buffer;
3267 int cpu = raw_smp_processor_id();
3268
3269 cpu_buffer = buffer->buffers[cpu];
3270
3271 rb_commit(cpu_buffer, event);
3272
3273 rb_wakeups(buffer, cpu_buffer);
3274
3275 trace_recursive_unlock(cpu_buffer);
3276
3277 preempt_enable_notrace();
3278
3279 return 0;
3280}
3281EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
3282
3283/* Special value to validate all deltas on a page. */
3284#define CHECK_FULL_PAGE 1L
3285
3286#ifdef CONFIG_RING_BUFFER_VALIDATE_TIME_DELTAS
3287static void dump_buffer_page(struct buffer_data_page *bpage,
3288 struct rb_event_info *info,
3289 unsigned long tail)
3290{
3291 struct ring_buffer_event *event;
3292 u64 ts, delta;
3293 int e;
3294
3295 ts = bpage->time_stamp;
3296 pr_warn(" [%lld] PAGE TIME STAMP\n", ts);
3297
3298 for (e = 0; e < tail; e += rb_event_length(event)) {
3299
3300 event = (struct ring_buffer_event *)(bpage->data + e);
3301
3302 switch (event->type_len) {
3303
3304 case RINGBUF_TYPE_TIME_EXTEND:
3305 delta = rb_event_time_stamp(event);
3306 ts += delta;
3307 pr_warn(" [%lld] delta:%lld TIME EXTEND\n", ts, delta);
3308 break;
3309
3310 case RINGBUF_TYPE_TIME_STAMP:
3311 delta = rb_event_time_stamp(event);
3312 ts = delta;
3313 pr_warn(" [%lld] absolute:%lld TIME STAMP\n", ts, delta);
3314 break;
3315
3316 case RINGBUF_TYPE_PADDING:
3317 ts += event->time_delta;
3318 pr_warn(" [%lld] delta:%d PADDING\n", ts, event->time_delta);
3319 break;
3320
3321 case RINGBUF_TYPE_DATA:
3322 ts += event->time_delta;
3323 pr_warn(" [%lld] delta:%d\n", ts, event->time_delta);
3324 break;
3325
3326 default:
3327 break;
3328 }
3329 }
3330}
3331
3332static DEFINE_PER_CPU(atomic_t, checking);
3333static atomic_t ts_dump;
3334
3335/*
3336 * Check if the current event time stamp matches the deltas on
3337 * the buffer page.
3338 */
3339static void check_buffer(struct ring_buffer_per_cpu *cpu_buffer,
3340 struct rb_event_info *info,
3341 unsigned long tail)
3342{
3343 struct ring_buffer_event *event;
3344 struct buffer_data_page *bpage;
3345 u64 ts, delta;
3346 bool full = false;
3347 int e;
3348
3349 bpage = info->tail_page->page;
3350
3351 if (tail == CHECK_FULL_PAGE) {
3352 full = true;
3353 tail = local_read(&bpage->commit);
3354 } else if (info->add_timestamp &
3355 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE)) {
3356 /* Ignore events with absolute time stamps */
3357 return;
3358 }
3359
3360 /*
3361 * Do not check the first event (skip possible extends too).
3362 * Also do not check if previous events have not been committed.
3363 */
3364 if (tail <= 8 || tail > local_read(&bpage->commit))
3365 return;
3366
3367 /*
3368 * If this interrupted another event,
3369 */
3370 if (atomic_inc_return(this_cpu_ptr(&checking)) != 1)
3371 goto out;
3372
3373 ts = bpage->time_stamp;
3374
3375 for (e = 0; e < tail; e += rb_event_length(event)) {
3376
3377 event = (struct ring_buffer_event *)(bpage->data + e);
3378
3379 switch (event->type_len) {
3380
3381 case RINGBUF_TYPE_TIME_EXTEND:
3382 delta = rb_event_time_stamp(event);
3383 ts += delta;
3384 break;
3385
3386 case RINGBUF_TYPE_TIME_STAMP:
3387 delta = rb_event_time_stamp(event);
3388 ts = delta;
3389 break;
3390
3391 case RINGBUF_TYPE_PADDING:
3392 if (event->time_delta == 1)
3393 break;
3394 fallthrough;
3395 case RINGBUF_TYPE_DATA:
3396 ts += event->time_delta;
3397 break;
3398
3399 default:
3400 RB_WARN_ON(cpu_buffer, 1);
3401 }
3402 }
3403 if ((full && ts > info->ts) ||
3404 (!full && ts + info->delta != info->ts)) {
3405 /* If another report is happening, ignore this one */
3406 if (atomic_inc_return(&ts_dump) != 1) {
3407 atomic_dec(&ts_dump);
3408 goto out;
3409 }
3410 atomic_inc(&cpu_buffer->record_disabled);
3411 /* There's some cases in boot up that this can happen */
3412 WARN_ON_ONCE(system_state != SYSTEM_BOOTING);
3413 pr_warn("[CPU: %d]TIME DOES NOT MATCH expected:%lld actual:%lld delta:%lld before:%lld after:%lld%s\n",
3414 cpu_buffer->cpu,
3415 ts + info->delta, info->ts, info->delta,
3416 info->before, info->after,
3417 full ? " (full)" : "");
3418 dump_buffer_page(bpage, info, tail);
3419 atomic_dec(&ts_dump);
3420 /* Do not re-enable checking */
3421 return;
3422 }
3423out:
3424 atomic_dec(this_cpu_ptr(&checking));
3425}
3426#else
3427static inline void check_buffer(struct ring_buffer_per_cpu *cpu_buffer,
3428 struct rb_event_info *info,
3429 unsigned long tail)
3430{
3431}
3432#endif /* CONFIG_RING_BUFFER_VALIDATE_TIME_DELTAS */
3433
3434static struct ring_buffer_event *
3435__rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
3436 struct rb_event_info *info)
3437{
3438 struct ring_buffer_event *event;
3439 struct buffer_page *tail_page;
3440 unsigned long tail, write, w;
3441 bool a_ok;
3442 bool b_ok;
3443
3444 /* Don't let the compiler play games with cpu_buffer->tail_page */
3445 tail_page = info->tail_page = READ_ONCE(cpu_buffer->tail_page);
3446
3447 /*A*/ w = local_read(&tail_page->write) & RB_WRITE_MASK;
3448 barrier();
3449 b_ok = rb_time_read(&cpu_buffer->before_stamp, &info->before);
3450 a_ok = rb_time_read(&cpu_buffer->write_stamp, &info->after);
3451 barrier();
3452 info->ts = rb_time_stamp(cpu_buffer->buffer);
3453
3454 if ((info->add_timestamp & RB_ADD_STAMP_ABSOLUTE)) {
3455 info->delta = info->ts;
3456 } else {
3457 /*
3458 * If interrupting an event time update, we may need an
3459 * absolute timestamp.
3460 * Don't bother if this is the start of a new page (w == 0).
3461 */
3462 if (unlikely(!a_ok || !b_ok || (info->before != info->after && w))) {
3463 info->add_timestamp |= RB_ADD_STAMP_FORCE | RB_ADD_STAMP_EXTEND;
3464 info->length += RB_LEN_TIME_EXTEND;
3465 } else {
3466 info->delta = info->ts - info->after;
3467 if (unlikely(test_time_stamp(info->delta))) {
3468 info->add_timestamp |= RB_ADD_STAMP_EXTEND;
3469 info->length += RB_LEN_TIME_EXTEND;
3470 }
3471 }
3472 }
3473
3474 /*B*/ rb_time_set(&cpu_buffer->before_stamp, info->ts);
3475
3476 /*C*/ write = local_add_return(info->length, &tail_page->write);
3477
3478 /* set write to only the index of the write */
3479 write &= RB_WRITE_MASK;
3480
3481 tail = write - info->length;
3482
3483 /* See if we shot pass the end of this buffer page */
3484 if (unlikely(write > BUF_PAGE_SIZE)) {
3485 /* before and after may now different, fix it up*/
3486 b_ok = rb_time_read(&cpu_buffer->before_stamp, &info->before);
3487 a_ok = rb_time_read(&cpu_buffer->write_stamp, &info->after);
3488 if (a_ok && b_ok && info->before != info->after)
3489 (void)rb_time_cmpxchg(&cpu_buffer->before_stamp,
3490 info->before, info->after);
3491 if (a_ok && b_ok)
3492 check_buffer(cpu_buffer, info, CHECK_FULL_PAGE);
3493 return rb_move_tail(cpu_buffer, tail, info);
3494 }
3495
3496 if (likely(tail == w)) {
3497 u64 save_before;
3498 bool s_ok;
3499
3500 /* Nothing interrupted us between A and C */
3501 /*D*/ rb_time_set(&cpu_buffer->write_stamp, info->ts);
3502 barrier();
3503 /*E*/ s_ok = rb_time_read(&cpu_buffer->before_stamp, &save_before);
3504 RB_WARN_ON(cpu_buffer, !s_ok);
3505 if (likely(!(info->add_timestamp &
3506 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE))))
3507 /* This did not interrupt any time update */
3508 info->delta = info->ts - info->after;
3509 else
3510 /* Just use full timestamp for interrupting event */
3511 info->delta = info->ts;
3512 barrier();
3513 check_buffer(cpu_buffer, info, tail);
3514 if (unlikely(info->ts != save_before)) {
3515 /* SLOW PATH - Interrupted between C and E */
3516
3517 a_ok = rb_time_read(&cpu_buffer->write_stamp, &info->after);
3518 RB_WARN_ON(cpu_buffer, !a_ok);
3519
3520 /* Write stamp must only go forward */
3521 if (save_before > info->after) {
3522 /*
3523 * We do not care about the result, only that
3524 * it gets updated atomically.
3525 */
3526 (void)rb_time_cmpxchg(&cpu_buffer->write_stamp,
3527 info->after, save_before);
3528 }
3529 }
3530 } else {
3531 u64 ts;
3532 /* SLOW PATH - Interrupted between A and C */
3533 a_ok = rb_time_read(&cpu_buffer->write_stamp, &info->after);
3534 /* Was interrupted before here, write_stamp must be valid */
3535 RB_WARN_ON(cpu_buffer, !a_ok);
3536 ts = rb_time_stamp(cpu_buffer->buffer);
3537 barrier();
3538 /*E*/ if (write == (local_read(&tail_page->write) & RB_WRITE_MASK) &&
3539 info->after < ts &&
3540 rb_time_cmpxchg(&cpu_buffer->write_stamp,
3541 info->after, ts)) {
3542 /* Nothing came after this event between C and E */
3543 info->delta = ts - info->after;
3544 } else {
3545 /*
3546 * Interrupted between C and E:
3547 * Lost the previous events time stamp. Just set the
3548 * delta to zero, and this will be the same time as
3549 * the event this event interrupted. And the events that
3550 * came after this will still be correct (as they would
3551 * have built their delta on the previous event.
3552 */
3553 info->delta = 0;
3554 }
3555 info->ts = ts;
3556 info->add_timestamp &= ~RB_ADD_STAMP_FORCE;
3557 }
3558
3559 /*
3560 * If this is the first commit on the page, then it has the same
3561 * timestamp as the page itself.
3562 */
3563 if (unlikely(!tail && !(info->add_timestamp &
3564 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE))))
3565 info->delta = 0;
3566
3567 /* We reserved something on the buffer */
3568
3569 event = __rb_page_index(tail_page, tail);
3570 rb_update_event(cpu_buffer, event, info);
3571
3572 local_inc(&tail_page->entries);
3573
3574 /*
3575 * If this is the first commit on the page, then update
3576 * its timestamp.
3577 */
3578 if (unlikely(!tail))
3579 tail_page->page->time_stamp = info->ts;
3580
3581 /* account for these added bytes */
3582 local_add(info->length, &cpu_buffer->entries_bytes);
3583
3584 return event;
3585}
3586
3587static __always_inline struct ring_buffer_event *
3588rb_reserve_next_event(struct trace_buffer *buffer,
3589 struct ring_buffer_per_cpu *cpu_buffer,
3590 unsigned long length)
3591{
3592 struct ring_buffer_event *event;
3593 struct rb_event_info info;
3594 int nr_loops = 0;
3595 int add_ts_default;
3596
3597 rb_start_commit(cpu_buffer);
3598 /* The commit page can not change after this */
3599
3600#ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
3601 /*
3602 * Due to the ability to swap a cpu buffer from a buffer
3603 * it is possible it was swapped before we committed.
3604 * (committing stops a swap). We check for it here and
3605 * if it happened, we have to fail the write.
3606 */
3607 barrier();
3608 if (unlikely(READ_ONCE(cpu_buffer->buffer) != buffer)) {
3609 local_dec(&cpu_buffer->committing);
3610 local_dec(&cpu_buffer->commits);
3611 return NULL;
3612 }
3613#endif
3614
3615 info.length = rb_calculate_event_length(length);
3616
3617 if (ring_buffer_time_stamp_abs(cpu_buffer->buffer)) {
3618 add_ts_default = RB_ADD_STAMP_ABSOLUTE;
3619 info.length += RB_LEN_TIME_EXTEND;
3620 } else {
3621 add_ts_default = RB_ADD_STAMP_NONE;
3622 }
3623
3624 again:
3625 info.add_timestamp = add_ts_default;
3626 info.delta = 0;
3627
3628 /*
3629 * We allow for interrupts to reenter here and do a trace.
3630 * If one does, it will cause this original code to loop
3631 * back here. Even with heavy interrupts happening, this
3632 * should only happen a few times in a row. If this happens
3633 * 1000 times in a row, there must be either an interrupt
3634 * storm or we have something buggy.
3635 * Bail!
3636 */
3637 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
3638 goto out_fail;
3639
3640 event = __rb_reserve_next(cpu_buffer, &info);
3641
3642 if (unlikely(PTR_ERR(event) == -EAGAIN)) {
3643 if (info.add_timestamp & (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_EXTEND))
3644 info.length -= RB_LEN_TIME_EXTEND;
3645 goto again;
3646 }
3647
3648 if (likely(event))
3649 return event;
3650 out_fail:
3651 rb_end_commit(cpu_buffer);
3652 return NULL;
3653}
3654
3655/**
3656 * ring_buffer_lock_reserve - reserve a part of the buffer
3657 * @buffer: the ring buffer to reserve from
3658 * @length: the length of the data to reserve (excluding event header)
3659 *
3660 * Returns a reserved event on the ring buffer to copy directly to.
3661 * The user of this interface will need to get the body to write into
3662 * and can use the ring_buffer_event_data() interface.
3663 *
3664 * The length is the length of the data needed, not the event length
3665 * which also includes the event header.
3666 *
3667 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
3668 * If NULL is returned, then nothing has been allocated or locked.
3669 */
3670struct ring_buffer_event *
3671ring_buffer_lock_reserve(struct trace_buffer *buffer, unsigned long length)
3672{
3673 struct ring_buffer_per_cpu *cpu_buffer;
3674 struct ring_buffer_event *event;
3675 int cpu;
3676
3677 /* If we are tracing schedule, we don't want to recurse */
3678 preempt_disable_notrace();
3679
3680 if (unlikely(atomic_read(&buffer->record_disabled)))
3681 goto out;
3682
3683 cpu = raw_smp_processor_id();
3684
3685 if (unlikely(!cpumask_test_cpu(cpu, buffer->cpumask)))
3686 goto out;
3687
3688 cpu_buffer = buffer->buffers[cpu];
3689
3690 if (unlikely(atomic_read(&cpu_buffer->record_disabled)))
3691 goto out;
3692
3693 if (unlikely(length > BUF_MAX_DATA_SIZE))
3694 goto out;
3695
3696 if (unlikely(trace_recursive_lock(cpu_buffer)))
3697 goto out;
3698
3699 event = rb_reserve_next_event(buffer, cpu_buffer, length);
3700 if (!event)
3701 goto out_unlock;
3702
3703 return event;
3704
3705 out_unlock:
3706 trace_recursive_unlock(cpu_buffer);
3707 out:
3708 preempt_enable_notrace();
3709 return NULL;
3710}
3711EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
3712
3713/*
3714 * Decrement the entries to the page that an event is on.
3715 * The event does not even need to exist, only the pointer
3716 * to the page it is on. This may only be called before the commit
3717 * takes place.
3718 */
3719static inline void
3720rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
3721 struct ring_buffer_event *event)
3722{
3723 unsigned long addr = (unsigned long)event;
3724 struct buffer_page *bpage = cpu_buffer->commit_page;
3725 struct buffer_page *start;
3726
3727 addr &= PAGE_MASK;
3728
3729 /* Do the likely case first */
3730 if (likely(bpage->page == (void *)addr)) {
3731 local_dec(&bpage->entries);
3732 return;
3733 }
3734
3735 /*
3736 * Because the commit page may be on the reader page we
3737 * start with the next page and check the end loop there.
3738 */
3739 rb_inc_page(&bpage);
3740 start = bpage;
3741 do {
3742 if (bpage->page == (void *)addr) {
3743 local_dec(&bpage->entries);
3744 return;
3745 }
3746 rb_inc_page(&bpage);
3747 } while (bpage != start);
3748
3749 /* commit not part of this buffer?? */
3750 RB_WARN_ON(cpu_buffer, 1);
3751}
3752
3753/**
3754 * ring_buffer_discard_commit - discard an event that has not been committed
3755 * @buffer: the ring buffer
3756 * @event: non committed event to discard
3757 *
3758 * Sometimes an event that is in the ring buffer needs to be ignored.
3759 * This function lets the user discard an event in the ring buffer
3760 * and then that event will not be read later.
3761 *
3762 * This function only works if it is called before the item has been
3763 * committed. It will try to free the event from the ring buffer
3764 * if another event has not been added behind it.
3765 *
3766 * If another event has been added behind it, it will set the event
3767 * up as discarded, and perform the commit.
3768 *
3769 * If this function is called, do not call ring_buffer_unlock_commit on
3770 * the event.
3771 */
3772void ring_buffer_discard_commit(struct trace_buffer *buffer,
3773 struct ring_buffer_event *event)
3774{
3775 struct ring_buffer_per_cpu *cpu_buffer;
3776 int cpu;
3777
3778 /* The event is discarded regardless */
3779 rb_event_discard(event);
3780
3781 cpu = smp_processor_id();
3782 cpu_buffer = buffer->buffers[cpu];
3783
3784 /*
3785 * This must only be called if the event has not been
3786 * committed yet. Thus we can assume that preemption
3787 * is still disabled.
3788 */
3789 RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
3790
3791 rb_decrement_entry(cpu_buffer, event);
3792 if (rb_try_to_discard(cpu_buffer, event))
3793 goto out;
3794
3795 out:
3796 rb_end_commit(cpu_buffer);
3797
3798 trace_recursive_unlock(cpu_buffer);
3799
3800 preempt_enable_notrace();
3801
3802}
3803EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
3804
3805/**
3806 * ring_buffer_write - write data to the buffer without reserving
3807 * @buffer: The ring buffer to write to.
3808 * @length: The length of the data being written (excluding the event header)
3809 * @data: The data to write to the buffer.
3810 *
3811 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
3812 * one function. If you already have the data to write to the buffer, it
3813 * may be easier to simply call this function.
3814 *
3815 * Note, like ring_buffer_lock_reserve, the length is the length of the data
3816 * and not the length of the event which would hold the header.
3817 */
3818int ring_buffer_write(struct trace_buffer *buffer,
3819 unsigned long length,
3820 void *data)
3821{
3822 struct ring_buffer_per_cpu *cpu_buffer;
3823 struct ring_buffer_event *event;
3824 void *body;
3825 int ret = -EBUSY;
3826 int cpu;
3827
3828 preempt_disable_notrace();
3829
3830 if (atomic_read(&buffer->record_disabled))
3831 goto out;
3832
3833 cpu = raw_smp_processor_id();
3834
3835 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3836 goto out;
3837
3838 cpu_buffer = buffer->buffers[cpu];
3839
3840 if (atomic_read(&cpu_buffer->record_disabled))
3841 goto out;
3842
3843 if (length > BUF_MAX_DATA_SIZE)
3844 goto out;
3845
3846 if (unlikely(trace_recursive_lock(cpu_buffer)))
3847 goto out;
3848
3849 event = rb_reserve_next_event(buffer, cpu_buffer, length);
3850 if (!event)
3851 goto out_unlock;
3852
3853 body = rb_event_data(event);
3854
3855 memcpy(body, data, length);
3856
3857 rb_commit(cpu_buffer, event);
3858
3859 rb_wakeups(buffer, cpu_buffer);
3860
3861 ret = 0;
3862
3863 out_unlock:
3864 trace_recursive_unlock(cpu_buffer);
3865
3866 out:
3867 preempt_enable_notrace();
3868
3869 return ret;
3870}
3871EXPORT_SYMBOL_GPL(ring_buffer_write);
3872
3873static bool rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
3874{
3875 struct buffer_page *reader = cpu_buffer->reader_page;
3876 struct buffer_page *head = rb_set_head_page(cpu_buffer);
3877 struct buffer_page *commit = cpu_buffer->commit_page;
3878
3879 /* In case of error, head will be NULL */
3880 if (unlikely(!head))
3881 return true;
3882
3883 /* Reader should exhaust content in reader page */
3884 if (reader->read != rb_page_commit(reader))
3885 return false;
3886
3887 /*
3888 * If writers are committing on the reader page, knowing all
3889 * committed content has been read, the ring buffer is empty.
3890 */
3891 if (commit == reader)
3892 return true;
3893
3894 /*
3895 * If writers are committing on a page other than reader page
3896 * and head page, there should always be content to read.
3897 */
3898 if (commit != head)
3899 return false;
3900
3901 /*
3902 * Writers are committing on the head page, we just need
3903 * to care about there're committed data, and the reader will
3904 * swap reader page with head page when it is to read data.
3905 */
3906 return rb_page_commit(commit) == 0;
3907}
3908
3909/**
3910 * ring_buffer_record_disable - stop all writes into the buffer
3911 * @buffer: The ring buffer to stop writes to.
3912 *
3913 * This prevents all writes to the buffer. Any attempt to write
3914 * to the buffer after this will fail and return NULL.
3915 *
3916 * The caller should call synchronize_rcu() after this.
3917 */
3918void ring_buffer_record_disable(struct trace_buffer *buffer)
3919{
3920 atomic_inc(&buffer->record_disabled);
3921}
3922EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
3923
3924/**
3925 * ring_buffer_record_enable - enable writes to the buffer
3926 * @buffer: The ring buffer to enable writes
3927 *
3928 * Note, multiple disables will need the same number of enables
3929 * to truly enable the writing (much like preempt_disable).
3930 */
3931void ring_buffer_record_enable(struct trace_buffer *buffer)
3932{
3933 atomic_dec(&buffer->record_disabled);
3934}
3935EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
3936
3937/**
3938 * ring_buffer_record_off - stop all writes into the buffer
3939 * @buffer: The ring buffer to stop writes to.
3940 *
3941 * This prevents all writes to the buffer. Any attempt to write
3942 * to the buffer after this will fail and return NULL.
3943 *
3944 * This is different than ring_buffer_record_disable() as
3945 * it works like an on/off switch, where as the disable() version
3946 * must be paired with a enable().
3947 */
3948void ring_buffer_record_off(struct trace_buffer *buffer)
3949{
3950 unsigned int rd;
3951 unsigned int new_rd;
3952
3953 do {
3954 rd = atomic_read(&buffer->record_disabled);
3955 new_rd = rd | RB_BUFFER_OFF;
3956 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3957}
3958EXPORT_SYMBOL_GPL(ring_buffer_record_off);
3959
3960/**
3961 * ring_buffer_record_on - restart writes into the buffer
3962 * @buffer: The ring buffer to start writes to.
3963 *
3964 * This enables all writes to the buffer that was disabled by
3965 * ring_buffer_record_off().
3966 *
3967 * This is different than ring_buffer_record_enable() as
3968 * it works like an on/off switch, where as the enable() version
3969 * must be paired with a disable().
3970 */
3971void ring_buffer_record_on(struct trace_buffer *buffer)
3972{
3973 unsigned int rd;
3974 unsigned int new_rd;
3975
3976 do {
3977 rd = atomic_read(&buffer->record_disabled);
3978 new_rd = rd & ~RB_BUFFER_OFF;
3979 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3980}
3981EXPORT_SYMBOL_GPL(ring_buffer_record_on);
3982
3983/**
3984 * ring_buffer_record_is_on - return true if the ring buffer can write
3985 * @buffer: The ring buffer to see if write is enabled
3986 *
3987 * Returns true if the ring buffer is in a state that it accepts writes.
3988 */
3989bool ring_buffer_record_is_on(struct trace_buffer *buffer)
3990{
3991 return !atomic_read(&buffer->record_disabled);
3992}
3993
3994/**
3995 * ring_buffer_record_is_set_on - return true if the ring buffer is set writable
3996 * @buffer: The ring buffer to see if write is set enabled
3997 *
3998 * Returns true if the ring buffer is set writable by ring_buffer_record_on().
3999 * Note that this does NOT mean it is in a writable state.
4000 *
4001 * It may return true when the ring buffer has been disabled by
4002 * ring_buffer_record_disable(), as that is a temporary disabling of
4003 * the ring buffer.
4004 */
4005bool ring_buffer_record_is_set_on(struct trace_buffer *buffer)
4006{
4007 return !(atomic_read(&buffer->record_disabled) & RB_BUFFER_OFF);
4008}
4009
4010/**
4011 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
4012 * @buffer: The ring buffer to stop writes to.
4013 * @cpu: The CPU buffer to stop
4014 *
4015 * This prevents all writes to the buffer. Any attempt to write
4016 * to the buffer after this will fail and return NULL.
4017 *
4018 * The caller should call synchronize_rcu() after this.
4019 */
4020void ring_buffer_record_disable_cpu(struct trace_buffer *buffer, int cpu)
4021{
4022 struct ring_buffer_per_cpu *cpu_buffer;
4023
4024 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4025 return;
4026
4027 cpu_buffer = buffer->buffers[cpu];
4028 atomic_inc(&cpu_buffer->record_disabled);
4029}
4030EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
4031
4032/**
4033 * ring_buffer_record_enable_cpu - enable writes to the buffer
4034 * @buffer: The ring buffer to enable writes
4035 * @cpu: The CPU to enable.
4036 *
4037 * Note, multiple disables will need the same number of enables
4038 * to truly enable the writing (much like preempt_disable).
4039 */
4040void ring_buffer_record_enable_cpu(struct trace_buffer *buffer, int cpu)
4041{
4042 struct ring_buffer_per_cpu *cpu_buffer;
4043
4044 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4045 return;
4046
4047 cpu_buffer = buffer->buffers[cpu];
4048 atomic_dec(&cpu_buffer->record_disabled);
4049}
4050EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
4051
4052/*
4053 * The total entries in the ring buffer is the running counter
4054 * of entries entered into the ring buffer, minus the sum of
4055 * the entries read from the ring buffer and the number of
4056 * entries that were overwritten.
4057 */
4058static inline unsigned long
4059rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
4060{
4061 return local_read(&cpu_buffer->entries) -
4062 (local_read(&cpu_buffer->overrun) + cpu_buffer->read);
4063}
4064
4065/**
4066 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
4067 * @buffer: The ring buffer
4068 * @cpu: The per CPU buffer to read from.
4069 */
4070u64 ring_buffer_oldest_event_ts(struct trace_buffer *buffer, int cpu)
4071{
4072 unsigned long flags;
4073 struct ring_buffer_per_cpu *cpu_buffer;
4074 struct buffer_page *bpage;
4075 u64 ret = 0;
4076
4077 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4078 return 0;
4079
4080 cpu_buffer = buffer->buffers[cpu];
4081 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4082 /*
4083 * if the tail is on reader_page, oldest time stamp is on the reader
4084 * page
4085 */
4086 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
4087 bpage = cpu_buffer->reader_page;
4088 else
4089 bpage = rb_set_head_page(cpu_buffer);
4090 if (bpage)
4091 ret = bpage->page->time_stamp;
4092 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4093
4094 return ret;
4095}
4096EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
4097
4098/**
4099 * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
4100 * @buffer: The ring buffer
4101 * @cpu: The per CPU buffer to read from.
4102 */
4103unsigned long ring_buffer_bytes_cpu(struct trace_buffer *buffer, int cpu)
4104{
4105 struct ring_buffer_per_cpu *cpu_buffer;
4106 unsigned long ret;
4107
4108 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4109 return 0;
4110
4111 cpu_buffer = buffer->buffers[cpu];
4112 ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
4113
4114 return ret;
4115}
4116EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
4117
4118/**
4119 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
4120 * @buffer: The ring buffer
4121 * @cpu: The per CPU buffer to get the entries from.
4122 */
4123unsigned long ring_buffer_entries_cpu(struct trace_buffer *buffer, int cpu)
4124{
4125 struct ring_buffer_per_cpu *cpu_buffer;
4126
4127 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4128 return 0;
4129
4130 cpu_buffer = buffer->buffers[cpu];
4131
4132 return rb_num_of_entries(cpu_buffer);
4133}
4134EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
4135
4136/**
4137 * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
4138 * buffer wrapping around (only if RB_FL_OVERWRITE is on).
4139 * @buffer: The ring buffer
4140 * @cpu: The per CPU buffer to get the number of overruns from
4141 */
4142unsigned long ring_buffer_overrun_cpu(struct trace_buffer *buffer, int cpu)
4143{
4144 struct ring_buffer_per_cpu *cpu_buffer;
4145 unsigned long ret;
4146
4147 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4148 return 0;
4149
4150 cpu_buffer = buffer->buffers[cpu];
4151 ret = local_read(&cpu_buffer->overrun);
4152
4153 return ret;
4154}
4155EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
4156
4157/**
4158 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
4159 * commits failing due to the buffer wrapping around while there are uncommitted
4160 * events, such as during an interrupt storm.
4161 * @buffer: The ring buffer
4162 * @cpu: The per CPU buffer to get the number of overruns from
4163 */
4164unsigned long
4165ring_buffer_commit_overrun_cpu(struct trace_buffer *buffer, int cpu)
4166{
4167 struct ring_buffer_per_cpu *cpu_buffer;
4168 unsigned long ret;
4169
4170 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4171 return 0;
4172
4173 cpu_buffer = buffer->buffers[cpu];
4174 ret = local_read(&cpu_buffer->commit_overrun);
4175
4176 return ret;
4177}
4178EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
4179
4180/**
4181 * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
4182 * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
4183 * @buffer: The ring buffer
4184 * @cpu: The per CPU buffer to get the number of overruns from
4185 */
4186unsigned long
4187ring_buffer_dropped_events_cpu(struct trace_buffer *buffer, int cpu)
4188{
4189 struct ring_buffer_per_cpu *cpu_buffer;
4190 unsigned long ret;
4191
4192 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4193 return 0;
4194
4195 cpu_buffer = buffer->buffers[cpu];
4196 ret = local_read(&cpu_buffer->dropped_events);
4197
4198 return ret;
4199}
4200EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
4201
4202/**
4203 * ring_buffer_read_events_cpu - get the number of events successfully read
4204 * @buffer: The ring buffer
4205 * @cpu: The per CPU buffer to get the number of events read
4206 */
4207unsigned long
4208ring_buffer_read_events_cpu(struct trace_buffer *buffer, int cpu)
4209{
4210 struct ring_buffer_per_cpu *cpu_buffer;
4211
4212 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4213 return 0;
4214
4215 cpu_buffer = buffer->buffers[cpu];
4216 return cpu_buffer->read;
4217}
4218EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu);
4219
4220/**
4221 * ring_buffer_entries - get the number of entries in a buffer
4222 * @buffer: The ring buffer
4223 *
4224 * Returns the total number of entries in the ring buffer
4225 * (all CPU entries)
4226 */
4227unsigned long ring_buffer_entries(struct trace_buffer *buffer)
4228{
4229 struct ring_buffer_per_cpu *cpu_buffer;
4230 unsigned long entries = 0;
4231 int cpu;
4232
4233 /* if you care about this being correct, lock the buffer */
4234 for_each_buffer_cpu(buffer, cpu) {
4235 cpu_buffer = buffer->buffers[cpu];
4236 entries += rb_num_of_entries(cpu_buffer);
4237 }
4238
4239 return entries;
4240}
4241EXPORT_SYMBOL_GPL(ring_buffer_entries);
4242
4243/**
4244 * ring_buffer_overruns - get the number of overruns in buffer
4245 * @buffer: The ring buffer
4246 *
4247 * Returns the total number of overruns in the ring buffer
4248 * (all CPU entries)
4249 */
4250unsigned long ring_buffer_overruns(struct trace_buffer *buffer)
4251{
4252 struct ring_buffer_per_cpu *cpu_buffer;
4253 unsigned long overruns = 0;
4254 int cpu;
4255
4256 /* if you care about this being correct, lock the buffer */
4257 for_each_buffer_cpu(buffer, cpu) {
4258 cpu_buffer = buffer->buffers[cpu];
4259 overruns += local_read(&cpu_buffer->overrun);
4260 }
4261
4262 return overruns;
4263}
4264EXPORT_SYMBOL_GPL(ring_buffer_overruns);
4265
4266static void rb_iter_reset(struct ring_buffer_iter *iter)
4267{
4268 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4269
4270 /* Iterator usage is expected to have record disabled */
4271 iter->head_page = cpu_buffer->reader_page;
4272 iter->head = cpu_buffer->reader_page->read;
4273 iter->next_event = iter->head;
4274
4275 iter->cache_reader_page = iter->head_page;
4276 iter->cache_read = cpu_buffer->read;
4277
4278 if (iter->head) {
4279 iter->read_stamp = cpu_buffer->read_stamp;
4280 iter->page_stamp = cpu_buffer->reader_page->page->time_stamp;
4281 } else {
4282 iter->read_stamp = iter->head_page->page->time_stamp;
4283 iter->page_stamp = iter->read_stamp;
4284 }
4285}
4286
4287/**
4288 * ring_buffer_iter_reset - reset an iterator
4289 * @iter: The iterator to reset
4290 *
4291 * Resets the iterator, so that it will start from the beginning
4292 * again.
4293 */
4294void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
4295{
4296 struct ring_buffer_per_cpu *cpu_buffer;
4297 unsigned long flags;
4298
4299 if (!iter)
4300 return;
4301
4302 cpu_buffer = iter->cpu_buffer;
4303
4304 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4305 rb_iter_reset(iter);
4306 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4307}
4308EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
4309
4310/**
4311 * ring_buffer_iter_empty - check if an iterator has no more to read
4312 * @iter: The iterator to check
4313 */
4314int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
4315{
4316 struct ring_buffer_per_cpu *cpu_buffer;
4317 struct buffer_page *reader;
4318 struct buffer_page *head_page;
4319 struct buffer_page *commit_page;
4320 struct buffer_page *curr_commit_page;
4321 unsigned commit;
4322 u64 curr_commit_ts;
4323 u64 commit_ts;
4324
4325 cpu_buffer = iter->cpu_buffer;
4326 reader = cpu_buffer->reader_page;
4327 head_page = cpu_buffer->head_page;
4328 commit_page = cpu_buffer->commit_page;
4329 commit_ts = commit_page->page->time_stamp;
4330
4331 /*
4332 * When the writer goes across pages, it issues a cmpxchg which
4333 * is a mb(), which will synchronize with the rmb here.
4334 * (see rb_tail_page_update())
4335 */
4336 smp_rmb();
4337 commit = rb_page_commit(commit_page);
4338 /* We want to make sure that the commit page doesn't change */
4339 smp_rmb();
4340
4341 /* Make sure commit page didn't change */
4342 curr_commit_page = READ_ONCE(cpu_buffer->commit_page);
4343 curr_commit_ts = READ_ONCE(curr_commit_page->page->time_stamp);
4344
4345 /* If the commit page changed, then there's more data */
4346 if (curr_commit_page != commit_page ||
4347 curr_commit_ts != commit_ts)
4348 return 0;
4349
4350 /* Still racy, as it may return a false positive, but that's OK */
4351 return ((iter->head_page == commit_page && iter->head >= commit) ||
4352 (iter->head_page == reader && commit_page == head_page &&
4353 head_page->read == commit &&
4354 iter->head == rb_page_commit(cpu_buffer->reader_page)));
4355}
4356EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
4357
4358static void
4359rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
4360 struct ring_buffer_event *event)
4361{
4362 u64 delta;
4363
4364 switch (event->type_len) {
4365 case RINGBUF_TYPE_PADDING:
4366 return;
4367
4368 case RINGBUF_TYPE_TIME_EXTEND:
4369 delta = rb_event_time_stamp(event);
4370 cpu_buffer->read_stamp += delta;
4371 return;
4372
4373 case RINGBUF_TYPE_TIME_STAMP:
4374 delta = rb_event_time_stamp(event);
4375 cpu_buffer->read_stamp = delta;
4376 return;
4377
4378 case RINGBUF_TYPE_DATA:
4379 cpu_buffer->read_stamp += event->time_delta;
4380 return;
4381
4382 default:
4383 RB_WARN_ON(cpu_buffer, 1);
4384 }
4385 return;
4386}
4387
4388static void
4389rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
4390 struct ring_buffer_event *event)
4391{
4392 u64 delta;
4393
4394 switch (event->type_len) {
4395 case RINGBUF_TYPE_PADDING:
4396 return;
4397
4398 case RINGBUF_TYPE_TIME_EXTEND:
4399 delta = rb_event_time_stamp(event);
4400 iter->read_stamp += delta;
4401 return;
4402
4403 case RINGBUF_TYPE_TIME_STAMP:
4404 delta = rb_event_time_stamp(event);
4405 iter->read_stamp = delta;
4406 return;
4407
4408 case RINGBUF_TYPE_DATA:
4409 iter->read_stamp += event->time_delta;
4410 return;
4411
4412 default:
4413 RB_WARN_ON(iter->cpu_buffer, 1);
4414 }
4415 return;
4416}
4417
4418static struct buffer_page *
4419rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
4420{
4421 struct buffer_page *reader = NULL;
4422 unsigned long overwrite;
4423 unsigned long flags;
4424 int nr_loops = 0;
4425 int ret;
4426
4427 local_irq_save(flags);
4428 arch_spin_lock(&cpu_buffer->lock);
4429
4430 again:
4431 /*
4432 * This should normally only loop twice. But because the
4433 * start of the reader inserts an empty page, it causes
4434 * a case where we will loop three times. There should be no
4435 * reason to loop four times (that I know of).
4436 */
4437 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
4438 reader = NULL;
4439 goto out;
4440 }
4441
4442 reader = cpu_buffer->reader_page;
4443
4444 /* If there's more to read, return this page */
4445 if (cpu_buffer->reader_page->read < rb_page_size(reader))
4446 goto out;
4447
4448 /* Never should we have an index greater than the size */
4449 if (RB_WARN_ON(cpu_buffer,
4450 cpu_buffer->reader_page->read > rb_page_size(reader)))
4451 goto out;
4452
4453 /* check if we caught up to the tail */
4454 reader = NULL;
4455 if (cpu_buffer->commit_page == cpu_buffer->reader_page)
4456 goto out;
4457
4458 /* Don't bother swapping if the ring buffer is empty */
4459 if (rb_num_of_entries(cpu_buffer) == 0)
4460 goto out;
4461
4462 /*
4463 * Reset the reader page to size zero.
4464 */
4465 local_set(&cpu_buffer->reader_page->write, 0);
4466 local_set(&cpu_buffer->reader_page->entries, 0);
4467 local_set(&cpu_buffer->reader_page->page->commit, 0);
4468 cpu_buffer->reader_page->real_end = 0;
4469
4470 spin:
4471 /*
4472 * Splice the empty reader page into the list around the head.
4473 */
4474 reader = rb_set_head_page(cpu_buffer);
4475 if (!reader)
4476 goto out;
4477 cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
4478 cpu_buffer->reader_page->list.prev = reader->list.prev;
4479
4480 /*
4481 * cpu_buffer->pages just needs to point to the buffer, it
4482 * has no specific buffer page to point to. Lets move it out
4483 * of our way so we don't accidentally swap it.
4484 */
4485 cpu_buffer->pages = reader->list.prev;
4486
4487 /* The reader page will be pointing to the new head */
4488 rb_set_list_to_head(&cpu_buffer->reader_page->list);
4489
4490 /*
4491 * We want to make sure we read the overruns after we set up our
4492 * pointers to the next object. The writer side does a
4493 * cmpxchg to cross pages which acts as the mb on the writer
4494 * side. Note, the reader will constantly fail the swap
4495 * while the writer is updating the pointers, so this
4496 * guarantees that the overwrite recorded here is the one we
4497 * want to compare with the last_overrun.
4498 */
4499 smp_mb();
4500 overwrite = local_read(&(cpu_buffer->overrun));
4501
4502 /*
4503 * Here's the tricky part.
4504 *
4505 * We need to move the pointer past the header page.
4506 * But we can only do that if a writer is not currently
4507 * moving it. The page before the header page has the
4508 * flag bit '1' set if it is pointing to the page we want.
4509 * but if the writer is in the process of moving it
4510 * than it will be '2' or already moved '0'.
4511 */
4512
4513 ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
4514
4515 /*
4516 * If we did not convert it, then we must try again.
4517 */
4518 if (!ret)
4519 goto spin;
4520
4521 /*
4522 * Yay! We succeeded in replacing the page.
4523 *
4524 * Now make the new head point back to the reader page.
4525 */
4526 rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
4527 rb_inc_page(&cpu_buffer->head_page);
4528
4529 local_inc(&cpu_buffer->pages_read);
4530
4531 /* Finally update the reader page to the new head */
4532 cpu_buffer->reader_page = reader;
4533 cpu_buffer->reader_page->read = 0;
4534
4535 if (overwrite != cpu_buffer->last_overrun) {
4536 cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
4537 cpu_buffer->last_overrun = overwrite;
4538 }
4539
4540 goto again;
4541
4542 out:
4543 /* Update the read_stamp on the first event */
4544 if (reader && reader->read == 0)
4545 cpu_buffer->read_stamp = reader->page->time_stamp;
4546
4547 arch_spin_unlock(&cpu_buffer->lock);
4548 local_irq_restore(flags);
4549
4550 return reader;
4551}
4552
4553static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
4554{
4555 struct ring_buffer_event *event;
4556 struct buffer_page *reader;
4557 unsigned length;
4558
4559 reader = rb_get_reader_page(cpu_buffer);
4560
4561 /* This function should not be called when buffer is empty */
4562 if (RB_WARN_ON(cpu_buffer, !reader))
4563 return;
4564
4565 event = rb_reader_event(cpu_buffer);
4566
4567 if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
4568 cpu_buffer->read++;
4569
4570 rb_update_read_stamp(cpu_buffer, event);
4571
4572 length = rb_event_length(event);
4573 cpu_buffer->reader_page->read += length;
4574}
4575
4576static void rb_advance_iter(struct ring_buffer_iter *iter)
4577{
4578 struct ring_buffer_per_cpu *cpu_buffer;
4579
4580 cpu_buffer = iter->cpu_buffer;
4581
4582 /* If head == next_event then we need to jump to the next event */
4583 if (iter->head == iter->next_event) {
4584 /* If the event gets overwritten again, there's nothing to do */
4585 if (rb_iter_head_event(iter) == NULL)
4586 return;
4587 }
4588
4589 iter->head = iter->next_event;
4590
4591 /*
4592 * Check if we are at the end of the buffer.
4593 */
4594 if (iter->next_event >= rb_page_size(iter->head_page)) {
4595 /* discarded commits can make the page empty */
4596 if (iter->head_page == cpu_buffer->commit_page)
4597 return;
4598 rb_inc_iter(iter);
4599 return;
4600 }
4601
4602 rb_update_iter_read_stamp(iter, iter->event);
4603}
4604
4605static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
4606{
4607 return cpu_buffer->lost_events;
4608}
4609
4610static struct ring_buffer_event *
4611rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
4612 unsigned long *lost_events)
4613{
4614 struct ring_buffer_event *event;
4615 struct buffer_page *reader;
4616 int nr_loops = 0;
4617
4618 if (ts)
4619 *ts = 0;
4620 again:
4621 /*
4622 * We repeat when a time extend is encountered.
4623 * Since the time extend is always attached to a data event,
4624 * we should never loop more than once.
4625 * (We never hit the following condition more than twice).
4626 */
4627 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
4628 return NULL;
4629
4630 reader = rb_get_reader_page(cpu_buffer);
4631 if (!reader)
4632 return NULL;
4633
4634 event = rb_reader_event(cpu_buffer);
4635
4636 switch (event->type_len) {
4637 case RINGBUF_TYPE_PADDING:
4638 if (rb_null_event(event))
4639 RB_WARN_ON(cpu_buffer, 1);
4640 /*
4641 * Because the writer could be discarding every
4642 * event it creates (which would probably be bad)
4643 * if we were to go back to "again" then we may never
4644 * catch up, and will trigger the warn on, or lock
4645 * the box. Return the padding, and we will release
4646 * the current locks, and try again.
4647 */
4648 return event;
4649
4650 case RINGBUF_TYPE_TIME_EXTEND:
4651 /* Internal data, OK to advance */
4652 rb_advance_reader(cpu_buffer);
4653 goto again;
4654
4655 case RINGBUF_TYPE_TIME_STAMP:
4656 if (ts) {
4657 *ts = rb_event_time_stamp(event);
4658 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
4659 cpu_buffer->cpu, ts);
4660 }
4661 /* Internal data, OK to advance */
4662 rb_advance_reader(cpu_buffer);
4663 goto again;
4664
4665 case RINGBUF_TYPE_DATA:
4666 if (ts && !(*ts)) {
4667 *ts = cpu_buffer->read_stamp + event->time_delta;
4668 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
4669 cpu_buffer->cpu, ts);
4670 }
4671 if (lost_events)
4672 *lost_events = rb_lost_events(cpu_buffer);
4673 return event;
4674
4675 default:
4676 RB_WARN_ON(cpu_buffer, 1);
4677 }
4678
4679 return NULL;
4680}
4681EXPORT_SYMBOL_GPL(ring_buffer_peek);
4682
4683static struct ring_buffer_event *
4684rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
4685{
4686 struct trace_buffer *buffer;
4687 struct ring_buffer_per_cpu *cpu_buffer;
4688 struct ring_buffer_event *event;
4689 int nr_loops = 0;
4690
4691 if (ts)
4692 *ts = 0;
4693
4694 cpu_buffer = iter->cpu_buffer;
4695 buffer = cpu_buffer->buffer;
4696
4697 /*
4698 * Check if someone performed a consuming read to
4699 * the buffer. A consuming read invalidates the iterator
4700 * and we need to reset the iterator in this case.
4701 */
4702 if (unlikely(iter->cache_read != cpu_buffer->read ||
4703 iter->cache_reader_page != cpu_buffer->reader_page))
4704 rb_iter_reset(iter);
4705
4706 again:
4707 if (ring_buffer_iter_empty(iter))
4708 return NULL;
4709
4710 /*
4711 * As the writer can mess with what the iterator is trying
4712 * to read, just give up if we fail to get an event after
4713 * three tries. The iterator is not as reliable when reading
4714 * the ring buffer with an active write as the consumer is.
4715 * Do not warn if the three failures is reached.
4716 */
4717 if (++nr_loops > 3)
4718 return NULL;
4719
4720 if (rb_per_cpu_empty(cpu_buffer))
4721 return NULL;
4722
4723 if (iter->head >= rb_page_size(iter->head_page)) {
4724 rb_inc_iter(iter);
4725 goto again;
4726 }
4727
4728 event = rb_iter_head_event(iter);
4729 if (!event)
4730 goto again;
4731
4732 switch (event->type_len) {
4733 case RINGBUF_TYPE_PADDING:
4734 if (rb_null_event(event)) {
4735 rb_inc_iter(iter);
4736 goto again;
4737 }
4738 rb_advance_iter(iter);
4739 return event;
4740
4741 case RINGBUF_TYPE_TIME_EXTEND:
4742 /* Internal data, OK to advance */
4743 rb_advance_iter(iter);
4744 goto again;
4745
4746 case RINGBUF_TYPE_TIME_STAMP:
4747 if (ts) {
4748 *ts = rb_event_time_stamp(event);
4749 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
4750 cpu_buffer->cpu, ts);
4751 }
4752 /* Internal data, OK to advance */
4753 rb_advance_iter(iter);
4754 goto again;
4755
4756 case RINGBUF_TYPE_DATA:
4757 if (ts && !(*ts)) {
4758 *ts = iter->read_stamp + event->time_delta;
4759 ring_buffer_normalize_time_stamp(buffer,
4760 cpu_buffer->cpu, ts);
4761 }
4762 return event;
4763
4764 default:
4765 RB_WARN_ON(cpu_buffer, 1);
4766 }
4767
4768 return NULL;
4769}
4770EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
4771
4772static inline bool rb_reader_lock(struct ring_buffer_per_cpu *cpu_buffer)
4773{
4774 if (likely(!in_nmi())) {
4775 raw_spin_lock(&cpu_buffer->reader_lock);
4776 return true;
4777 }
4778
4779 /*
4780 * If an NMI die dumps out the content of the ring buffer
4781 * trylock must be used to prevent a deadlock if the NMI
4782 * preempted a task that holds the ring buffer locks. If
4783 * we get the lock then all is fine, if not, then continue
4784 * to do the read, but this can corrupt the ring buffer,
4785 * so it must be permanently disabled from future writes.
4786 * Reading from NMI is a oneshot deal.
4787 */
4788 if (raw_spin_trylock(&cpu_buffer->reader_lock))
4789 return true;
4790
4791 /* Continue without locking, but disable the ring buffer */
4792 atomic_inc(&cpu_buffer->record_disabled);
4793 return false;
4794}
4795
4796static inline void
4797rb_reader_unlock(struct ring_buffer_per_cpu *cpu_buffer, bool locked)
4798{
4799 if (likely(locked))
4800 raw_spin_unlock(&cpu_buffer->reader_lock);
4801 return;
4802}
4803
4804/**
4805 * ring_buffer_peek - peek at the next event to be read
4806 * @buffer: The ring buffer to read
4807 * @cpu: The cpu to peak at
4808 * @ts: The timestamp counter of this event.
4809 * @lost_events: a variable to store if events were lost (may be NULL)
4810 *
4811 * This will return the event that will be read next, but does
4812 * not consume the data.
4813 */
4814struct ring_buffer_event *
4815ring_buffer_peek(struct trace_buffer *buffer, int cpu, u64 *ts,
4816 unsigned long *lost_events)
4817{
4818 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4819 struct ring_buffer_event *event;
4820 unsigned long flags;
4821 bool dolock;
4822
4823 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4824 return NULL;
4825
4826 again:
4827 local_irq_save(flags);
4828 dolock = rb_reader_lock(cpu_buffer);
4829 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
4830 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4831 rb_advance_reader(cpu_buffer);
4832 rb_reader_unlock(cpu_buffer, dolock);
4833 local_irq_restore(flags);
4834
4835 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4836 goto again;
4837
4838 return event;
4839}
4840
4841/** ring_buffer_iter_dropped - report if there are dropped events
4842 * @iter: The ring buffer iterator
4843 *
4844 * Returns true if there was dropped events since the last peek.
4845 */
4846bool ring_buffer_iter_dropped(struct ring_buffer_iter *iter)
4847{
4848 bool ret = iter->missed_events != 0;
4849
4850 iter->missed_events = 0;
4851 return ret;
4852}
4853EXPORT_SYMBOL_GPL(ring_buffer_iter_dropped);
4854
4855/**
4856 * ring_buffer_iter_peek - peek at the next event to be read
4857 * @iter: The ring buffer iterator
4858 * @ts: The timestamp counter of this event.
4859 *
4860 * This will return the event that will be read next, but does
4861 * not increment the iterator.
4862 */
4863struct ring_buffer_event *
4864ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
4865{
4866 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4867 struct ring_buffer_event *event;
4868 unsigned long flags;
4869
4870 again:
4871 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4872 event = rb_iter_peek(iter, ts);
4873 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4874
4875 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4876 goto again;
4877
4878 return event;
4879}
4880
4881/**
4882 * ring_buffer_consume - return an event and consume it
4883 * @buffer: The ring buffer to get the next event from
4884 * @cpu: the cpu to read the buffer from
4885 * @ts: a variable to store the timestamp (may be NULL)
4886 * @lost_events: a variable to store if events were lost (may be NULL)
4887 *
4888 * Returns the next event in the ring buffer, and that event is consumed.
4889 * Meaning, that sequential reads will keep returning a different event,
4890 * and eventually empty the ring buffer if the producer is slower.
4891 */
4892struct ring_buffer_event *
4893ring_buffer_consume(struct trace_buffer *buffer, int cpu, u64 *ts,
4894 unsigned long *lost_events)
4895{
4896 struct ring_buffer_per_cpu *cpu_buffer;
4897 struct ring_buffer_event *event = NULL;
4898 unsigned long flags;
4899 bool dolock;
4900
4901 again:
4902 /* might be called in atomic */
4903 preempt_disable();
4904
4905 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4906 goto out;
4907
4908 cpu_buffer = buffer->buffers[cpu];
4909 local_irq_save(flags);
4910 dolock = rb_reader_lock(cpu_buffer);
4911
4912 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
4913 if (event) {
4914 cpu_buffer->lost_events = 0;
4915 rb_advance_reader(cpu_buffer);
4916 }
4917
4918 rb_reader_unlock(cpu_buffer, dolock);
4919 local_irq_restore(flags);
4920
4921 out:
4922 preempt_enable();
4923
4924 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4925 goto again;
4926
4927 return event;
4928}
4929EXPORT_SYMBOL_GPL(ring_buffer_consume);
4930
4931/**
4932 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
4933 * @buffer: The ring buffer to read from
4934 * @cpu: The cpu buffer to iterate over
4935 * @flags: gfp flags to use for memory allocation
4936 *
4937 * This performs the initial preparations necessary to iterate
4938 * through the buffer. Memory is allocated, buffer recording
4939 * is disabled, and the iterator pointer is returned to the caller.
4940 *
4941 * Disabling buffer recording prevents the reading from being
4942 * corrupted. This is not a consuming read, so a producer is not
4943 * expected.
4944 *
4945 * After a sequence of ring_buffer_read_prepare calls, the user is
4946 * expected to make at least one call to ring_buffer_read_prepare_sync.
4947 * Afterwards, ring_buffer_read_start is invoked to get things going
4948 * for real.
4949 *
4950 * This overall must be paired with ring_buffer_read_finish.
4951 */
4952struct ring_buffer_iter *
4953ring_buffer_read_prepare(struct trace_buffer *buffer, int cpu, gfp_t flags)
4954{
4955 struct ring_buffer_per_cpu *cpu_buffer;
4956 struct ring_buffer_iter *iter;
4957
4958 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4959 return NULL;
4960
4961 iter = kzalloc(sizeof(*iter), flags);
4962 if (!iter)
4963 return NULL;
4964
4965 iter->event = kmalloc(BUF_MAX_DATA_SIZE, flags);
4966 if (!iter->event) {
4967 kfree(iter);
4968 return NULL;
4969 }
4970
4971 cpu_buffer = buffer->buffers[cpu];
4972
4973 iter->cpu_buffer = cpu_buffer;
4974
4975 atomic_inc(&cpu_buffer->resize_disabled);
4976
4977 return iter;
4978}
4979EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
4980
4981/**
4982 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
4983 *
4984 * All previously invoked ring_buffer_read_prepare calls to prepare
4985 * iterators will be synchronized. Afterwards, read_buffer_read_start
4986 * calls on those iterators are allowed.
4987 */
4988void
4989ring_buffer_read_prepare_sync(void)
4990{
4991 synchronize_rcu();
4992}
4993EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
4994
4995/**
4996 * ring_buffer_read_start - start a non consuming read of the buffer
4997 * @iter: The iterator returned by ring_buffer_read_prepare
4998 *
4999 * This finalizes the startup of an iteration through the buffer.
5000 * The iterator comes from a call to ring_buffer_read_prepare and
5001 * an intervening ring_buffer_read_prepare_sync must have been
5002 * performed.
5003 *
5004 * Must be paired with ring_buffer_read_finish.
5005 */
5006void
5007ring_buffer_read_start(struct ring_buffer_iter *iter)
5008{
5009 struct ring_buffer_per_cpu *cpu_buffer;
5010 unsigned long flags;
5011
5012 if (!iter)
5013 return;
5014
5015 cpu_buffer = iter->cpu_buffer;
5016
5017 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5018 arch_spin_lock(&cpu_buffer->lock);
5019 rb_iter_reset(iter);
5020 arch_spin_unlock(&cpu_buffer->lock);
5021 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5022}
5023EXPORT_SYMBOL_GPL(ring_buffer_read_start);
5024
5025/**
5026 * ring_buffer_read_finish - finish reading the iterator of the buffer
5027 * @iter: The iterator retrieved by ring_buffer_start
5028 *
5029 * This re-enables the recording to the buffer, and frees the
5030 * iterator.
5031 */
5032void
5033ring_buffer_read_finish(struct ring_buffer_iter *iter)
5034{
5035 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
5036 unsigned long flags;
5037
5038 /*
5039 * Ring buffer is disabled from recording, here's a good place
5040 * to check the integrity of the ring buffer.
5041 * Must prevent readers from trying to read, as the check
5042 * clears the HEAD page and readers require it.
5043 */
5044 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5045 rb_check_pages(cpu_buffer);
5046 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5047
5048 atomic_dec(&cpu_buffer->resize_disabled);
5049 kfree(iter->event);
5050 kfree(iter);
5051}
5052EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
5053
5054/**
5055 * ring_buffer_iter_advance - advance the iterator to the next location
5056 * @iter: The ring buffer iterator
5057 *
5058 * Move the location of the iterator such that the next read will
5059 * be the next location of the iterator.
5060 */
5061void ring_buffer_iter_advance(struct ring_buffer_iter *iter)
5062{
5063 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
5064 unsigned long flags;
5065
5066 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5067
5068 rb_advance_iter(iter);
5069
5070 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5071}
5072EXPORT_SYMBOL_GPL(ring_buffer_iter_advance);
5073
5074/**
5075 * ring_buffer_size - return the size of the ring buffer (in bytes)
5076 * @buffer: The ring buffer.
5077 * @cpu: The CPU to get ring buffer size from.
5078 */
5079unsigned long ring_buffer_size(struct trace_buffer *buffer, int cpu)
5080{
5081 /*
5082 * Earlier, this method returned
5083 * BUF_PAGE_SIZE * buffer->nr_pages
5084 * Since the nr_pages field is now removed, we have converted this to
5085 * return the per cpu buffer value.
5086 */
5087 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5088 return 0;
5089
5090 return BUF_PAGE_SIZE * buffer->buffers[cpu]->nr_pages;
5091}
5092EXPORT_SYMBOL_GPL(ring_buffer_size);
5093
5094static void
5095rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
5096{
5097 rb_head_page_deactivate(cpu_buffer);
5098
5099 cpu_buffer->head_page
5100 = list_entry(cpu_buffer->pages, struct buffer_page, list);
5101 local_set(&cpu_buffer->head_page->write, 0);
5102 local_set(&cpu_buffer->head_page->entries, 0);
5103 local_set(&cpu_buffer->head_page->page->commit, 0);
5104
5105 cpu_buffer->head_page->read = 0;
5106
5107 cpu_buffer->tail_page = cpu_buffer->head_page;
5108 cpu_buffer->commit_page = cpu_buffer->head_page;
5109
5110 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
5111 INIT_LIST_HEAD(&cpu_buffer->new_pages);
5112 local_set(&cpu_buffer->reader_page->write, 0);
5113 local_set(&cpu_buffer->reader_page->entries, 0);
5114 local_set(&cpu_buffer->reader_page->page->commit, 0);
5115 cpu_buffer->reader_page->read = 0;
5116
5117 local_set(&cpu_buffer->entries_bytes, 0);
5118 local_set(&cpu_buffer->overrun, 0);
5119 local_set(&cpu_buffer->commit_overrun, 0);
5120 local_set(&cpu_buffer->dropped_events, 0);
5121 local_set(&cpu_buffer->entries, 0);
5122 local_set(&cpu_buffer->committing, 0);
5123 local_set(&cpu_buffer->commits, 0);
5124 local_set(&cpu_buffer->pages_touched, 0);
5125 local_set(&cpu_buffer->pages_read, 0);
5126 cpu_buffer->last_pages_touch = 0;
5127 cpu_buffer->shortest_full = 0;
5128 cpu_buffer->read = 0;
5129 cpu_buffer->read_bytes = 0;
5130
5131 rb_time_set(&cpu_buffer->write_stamp, 0);
5132 rb_time_set(&cpu_buffer->before_stamp, 0);
5133
5134 memset(cpu_buffer->event_stamp, 0, sizeof(cpu_buffer->event_stamp));
5135
5136 cpu_buffer->lost_events = 0;
5137 cpu_buffer->last_overrun = 0;
5138
5139 rb_head_page_activate(cpu_buffer);
5140}
5141
5142/* Must have disabled the cpu buffer then done a synchronize_rcu */
5143static void reset_disabled_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
5144{
5145 unsigned long flags;
5146
5147 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5148
5149 if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
5150 goto out;
5151
5152 arch_spin_lock(&cpu_buffer->lock);
5153
5154 rb_reset_cpu(cpu_buffer);
5155
5156 arch_spin_unlock(&cpu_buffer->lock);
5157
5158 out:
5159 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5160}
5161
5162/**
5163 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
5164 * @buffer: The ring buffer to reset a per cpu buffer of
5165 * @cpu: The CPU buffer to be reset
5166 */
5167void ring_buffer_reset_cpu(struct trace_buffer *buffer, int cpu)
5168{
5169 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
5170
5171 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5172 return;
5173
5174 /* prevent another thread from changing buffer sizes */
5175 mutex_lock(&buffer->mutex);
5176
5177 atomic_inc(&cpu_buffer->resize_disabled);
5178 atomic_inc(&cpu_buffer->record_disabled);
5179
5180 /* Make sure all commits have finished */
5181 synchronize_rcu();
5182
5183 reset_disabled_cpu_buffer(cpu_buffer);
5184
5185 atomic_dec(&cpu_buffer->record_disabled);
5186 atomic_dec(&cpu_buffer->resize_disabled);
5187
5188 mutex_unlock(&buffer->mutex);
5189}
5190EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
5191
5192/**
5193 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
5194 * @buffer: The ring buffer to reset a per cpu buffer of
5195 * @cpu: The CPU buffer to be reset
5196 */
5197void ring_buffer_reset_online_cpus(struct trace_buffer *buffer)
5198{
5199 struct ring_buffer_per_cpu *cpu_buffer;
5200 int cpu;
5201
5202 /* prevent another thread from changing buffer sizes */
5203 mutex_lock(&buffer->mutex);
5204
5205 for_each_online_buffer_cpu(buffer, cpu) {
5206 cpu_buffer = buffer->buffers[cpu];
5207
5208 atomic_inc(&cpu_buffer->resize_disabled);
5209 atomic_inc(&cpu_buffer->record_disabled);
5210 }
5211
5212 /* Make sure all commits have finished */
5213 synchronize_rcu();
5214
5215 for_each_online_buffer_cpu(buffer, cpu) {
5216 cpu_buffer = buffer->buffers[cpu];
5217
5218 reset_disabled_cpu_buffer(cpu_buffer);
5219
5220 atomic_dec(&cpu_buffer->record_disabled);
5221 atomic_dec(&cpu_buffer->resize_disabled);
5222 }
5223
5224 mutex_unlock(&buffer->mutex);
5225}
5226
5227/**
5228 * ring_buffer_reset - reset a ring buffer
5229 * @buffer: The ring buffer to reset all cpu buffers
5230 */
5231void ring_buffer_reset(struct trace_buffer *buffer)
5232{
5233 struct ring_buffer_per_cpu *cpu_buffer;
5234 int cpu;
5235
5236 for_each_buffer_cpu(buffer, cpu) {
5237 cpu_buffer = buffer->buffers[cpu];
5238
5239 atomic_inc(&cpu_buffer->resize_disabled);
5240 atomic_inc(&cpu_buffer->record_disabled);
5241 }
5242
5243 /* Make sure all commits have finished */
5244 synchronize_rcu();
5245
5246 for_each_buffer_cpu(buffer, cpu) {
5247 cpu_buffer = buffer->buffers[cpu];
5248
5249 reset_disabled_cpu_buffer(cpu_buffer);
5250
5251 atomic_dec(&cpu_buffer->record_disabled);
5252 atomic_dec(&cpu_buffer->resize_disabled);
5253 }
5254}
5255EXPORT_SYMBOL_GPL(ring_buffer_reset);
5256
5257/**
5258 * rind_buffer_empty - is the ring buffer empty?
5259 * @buffer: The ring buffer to test
5260 */
5261bool ring_buffer_empty(struct trace_buffer *buffer)
5262{
5263 struct ring_buffer_per_cpu *cpu_buffer;
5264 unsigned long flags;
5265 bool dolock;
5266 int cpu;
5267 int ret;
5268
5269 /* yes this is racy, but if you don't like the race, lock the buffer */
5270 for_each_buffer_cpu(buffer, cpu) {
5271 cpu_buffer = buffer->buffers[cpu];
5272 local_irq_save(flags);
5273 dolock = rb_reader_lock(cpu_buffer);
5274 ret = rb_per_cpu_empty(cpu_buffer);
5275 rb_reader_unlock(cpu_buffer, dolock);
5276 local_irq_restore(flags);
5277
5278 if (!ret)
5279 return false;
5280 }
5281
5282 return true;
5283}
5284EXPORT_SYMBOL_GPL(ring_buffer_empty);
5285
5286/**
5287 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
5288 * @buffer: The ring buffer
5289 * @cpu: The CPU buffer to test
5290 */
5291bool ring_buffer_empty_cpu(struct trace_buffer *buffer, int cpu)
5292{
5293 struct ring_buffer_per_cpu *cpu_buffer;
5294 unsigned long flags;
5295 bool dolock;
5296 int ret;
5297
5298 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5299 return true;
5300
5301 cpu_buffer = buffer->buffers[cpu];
5302 local_irq_save(flags);
5303 dolock = rb_reader_lock(cpu_buffer);
5304 ret = rb_per_cpu_empty(cpu_buffer);
5305 rb_reader_unlock(cpu_buffer, dolock);
5306 local_irq_restore(flags);
5307
5308 return ret;
5309}
5310EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
5311
5312#ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
5313/**
5314 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
5315 * @buffer_a: One buffer to swap with
5316 * @buffer_b: The other buffer to swap with
5317 * @cpu: the CPU of the buffers to swap
5318 *
5319 * This function is useful for tracers that want to take a "snapshot"
5320 * of a CPU buffer and has another back up buffer lying around.
5321 * it is expected that the tracer handles the cpu buffer not being
5322 * used at the moment.
5323 */
5324int ring_buffer_swap_cpu(struct trace_buffer *buffer_a,
5325 struct trace_buffer *buffer_b, int cpu)
5326{
5327 struct ring_buffer_per_cpu *cpu_buffer_a;
5328 struct ring_buffer_per_cpu *cpu_buffer_b;
5329 int ret = -EINVAL;
5330
5331 if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
5332 !cpumask_test_cpu(cpu, buffer_b->cpumask))
5333 goto out;
5334
5335 cpu_buffer_a = buffer_a->buffers[cpu];
5336 cpu_buffer_b = buffer_b->buffers[cpu];
5337
5338 /* At least make sure the two buffers are somewhat the same */
5339 if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
5340 goto out;
5341
5342 ret = -EAGAIN;
5343
5344 if (atomic_read(&buffer_a->record_disabled))
5345 goto out;
5346
5347 if (atomic_read(&buffer_b->record_disabled))
5348 goto out;
5349
5350 if (atomic_read(&cpu_buffer_a->record_disabled))
5351 goto out;
5352
5353 if (atomic_read(&cpu_buffer_b->record_disabled))
5354 goto out;
5355
5356 /*
5357 * We can't do a synchronize_rcu here because this
5358 * function can be called in atomic context.
5359 * Normally this will be called from the same CPU as cpu.
5360 * If not it's up to the caller to protect this.
5361 */
5362 atomic_inc(&cpu_buffer_a->record_disabled);
5363 atomic_inc(&cpu_buffer_b->record_disabled);
5364
5365 ret = -EBUSY;
5366 if (local_read(&cpu_buffer_a->committing))
5367 goto out_dec;
5368 if (local_read(&cpu_buffer_b->committing))
5369 goto out_dec;
5370
5371 buffer_a->buffers[cpu] = cpu_buffer_b;
5372 buffer_b->buffers[cpu] = cpu_buffer_a;
5373
5374 cpu_buffer_b->buffer = buffer_a;
5375 cpu_buffer_a->buffer = buffer_b;
5376
5377 ret = 0;
5378
5379out_dec:
5380 atomic_dec(&cpu_buffer_a->record_disabled);
5381 atomic_dec(&cpu_buffer_b->record_disabled);
5382out:
5383 return ret;
5384}
5385EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
5386#endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
5387
5388/**
5389 * ring_buffer_alloc_read_page - allocate a page to read from buffer
5390 * @buffer: the buffer to allocate for.
5391 * @cpu: the cpu buffer to allocate.
5392 *
5393 * This function is used in conjunction with ring_buffer_read_page.
5394 * When reading a full page from the ring buffer, these functions
5395 * can be used to speed up the process. The calling function should
5396 * allocate a few pages first with this function. Then when it
5397 * needs to get pages from the ring buffer, it passes the result
5398 * of this function into ring_buffer_read_page, which will swap
5399 * the page that was allocated, with the read page of the buffer.
5400 *
5401 * Returns:
5402 * The page allocated, or ERR_PTR
5403 */
5404void *ring_buffer_alloc_read_page(struct trace_buffer *buffer, int cpu)
5405{
5406 struct ring_buffer_per_cpu *cpu_buffer;
5407 struct buffer_data_page *bpage = NULL;
5408 unsigned long flags;
5409 struct page *page;
5410
5411 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5412 return ERR_PTR(-ENODEV);
5413
5414 cpu_buffer = buffer->buffers[cpu];
5415 local_irq_save(flags);
5416 arch_spin_lock(&cpu_buffer->lock);
5417
5418 if (cpu_buffer->free_page) {
5419 bpage = cpu_buffer->free_page;
5420 cpu_buffer->free_page = NULL;
5421 }
5422
5423 arch_spin_unlock(&cpu_buffer->lock);
5424 local_irq_restore(flags);
5425
5426 if (bpage)
5427 goto out;
5428
5429 page = alloc_pages_node(cpu_to_node(cpu),
5430 GFP_KERNEL | __GFP_NORETRY, 0);
5431 if (!page)
5432 return ERR_PTR(-ENOMEM);
5433
5434 bpage = page_address(page);
5435
5436 out:
5437 rb_init_page(bpage);
5438
5439 return bpage;
5440}
5441EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
5442
5443/**
5444 * ring_buffer_free_read_page - free an allocated read page
5445 * @buffer: the buffer the page was allocate for
5446 * @cpu: the cpu buffer the page came from
5447 * @data: the page to free
5448 *
5449 * Free a page allocated from ring_buffer_alloc_read_page.
5450 */
5451void ring_buffer_free_read_page(struct trace_buffer *buffer, int cpu, void *data)
5452{
5453 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
5454 struct buffer_data_page *bpage = data;
5455 struct page *page = virt_to_page(bpage);
5456 unsigned long flags;
5457
5458 /* If the page is still in use someplace else, we can't reuse it */
5459 if (page_ref_count(page) > 1)
5460 goto out;
5461
5462 local_irq_save(flags);
5463 arch_spin_lock(&cpu_buffer->lock);
5464
5465 if (!cpu_buffer->free_page) {
5466 cpu_buffer->free_page = bpage;
5467 bpage = NULL;
5468 }
5469
5470 arch_spin_unlock(&cpu_buffer->lock);
5471 local_irq_restore(flags);
5472
5473 out:
5474 free_page((unsigned long)bpage);
5475}
5476EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
5477
5478/**
5479 * ring_buffer_read_page - extract a page from the ring buffer
5480 * @buffer: buffer to extract from
5481 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
5482 * @len: amount to extract
5483 * @cpu: the cpu of the buffer to extract
5484 * @full: should the extraction only happen when the page is full.
5485 *
5486 * This function will pull out a page from the ring buffer and consume it.
5487 * @data_page must be the address of the variable that was returned
5488 * from ring_buffer_alloc_read_page. This is because the page might be used
5489 * to swap with a page in the ring buffer.
5490 *
5491 * for example:
5492 * rpage = ring_buffer_alloc_read_page(buffer, cpu);
5493 * if (IS_ERR(rpage))
5494 * return PTR_ERR(rpage);
5495 * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
5496 * if (ret >= 0)
5497 * process_page(rpage, ret);
5498 *
5499 * When @full is set, the function will not return true unless
5500 * the writer is off the reader page.
5501 *
5502 * Note: it is up to the calling functions to handle sleeps and wakeups.
5503 * The ring buffer can be used anywhere in the kernel and can not
5504 * blindly call wake_up. The layer that uses the ring buffer must be
5505 * responsible for that.
5506 *
5507 * Returns:
5508 * >=0 if data has been transferred, returns the offset of consumed data.
5509 * <0 if no data has been transferred.
5510 */
5511int ring_buffer_read_page(struct trace_buffer *buffer,
5512 void **data_page, size_t len, int cpu, int full)
5513{
5514 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
5515 struct ring_buffer_event *event;
5516 struct buffer_data_page *bpage;
5517 struct buffer_page *reader;
5518 unsigned long missed_events;
5519 unsigned long flags;
5520 unsigned int commit;
5521 unsigned int read;
5522 u64 save_timestamp;
5523 int ret = -1;
5524
5525 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5526 goto out;
5527
5528 /*
5529 * If len is not big enough to hold the page header, then
5530 * we can not copy anything.
5531 */
5532 if (len <= BUF_PAGE_HDR_SIZE)
5533 goto out;
5534
5535 len -= BUF_PAGE_HDR_SIZE;
5536
5537 if (!data_page)
5538 goto out;
5539
5540 bpage = *data_page;
5541 if (!bpage)
5542 goto out;
5543
5544 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5545
5546 reader = rb_get_reader_page(cpu_buffer);
5547 if (!reader)
5548 goto out_unlock;
5549
5550 event = rb_reader_event(cpu_buffer);
5551
5552 read = reader->read;
5553 commit = rb_page_commit(reader);
5554
5555 /* Check if any events were dropped */
5556 missed_events = cpu_buffer->lost_events;
5557
5558 /*
5559 * If this page has been partially read or
5560 * if len is not big enough to read the rest of the page or
5561 * a writer is still on the page, then
5562 * we must copy the data from the page to the buffer.
5563 * Otherwise, we can simply swap the page with the one passed in.
5564 */
5565 if (read || (len < (commit - read)) ||
5566 cpu_buffer->reader_page == cpu_buffer->commit_page) {
5567 struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
5568 unsigned int rpos = read;
5569 unsigned int pos = 0;
5570 unsigned int size;
5571
5572 if (full)
5573 goto out_unlock;
5574
5575 if (len > (commit - read))
5576 len = (commit - read);
5577
5578 /* Always keep the time extend and data together */
5579 size = rb_event_ts_length(event);
5580
5581 if (len < size)
5582 goto out_unlock;
5583
5584 /* save the current timestamp, since the user will need it */
5585 save_timestamp = cpu_buffer->read_stamp;
5586
5587 /* Need to copy one event at a time */
5588 do {
5589 /* We need the size of one event, because
5590 * rb_advance_reader only advances by one event,
5591 * whereas rb_event_ts_length may include the size of
5592 * one or two events.
5593 * We have already ensured there's enough space if this
5594 * is a time extend. */
5595 size = rb_event_length(event);
5596 memcpy(bpage->data + pos, rpage->data + rpos, size);
5597
5598 len -= size;
5599
5600 rb_advance_reader(cpu_buffer);
5601 rpos = reader->read;
5602 pos += size;
5603
5604 if (rpos >= commit)
5605 break;
5606
5607 event = rb_reader_event(cpu_buffer);
5608 /* Always keep the time extend and data together */
5609 size = rb_event_ts_length(event);
5610 } while (len >= size);
5611
5612 /* update bpage */
5613 local_set(&bpage->commit, pos);
5614 bpage->time_stamp = save_timestamp;
5615
5616 /* we copied everything to the beginning */
5617 read = 0;
5618 } else {
5619 /* update the entry counter */
5620 cpu_buffer->read += rb_page_entries(reader);
5621 cpu_buffer->read_bytes += BUF_PAGE_SIZE;
5622
5623 /* swap the pages */
5624 rb_init_page(bpage);
5625 bpage = reader->page;
5626 reader->page = *data_page;
5627 local_set(&reader->write, 0);
5628 local_set(&reader->entries, 0);
5629 reader->read = 0;
5630 *data_page = bpage;
5631
5632 /*
5633 * Use the real_end for the data size,
5634 * This gives us a chance to store the lost events
5635 * on the page.
5636 */
5637 if (reader->real_end)
5638 local_set(&bpage->commit, reader->real_end);
5639 }
5640 ret = read;
5641
5642 cpu_buffer->lost_events = 0;
5643
5644 commit = local_read(&bpage->commit);
5645 /*
5646 * Set a flag in the commit field if we lost events
5647 */
5648 if (missed_events) {
5649 /* If there is room at the end of the page to save the
5650 * missed events, then record it there.
5651 */
5652 if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
5653 memcpy(&bpage->data[commit], &missed_events,
5654 sizeof(missed_events));
5655 local_add(RB_MISSED_STORED, &bpage->commit);
5656 commit += sizeof(missed_events);
5657 }
5658 local_add(RB_MISSED_EVENTS, &bpage->commit);
5659 }
5660
5661 /*
5662 * This page may be off to user land. Zero it out here.
5663 */
5664 if (commit < BUF_PAGE_SIZE)
5665 memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
5666
5667 out_unlock:
5668 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5669
5670 out:
5671 return ret;
5672}
5673EXPORT_SYMBOL_GPL(ring_buffer_read_page);
5674
5675/*
5676 * We only allocate new buffers, never free them if the CPU goes down.
5677 * If we were to free the buffer, then the user would lose any trace that was in
5678 * the buffer.
5679 */
5680int trace_rb_cpu_prepare(unsigned int cpu, struct hlist_node *node)
5681{
5682 struct trace_buffer *buffer;
5683 long nr_pages_same;
5684 int cpu_i;
5685 unsigned long nr_pages;
5686
5687 buffer = container_of(node, struct trace_buffer, node);
5688 if (cpumask_test_cpu(cpu, buffer->cpumask))
5689 return 0;
5690
5691 nr_pages = 0;
5692 nr_pages_same = 1;
5693 /* check if all cpu sizes are same */
5694 for_each_buffer_cpu(buffer, cpu_i) {
5695 /* fill in the size from first enabled cpu */
5696 if (nr_pages == 0)
5697 nr_pages = buffer->buffers[cpu_i]->nr_pages;
5698 if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
5699 nr_pages_same = 0;
5700 break;
5701 }
5702 }
5703 /* allocate minimum pages, user can later expand it */
5704 if (!nr_pages_same)
5705 nr_pages = 2;
5706 buffer->buffers[cpu] =
5707 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
5708 if (!buffer->buffers[cpu]) {
5709 WARN(1, "failed to allocate ring buffer on CPU %u\n",
5710 cpu);
5711 return -ENOMEM;
5712 }
5713 smp_wmb();
5714 cpumask_set_cpu(cpu, buffer->cpumask);
5715 return 0;
5716}
5717
5718#ifdef CONFIG_RING_BUFFER_STARTUP_TEST
5719/*
5720 * This is a basic integrity check of the ring buffer.
5721 * Late in the boot cycle this test will run when configured in.
5722 * It will kick off a thread per CPU that will go into a loop
5723 * writing to the per cpu ring buffer various sizes of data.
5724 * Some of the data will be large items, some small.
5725 *
5726 * Another thread is created that goes into a spin, sending out
5727 * IPIs to the other CPUs to also write into the ring buffer.
5728 * this is to test the nesting ability of the buffer.
5729 *
5730 * Basic stats are recorded and reported. If something in the
5731 * ring buffer should happen that's not expected, a big warning
5732 * is displayed and all ring buffers are disabled.
5733 */
5734static struct task_struct *rb_threads[NR_CPUS] __initdata;
5735
5736struct rb_test_data {
5737 struct trace_buffer *buffer;
5738 unsigned long events;
5739 unsigned long bytes_written;
5740 unsigned long bytes_alloc;
5741 unsigned long bytes_dropped;
5742 unsigned long events_nested;
5743 unsigned long bytes_written_nested;
5744 unsigned long bytes_alloc_nested;
5745 unsigned long bytes_dropped_nested;
5746 int min_size_nested;
5747 int max_size_nested;
5748 int max_size;
5749 int min_size;
5750 int cpu;
5751 int cnt;
5752};
5753
5754static struct rb_test_data rb_data[NR_CPUS] __initdata;
5755
5756/* 1 meg per cpu */
5757#define RB_TEST_BUFFER_SIZE 1048576
5758
5759static char rb_string[] __initdata =
5760 "abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
5761 "?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
5762 "!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
5763
5764static bool rb_test_started __initdata;
5765
5766struct rb_item {
5767 int size;
5768 char str[];
5769};
5770
5771static __init int rb_write_something(struct rb_test_data *data, bool nested)
5772{
5773 struct ring_buffer_event *event;
5774 struct rb_item *item;
5775 bool started;
5776 int event_len;
5777 int size;
5778 int len;
5779 int cnt;
5780
5781 /* Have nested writes different that what is written */
5782 cnt = data->cnt + (nested ? 27 : 0);
5783
5784 /* Multiply cnt by ~e, to make some unique increment */
5785 size = (cnt * 68 / 25) % (sizeof(rb_string) - 1);
5786
5787 len = size + sizeof(struct rb_item);
5788
5789 started = rb_test_started;
5790 /* read rb_test_started before checking buffer enabled */
5791 smp_rmb();
5792
5793 event = ring_buffer_lock_reserve(data->buffer, len);
5794 if (!event) {
5795 /* Ignore dropped events before test starts. */
5796 if (started) {
5797 if (nested)
5798 data->bytes_dropped += len;
5799 else
5800 data->bytes_dropped_nested += len;
5801 }
5802 return len;
5803 }
5804
5805 event_len = ring_buffer_event_length(event);
5806
5807 if (RB_WARN_ON(data->buffer, event_len < len))
5808 goto out;
5809
5810 item = ring_buffer_event_data(event);
5811 item->size = size;
5812 memcpy(item->str, rb_string, size);
5813
5814 if (nested) {
5815 data->bytes_alloc_nested += event_len;
5816 data->bytes_written_nested += len;
5817 data->events_nested++;
5818 if (!data->min_size_nested || len < data->min_size_nested)
5819 data->min_size_nested = len;
5820 if (len > data->max_size_nested)
5821 data->max_size_nested = len;
5822 } else {
5823 data->bytes_alloc += event_len;
5824 data->bytes_written += len;
5825 data->events++;
5826 if (!data->min_size || len < data->min_size)
5827 data->max_size = len;
5828 if (len > data->max_size)
5829 data->max_size = len;
5830 }
5831
5832 out:
5833 ring_buffer_unlock_commit(data->buffer, event);
5834
5835 return 0;
5836}
5837
5838static __init int rb_test(void *arg)
5839{
5840 struct rb_test_data *data = arg;
5841
5842 while (!kthread_should_stop()) {
5843 rb_write_something(data, false);
5844 data->cnt++;
5845
5846 set_current_state(TASK_INTERRUPTIBLE);
5847 /* Now sleep between a min of 100-300us and a max of 1ms */
5848 usleep_range(((data->cnt % 3) + 1) * 100, 1000);
5849 }
5850
5851 return 0;
5852}
5853
5854static __init void rb_ipi(void *ignore)
5855{
5856 struct rb_test_data *data;
5857 int cpu = smp_processor_id();
5858
5859 data = &rb_data[cpu];
5860 rb_write_something(data, true);
5861}
5862
5863static __init int rb_hammer_test(void *arg)
5864{
5865 while (!kthread_should_stop()) {
5866
5867 /* Send an IPI to all cpus to write data! */
5868 smp_call_function(rb_ipi, NULL, 1);
5869 /* No sleep, but for non preempt, let others run */
5870 schedule();
5871 }
5872
5873 return 0;
5874}
5875
5876static __init int test_ringbuffer(void)
5877{
5878 struct task_struct *rb_hammer;
5879 struct trace_buffer *buffer;
5880 int cpu;
5881 int ret = 0;
5882
5883 if (security_locked_down(LOCKDOWN_TRACEFS)) {
5884 pr_warn("Lockdown is enabled, skipping ring buffer tests\n");
5885 return 0;
5886 }
5887
5888 pr_info("Running ring buffer tests...\n");
5889
5890 buffer = ring_buffer_alloc(RB_TEST_BUFFER_SIZE, RB_FL_OVERWRITE);
5891 if (WARN_ON(!buffer))
5892 return 0;
5893
5894 /* Disable buffer so that threads can't write to it yet */
5895 ring_buffer_record_off(buffer);
5896
5897 for_each_online_cpu(cpu) {
5898 rb_data[cpu].buffer = buffer;
5899 rb_data[cpu].cpu = cpu;
5900 rb_data[cpu].cnt = cpu;
5901 rb_threads[cpu] = kthread_create(rb_test, &rb_data[cpu],
5902 "rbtester/%d", cpu);
5903 if (WARN_ON(IS_ERR(rb_threads[cpu]))) {
5904 pr_cont("FAILED\n");
5905 ret = PTR_ERR(rb_threads[cpu]);
5906 goto out_free;
5907 }
5908
5909 kthread_bind(rb_threads[cpu], cpu);
5910 wake_up_process(rb_threads[cpu]);
5911 }
5912
5913 /* Now create the rb hammer! */
5914 rb_hammer = kthread_run(rb_hammer_test, NULL, "rbhammer");
5915 if (WARN_ON(IS_ERR(rb_hammer))) {
5916 pr_cont("FAILED\n");
5917 ret = PTR_ERR(rb_hammer);
5918 goto out_free;
5919 }
5920
5921 ring_buffer_record_on(buffer);
5922 /*
5923 * Show buffer is enabled before setting rb_test_started.
5924 * Yes there's a small race window where events could be
5925 * dropped and the thread wont catch it. But when a ring
5926 * buffer gets enabled, there will always be some kind of
5927 * delay before other CPUs see it. Thus, we don't care about
5928 * those dropped events. We care about events dropped after
5929 * the threads see that the buffer is active.
5930 */
5931 smp_wmb();
5932 rb_test_started = true;
5933
5934 set_current_state(TASK_INTERRUPTIBLE);
5935 /* Just run for 10 seconds */;
5936 schedule_timeout(10 * HZ);
5937
5938 kthread_stop(rb_hammer);
5939
5940 out_free:
5941 for_each_online_cpu(cpu) {
5942 if (!rb_threads[cpu])
5943 break;
5944 kthread_stop(rb_threads[cpu]);
5945 }
5946 if (ret) {
5947 ring_buffer_free(buffer);
5948 return ret;
5949 }
5950
5951 /* Report! */
5952 pr_info("finished\n");
5953 for_each_online_cpu(cpu) {
5954 struct ring_buffer_event *event;
5955 struct rb_test_data *data = &rb_data[cpu];
5956 struct rb_item *item;
5957 unsigned long total_events;
5958 unsigned long total_dropped;
5959 unsigned long total_written;
5960 unsigned long total_alloc;
5961 unsigned long total_read = 0;
5962 unsigned long total_size = 0;
5963 unsigned long total_len = 0;
5964 unsigned long total_lost = 0;
5965 unsigned long lost;
5966 int big_event_size;
5967 int small_event_size;
5968
5969 ret = -1;
5970
5971 total_events = data->events + data->events_nested;
5972 total_written = data->bytes_written + data->bytes_written_nested;
5973 total_alloc = data->bytes_alloc + data->bytes_alloc_nested;
5974 total_dropped = data->bytes_dropped + data->bytes_dropped_nested;
5975
5976 big_event_size = data->max_size + data->max_size_nested;
5977 small_event_size = data->min_size + data->min_size_nested;
5978
5979 pr_info("CPU %d:\n", cpu);
5980 pr_info(" events: %ld\n", total_events);
5981 pr_info(" dropped bytes: %ld\n", total_dropped);
5982 pr_info(" alloced bytes: %ld\n", total_alloc);
5983 pr_info(" written bytes: %ld\n", total_written);
5984 pr_info(" biggest event: %d\n", big_event_size);
5985 pr_info(" smallest event: %d\n", small_event_size);
5986
5987 if (RB_WARN_ON(buffer, total_dropped))
5988 break;
5989
5990 ret = 0;
5991
5992 while ((event = ring_buffer_consume(buffer, cpu, NULL, &lost))) {
5993 total_lost += lost;
5994 item = ring_buffer_event_data(event);
5995 total_len += ring_buffer_event_length(event);
5996 total_size += item->size + sizeof(struct rb_item);
5997 if (memcmp(&item->str[0], rb_string, item->size) != 0) {
5998 pr_info("FAILED!\n");
5999 pr_info("buffer had: %.*s\n", item->size, item->str);
6000 pr_info("expected: %.*s\n", item->size, rb_string);
6001 RB_WARN_ON(buffer, 1);
6002 ret = -1;
6003 break;
6004 }
6005 total_read++;
6006 }
6007 if (ret)
6008 break;
6009
6010 ret = -1;
6011
6012 pr_info(" read events: %ld\n", total_read);
6013 pr_info(" lost events: %ld\n", total_lost);
6014 pr_info(" total events: %ld\n", total_lost + total_read);
6015 pr_info(" recorded len bytes: %ld\n", total_len);
6016 pr_info(" recorded size bytes: %ld\n", total_size);
6017 if (total_lost)
6018 pr_info(" With dropped events, record len and size may not match\n"
6019 " alloced and written from above\n");
6020 if (!total_lost) {
6021 if (RB_WARN_ON(buffer, total_len != total_alloc ||
6022 total_size != total_written))
6023 break;
6024 }
6025 if (RB_WARN_ON(buffer, total_lost + total_read != total_events))
6026 break;
6027
6028 ret = 0;
6029 }
6030 if (!ret)
6031 pr_info("Ring buffer PASSED!\n");
6032
6033 ring_buffer_free(buffer);
6034 return 0;
6035}
6036
6037late_initcall(test_ringbuffer);
6038#endif /* CONFIG_RING_BUFFER_STARTUP_TEST */
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/local64.h>
31#include <asm/local.h>
32
33/*
34 * The "absolute" timestamp in the buffer is only 59 bits.
35 * If a clock has the 5 MSBs set, it needs to be saved and
36 * reinserted.
37 */
38#define TS_MSB (0xf8ULL << 56)
39#define ABS_TS_MASK (~TS_MSB)
40
41static void update_pages_handler(struct work_struct *work);
42
43/*
44 * The ring buffer header is special. We must manually up keep it.
45 */
46int ring_buffer_print_entry_header(struct trace_seq *s)
47{
48 trace_seq_puts(s, "# compressed entry header\n");
49 trace_seq_puts(s, "\ttype_len : 5 bits\n");
50 trace_seq_puts(s, "\ttime_delta : 27 bits\n");
51 trace_seq_puts(s, "\tarray : 32 bits\n");
52 trace_seq_putc(s, '\n');
53 trace_seq_printf(s, "\tpadding : type == %d\n",
54 RINGBUF_TYPE_PADDING);
55 trace_seq_printf(s, "\ttime_extend : type == %d\n",
56 RINGBUF_TYPE_TIME_EXTEND);
57 trace_seq_printf(s, "\ttime_stamp : type == %d\n",
58 RINGBUF_TYPE_TIME_STAMP);
59 trace_seq_printf(s, "\tdata max type_len == %d\n",
60 RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
61
62 return !trace_seq_has_overflowed(s);
63}
64
65/*
66 * The ring buffer is made up of a list of pages. A separate list of pages is
67 * allocated for each CPU. A writer may only write to a buffer that is
68 * associated with the CPU it is currently executing on. A reader may read
69 * from any per cpu buffer.
70 *
71 * The reader is special. For each per cpu buffer, the reader has its own
72 * reader page. When a reader has read the entire reader page, this reader
73 * page is swapped with another page in the ring buffer.
74 *
75 * Now, as long as the writer is off the reader page, the reader can do what
76 * ever it wants with that page. The writer will never write to that page
77 * again (as long as it is out of the ring buffer).
78 *
79 * Here's some silly ASCII art.
80 *
81 * +------+
82 * |reader| RING BUFFER
83 * |page |
84 * +------+ +---+ +---+ +---+
85 * | |-->| |-->| |
86 * +---+ +---+ +---+
87 * ^ |
88 * | |
89 * +---------------+
90 *
91 *
92 * +------+
93 * |reader| RING BUFFER
94 * |page |------------------v
95 * +------+ +---+ +---+ +---+
96 * | |-->| |-->| |
97 * +---+ +---+ +---+
98 * ^ |
99 * | |
100 * +---------------+
101 *
102 *
103 * +------+
104 * |reader| RING BUFFER
105 * |page |------------------v
106 * +------+ +---+ +---+ +---+
107 * ^ | |-->| |-->| |
108 * | +---+ +---+ +---+
109 * | |
110 * | |
111 * +------------------------------+
112 *
113 *
114 * +------+
115 * |buffer| RING BUFFER
116 * |page |------------------v
117 * +------+ +---+ +---+ +---+
118 * ^ | | | |-->| |
119 * | New +---+ +---+ +---+
120 * | Reader------^ |
121 * | page |
122 * +------------------------------+
123 *
124 *
125 * After we make this swap, the reader can hand this page off to the splice
126 * code and be done with it. It can even allocate a new page if it needs to
127 * and swap that into the ring buffer.
128 *
129 * We will be using cmpxchg soon to make all this lockless.
130 *
131 */
132
133/* Used for individual buffers (after the counter) */
134#define RB_BUFFER_OFF (1 << 20)
135
136#define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
137
138#define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
139#define RB_ALIGNMENT 4U
140#define RB_MAX_SMALL_DATA (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
141#define RB_EVNT_MIN_SIZE 8U /* two 32bit words */
142
143#ifndef CONFIG_HAVE_64BIT_ALIGNED_ACCESS
144# define RB_FORCE_8BYTE_ALIGNMENT 0
145# define RB_ARCH_ALIGNMENT RB_ALIGNMENT
146#else
147# define RB_FORCE_8BYTE_ALIGNMENT 1
148# define RB_ARCH_ALIGNMENT 8U
149#endif
150
151#define RB_ALIGN_DATA __aligned(RB_ARCH_ALIGNMENT)
152
153/* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
154#define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
155
156enum {
157 RB_LEN_TIME_EXTEND = 8,
158 RB_LEN_TIME_STAMP = 8,
159};
160
161#define skip_time_extend(event) \
162 ((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
163
164#define extended_time(event) \
165 (event->type_len >= RINGBUF_TYPE_TIME_EXTEND)
166
167static inline bool rb_null_event(struct ring_buffer_event *event)
168{
169 return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
170}
171
172static void rb_event_set_padding(struct ring_buffer_event *event)
173{
174 /* padding has a NULL time_delta */
175 event->type_len = RINGBUF_TYPE_PADDING;
176 event->time_delta = 0;
177}
178
179static unsigned
180rb_event_data_length(struct ring_buffer_event *event)
181{
182 unsigned length;
183
184 if (event->type_len)
185 length = event->type_len * RB_ALIGNMENT;
186 else
187 length = event->array[0];
188 return length + RB_EVNT_HDR_SIZE;
189}
190
191/*
192 * Return the length of the given event. Will return
193 * the length of the time extend if the event is a
194 * time extend.
195 */
196static inline unsigned
197rb_event_length(struct ring_buffer_event *event)
198{
199 switch (event->type_len) {
200 case RINGBUF_TYPE_PADDING:
201 if (rb_null_event(event))
202 /* undefined */
203 return -1;
204 return event->array[0] + RB_EVNT_HDR_SIZE;
205
206 case RINGBUF_TYPE_TIME_EXTEND:
207 return RB_LEN_TIME_EXTEND;
208
209 case RINGBUF_TYPE_TIME_STAMP:
210 return RB_LEN_TIME_STAMP;
211
212 case RINGBUF_TYPE_DATA:
213 return rb_event_data_length(event);
214 default:
215 WARN_ON_ONCE(1);
216 }
217 /* not hit */
218 return 0;
219}
220
221/*
222 * Return total length of time extend and data,
223 * or just the event length for all other events.
224 */
225static inline unsigned
226rb_event_ts_length(struct ring_buffer_event *event)
227{
228 unsigned len = 0;
229
230 if (extended_time(event)) {
231 /* time extends include the data event after it */
232 len = RB_LEN_TIME_EXTEND;
233 event = skip_time_extend(event);
234 }
235 return len + rb_event_length(event);
236}
237
238/**
239 * ring_buffer_event_length - return the length of the event
240 * @event: the event to get the length of
241 *
242 * Returns the size of the data load of a data event.
243 * If the event is something other than a data event, it
244 * returns the size of the event itself. With the exception
245 * of a TIME EXTEND, where it still returns the size of the
246 * data load of the data event after it.
247 */
248unsigned ring_buffer_event_length(struct ring_buffer_event *event)
249{
250 unsigned length;
251
252 if (extended_time(event))
253 event = skip_time_extend(event);
254
255 length = rb_event_length(event);
256 if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
257 return length;
258 length -= RB_EVNT_HDR_SIZE;
259 if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
260 length -= sizeof(event->array[0]);
261 return length;
262}
263EXPORT_SYMBOL_GPL(ring_buffer_event_length);
264
265/* inline for ring buffer fast paths */
266static __always_inline void *
267rb_event_data(struct ring_buffer_event *event)
268{
269 if (extended_time(event))
270 event = skip_time_extend(event);
271 WARN_ON_ONCE(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
272 /* If length is in len field, then array[0] has the data */
273 if (event->type_len)
274 return (void *)&event->array[0];
275 /* Otherwise length is in array[0] and array[1] has the data */
276 return (void *)&event->array[1];
277}
278
279/**
280 * ring_buffer_event_data - return the data of the event
281 * @event: the event to get the data from
282 */
283void *ring_buffer_event_data(struct ring_buffer_event *event)
284{
285 return rb_event_data(event);
286}
287EXPORT_SYMBOL_GPL(ring_buffer_event_data);
288
289#define for_each_buffer_cpu(buffer, cpu) \
290 for_each_cpu(cpu, buffer->cpumask)
291
292#define for_each_online_buffer_cpu(buffer, cpu) \
293 for_each_cpu_and(cpu, buffer->cpumask, cpu_online_mask)
294
295#define TS_SHIFT 27
296#define TS_MASK ((1ULL << TS_SHIFT) - 1)
297#define TS_DELTA_TEST (~TS_MASK)
298
299static u64 rb_event_time_stamp(struct ring_buffer_event *event)
300{
301 u64 ts;
302
303 ts = event->array[0];
304 ts <<= TS_SHIFT;
305 ts += event->time_delta;
306
307 return ts;
308}
309
310/* Flag when events were overwritten */
311#define RB_MISSED_EVENTS (1 << 31)
312/* Missed count stored at end */
313#define RB_MISSED_STORED (1 << 30)
314
315struct buffer_data_page {
316 u64 time_stamp; /* page time stamp */
317 local_t commit; /* write committed index */
318 unsigned char data[] RB_ALIGN_DATA; /* data of buffer page */
319};
320
321struct buffer_data_read_page {
322 unsigned order; /* order of the page */
323 struct buffer_data_page *data; /* actual data, stored in this page */
324};
325
326/*
327 * Note, the buffer_page list must be first. The buffer pages
328 * are allocated in cache lines, which means that each buffer
329 * page will be at the beginning of a cache line, and thus
330 * the least significant bits will be zero. We use this to
331 * add flags in the list struct pointers, to make the ring buffer
332 * lockless.
333 */
334struct buffer_page {
335 struct list_head list; /* list of buffer pages */
336 local_t write; /* index for next write */
337 unsigned read; /* index for next read */
338 local_t entries; /* entries on this page */
339 unsigned long real_end; /* real end of data */
340 unsigned order; /* order of the page */
341 struct buffer_data_page *page; /* Actual data page */
342};
343
344/*
345 * The buffer page counters, write and entries, must be reset
346 * atomically when crossing page boundaries. To synchronize this
347 * update, two counters are inserted into the number. One is
348 * the actual counter for the write position or count on the page.
349 *
350 * The other is a counter of updaters. Before an update happens
351 * the update partition of the counter is incremented. This will
352 * allow the updater to update the counter atomically.
353 *
354 * The counter is 20 bits, and the state data is 12.
355 */
356#define RB_WRITE_MASK 0xfffff
357#define RB_WRITE_INTCNT (1 << 20)
358
359static void rb_init_page(struct buffer_data_page *bpage)
360{
361 local_set(&bpage->commit, 0);
362}
363
364static __always_inline unsigned int rb_page_commit(struct buffer_page *bpage)
365{
366 return local_read(&bpage->page->commit);
367}
368
369static void free_buffer_page(struct buffer_page *bpage)
370{
371 free_pages((unsigned long)bpage->page, bpage->order);
372 kfree(bpage);
373}
374
375/*
376 * We need to fit the time_stamp delta into 27 bits.
377 */
378static inline bool test_time_stamp(u64 delta)
379{
380 return !!(delta & TS_DELTA_TEST);
381}
382
383struct rb_irq_work {
384 struct irq_work work;
385 wait_queue_head_t waiters;
386 wait_queue_head_t full_waiters;
387 bool waiters_pending;
388 bool full_waiters_pending;
389 bool wakeup_full;
390};
391
392/*
393 * Structure to hold event state and handle nested events.
394 */
395struct rb_event_info {
396 u64 ts;
397 u64 delta;
398 u64 before;
399 u64 after;
400 unsigned long length;
401 struct buffer_page *tail_page;
402 int add_timestamp;
403};
404
405/*
406 * Used for the add_timestamp
407 * NONE
408 * EXTEND - wants a time extend
409 * ABSOLUTE - the buffer requests all events to have absolute time stamps
410 * FORCE - force a full time stamp.
411 */
412enum {
413 RB_ADD_STAMP_NONE = 0,
414 RB_ADD_STAMP_EXTEND = BIT(1),
415 RB_ADD_STAMP_ABSOLUTE = BIT(2),
416 RB_ADD_STAMP_FORCE = BIT(3)
417};
418/*
419 * Used for which event context the event is in.
420 * TRANSITION = 0
421 * NMI = 1
422 * IRQ = 2
423 * SOFTIRQ = 3
424 * NORMAL = 4
425 *
426 * See trace_recursive_lock() comment below for more details.
427 */
428enum {
429 RB_CTX_TRANSITION,
430 RB_CTX_NMI,
431 RB_CTX_IRQ,
432 RB_CTX_SOFTIRQ,
433 RB_CTX_NORMAL,
434 RB_CTX_MAX
435};
436
437struct rb_time_struct {
438 local64_t time;
439};
440typedef struct rb_time_struct rb_time_t;
441
442#define MAX_NEST 5
443
444/*
445 * head_page == tail_page && head == tail then buffer is empty.
446 */
447struct ring_buffer_per_cpu {
448 int cpu;
449 atomic_t record_disabled;
450 atomic_t resize_disabled;
451 struct trace_buffer *buffer;
452 raw_spinlock_t reader_lock; /* serialize readers */
453 arch_spinlock_t lock;
454 struct lock_class_key lock_key;
455 struct buffer_data_page *free_page;
456 unsigned long nr_pages;
457 unsigned int current_context;
458 struct list_head *pages;
459 struct buffer_page *head_page; /* read from head */
460 struct buffer_page *tail_page; /* write to tail */
461 struct buffer_page *commit_page; /* committed pages */
462 struct buffer_page *reader_page;
463 unsigned long lost_events;
464 unsigned long last_overrun;
465 unsigned long nest;
466 local_t entries_bytes;
467 local_t entries;
468 local_t overrun;
469 local_t commit_overrun;
470 local_t dropped_events;
471 local_t committing;
472 local_t commits;
473 local_t pages_touched;
474 local_t pages_lost;
475 local_t pages_read;
476 long last_pages_touch;
477 size_t shortest_full;
478 unsigned long read;
479 unsigned long read_bytes;
480 rb_time_t write_stamp;
481 rb_time_t before_stamp;
482 u64 event_stamp[MAX_NEST];
483 u64 read_stamp;
484 /* pages removed since last reset */
485 unsigned long pages_removed;
486 /* ring buffer pages to update, > 0 to add, < 0 to remove */
487 long nr_pages_to_update;
488 struct list_head new_pages; /* new pages to add */
489 struct work_struct update_pages_work;
490 struct completion update_done;
491
492 struct rb_irq_work irq_work;
493};
494
495struct trace_buffer {
496 unsigned flags;
497 int cpus;
498 atomic_t record_disabled;
499 atomic_t resizing;
500 cpumask_var_t cpumask;
501
502 struct lock_class_key *reader_lock_key;
503
504 struct mutex mutex;
505
506 struct ring_buffer_per_cpu **buffers;
507
508 struct hlist_node node;
509 u64 (*clock)(void);
510
511 struct rb_irq_work irq_work;
512 bool time_stamp_abs;
513
514 unsigned int subbuf_size;
515 unsigned int subbuf_order;
516 unsigned int max_data_size;
517};
518
519struct ring_buffer_iter {
520 struct ring_buffer_per_cpu *cpu_buffer;
521 unsigned long head;
522 unsigned long next_event;
523 struct buffer_page *head_page;
524 struct buffer_page *cache_reader_page;
525 unsigned long cache_read;
526 unsigned long cache_pages_removed;
527 u64 read_stamp;
528 u64 page_stamp;
529 struct ring_buffer_event *event;
530 size_t event_size;
531 int missed_events;
532};
533
534int ring_buffer_print_page_header(struct trace_buffer *buffer, struct trace_seq *s)
535{
536 struct buffer_data_page field;
537
538 trace_seq_printf(s, "\tfield: u64 timestamp;\t"
539 "offset:0;\tsize:%u;\tsigned:%u;\n",
540 (unsigned int)sizeof(field.time_stamp),
541 (unsigned int)is_signed_type(u64));
542
543 trace_seq_printf(s, "\tfield: local_t commit;\t"
544 "offset:%u;\tsize:%u;\tsigned:%u;\n",
545 (unsigned int)offsetof(typeof(field), commit),
546 (unsigned int)sizeof(field.commit),
547 (unsigned int)is_signed_type(long));
548
549 trace_seq_printf(s, "\tfield: int overwrite;\t"
550 "offset:%u;\tsize:%u;\tsigned:%u;\n",
551 (unsigned int)offsetof(typeof(field), commit),
552 1,
553 (unsigned int)is_signed_type(long));
554
555 trace_seq_printf(s, "\tfield: char data;\t"
556 "offset:%u;\tsize:%u;\tsigned:%u;\n",
557 (unsigned int)offsetof(typeof(field), data),
558 (unsigned int)buffer->subbuf_size,
559 (unsigned int)is_signed_type(char));
560
561 return !trace_seq_has_overflowed(s);
562}
563
564static inline void rb_time_read(rb_time_t *t, u64 *ret)
565{
566 *ret = local64_read(&t->time);
567}
568static void rb_time_set(rb_time_t *t, u64 val)
569{
570 local64_set(&t->time, val);
571}
572
573/*
574 * Enable this to make sure that the event passed to
575 * ring_buffer_event_time_stamp() is not committed and also
576 * is on the buffer that it passed in.
577 */
578//#define RB_VERIFY_EVENT
579#ifdef RB_VERIFY_EVENT
580static struct list_head *rb_list_head(struct list_head *list);
581static void verify_event(struct ring_buffer_per_cpu *cpu_buffer,
582 void *event)
583{
584 struct buffer_page *page = cpu_buffer->commit_page;
585 struct buffer_page *tail_page = READ_ONCE(cpu_buffer->tail_page);
586 struct list_head *next;
587 long commit, write;
588 unsigned long addr = (unsigned long)event;
589 bool done = false;
590 int stop = 0;
591
592 /* Make sure the event exists and is not committed yet */
593 do {
594 if (page == tail_page || WARN_ON_ONCE(stop++ > 100))
595 done = true;
596 commit = local_read(&page->page->commit);
597 write = local_read(&page->write);
598 if (addr >= (unsigned long)&page->page->data[commit] &&
599 addr < (unsigned long)&page->page->data[write])
600 return;
601
602 next = rb_list_head(page->list.next);
603 page = list_entry(next, struct buffer_page, list);
604 } while (!done);
605 WARN_ON_ONCE(1);
606}
607#else
608static inline void verify_event(struct ring_buffer_per_cpu *cpu_buffer,
609 void *event)
610{
611}
612#endif
613
614/*
615 * The absolute time stamp drops the 5 MSBs and some clocks may
616 * require them. The rb_fix_abs_ts() will take a previous full
617 * time stamp, and add the 5 MSB of that time stamp on to the
618 * saved absolute time stamp. Then they are compared in case of
619 * the unlikely event that the latest time stamp incremented
620 * the 5 MSB.
621 */
622static inline u64 rb_fix_abs_ts(u64 abs, u64 save_ts)
623{
624 if (save_ts & TS_MSB) {
625 abs |= save_ts & TS_MSB;
626 /* Check for overflow */
627 if (unlikely(abs < save_ts))
628 abs += 1ULL << 59;
629 }
630 return abs;
631}
632
633static inline u64 rb_time_stamp(struct trace_buffer *buffer);
634
635/**
636 * ring_buffer_event_time_stamp - return the event's current time stamp
637 * @buffer: The buffer that the event is on
638 * @event: the event to get the time stamp of
639 *
640 * Note, this must be called after @event is reserved, and before it is
641 * committed to the ring buffer. And must be called from the same
642 * context where the event was reserved (normal, softirq, irq, etc).
643 *
644 * Returns the time stamp associated with the current event.
645 * If the event has an extended time stamp, then that is used as
646 * the time stamp to return.
647 * In the highly unlikely case that the event was nested more than
648 * the max nesting, then the write_stamp of the buffer is returned,
649 * otherwise current time is returned, but that really neither of
650 * the last two cases should ever happen.
651 */
652u64 ring_buffer_event_time_stamp(struct trace_buffer *buffer,
653 struct ring_buffer_event *event)
654{
655 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[smp_processor_id()];
656 unsigned int nest;
657 u64 ts;
658
659 /* If the event includes an absolute time, then just use that */
660 if (event->type_len == RINGBUF_TYPE_TIME_STAMP) {
661 ts = rb_event_time_stamp(event);
662 return rb_fix_abs_ts(ts, cpu_buffer->tail_page->page->time_stamp);
663 }
664
665 nest = local_read(&cpu_buffer->committing);
666 verify_event(cpu_buffer, event);
667 if (WARN_ON_ONCE(!nest))
668 goto fail;
669
670 /* Read the current saved nesting level time stamp */
671 if (likely(--nest < MAX_NEST))
672 return cpu_buffer->event_stamp[nest];
673
674 /* Shouldn't happen, warn if it does */
675 WARN_ONCE(1, "nest (%d) greater than max", nest);
676
677 fail:
678 rb_time_read(&cpu_buffer->write_stamp, &ts);
679
680 return ts;
681}
682
683/**
684 * ring_buffer_nr_pages - get the number of buffer pages in the ring buffer
685 * @buffer: The ring_buffer to get the number of pages from
686 * @cpu: The cpu of the ring_buffer to get the number of pages from
687 *
688 * Returns the number of pages used by a per_cpu buffer of the ring buffer.
689 */
690size_t ring_buffer_nr_pages(struct trace_buffer *buffer, int cpu)
691{
692 return buffer->buffers[cpu]->nr_pages;
693}
694
695/**
696 * ring_buffer_nr_dirty_pages - get the number of used pages in the ring buffer
697 * @buffer: The ring_buffer to get the number of pages from
698 * @cpu: The cpu of the ring_buffer to get the number of pages from
699 *
700 * Returns the number of pages that have content in the ring buffer.
701 */
702size_t ring_buffer_nr_dirty_pages(struct trace_buffer *buffer, int cpu)
703{
704 size_t read;
705 size_t lost;
706 size_t cnt;
707
708 read = local_read(&buffer->buffers[cpu]->pages_read);
709 lost = local_read(&buffer->buffers[cpu]->pages_lost);
710 cnt = local_read(&buffer->buffers[cpu]->pages_touched);
711
712 if (WARN_ON_ONCE(cnt < lost))
713 return 0;
714
715 cnt -= lost;
716
717 /* The reader can read an empty page, but not more than that */
718 if (cnt < read) {
719 WARN_ON_ONCE(read > cnt + 1);
720 return 0;
721 }
722
723 return cnt - read;
724}
725
726static __always_inline bool full_hit(struct trace_buffer *buffer, int cpu, int full)
727{
728 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
729 size_t nr_pages;
730 size_t dirty;
731
732 nr_pages = cpu_buffer->nr_pages;
733 if (!nr_pages || !full)
734 return true;
735
736 /*
737 * Add one as dirty will never equal nr_pages, as the sub-buffer
738 * that the writer is on is not counted as dirty.
739 * This is needed if "buffer_percent" is set to 100.
740 */
741 dirty = ring_buffer_nr_dirty_pages(buffer, cpu) + 1;
742
743 return (dirty * 100) >= (full * nr_pages);
744}
745
746/*
747 * rb_wake_up_waiters - wake up tasks waiting for ring buffer input
748 *
749 * Schedules a delayed work to wake up any task that is blocked on the
750 * ring buffer waiters queue.
751 */
752static void rb_wake_up_waiters(struct irq_work *work)
753{
754 struct rb_irq_work *rbwork = container_of(work, struct rb_irq_work, work);
755
756 wake_up_all(&rbwork->waiters);
757 if (rbwork->full_waiters_pending || rbwork->wakeup_full) {
758 /* Only cpu_buffer sets the above flags */
759 struct ring_buffer_per_cpu *cpu_buffer =
760 container_of(rbwork, struct ring_buffer_per_cpu, irq_work);
761
762 /* Called from interrupt context */
763 raw_spin_lock(&cpu_buffer->reader_lock);
764 rbwork->wakeup_full = false;
765 rbwork->full_waiters_pending = false;
766
767 /* Waking up all waiters, they will reset the shortest full */
768 cpu_buffer->shortest_full = 0;
769 raw_spin_unlock(&cpu_buffer->reader_lock);
770
771 wake_up_all(&rbwork->full_waiters);
772 }
773}
774
775/**
776 * ring_buffer_wake_waiters - wake up any waiters on this ring buffer
777 * @buffer: The ring buffer to wake waiters on
778 * @cpu: The CPU buffer to wake waiters on
779 *
780 * In the case of a file that represents a ring buffer is closing,
781 * it is prudent to wake up any waiters that are on this.
782 */
783void ring_buffer_wake_waiters(struct trace_buffer *buffer, int cpu)
784{
785 struct ring_buffer_per_cpu *cpu_buffer;
786 struct rb_irq_work *rbwork;
787
788 if (!buffer)
789 return;
790
791 if (cpu == RING_BUFFER_ALL_CPUS) {
792
793 /* Wake up individual ones too. One level recursion */
794 for_each_buffer_cpu(buffer, cpu)
795 ring_buffer_wake_waiters(buffer, cpu);
796
797 rbwork = &buffer->irq_work;
798 } else {
799 if (WARN_ON_ONCE(!buffer->buffers))
800 return;
801 if (WARN_ON_ONCE(cpu >= nr_cpu_ids))
802 return;
803
804 cpu_buffer = buffer->buffers[cpu];
805 /* The CPU buffer may not have been initialized yet */
806 if (!cpu_buffer)
807 return;
808 rbwork = &cpu_buffer->irq_work;
809 }
810
811 /* This can be called in any context */
812 irq_work_queue(&rbwork->work);
813}
814
815static bool rb_watermark_hit(struct trace_buffer *buffer, int cpu, int full)
816{
817 struct ring_buffer_per_cpu *cpu_buffer;
818 bool ret = false;
819
820 /* Reads of all CPUs always waits for any data */
821 if (cpu == RING_BUFFER_ALL_CPUS)
822 return !ring_buffer_empty(buffer);
823
824 cpu_buffer = buffer->buffers[cpu];
825
826 if (!ring_buffer_empty_cpu(buffer, cpu)) {
827 unsigned long flags;
828 bool pagebusy;
829
830 if (!full)
831 return true;
832
833 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
834 pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
835 ret = !pagebusy && full_hit(buffer, cpu, full);
836
837 if (!cpu_buffer->shortest_full ||
838 cpu_buffer->shortest_full > full)
839 cpu_buffer->shortest_full = full;
840 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
841 }
842 return ret;
843}
844
845/**
846 * ring_buffer_wait - wait for input to the ring buffer
847 * @buffer: buffer to wait on
848 * @cpu: the cpu buffer to wait on
849 * @full: wait until the percentage of pages are available, if @cpu != RING_BUFFER_ALL_CPUS
850 *
851 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
852 * as data is added to any of the @buffer's cpu buffers. Otherwise
853 * it will wait for data to be added to a specific cpu buffer.
854 */
855int ring_buffer_wait(struct trace_buffer *buffer, int cpu, int full)
856{
857 struct ring_buffer_per_cpu *cpu_buffer;
858 DEFINE_WAIT(wait);
859 struct rb_irq_work *work;
860 int ret = 0;
861
862 /*
863 * Depending on what the caller is waiting for, either any
864 * data in any cpu buffer, or a specific buffer, put the
865 * caller on the appropriate wait queue.
866 */
867 if (cpu == RING_BUFFER_ALL_CPUS) {
868 work = &buffer->irq_work;
869 /* Full only makes sense on per cpu reads */
870 full = 0;
871 } else {
872 if (!cpumask_test_cpu(cpu, buffer->cpumask))
873 return -ENODEV;
874 cpu_buffer = buffer->buffers[cpu];
875 work = &cpu_buffer->irq_work;
876 }
877
878 if (full)
879 prepare_to_wait(&work->full_waiters, &wait, TASK_INTERRUPTIBLE);
880 else
881 prepare_to_wait(&work->waiters, &wait, TASK_INTERRUPTIBLE);
882
883 /*
884 * The events can happen in critical sections where
885 * checking a work queue can cause deadlocks.
886 * After adding a task to the queue, this flag is set
887 * only to notify events to try to wake up the queue
888 * using irq_work.
889 *
890 * We don't clear it even if the buffer is no longer
891 * empty. The flag only causes the next event to run
892 * irq_work to do the work queue wake up. The worse
893 * that can happen if we race with !trace_empty() is that
894 * an event will cause an irq_work to try to wake up
895 * an empty queue.
896 *
897 * There's no reason to protect this flag either, as
898 * the work queue and irq_work logic will do the necessary
899 * synchronization for the wake ups. The only thing
900 * that is necessary is that the wake up happens after
901 * a task has been queued. It's OK for spurious wake ups.
902 */
903 if (full)
904 work->full_waiters_pending = true;
905 else
906 work->waiters_pending = true;
907
908 if (rb_watermark_hit(buffer, cpu, full))
909 goto out;
910
911 if (signal_pending(current)) {
912 ret = -EINTR;
913 goto out;
914 }
915
916 schedule();
917 out:
918 if (full)
919 finish_wait(&work->full_waiters, &wait);
920 else
921 finish_wait(&work->waiters, &wait);
922
923 if (!ret && !rb_watermark_hit(buffer, cpu, full) && signal_pending(current))
924 ret = -EINTR;
925
926 return ret;
927}
928
929/**
930 * ring_buffer_poll_wait - poll on buffer input
931 * @buffer: buffer to wait on
932 * @cpu: the cpu buffer to wait on
933 * @filp: the file descriptor
934 * @poll_table: The poll descriptor
935 * @full: wait until the percentage of pages are available, if @cpu != RING_BUFFER_ALL_CPUS
936 *
937 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
938 * as data is added to any of the @buffer's cpu buffers. Otherwise
939 * it will wait for data to be added to a specific cpu buffer.
940 *
941 * Returns EPOLLIN | EPOLLRDNORM if data exists in the buffers,
942 * zero otherwise.
943 */
944__poll_t ring_buffer_poll_wait(struct trace_buffer *buffer, int cpu,
945 struct file *filp, poll_table *poll_table, int full)
946{
947 struct ring_buffer_per_cpu *cpu_buffer;
948 struct rb_irq_work *rbwork;
949
950 if (cpu == RING_BUFFER_ALL_CPUS) {
951 rbwork = &buffer->irq_work;
952 full = 0;
953 } else {
954 if (!cpumask_test_cpu(cpu, buffer->cpumask))
955 return EPOLLERR;
956
957 cpu_buffer = buffer->buffers[cpu];
958 rbwork = &cpu_buffer->irq_work;
959 }
960
961 if (full) {
962 unsigned long flags;
963
964 poll_wait(filp, &rbwork->full_waiters, poll_table);
965
966 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
967 rbwork->full_waiters_pending = true;
968 if (!cpu_buffer->shortest_full ||
969 cpu_buffer->shortest_full > full)
970 cpu_buffer->shortest_full = full;
971 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
972 } else {
973 poll_wait(filp, &rbwork->waiters, poll_table);
974 rbwork->waiters_pending = true;
975 }
976
977 /*
978 * There's a tight race between setting the waiters_pending and
979 * checking if the ring buffer is empty. Once the waiters_pending bit
980 * is set, the next event will wake the task up, but we can get stuck
981 * if there's only a single event in.
982 *
983 * FIXME: Ideally, we need a memory barrier on the writer side as well,
984 * but adding a memory barrier to all events will cause too much of a
985 * performance hit in the fast path. We only need a memory barrier when
986 * the buffer goes from empty to having content. But as this race is
987 * extremely small, and it's not a problem if another event comes in, we
988 * will fix it later.
989 */
990 smp_mb();
991
992 if (full)
993 return full_hit(buffer, cpu, full) ? EPOLLIN | EPOLLRDNORM : 0;
994
995 if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
996 (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
997 return EPOLLIN | EPOLLRDNORM;
998 return 0;
999}
1000
1001/* buffer may be either ring_buffer or ring_buffer_per_cpu */
1002#define RB_WARN_ON(b, cond) \
1003 ({ \
1004 int _____ret = unlikely(cond); \
1005 if (_____ret) { \
1006 if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
1007 struct ring_buffer_per_cpu *__b = \
1008 (void *)b; \
1009 atomic_inc(&__b->buffer->record_disabled); \
1010 } else \
1011 atomic_inc(&b->record_disabled); \
1012 WARN_ON(1); \
1013 } \
1014 _____ret; \
1015 })
1016
1017/* Up this if you want to test the TIME_EXTENTS and normalization */
1018#define DEBUG_SHIFT 0
1019
1020static inline u64 rb_time_stamp(struct trace_buffer *buffer)
1021{
1022 u64 ts;
1023
1024 /* Skip retpolines :-( */
1025 if (IS_ENABLED(CONFIG_RETPOLINE) && likely(buffer->clock == trace_clock_local))
1026 ts = trace_clock_local();
1027 else
1028 ts = buffer->clock();
1029
1030 /* shift to debug/test normalization and TIME_EXTENTS */
1031 return ts << DEBUG_SHIFT;
1032}
1033
1034u64 ring_buffer_time_stamp(struct trace_buffer *buffer)
1035{
1036 u64 time;
1037
1038 preempt_disable_notrace();
1039 time = rb_time_stamp(buffer);
1040 preempt_enable_notrace();
1041
1042 return time;
1043}
1044EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
1045
1046void ring_buffer_normalize_time_stamp(struct trace_buffer *buffer,
1047 int cpu, u64 *ts)
1048{
1049 /* Just stupid testing the normalize function and deltas */
1050 *ts >>= DEBUG_SHIFT;
1051}
1052EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
1053
1054/*
1055 * Making the ring buffer lockless makes things tricky.
1056 * Although writes only happen on the CPU that they are on,
1057 * and they only need to worry about interrupts. Reads can
1058 * happen on any CPU.
1059 *
1060 * The reader page is always off the ring buffer, but when the
1061 * reader finishes with a page, it needs to swap its page with
1062 * a new one from the buffer. The reader needs to take from
1063 * the head (writes go to the tail). But if a writer is in overwrite
1064 * mode and wraps, it must push the head page forward.
1065 *
1066 * Here lies the problem.
1067 *
1068 * The reader must be careful to replace only the head page, and
1069 * not another one. As described at the top of the file in the
1070 * ASCII art, the reader sets its old page to point to the next
1071 * page after head. It then sets the page after head to point to
1072 * the old reader page. But if the writer moves the head page
1073 * during this operation, the reader could end up with the tail.
1074 *
1075 * We use cmpxchg to help prevent this race. We also do something
1076 * special with the page before head. We set the LSB to 1.
1077 *
1078 * When the writer must push the page forward, it will clear the
1079 * bit that points to the head page, move the head, and then set
1080 * the bit that points to the new head page.
1081 *
1082 * We also don't want an interrupt coming in and moving the head
1083 * page on another writer. Thus we use the second LSB to catch
1084 * that too. Thus:
1085 *
1086 * head->list->prev->next bit 1 bit 0
1087 * ------- -------
1088 * Normal page 0 0
1089 * Points to head page 0 1
1090 * New head page 1 0
1091 *
1092 * Note we can not trust the prev pointer of the head page, because:
1093 *
1094 * +----+ +-----+ +-----+
1095 * | |------>| T |---X--->| N |
1096 * | |<------| | | |
1097 * +----+ +-----+ +-----+
1098 * ^ ^ |
1099 * | +-----+ | |
1100 * +----------| R |----------+ |
1101 * | |<-----------+
1102 * +-----+
1103 *
1104 * Key: ---X--> HEAD flag set in pointer
1105 * T Tail page
1106 * R Reader page
1107 * N Next page
1108 *
1109 * (see __rb_reserve_next() to see where this happens)
1110 *
1111 * What the above shows is that the reader just swapped out
1112 * the reader page with a page in the buffer, but before it
1113 * could make the new header point back to the new page added
1114 * it was preempted by a writer. The writer moved forward onto
1115 * the new page added by the reader and is about to move forward
1116 * again.
1117 *
1118 * You can see, it is legitimate for the previous pointer of
1119 * the head (or any page) not to point back to itself. But only
1120 * temporarily.
1121 */
1122
1123#define RB_PAGE_NORMAL 0UL
1124#define RB_PAGE_HEAD 1UL
1125#define RB_PAGE_UPDATE 2UL
1126
1127
1128#define RB_FLAG_MASK 3UL
1129
1130/* PAGE_MOVED is not part of the mask */
1131#define RB_PAGE_MOVED 4UL
1132
1133/*
1134 * rb_list_head - remove any bit
1135 */
1136static struct list_head *rb_list_head(struct list_head *list)
1137{
1138 unsigned long val = (unsigned long)list;
1139
1140 return (struct list_head *)(val & ~RB_FLAG_MASK);
1141}
1142
1143/*
1144 * rb_is_head_page - test if the given page is the head page
1145 *
1146 * Because the reader may move the head_page pointer, we can
1147 * not trust what the head page is (it may be pointing to
1148 * the reader page). But if the next page is a header page,
1149 * its flags will be non zero.
1150 */
1151static inline int
1152rb_is_head_page(struct buffer_page *page, struct list_head *list)
1153{
1154 unsigned long val;
1155
1156 val = (unsigned long)list->next;
1157
1158 if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
1159 return RB_PAGE_MOVED;
1160
1161 return val & RB_FLAG_MASK;
1162}
1163
1164/*
1165 * rb_is_reader_page
1166 *
1167 * The unique thing about the reader page, is that, if the
1168 * writer is ever on it, the previous pointer never points
1169 * back to the reader page.
1170 */
1171static bool rb_is_reader_page(struct buffer_page *page)
1172{
1173 struct list_head *list = page->list.prev;
1174
1175 return rb_list_head(list->next) != &page->list;
1176}
1177
1178/*
1179 * rb_set_list_to_head - set a list_head to be pointing to head.
1180 */
1181static void rb_set_list_to_head(struct list_head *list)
1182{
1183 unsigned long *ptr;
1184
1185 ptr = (unsigned long *)&list->next;
1186 *ptr |= RB_PAGE_HEAD;
1187 *ptr &= ~RB_PAGE_UPDATE;
1188}
1189
1190/*
1191 * rb_head_page_activate - sets up head page
1192 */
1193static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
1194{
1195 struct buffer_page *head;
1196
1197 head = cpu_buffer->head_page;
1198 if (!head)
1199 return;
1200
1201 /*
1202 * Set the previous list pointer to have the HEAD flag.
1203 */
1204 rb_set_list_to_head(head->list.prev);
1205}
1206
1207static void rb_list_head_clear(struct list_head *list)
1208{
1209 unsigned long *ptr = (unsigned long *)&list->next;
1210
1211 *ptr &= ~RB_FLAG_MASK;
1212}
1213
1214/*
1215 * rb_head_page_deactivate - clears head page ptr (for free list)
1216 */
1217static void
1218rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
1219{
1220 struct list_head *hd;
1221
1222 /* Go through the whole list and clear any pointers found. */
1223 rb_list_head_clear(cpu_buffer->pages);
1224
1225 list_for_each(hd, cpu_buffer->pages)
1226 rb_list_head_clear(hd);
1227}
1228
1229static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
1230 struct buffer_page *head,
1231 struct buffer_page *prev,
1232 int old_flag, int new_flag)
1233{
1234 struct list_head *list;
1235 unsigned long val = (unsigned long)&head->list;
1236 unsigned long ret;
1237
1238 list = &prev->list;
1239
1240 val &= ~RB_FLAG_MASK;
1241
1242 ret = cmpxchg((unsigned long *)&list->next,
1243 val | old_flag, val | new_flag);
1244
1245 /* check if the reader took the page */
1246 if ((ret & ~RB_FLAG_MASK) != val)
1247 return RB_PAGE_MOVED;
1248
1249 return ret & RB_FLAG_MASK;
1250}
1251
1252static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
1253 struct buffer_page *head,
1254 struct buffer_page *prev,
1255 int old_flag)
1256{
1257 return rb_head_page_set(cpu_buffer, head, prev,
1258 old_flag, RB_PAGE_UPDATE);
1259}
1260
1261static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
1262 struct buffer_page *head,
1263 struct buffer_page *prev,
1264 int old_flag)
1265{
1266 return rb_head_page_set(cpu_buffer, head, prev,
1267 old_flag, RB_PAGE_HEAD);
1268}
1269
1270static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
1271 struct buffer_page *head,
1272 struct buffer_page *prev,
1273 int old_flag)
1274{
1275 return rb_head_page_set(cpu_buffer, head, prev,
1276 old_flag, RB_PAGE_NORMAL);
1277}
1278
1279static inline void rb_inc_page(struct buffer_page **bpage)
1280{
1281 struct list_head *p = rb_list_head((*bpage)->list.next);
1282
1283 *bpage = list_entry(p, struct buffer_page, list);
1284}
1285
1286static struct buffer_page *
1287rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
1288{
1289 struct buffer_page *head;
1290 struct buffer_page *page;
1291 struct list_head *list;
1292 int i;
1293
1294 if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
1295 return NULL;
1296
1297 /* sanity check */
1298 list = cpu_buffer->pages;
1299 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
1300 return NULL;
1301
1302 page = head = cpu_buffer->head_page;
1303 /*
1304 * It is possible that the writer moves the header behind
1305 * where we started, and we miss in one loop.
1306 * A second loop should grab the header, but we'll do
1307 * three loops just because I'm paranoid.
1308 */
1309 for (i = 0; i < 3; i++) {
1310 do {
1311 if (rb_is_head_page(page, page->list.prev)) {
1312 cpu_buffer->head_page = page;
1313 return page;
1314 }
1315 rb_inc_page(&page);
1316 } while (page != head);
1317 }
1318
1319 RB_WARN_ON(cpu_buffer, 1);
1320
1321 return NULL;
1322}
1323
1324static bool rb_head_page_replace(struct buffer_page *old,
1325 struct buffer_page *new)
1326{
1327 unsigned long *ptr = (unsigned long *)&old->list.prev->next;
1328 unsigned long val;
1329
1330 val = *ptr & ~RB_FLAG_MASK;
1331 val |= RB_PAGE_HEAD;
1332
1333 return try_cmpxchg(ptr, &val, (unsigned long)&new->list);
1334}
1335
1336/*
1337 * rb_tail_page_update - move the tail page forward
1338 */
1339static void rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
1340 struct buffer_page *tail_page,
1341 struct buffer_page *next_page)
1342{
1343 unsigned long old_entries;
1344 unsigned long old_write;
1345
1346 /*
1347 * The tail page now needs to be moved forward.
1348 *
1349 * We need to reset the tail page, but without messing
1350 * with possible erasing of data brought in by interrupts
1351 * that have moved the tail page and are currently on it.
1352 *
1353 * We add a counter to the write field to denote this.
1354 */
1355 old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
1356 old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
1357
1358 local_inc(&cpu_buffer->pages_touched);
1359 /*
1360 * Just make sure we have seen our old_write and synchronize
1361 * with any interrupts that come in.
1362 */
1363 barrier();
1364
1365 /*
1366 * If the tail page is still the same as what we think
1367 * it is, then it is up to us to update the tail
1368 * pointer.
1369 */
1370 if (tail_page == READ_ONCE(cpu_buffer->tail_page)) {
1371 /* Zero the write counter */
1372 unsigned long val = old_write & ~RB_WRITE_MASK;
1373 unsigned long eval = old_entries & ~RB_WRITE_MASK;
1374
1375 /*
1376 * This will only succeed if an interrupt did
1377 * not come in and change it. In which case, we
1378 * do not want to modify it.
1379 *
1380 * We add (void) to let the compiler know that we do not care
1381 * about the return value of these functions. We use the
1382 * cmpxchg to only update if an interrupt did not already
1383 * do it for us. If the cmpxchg fails, we don't care.
1384 */
1385 (void)local_cmpxchg(&next_page->write, old_write, val);
1386 (void)local_cmpxchg(&next_page->entries, old_entries, eval);
1387
1388 /*
1389 * No need to worry about races with clearing out the commit.
1390 * it only can increment when a commit takes place. But that
1391 * only happens in the outer most nested commit.
1392 */
1393 local_set(&next_page->page->commit, 0);
1394
1395 /* Again, either we update tail_page or an interrupt does */
1396 (void)cmpxchg(&cpu_buffer->tail_page, tail_page, next_page);
1397 }
1398}
1399
1400static void rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
1401 struct buffer_page *bpage)
1402{
1403 unsigned long val = (unsigned long)bpage;
1404
1405 RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK);
1406}
1407
1408/**
1409 * rb_check_pages - integrity check of buffer pages
1410 * @cpu_buffer: CPU buffer with pages to test
1411 *
1412 * As a safety measure we check to make sure the data pages have not
1413 * been corrupted.
1414 */
1415static void rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
1416{
1417 struct list_head *head = rb_list_head(cpu_buffer->pages);
1418 struct list_head *tmp;
1419
1420 if (RB_WARN_ON(cpu_buffer,
1421 rb_list_head(rb_list_head(head->next)->prev) != head))
1422 return;
1423
1424 if (RB_WARN_ON(cpu_buffer,
1425 rb_list_head(rb_list_head(head->prev)->next) != head))
1426 return;
1427
1428 for (tmp = rb_list_head(head->next); tmp != head; tmp = rb_list_head(tmp->next)) {
1429 if (RB_WARN_ON(cpu_buffer,
1430 rb_list_head(rb_list_head(tmp->next)->prev) != tmp))
1431 return;
1432
1433 if (RB_WARN_ON(cpu_buffer,
1434 rb_list_head(rb_list_head(tmp->prev)->next) != tmp))
1435 return;
1436 }
1437}
1438
1439static int __rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1440 long nr_pages, struct list_head *pages)
1441{
1442 struct buffer_page *bpage, *tmp;
1443 bool user_thread = current->mm != NULL;
1444 gfp_t mflags;
1445 long i;
1446
1447 /*
1448 * Check if the available memory is there first.
1449 * Note, si_mem_available() only gives us a rough estimate of available
1450 * memory. It may not be accurate. But we don't care, we just want
1451 * to prevent doing any allocation when it is obvious that it is
1452 * not going to succeed.
1453 */
1454 i = si_mem_available();
1455 if (i < nr_pages)
1456 return -ENOMEM;
1457
1458 /*
1459 * __GFP_RETRY_MAYFAIL flag makes sure that the allocation fails
1460 * gracefully without invoking oom-killer and the system is not
1461 * destabilized.
1462 */
1463 mflags = GFP_KERNEL | __GFP_RETRY_MAYFAIL;
1464
1465 /*
1466 * If a user thread allocates too much, and si_mem_available()
1467 * reports there's enough memory, even though there is not.
1468 * Make sure the OOM killer kills this thread. This can happen
1469 * even with RETRY_MAYFAIL because another task may be doing
1470 * an allocation after this task has taken all memory.
1471 * This is the task the OOM killer needs to take out during this
1472 * loop, even if it was triggered by an allocation somewhere else.
1473 */
1474 if (user_thread)
1475 set_current_oom_origin();
1476 for (i = 0; i < nr_pages; i++) {
1477 struct page *page;
1478
1479 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1480 mflags, cpu_to_node(cpu_buffer->cpu));
1481 if (!bpage)
1482 goto free_pages;
1483
1484 rb_check_bpage(cpu_buffer, bpage);
1485
1486 list_add(&bpage->list, pages);
1487
1488 page = alloc_pages_node(cpu_to_node(cpu_buffer->cpu), mflags,
1489 cpu_buffer->buffer->subbuf_order);
1490 if (!page)
1491 goto free_pages;
1492 bpage->page = page_address(page);
1493 bpage->order = cpu_buffer->buffer->subbuf_order;
1494 rb_init_page(bpage->page);
1495
1496 if (user_thread && fatal_signal_pending(current))
1497 goto free_pages;
1498 }
1499 if (user_thread)
1500 clear_current_oom_origin();
1501
1502 return 0;
1503
1504free_pages:
1505 list_for_each_entry_safe(bpage, tmp, pages, list) {
1506 list_del_init(&bpage->list);
1507 free_buffer_page(bpage);
1508 }
1509 if (user_thread)
1510 clear_current_oom_origin();
1511
1512 return -ENOMEM;
1513}
1514
1515static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1516 unsigned long nr_pages)
1517{
1518 LIST_HEAD(pages);
1519
1520 WARN_ON(!nr_pages);
1521
1522 if (__rb_allocate_pages(cpu_buffer, nr_pages, &pages))
1523 return -ENOMEM;
1524
1525 /*
1526 * The ring buffer page list is a circular list that does not
1527 * start and end with a list head. All page list items point to
1528 * other pages.
1529 */
1530 cpu_buffer->pages = pages.next;
1531 list_del(&pages);
1532
1533 cpu_buffer->nr_pages = nr_pages;
1534
1535 rb_check_pages(cpu_buffer);
1536
1537 return 0;
1538}
1539
1540static struct ring_buffer_per_cpu *
1541rb_allocate_cpu_buffer(struct trace_buffer *buffer, long nr_pages, int cpu)
1542{
1543 struct ring_buffer_per_cpu *cpu_buffer;
1544 struct buffer_page *bpage;
1545 struct page *page;
1546 int ret;
1547
1548 cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1549 GFP_KERNEL, cpu_to_node(cpu));
1550 if (!cpu_buffer)
1551 return NULL;
1552
1553 cpu_buffer->cpu = cpu;
1554 cpu_buffer->buffer = buffer;
1555 raw_spin_lock_init(&cpu_buffer->reader_lock);
1556 lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1557 cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1558 INIT_WORK(&cpu_buffer->update_pages_work, update_pages_handler);
1559 init_completion(&cpu_buffer->update_done);
1560 init_irq_work(&cpu_buffer->irq_work.work, rb_wake_up_waiters);
1561 init_waitqueue_head(&cpu_buffer->irq_work.waiters);
1562 init_waitqueue_head(&cpu_buffer->irq_work.full_waiters);
1563
1564 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1565 GFP_KERNEL, cpu_to_node(cpu));
1566 if (!bpage)
1567 goto fail_free_buffer;
1568
1569 rb_check_bpage(cpu_buffer, bpage);
1570
1571 cpu_buffer->reader_page = bpage;
1572
1573 page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, cpu_buffer->buffer->subbuf_order);
1574 if (!page)
1575 goto fail_free_reader;
1576 bpage->page = page_address(page);
1577 rb_init_page(bpage->page);
1578
1579 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1580 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1581
1582 ret = rb_allocate_pages(cpu_buffer, nr_pages);
1583 if (ret < 0)
1584 goto fail_free_reader;
1585
1586 cpu_buffer->head_page
1587 = list_entry(cpu_buffer->pages, struct buffer_page, list);
1588 cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1589
1590 rb_head_page_activate(cpu_buffer);
1591
1592 return cpu_buffer;
1593
1594 fail_free_reader:
1595 free_buffer_page(cpu_buffer->reader_page);
1596
1597 fail_free_buffer:
1598 kfree(cpu_buffer);
1599 return NULL;
1600}
1601
1602static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1603{
1604 struct list_head *head = cpu_buffer->pages;
1605 struct buffer_page *bpage, *tmp;
1606
1607 irq_work_sync(&cpu_buffer->irq_work.work);
1608
1609 free_buffer_page(cpu_buffer->reader_page);
1610
1611 if (head) {
1612 rb_head_page_deactivate(cpu_buffer);
1613
1614 list_for_each_entry_safe(bpage, tmp, head, list) {
1615 list_del_init(&bpage->list);
1616 free_buffer_page(bpage);
1617 }
1618 bpage = list_entry(head, struct buffer_page, list);
1619 free_buffer_page(bpage);
1620 }
1621
1622 free_page((unsigned long)cpu_buffer->free_page);
1623
1624 kfree(cpu_buffer);
1625}
1626
1627/**
1628 * __ring_buffer_alloc - allocate a new ring_buffer
1629 * @size: the size in bytes per cpu that is needed.
1630 * @flags: attributes to set for the ring buffer.
1631 * @key: ring buffer reader_lock_key.
1632 *
1633 * Currently the only flag that is available is the RB_FL_OVERWRITE
1634 * flag. This flag means that the buffer will overwrite old data
1635 * when the buffer wraps. If this flag is not set, the buffer will
1636 * drop data when the tail hits the head.
1637 */
1638struct trace_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1639 struct lock_class_key *key)
1640{
1641 struct trace_buffer *buffer;
1642 long nr_pages;
1643 int bsize;
1644 int cpu;
1645 int ret;
1646
1647 /* keep it in its own cache line */
1648 buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1649 GFP_KERNEL);
1650 if (!buffer)
1651 return NULL;
1652
1653 if (!zalloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1654 goto fail_free_buffer;
1655
1656 /* Default buffer page size - one system page */
1657 buffer->subbuf_order = 0;
1658 buffer->subbuf_size = PAGE_SIZE - BUF_PAGE_HDR_SIZE;
1659
1660 /* Max payload is buffer page size - header (8bytes) */
1661 buffer->max_data_size = buffer->subbuf_size - (sizeof(u32) * 2);
1662
1663 nr_pages = DIV_ROUND_UP(size, buffer->subbuf_size);
1664 buffer->flags = flags;
1665 buffer->clock = trace_clock_local;
1666 buffer->reader_lock_key = key;
1667
1668 init_irq_work(&buffer->irq_work.work, rb_wake_up_waiters);
1669 init_waitqueue_head(&buffer->irq_work.waiters);
1670
1671 /* need at least two pages */
1672 if (nr_pages < 2)
1673 nr_pages = 2;
1674
1675 buffer->cpus = nr_cpu_ids;
1676
1677 bsize = sizeof(void *) * nr_cpu_ids;
1678 buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1679 GFP_KERNEL);
1680 if (!buffer->buffers)
1681 goto fail_free_cpumask;
1682
1683 cpu = raw_smp_processor_id();
1684 cpumask_set_cpu(cpu, buffer->cpumask);
1685 buffer->buffers[cpu] = rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
1686 if (!buffer->buffers[cpu])
1687 goto fail_free_buffers;
1688
1689 ret = cpuhp_state_add_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1690 if (ret < 0)
1691 goto fail_free_buffers;
1692
1693 mutex_init(&buffer->mutex);
1694
1695 return buffer;
1696
1697 fail_free_buffers:
1698 for_each_buffer_cpu(buffer, cpu) {
1699 if (buffer->buffers[cpu])
1700 rb_free_cpu_buffer(buffer->buffers[cpu]);
1701 }
1702 kfree(buffer->buffers);
1703
1704 fail_free_cpumask:
1705 free_cpumask_var(buffer->cpumask);
1706
1707 fail_free_buffer:
1708 kfree(buffer);
1709 return NULL;
1710}
1711EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1712
1713/**
1714 * ring_buffer_free - free a ring buffer.
1715 * @buffer: the buffer to free.
1716 */
1717void
1718ring_buffer_free(struct trace_buffer *buffer)
1719{
1720 int cpu;
1721
1722 cpuhp_state_remove_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1723
1724 irq_work_sync(&buffer->irq_work.work);
1725
1726 for_each_buffer_cpu(buffer, cpu)
1727 rb_free_cpu_buffer(buffer->buffers[cpu]);
1728
1729 kfree(buffer->buffers);
1730 free_cpumask_var(buffer->cpumask);
1731
1732 kfree(buffer);
1733}
1734EXPORT_SYMBOL_GPL(ring_buffer_free);
1735
1736void ring_buffer_set_clock(struct trace_buffer *buffer,
1737 u64 (*clock)(void))
1738{
1739 buffer->clock = clock;
1740}
1741
1742void ring_buffer_set_time_stamp_abs(struct trace_buffer *buffer, bool abs)
1743{
1744 buffer->time_stamp_abs = abs;
1745}
1746
1747bool ring_buffer_time_stamp_abs(struct trace_buffer *buffer)
1748{
1749 return buffer->time_stamp_abs;
1750}
1751
1752static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1753
1754static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1755{
1756 return local_read(&bpage->entries) & RB_WRITE_MASK;
1757}
1758
1759static inline unsigned long rb_page_write(struct buffer_page *bpage)
1760{
1761 return local_read(&bpage->write) & RB_WRITE_MASK;
1762}
1763
1764static bool
1765rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned long nr_pages)
1766{
1767 struct list_head *tail_page, *to_remove, *next_page;
1768 struct buffer_page *to_remove_page, *tmp_iter_page;
1769 struct buffer_page *last_page, *first_page;
1770 unsigned long nr_removed;
1771 unsigned long head_bit;
1772 int page_entries;
1773
1774 head_bit = 0;
1775
1776 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1777 atomic_inc(&cpu_buffer->record_disabled);
1778 /*
1779 * We don't race with the readers since we have acquired the reader
1780 * lock. We also don't race with writers after disabling recording.
1781 * This makes it easy to figure out the first and the last page to be
1782 * removed from the list. We unlink all the pages in between including
1783 * the first and last pages. This is done in a busy loop so that we
1784 * lose the least number of traces.
1785 * The pages are freed after we restart recording and unlock readers.
1786 */
1787 tail_page = &cpu_buffer->tail_page->list;
1788
1789 /*
1790 * tail page might be on reader page, we remove the next page
1791 * from the ring buffer
1792 */
1793 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
1794 tail_page = rb_list_head(tail_page->next);
1795 to_remove = tail_page;
1796
1797 /* start of pages to remove */
1798 first_page = list_entry(rb_list_head(to_remove->next),
1799 struct buffer_page, list);
1800
1801 for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
1802 to_remove = rb_list_head(to_remove)->next;
1803 head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
1804 }
1805 /* Read iterators need to reset themselves when some pages removed */
1806 cpu_buffer->pages_removed += nr_removed;
1807
1808 next_page = rb_list_head(to_remove)->next;
1809
1810 /*
1811 * Now we remove all pages between tail_page and next_page.
1812 * Make sure that we have head_bit value preserved for the
1813 * next page
1814 */
1815 tail_page->next = (struct list_head *)((unsigned long)next_page |
1816 head_bit);
1817 next_page = rb_list_head(next_page);
1818 next_page->prev = tail_page;
1819
1820 /* make sure pages points to a valid page in the ring buffer */
1821 cpu_buffer->pages = next_page;
1822
1823 /* update head page */
1824 if (head_bit)
1825 cpu_buffer->head_page = list_entry(next_page,
1826 struct buffer_page, list);
1827
1828 /* pages are removed, resume tracing and then free the pages */
1829 atomic_dec(&cpu_buffer->record_disabled);
1830 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1831
1832 RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
1833
1834 /* last buffer page to remove */
1835 last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
1836 list);
1837 tmp_iter_page = first_page;
1838
1839 do {
1840 cond_resched();
1841
1842 to_remove_page = tmp_iter_page;
1843 rb_inc_page(&tmp_iter_page);
1844
1845 /* update the counters */
1846 page_entries = rb_page_entries(to_remove_page);
1847 if (page_entries) {
1848 /*
1849 * If something was added to this page, it was full
1850 * since it is not the tail page. So we deduct the
1851 * bytes consumed in ring buffer from here.
1852 * Increment overrun to account for the lost events.
1853 */
1854 local_add(page_entries, &cpu_buffer->overrun);
1855 local_sub(rb_page_commit(to_remove_page), &cpu_buffer->entries_bytes);
1856 local_inc(&cpu_buffer->pages_lost);
1857 }
1858
1859 /*
1860 * We have already removed references to this list item, just
1861 * free up the buffer_page and its page
1862 */
1863 free_buffer_page(to_remove_page);
1864 nr_removed--;
1865
1866 } while (to_remove_page != last_page);
1867
1868 RB_WARN_ON(cpu_buffer, nr_removed);
1869
1870 return nr_removed == 0;
1871}
1872
1873static bool
1874rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
1875{
1876 struct list_head *pages = &cpu_buffer->new_pages;
1877 unsigned long flags;
1878 bool success;
1879 int retries;
1880
1881 /* Can be called at early boot up, where interrupts must not been enabled */
1882 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
1883 /*
1884 * We are holding the reader lock, so the reader page won't be swapped
1885 * in the ring buffer. Now we are racing with the writer trying to
1886 * move head page and the tail page.
1887 * We are going to adapt the reader page update process where:
1888 * 1. We first splice the start and end of list of new pages between
1889 * the head page and its previous page.
1890 * 2. We cmpxchg the prev_page->next to point from head page to the
1891 * start of new pages list.
1892 * 3. Finally, we update the head->prev to the end of new list.
1893 *
1894 * We will try this process 10 times, to make sure that we don't keep
1895 * spinning.
1896 */
1897 retries = 10;
1898 success = false;
1899 while (retries--) {
1900 struct list_head *head_page, *prev_page;
1901 struct list_head *last_page, *first_page;
1902 struct list_head *head_page_with_bit;
1903 struct buffer_page *hpage = rb_set_head_page(cpu_buffer);
1904
1905 if (!hpage)
1906 break;
1907 head_page = &hpage->list;
1908 prev_page = head_page->prev;
1909
1910 first_page = pages->next;
1911 last_page = pages->prev;
1912
1913 head_page_with_bit = (struct list_head *)
1914 ((unsigned long)head_page | RB_PAGE_HEAD);
1915
1916 last_page->next = head_page_with_bit;
1917 first_page->prev = prev_page;
1918
1919 /* caution: head_page_with_bit gets updated on cmpxchg failure */
1920 if (try_cmpxchg(&prev_page->next,
1921 &head_page_with_bit, first_page)) {
1922 /*
1923 * yay, we replaced the page pointer to our new list,
1924 * now, we just have to update to head page's prev
1925 * pointer to point to end of list
1926 */
1927 head_page->prev = last_page;
1928 success = true;
1929 break;
1930 }
1931 }
1932
1933 if (success)
1934 INIT_LIST_HEAD(pages);
1935 /*
1936 * If we weren't successful in adding in new pages, warn and stop
1937 * tracing
1938 */
1939 RB_WARN_ON(cpu_buffer, !success);
1940 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
1941
1942 /* free pages if they weren't inserted */
1943 if (!success) {
1944 struct buffer_page *bpage, *tmp;
1945 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1946 list) {
1947 list_del_init(&bpage->list);
1948 free_buffer_page(bpage);
1949 }
1950 }
1951 return success;
1952}
1953
1954static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
1955{
1956 bool success;
1957
1958 if (cpu_buffer->nr_pages_to_update > 0)
1959 success = rb_insert_pages(cpu_buffer);
1960 else
1961 success = rb_remove_pages(cpu_buffer,
1962 -cpu_buffer->nr_pages_to_update);
1963
1964 if (success)
1965 cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
1966}
1967
1968static void update_pages_handler(struct work_struct *work)
1969{
1970 struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
1971 struct ring_buffer_per_cpu, update_pages_work);
1972 rb_update_pages(cpu_buffer);
1973 complete(&cpu_buffer->update_done);
1974}
1975
1976/**
1977 * ring_buffer_resize - resize the ring buffer
1978 * @buffer: the buffer to resize.
1979 * @size: the new size.
1980 * @cpu_id: the cpu buffer to resize
1981 *
1982 * Minimum size is 2 * buffer->subbuf_size.
1983 *
1984 * Returns 0 on success and < 0 on failure.
1985 */
1986int ring_buffer_resize(struct trace_buffer *buffer, unsigned long size,
1987 int cpu_id)
1988{
1989 struct ring_buffer_per_cpu *cpu_buffer;
1990 unsigned long nr_pages;
1991 int cpu, err;
1992
1993 /*
1994 * Always succeed at resizing a non-existent buffer:
1995 */
1996 if (!buffer)
1997 return 0;
1998
1999 /* Make sure the requested buffer exists */
2000 if (cpu_id != RING_BUFFER_ALL_CPUS &&
2001 !cpumask_test_cpu(cpu_id, buffer->cpumask))
2002 return 0;
2003
2004 nr_pages = DIV_ROUND_UP(size, buffer->subbuf_size);
2005
2006 /* we need a minimum of two pages */
2007 if (nr_pages < 2)
2008 nr_pages = 2;
2009
2010 /* prevent another thread from changing buffer sizes */
2011 mutex_lock(&buffer->mutex);
2012 atomic_inc(&buffer->resizing);
2013
2014 if (cpu_id == RING_BUFFER_ALL_CPUS) {
2015 /*
2016 * Don't succeed if resizing is disabled, as a reader might be
2017 * manipulating the ring buffer and is expecting a sane state while
2018 * this is true.
2019 */
2020 for_each_buffer_cpu(buffer, cpu) {
2021 cpu_buffer = buffer->buffers[cpu];
2022 if (atomic_read(&cpu_buffer->resize_disabled)) {
2023 err = -EBUSY;
2024 goto out_err_unlock;
2025 }
2026 }
2027
2028 /* calculate the pages to update */
2029 for_each_buffer_cpu(buffer, cpu) {
2030 cpu_buffer = buffer->buffers[cpu];
2031
2032 cpu_buffer->nr_pages_to_update = nr_pages -
2033 cpu_buffer->nr_pages;
2034 /*
2035 * nothing more to do for removing pages or no update
2036 */
2037 if (cpu_buffer->nr_pages_to_update <= 0)
2038 continue;
2039 /*
2040 * to add pages, make sure all new pages can be
2041 * allocated without receiving ENOMEM
2042 */
2043 INIT_LIST_HEAD(&cpu_buffer->new_pages);
2044 if (__rb_allocate_pages(cpu_buffer, cpu_buffer->nr_pages_to_update,
2045 &cpu_buffer->new_pages)) {
2046 /* not enough memory for new pages */
2047 err = -ENOMEM;
2048 goto out_err;
2049 }
2050
2051 cond_resched();
2052 }
2053
2054 cpus_read_lock();
2055 /*
2056 * Fire off all the required work handlers
2057 * We can't schedule on offline CPUs, but it's not necessary
2058 * since we can change their buffer sizes without any race.
2059 */
2060 for_each_buffer_cpu(buffer, cpu) {
2061 cpu_buffer = buffer->buffers[cpu];
2062 if (!cpu_buffer->nr_pages_to_update)
2063 continue;
2064
2065 /* Can't run something on an offline CPU. */
2066 if (!cpu_online(cpu)) {
2067 rb_update_pages(cpu_buffer);
2068 cpu_buffer->nr_pages_to_update = 0;
2069 } else {
2070 /* Run directly if possible. */
2071 migrate_disable();
2072 if (cpu != smp_processor_id()) {
2073 migrate_enable();
2074 schedule_work_on(cpu,
2075 &cpu_buffer->update_pages_work);
2076 } else {
2077 update_pages_handler(&cpu_buffer->update_pages_work);
2078 migrate_enable();
2079 }
2080 }
2081 }
2082
2083 /* wait for all the updates to complete */
2084 for_each_buffer_cpu(buffer, cpu) {
2085 cpu_buffer = buffer->buffers[cpu];
2086 if (!cpu_buffer->nr_pages_to_update)
2087 continue;
2088
2089 if (cpu_online(cpu))
2090 wait_for_completion(&cpu_buffer->update_done);
2091 cpu_buffer->nr_pages_to_update = 0;
2092 }
2093
2094 cpus_read_unlock();
2095 } else {
2096 cpu_buffer = buffer->buffers[cpu_id];
2097
2098 if (nr_pages == cpu_buffer->nr_pages)
2099 goto out;
2100
2101 /*
2102 * Don't succeed if resizing is disabled, as a reader might be
2103 * manipulating the ring buffer and is expecting a sane state while
2104 * this is true.
2105 */
2106 if (atomic_read(&cpu_buffer->resize_disabled)) {
2107 err = -EBUSY;
2108 goto out_err_unlock;
2109 }
2110
2111 cpu_buffer->nr_pages_to_update = nr_pages -
2112 cpu_buffer->nr_pages;
2113
2114 INIT_LIST_HEAD(&cpu_buffer->new_pages);
2115 if (cpu_buffer->nr_pages_to_update > 0 &&
2116 __rb_allocate_pages(cpu_buffer, cpu_buffer->nr_pages_to_update,
2117 &cpu_buffer->new_pages)) {
2118 err = -ENOMEM;
2119 goto out_err;
2120 }
2121
2122 cpus_read_lock();
2123
2124 /* Can't run something on an offline CPU. */
2125 if (!cpu_online(cpu_id))
2126 rb_update_pages(cpu_buffer);
2127 else {
2128 /* Run directly if possible. */
2129 migrate_disable();
2130 if (cpu_id == smp_processor_id()) {
2131 rb_update_pages(cpu_buffer);
2132 migrate_enable();
2133 } else {
2134 migrate_enable();
2135 schedule_work_on(cpu_id,
2136 &cpu_buffer->update_pages_work);
2137 wait_for_completion(&cpu_buffer->update_done);
2138 }
2139 }
2140
2141 cpu_buffer->nr_pages_to_update = 0;
2142 cpus_read_unlock();
2143 }
2144
2145 out:
2146 /*
2147 * The ring buffer resize can happen with the ring buffer
2148 * enabled, so that the update disturbs the tracing as little
2149 * as possible. But if the buffer is disabled, we do not need
2150 * to worry about that, and we can take the time to verify
2151 * that the buffer is not corrupt.
2152 */
2153 if (atomic_read(&buffer->record_disabled)) {
2154 atomic_inc(&buffer->record_disabled);
2155 /*
2156 * Even though the buffer was disabled, we must make sure
2157 * that it is truly disabled before calling rb_check_pages.
2158 * There could have been a race between checking
2159 * record_disable and incrementing it.
2160 */
2161 synchronize_rcu();
2162 for_each_buffer_cpu(buffer, cpu) {
2163 cpu_buffer = buffer->buffers[cpu];
2164 rb_check_pages(cpu_buffer);
2165 }
2166 atomic_dec(&buffer->record_disabled);
2167 }
2168
2169 atomic_dec(&buffer->resizing);
2170 mutex_unlock(&buffer->mutex);
2171 return 0;
2172
2173 out_err:
2174 for_each_buffer_cpu(buffer, cpu) {
2175 struct buffer_page *bpage, *tmp;
2176
2177 cpu_buffer = buffer->buffers[cpu];
2178 cpu_buffer->nr_pages_to_update = 0;
2179
2180 if (list_empty(&cpu_buffer->new_pages))
2181 continue;
2182
2183 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
2184 list) {
2185 list_del_init(&bpage->list);
2186 free_buffer_page(bpage);
2187 }
2188 }
2189 out_err_unlock:
2190 atomic_dec(&buffer->resizing);
2191 mutex_unlock(&buffer->mutex);
2192 return err;
2193}
2194EXPORT_SYMBOL_GPL(ring_buffer_resize);
2195
2196void ring_buffer_change_overwrite(struct trace_buffer *buffer, int val)
2197{
2198 mutex_lock(&buffer->mutex);
2199 if (val)
2200 buffer->flags |= RB_FL_OVERWRITE;
2201 else
2202 buffer->flags &= ~RB_FL_OVERWRITE;
2203 mutex_unlock(&buffer->mutex);
2204}
2205EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
2206
2207static __always_inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
2208{
2209 return bpage->page->data + index;
2210}
2211
2212static __always_inline struct ring_buffer_event *
2213rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
2214{
2215 return __rb_page_index(cpu_buffer->reader_page,
2216 cpu_buffer->reader_page->read);
2217}
2218
2219static struct ring_buffer_event *
2220rb_iter_head_event(struct ring_buffer_iter *iter)
2221{
2222 struct ring_buffer_event *event;
2223 struct buffer_page *iter_head_page = iter->head_page;
2224 unsigned long commit;
2225 unsigned length;
2226
2227 if (iter->head != iter->next_event)
2228 return iter->event;
2229
2230 /*
2231 * When the writer goes across pages, it issues a cmpxchg which
2232 * is a mb(), which will synchronize with the rmb here.
2233 * (see rb_tail_page_update() and __rb_reserve_next())
2234 */
2235 commit = rb_page_commit(iter_head_page);
2236 smp_rmb();
2237
2238 /* An event needs to be at least 8 bytes in size */
2239 if (iter->head > commit - 8)
2240 goto reset;
2241
2242 event = __rb_page_index(iter_head_page, iter->head);
2243 length = rb_event_length(event);
2244
2245 /*
2246 * READ_ONCE() doesn't work on functions and we don't want the
2247 * compiler doing any crazy optimizations with length.
2248 */
2249 barrier();
2250
2251 if ((iter->head + length) > commit || length > iter->event_size)
2252 /* Writer corrupted the read? */
2253 goto reset;
2254
2255 memcpy(iter->event, event, length);
2256 /*
2257 * If the page stamp is still the same after this rmb() then the
2258 * event was safely copied without the writer entering the page.
2259 */
2260 smp_rmb();
2261
2262 /* Make sure the page didn't change since we read this */
2263 if (iter->page_stamp != iter_head_page->page->time_stamp ||
2264 commit > rb_page_commit(iter_head_page))
2265 goto reset;
2266
2267 iter->next_event = iter->head + length;
2268 return iter->event;
2269 reset:
2270 /* Reset to the beginning */
2271 iter->page_stamp = iter->read_stamp = iter->head_page->page->time_stamp;
2272 iter->head = 0;
2273 iter->next_event = 0;
2274 iter->missed_events = 1;
2275 return NULL;
2276}
2277
2278/* Size is determined by what has been committed */
2279static __always_inline unsigned rb_page_size(struct buffer_page *bpage)
2280{
2281 return rb_page_commit(bpage);
2282}
2283
2284static __always_inline unsigned
2285rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
2286{
2287 return rb_page_commit(cpu_buffer->commit_page);
2288}
2289
2290static __always_inline unsigned
2291rb_event_index(struct ring_buffer_per_cpu *cpu_buffer, struct ring_buffer_event *event)
2292{
2293 unsigned long addr = (unsigned long)event;
2294
2295 addr &= (PAGE_SIZE << cpu_buffer->buffer->subbuf_order) - 1;
2296
2297 return addr - BUF_PAGE_HDR_SIZE;
2298}
2299
2300static void rb_inc_iter(struct ring_buffer_iter *iter)
2301{
2302 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
2303
2304 /*
2305 * The iterator could be on the reader page (it starts there).
2306 * But the head could have moved, since the reader was
2307 * found. Check for this case and assign the iterator
2308 * to the head page instead of next.
2309 */
2310 if (iter->head_page == cpu_buffer->reader_page)
2311 iter->head_page = rb_set_head_page(cpu_buffer);
2312 else
2313 rb_inc_page(&iter->head_page);
2314
2315 iter->page_stamp = iter->read_stamp = iter->head_page->page->time_stamp;
2316 iter->head = 0;
2317 iter->next_event = 0;
2318}
2319
2320/*
2321 * rb_handle_head_page - writer hit the head page
2322 *
2323 * Returns: +1 to retry page
2324 * 0 to continue
2325 * -1 on error
2326 */
2327static int
2328rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
2329 struct buffer_page *tail_page,
2330 struct buffer_page *next_page)
2331{
2332 struct buffer_page *new_head;
2333 int entries;
2334 int type;
2335 int ret;
2336
2337 entries = rb_page_entries(next_page);
2338
2339 /*
2340 * The hard part is here. We need to move the head
2341 * forward, and protect against both readers on
2342 * other CPUs and writers coming in via interrupts.
2343 */
2344 type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
2345 RB_PAGE_HEAD);
2346
2347 /*
2348 * type can be one of four:
2349 * NORMAL - an interrupt already moved it for us
2350 * HEAD - we are the first to get here.
2351 * UPDATE - we are the interrupt interrupting
2352 * a current move.
2353 * MOVED - a reader on another CPU moved the next
2354 * pointer to its reader page. Give up
2355 * and try again.
2356 */
2357
2358 switch (type) {
2359 case RB_PAGE_HEAD:
2360 /*
2361 * We changed the head to UPDATE, thus
2362 * it is our responsibility to update
2363 * the counters.
2364 */
2365 local_add(entries, &cpu_buffer->overrun);
2366 local_sub(rb_page_commit(next_page), &cpu_buffer->entries_bytes);
2367 local_inc(&cpu_buffer->pages_lost);
2368
2369 /*
2370 * The entries will be zeroed out when we move the
2371 * tail page.
2372 */
2373
2374 /* still more to do */
2375 break;
2376
2377 case RB_PAGE_UPDATE:
2378 /*
2379 * This is an interrupt that interrupt the
2380 * previous update. Still more to do.
2381 */
2382 break;
2383 case RB_PAGE_NORMAL:
2384 /*
2385 * An interrupt came in before the update
2386 * and processed this for us.
2387 * Nothing left to do.
2388 */
2389 return 1;
2390 case RB_PAGE_MOVED:
2391 /*
2392 * The reader is on another CPU and just did
2393 * a swap with our next_page.
2394 * Try again.
2395 */
2396 return 1;
2397 default:
2398 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
2399 return -1;
2400 }
2401
2402 /*
2403 * Now that we are here, the old head pointer is
2404 * set to UPDATE. This will keep the reader from
2405 * swapping the head page with the reader page.
2406 * The reader (on another CPU) will spin till
2407 * we are finished.
2408 *
2409 * We just need to protect against interrupts
2410 * doing the job. We will set the next pointer
2411 * to HEAD. After that, we set the old pointer
2412 * to NORMAL, but only if it was HEAD before.
2413 * otherwise we are an interrupt, and only
2414 * want the outer most commit to reset it.
2415 */
2416 new_head = next_page;
2417 rb_inc_page(&new_head);
2418
2419 ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
2420 RB_PAGE_NORMAL);
2421
2422 /*
2423 * Valid returns are:
2424 * HEAD - an interrupt came in and already set it.
2425 * NORMAL - One of two things:
2426 * 1) We really set it.
2427 * 2) A bunch of interrupts came in and moved
2428 * the page forward again.
2429 */
2430 switch (ret) {
2431 case RB_PAGE_HEAD:
2432 case RB_PAGE_NORMAL:
2433 /* OK */
2434 break;
2435 default:
2436 RB_WARN_ON(cpu_buffer, 1);
2437 return -1;
2438 }
2439
2440 /*
2441 * It is possible that an interrupt came in,
2442 * set the head up, then more interrupts came in
2443 * and moved it again. When we get back here,
2444 * the page would have been set to NORMAL but we
2445 * just set it back to HEAD.
2446 *
2447 * How do you detect this? Well, if that happened
2448 * the tail page would have moved.
2449 */
2450 if (ret == RB_PAGE_NORMAL) {
2451 struct buffer_page *buffer_tail_page;
2452
2453 buffer_tail_page = READ_ONCE(cpu_buffer->tail_page);
2454 /*
2455 * If the tail had moved passed next, then we need
2456 * to reset the pointer.
2457 */
2458 if (buffer_tail_page != tail_page &&
2459 buffer_tail_page != next_page)
2460 rb_head_page_set_normal(cpu_buffer, new_head,
2461 next_page,
2462 RB_PAGE_HEAD);
2463 }
2464
2465 /*
2466 * If this was the outer most commit (the one that
2467 * changed the original pointer from HEAD to UPDATE),
2468 * then it is up to us to reset it to NORMAL.
2469 */
2470 if (type == RB_PAGE_HEAD) {
2471 ret = rb_head_page_set_normal(cpu_buffer, next_page,
2472 tail_page,
2473 RB_PAGE_UPDATE);
2474 if (RB_WARN_ON(cpu_buffer,
2475 ret != RB_PAGE_UPDATE))
2476 return -1;
2477 }
2478
2479 return 0;
2480}
2481
2482static inline void
2483rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2484 unsigned long tail, struct rb_event_info *info)
2485{
2486 unsigned long bsize = READ_ONCE(cpu_buffer->buffer->subbuf_size);
2487 struct buffer_page *tail_page = info->tail_page;
2488 struct ring_buffer_event *event;
2489 unsigned long length = info->length;
2490
2491 /*
2492 * Only the event that crossed the page boundary
2493 * must fill the old tail_page with padding.
2494 */
2495 if (tail >= bsize) {
2496 /*
2497 * If the page was filled, then we still need
2498 * to update the real_end. Reset it to zero
2499 * and the reader will ignore it.
2500 */
2501 if (tail == bsize)
2502 tail_page->real_end = 0;
2503
2504 local_sub(length, &tail_page->write);
2505 return;
2506 }
2507
2508 event = __rb_page_index(tail_page, tail);
2509
2510 /*
2511 * Save the original length to the meta data.
2512 * This will be used by the reader to add lost event
2513 * counter.
2514 */
2515 tail_page->real_end = tail;
2516
2517 /*
2518 * If this event is bigger than the minimum size, then
2519 * we need to be careful that we don't subtract the
2520 * write counter enough to allow another writer to slip
2521 * in on this page.
2522 * We put in a discarded commit instead, to make sure
2523 * that this space is not used again, and this space will
2524 * not be accounted into 'entries_bytes'.
2525 *
2526 * If we are less than the minimum size, we don't need to
2527 * worry about it.
2528 */
2529 if (tail > (bsize - RB_EVNT_MIN_SIZE)) {
2530 /* No room for any events */
2531
2532 /* Mark the rest of the page with padding */
2533 rb_event_set_padding(event);
2534
2535 /* Make sure the padding is visible before the write update */
2536 smp_wmb();
2537
2538 /* Set the write back to the previous setting */
2539 local_sub(length, &tail_page->write);
2540 return;
2541 }
2542
2543 /* Put in a discarded event */
2544 event->array[0] = (bsize - tail) - RB_EVNT_HDR_SIZE;
2545 event->type_len = RINGBUF_TYPE_PADDING;
2546 /* time delta must be non zero */
2547 event->time_delta = 1;
2548
2549 /* account for padding bytes */
2550 local_add(bsize - tail, &cpu_buffer->entries_bytes);
2551
2552 /* Make sure the padding is visible before the tail_page->write update */
2553 smp_wmb();
2554
2555 /* Set write to end of buffer */
2556 length = (tail + length) - bsize;
2557 local_sub(length, &tail_page->write);
2558}
2559
2560static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer);
2561
2562/*
2563 * This is the slow path, force gcc not to inline it.
2564 */
2565static noinline struct ring_buffer_event *
2566rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2567 unsigned long tail, struct rb_event_info *info)
2568{
2569 struct buffer_page *tail_page = info->tail_page;
2570 struct buffer_page *commit_page = cpu_buffer->commit_page;
2571 struct trace_buffer *buffer = cpu_buffer->buffer;
2572 struct buffer_page *next_page;
2573 int ret;
2574
2575 next_page = tail_page;
2576
2577 rb_inc_page(&next_page);
2578
2579 /*
2580 * If for some reason, we had an interrupt storm that made
2581 * it all the way around the buffer, bail, and warn
2582 * about it.
2583 */
2584 if (unlikely(next_page == commit_page)) {
2585 local_inc(&cpu_buffer->commit_overrun);
2586 goto out_reset;
2587 }
2588
2589 /*
2590 * This is where the fun begins!
2591 *
2592 * We are fighting against races between a reader that
2593 * could be on another CPU trying to swap its reader
2594 * page with the buffer head.
2595 *
2596 * We are also fighting against interrupts coming in and
2597 * moving the head or tail on us as well.
2598 *
2599 * If the next page is the head page then we have filled
2600 * the buffer, unless the commit page is still on the
2601 * reader page.
2602 */
2603 if (rb_is_head_page(next_page, &tail_page->list)) {
2604
2605 /*
2606 * If the commit is not on the reader page, then
2607 * move the header page.
2608 */
2609 if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2610 /*
2611 * If we are not in overwrite mode,
2612 * this is easy, just stop here.
2613 */
2614 if (!(buffer->flags & RB_FL_OVERWRITE)) {
2615 local_inc(&cpu_buffer->dropped_events);
2616 goto out_reset;
2617 }
2618
2619 ret = rb_handle_head_page(cpu_buffer,
2620 tail_page,
2621 next_page);
2622 if (ret < 0)
2623 goto out_reset;
2624 if (ret)
2625 goto out_again;
2626 } else {
2627 /*
2628 * We need to be careful here too. The
2629 * commit page could still be on the reader
2630 * page. We could have a small buffer, and
2631 * have filled up the buffer with events
2632 * from interrupts and such, and wrapped.
2633 *
2634 * Note, if the tail page is also on the
2635 * reader_page, we let it move out.
2636 */
2637 if (unlikely((cpu_buffer->commit_page !=
2638 cpu_buffer->tail_page) &&
2639 (cpu_buffer->commit_page ==
2640 cpu_buffer->reader_page))) {
2641 local_inc(&cpu_buffer->commit_overrun);
2642 goto out_reset;
2643 }
2644 }
2645 }
2646
2647 rb_tail_page_update(cpu_buffer, tail_page, next_page);
2648
2649 out_again:
2650
2651 rb_reset_tail(cpu_buffer, tail, info);
2652
2653 /* Commit what we have for now. */
2654 rb_end_commit(cpu_buffer);
2655 /* rb_end_commit() decs committing */
2656 local_inc(&cpu_buffer->committing);
2657
2658 /* fail and let the caller try again */
2659 return ERR_PTR(-EAGAIN);
2660
2661 out_reset:
2662 /* reset write */
2663 rb_reset_tail(cpu_buffer, tail, info);
2664
2665 return NULL;
2666}
2667
2668/* Slow path */
2669static struct ring_buffer_event *
2670rb_add_time_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2671 struct ring_buffer_event *event, u64 delta, bool abs)
2672{
2673 if (abs)
2674 event->type_len = RINGBUF_TYPE_TIME_STAMP;
2675 else
2676 event->type_len = RINGBUF_TYPE_TIME_EXTEND;
2677
2678 /* Not the first event on the page, or not delta? */
2679 if (abs || rb_event_index(cpu_buffer, event)) {
2680 event->time_delta = delta & TS_MASK;
2681 event->array[0] = delta >> TS_SHIFT;
2682 } else {
2683 /* nope, just zero it */
2684 event->time_delta = 0;
2685 event->array[0] = 0;
2686 }
2687
2688 return skip_time_extend(event);
2689}
2690
2691#ifndef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2692static inline bool sched_clock_stable(void)
2693{
2694 return true;
2695}
2696#endif
2697
2698static void
2699rb_check_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2700 struct rb_event_info *info)
2701{
2702 u64 write_stamp;
2703
2704 WARN_ONCE(1, "Delta way too big! %llu ts=%llu before=%llu after=%llu write stamp=%llu\n%s",
2705 (unsigned long long)info->delta,
2706 (unsigned long long)info->ts,
2707 (unsigned long long)info->before,
2708 (unsigned long long)info->after,
2709 (unsigned long long)({rb_time_read(&cpu_buffer->write_stamp, &write_stamp); write_stamp;}),
2710 sched_clock_stable() ? "" :
2711 "If you just came from a suspend/resume,\n"
2712 "please switch to the trace global clock:\n"
2713 " echo global > /sys/kernel/tracing/trace_clock\n"
2714 "or add trace_clock=global to the kernel command line\n");
2715}
2716
2717static void rb_add_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2718 struct ring_buffer_event **event,
2719 struct rb_event_info *info,
2720 u64 *delta,
2721 unsigned int *length)
2722{
2723 bool abs = info->add_timestamp &
2724 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE);
2725
2726 if (unlikely(info->delta > (1ULL << 59))) {
2727 /*
2728 * Some timers can use more than 59 bits, and when a timestamp
2729 * is added to the buffer, it will lose those bits.
2730 */
2731 if (abs && (info->ts & TS_MSB)) {
2732 info->delta &= ABS_TS_MASK;
2733
2734 /* did the clock go backwards */
2735 } else if (info->before == info->after && info->before > info->ts) {
2736 /* not interrupted */
2737 static int once;
2738
2739 /*
2740 * This is possible with a recalibrating of the TSC.
2741 * Do not produce a call stack, but just report it.
2742 */
2743 if (!once) {
2744 once++;
2745 pr_warn("Ring buffer clock went backwards: %llu -> %llu\n",
2746 info->before, info->ts);
2747 }
2748 } else
2749 rb_check_timestamp(cpu_buffer, info);
2750 if (!abs)
2751 info->delta = 0;
2752 }
2753 *event = rb_add_time_stamp(cpu_buffer, *event, info->delta, abs);
2754 *length -= RB_LEN_TIME_EXTEND;
2755 *delta = 0;
2756}
2757
2758/**
2759 * rb_update_event - update event type and data
2760 * @cpu_buffer: The per cpu buffer of the @event
2761 * @event: the event to update
2762 * @info: The info to update the @event with (contains length and delta)
2763 *
2764 * Update the type and data fields of the @event. The length
2765 * is the actual size that is written to the ring buffer,
2766 * and with this, we can determine what to place into the
2767 * data field.
2768 */
2769static void
2770rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
2771 struct ring_buffer_event *event,
2772 struct rb_event_info *info)
2773{
2774 unsigned length = info->length;
2775 u64 delta = info->delta;
2776 unsigned int nest = local_read(&cpu_buffer->committing) - 1;
2777
2778 if (!WARN_ON_ONCE(nest >= MAX_NEST))
2779 cpu_buffer->event_stamp[nest] = info->ts;
2780
2781 /*
2782 * If we need to add a timestamp, then we
2783 * add it to the start of the reserved space.
2784 */
2785 if (unlikely(info->add_timestamp))
2786 rb_add_timestamp(cpu_buffer, &event, info, &delta, &length);
2787
2788 event->time_delta = delta;
2789 length -= RB_EVNT_HDR_SIZE;
2790 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
2791 event->type_len = 0;
2792 event->array[0] = length;
2793 } else
2794 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
2795}
2796
2797static unsigned rb_calculate_event_length(unsigned length)
2798{
2799 struct ring_buffer_event event; /* Used only for sizeof array */
2800
2801 /* zero length can cause confusions */
2802 if (!length)
2803 length++;
2804
2805 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
2806 length += sizeof(event.array[0]);
2807
2808 length += RB_EVNT_HDR_SIZE;
2809 length = ALIGN(length, RB_ARCH_ALIGNMENT);
2810
2811 /*
2812 * In case the time delta is larger than the 27 bits for it
2813 * in the header, we need to add a timestamp. If another
2814 * event comes in when trying to discard this one to increase
2815 * the length, then the timestamp will be added in the allocated
2816 * space of this event. If length is bigger than the size needed
2817 * for the TIME_EXTEND, then padding has to be used. The events
2818 * length must be either RB_LEN_TIME_EXTEND, or greater than or equal
2819 * to RB_LEN_TIME_EXTEND + 8, as 8 is the minimum size for padding.
2820 * As length is a multiple of 4, we only need to worry if it
2821 * is 12 (RB_LEN_TIME_EXTEND + 4).
2822 */
2823 if (length == RB_LEN_TIME_EXTEND + RB_ALIGNMENT)
2824 length += RB_ALIGNMENT;
2825
2826 return length;
2827}
2828
2829static inline bool
2830rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2831 struct ring_buffer_event *event)
2832{
2833 unsigned long new_index, old_index;
2834 struct buffer_page *bpage;
2835 unsigned long addr;
2836
2837 new_index = rb_event_index(cpu_buffer, event);
2838 old_index = new_index + rb_event_ts_length(event);
2839 addr = (unsigned long)event;
2840 addr &= ~((PAGE_SIZE << cpu_buffer->buffer->subbuf_order) - 1);
2841
2842 bpage = READ_ONCE(cpu_buffer->tail_page);
2843
2844 /*
2845 * Make sure the tail_page is still the same and
2846 * the next write location is the end of this event
2847 */
2848 if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2849 unsigned long write_mask =
2850 local_read(&bpage->write) & ~RB_WRITE_MASK;
2851 unsigned long event_length = rb_event_length(event);
2852
2853 /*
2854 * For the before_stamp to be different than the write_stamp
2855 * to make sure that the next event adds an absolute
2856 * value and does not rely on the saved write stamp, which
2857 * is now going to be bogus.
2858 *
2859 * By setting the before_stamp to zero, the next event
2860 * is not going to use the write_stamp and will instead
2861 * create an absolute timestamp. This means there's no
2862 * reason to update the wirte_stamp!
2863 */
2864 rb_time_set(&cpu_buffer->before_stamp, 0);
2865
2866 /*
2867 * If an event were to come in now, it would see that the
2868 * write_stamp and the before_stamp are different, and assume
2869 * that this event just added itself before updating
2870 * the write stamp. The interrupting event will fix the
2871 * write stamp for us, and use an absolute timestamp.
2872 */
2873
2874 /*
2875 * This is on the tail page. It is possible that
2876 * a write could come in and move the tail page
2877 * and write to the next page. That is fine
2878 * because we just shorten what is on this page.
2879 */
2880 old_index += write_mask;
2881 new_index += write_mask;
2882
2883 /* caution: old_index gets updated on cmpxchg failure */
2884 if (local_try_cmpxchg(&bpage->write, &old_index, new_index)) {
2885 /* update counters */
2886 local_sub(event_length, &cpu_buffer->entries_bytes);
2887 return true;
2888 }
2889 }
2890
2891 /* could not discard */
2892 return false;
2893}
2894
2895static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2896{
2897 local_inc(&cpu_buffer->committing);
2898 local_inc(&cpu_buffer->commits);
2899}
2900
2901static __always_inline void
2902rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
2903{
2904 unsigned long max_count;
2905
2906 /*
2907 * We only race with interrupts and NMIs on this CPU.
2908 * If we own the commit event, then we can commit
2909 * all others that interrupted us, since the interruptions
2910 * are in stack format (they finish before they come
2911 * back to us). This allows us to do a simple loop to
2912 * assign the commit to the tail.
2913 */
2914 again:
2915 max_count = cpu_buffer->nr_pages * 100;
2916
2917 while (cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)) {
2918 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
2919 return;
2920 if (RB_WARN_ON(cpu_buffer,
2921 rb_is_reader_page(cpu_buffer->tail_page)))
2922 return;
2923 /*
2924 * No need for a memory barrier here, as the update
2925 * of the tail_page did it for this page.
2926 */
2927 local_set(&cpu_buffer->commit_page->page->commit,
2928 rb_page_write(cpu_buffer->commit_page));
2929 rb_inc_page(&cpu_buffer->commit_page);
2930 /* add barrier to keep gcc from optimizing too much */
2931 barrier();
2932 }
2933 while (rb_commit_index(cpu_buffer) !=
2934 rb_page_write(cpu_buffer->commit_page)) {
2935
2936 /* Make sure the readers see the content of what is committed. */
2937 smp_wmb();
2938 local_set(&cpu_buffer->commit_page->page->commit,
2939 rb_page_write(cpu_buffer->commit_page));
2940 RB_WARN_ON(cpu_buffer,
2941 local_read(&cpu_buffer->commit_page->page->commit) &
2942 ~RB_WRITE_MASK);
2943 barrier();
2944 }
2945
2946 /* again, keep gcc from optimizing */
2947 barrier();
2948
2949 /*
2950 * If an interrupt came in just after the first while loop
2951 * and pushed the tail page forward, we will be left with
2952 * a dangling commit that will never go forward.
2953 */
2954 if (unlikely(cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)))
2955 goto again;
2956}
2957
2958static __always_inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
2959{
2960 unsigned long commits;
2961
2962 if (RB_WARN_ON(cpu_buffer,
2963 !local_read(&cpu_buffer->committing)))
2964 return;
2965
2966 again:
2967 commits = local_read(&cpu_buffer->commits);
2968 /* synchronize with interrupts */
2969 barrier();
2970 if (local_read(&cpu_buffer->committing) == 1)
2971 rb_set_commit_to_write(cpu_buffer);
2972
2973 local_dec(&cpu_buffer->committing);
2974
2975 /* synchronize with interrupts */
2976 barrier();
2977
2978 /*
2979 * Need to account for interrupts coming in between the
2980 * updating of the commit page and the clearing of the
2981 * committing counter.
2982 */
2983 if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
2984 !local_read(&cpu_buffer->committing)) {
2985 local_inc(&cpu_buffer->committing);
2986 goto again;
2987 }
2988}
2989
2990static inline void rb_event_discard(struct ring_buffer_event *event)
2991{
2992 if (extended_time(event))
2993 event = skip_time_extend(event);
2994
2995 /* array[0] holds the actual length for the discarded event */
2996 event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
2997 event->type_len = RINGBUF_TYPE_PADDING;
2998 /* time delta must be non zero */
2999 if (!event->time_delta)
3000 event->time_delta = 1;
3001}
3002
3003static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer)
3004{
3005 local_inc(&cpu_buffer->entries);
3006 rb_end_commit(cpu_buffer);
3007}
3008
3009static __always_inline void
3010rb_wakeups(struct trace_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer)
3011{
3012 if (buffer->irq_work.waiters_pending) {
3013 buffer->irq_work.waiters_pending = false;
3014 /* irq_work_queue() supplies it's own memory barriers */
3015 irq_work_queue(&buffer->irq_work.work);
3016 }
3017
3018 if (cpu_buffer->irq_work.waiters_pending) {
3019 cpu_buffer->irq_work.waiters_pending = false;
3020 /* irq_work_queue() supplies it's own memory barriers */
3021 irq_work_queue(&cpu_buffer->irq_work.work);
3022 }
3023
3024 if (cpu_buffer->last_pages_touch == local_read(&cpu_buffer->pages_touched))
3025 return;
3026
3027 if (cpu_buffer->reader_page == cpu_buffer->commit_page)
3028 return;
3029
3030 if (!cpu_buffer->irq_work.full_waiters_pending)
3031 return;
3032
3033 cpu_buffer->last_pages_touch = local_read(&cpu_buffer->pages_touched);
3034
3035 if (!full_hit(buffer, cpu_buffer->cpu, cpu_buffer->shortest_full))
3036 return;
3037
3038 cpu_buffer->irq_work.wakeup_full = true;
3039 cpu_buffer->irq_work.full_waiters_pending = false;
3040 /* irq_work_queue() supplies it's own memory barriers */
3041 irq_work_queue(&cpu_buffer->irq_work.work);
3042}
3043
3044#ifdef CONFIG_RING_BUFFER_RECORD_RECURSION
3045# define do_ring_buffer_record_recursion() \
3046 do_ftrace_record_recursion(_THIS_IP_, _RET_IP_)
3047#else
3048# define do_ring_buffer_record_recursion() do { } while (0)
3049#endif
3050
3051/*
3052 * The lock and unlock are done within a preempt disable section.
3053 * The current_context per_cpu variable can only be modified
3054 * by the current task between lock and unlock. But it can
3055 * be modified more than once via an interrupt. To pass this
3056 * information from the lock to the unlock without having to
3057 * access the 'in_interrupt()' functions again (which do show
3058 * a bit of overhead in something as critical as function tracing,
3059 * we use a bitmask trick.
3060 *
3061 * bit 1 = NMI context
3062 * bit 2 = IRQ context
3063 * bit 3 = SoftIRQ context
3064 * bit 4 = normal context.
3065 *
3066 * This works because this is the order of contexts that can
3067 * preempt other contexts. A SoftIRQ never preempts an IRQ
3068 * context.
3069 *
3070 * When the context is determined, the corresponding bit is
3071 * checked and set (if it was set, then a recursion of that context
3072 * happened).
3073 *
3074 * On unlock, we need to clear this bit. To do so, just subtract
3075 * 1 from the current_context and AND it to itself.
3076 *
3077 * (binary)
3078 * 101 - 1 = 100
3079 * 101 & 100 = 100 (clearing bit zero)
3080 *
3081 * 1010 - 1 = 1001
3082 * 1010 & 1001 = 1000 (clearing bit 1)
3083 *
3084 * The least significant bit can be cleared this way, and it
3085 * just so happens that it is the same bit corresponding to
3086 * the current context.
3087 *
3088 * Now the TRANSITION bit breaks the above slightly. The TRANSITION bit
3089 * is set when a recursion is detected at the current context, and if
3090 * the TRANSITION bit is already set, it will fail the recursion.
3091 * This is needed because there's a lag between the changing of
3092 * interrupt context and updating the preempt count. In this case,
3093 * a false positive will be found. To handle this, one extra recursion
3094 * is allowed, and this is done by the TRANSITION bit. If the TRANSITION
3095 * bit is already set, then it is considered a recursion and the function
3096 * ends. Otherwise, the TRANSITION bit is set, and that bit is returned.
3097 *
3098 * On the trace_recursive_unlock(), the TRANSITION bit will be the first
3099 * to be cleared. Even if it wasn't the context that set it. That is,
3100 * if an interrupt comes in while NORMAL bit is set and the ring buffer
3101 * is called before preempt_count() is updated, since the check will
3102 * be on the NORMAL bit, the TRANSITION bit will then be set. If an
3103 * NMI then comes in, it will set the NMI bit, but when the NMI code
3104 * does the trace_recursive_unlock() it will clear the TRANSITION bit
3105 * and leave the NMI bit set. But this is fine, because the interrupt
3106 * code that set the TRANSITION bit will then clear the NMI bit when it
3107 * calls trace_recursive_unlock(). If another NMI comes in, it will
3108 * set the TRANSITION bit and continue.
3109 *
3110 * Note: The TRANSITION bit only handles a single transition between context.
3111 */
3112
3113static __always_inline bool
3114trace_recursive_lock(struct ring_buffer_per_cpu *cpu_buffer)
3115{
3116 unsigned int val = cpu_buffer->current_context;
3117 int bit = interrupt_context_level();
3118
3119 bit = RB_CTX_NORMAL - bit;
3120
3121 if (unlikely(val & (1 << (bit + cpu_buffer->nest)))) {
3122 /*
3123 * It is possible that this was called by transitioning
3124 * between interrupt context, and preempt_count() has not
3125 * been updated yet. In this case, use the TRANSITION bit.
3126 */
3127 bit = RB_CTX_TRANSITION;
3128 if (val & (1 << (bit + cpu_buffer->nest))) {
3129 do_ring_buffer_record_recursion();
3130 return true;
3131 }
3132 }
3133
3134 val |= (1 << (bit + cpu_buffer->nest));
3135 cpu_buffer->current_context = val;
3136
3137 return false;
3138}
3139
3140static __always_inline void
3141trace_recursive_unlock(struct ring_buffer_per_cpu *cpu_buffer)
3142{
3143 cpu_buffer->current_context &=
3144 cpu_buffer->current_context - (1 << cpu_buffer->nest);
3145}
3146
3147/* The recursive locking above uses 5 bits */
3148#define NESTED_BITS 5
3149
3150/**
3151 * ring_buffer_nest_start - Allow to trace while nested
3152 * @buffer: The ring buffer to modify
3153 *
3154 * The ring buffer has a safety mechanism to prevent recursion.
3155 * But there may be a case where a trace needs to be done while
3156 * tracing something else. In this case, calling this function
3157 * will allow this function to nest within a currently active
3158 * ring_buffer_lock_reserve().
3159 *
3160 * Call this function before calling another ring_buffer_lock_reserve() and
3161 * call ring_buffer_nest_end() after the nested ring_buffer_unlock_commit().
3162 */
3163void ring_buffer_nest_start(struct trace_buffer *buffer)
3164{
3165 struct ring_buffer_per_cpu *cpu_buffer;
3166 int cpu;
3167
3168 /* Enabled by ring_buffer_nest_end() */
3169 preempt_disable_notrace();
3170 cpu = raw_smp_processor_id();
3171 cpu_buffer = buffer->buffers[cpu];
3172 /* This is the shift value for the above recursive locking */
3173 cpu_buffer->nest += NESTED_BITS;
3174}
3175
3176/**
3177 * ring_buffer_nest_end - Allow to trace while nested
3178 * @buffer: The ring buffer to modify
3179 *
3180 * Must be called after ring_buffer_nest_start() and after the
3181 * ring_buffer_unlock_commit().
3182 */
3183void ring_buffer_nest_end(struct trace_buffer *buffer)
3184{
3185 struct ring_buffer_per_cpu *cpu_buffer;
3186 int cpu;
3187
3188 /* disabled by ring_buffer_nest_start() */
3189 cpu = raw_smp_processor_id();
3190 cpu_buffer = buffer->buffers[cpu];
3191 /* This is the shift value for the above recursive locking */
3192 cpu_buffer->nest -= NESTED_BITS;
3193 preempt_enable_notrace();
3194}
3195
3196/**
3197 * ring_buffer_unlock_commit - commit a reserved
3198 * @buffer: The buffer to commit to
3199 *
3200 * This commits the data to the ring buffer, and releases any locks held.
3201 *
3202 * Must be paired with ring_buffer_lock_reserve.
3203 */
3204int ring_buffer_unlock_commit(struct trace_buffer *buffer)
3205{
3206 struct ring_buffer_per_cpu *cpu_buffer;
3207 int cpu = raw_smp_processor_id();
3208
3209 cpu_buffer = buffer->buffers[cpu];
3210
3211 rb_commit(cpu_buffer);
3212
3213 rb_wakeups(buffer, cpu_buffer);
3214
3215 trace_recursive_unlock(cpu_buffer);
3216
3217 preempt_enable_notrace();
3218
3219 return 0;
3220}
3221EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
3222
3223/* Special value to validate all deltas on a page. */
3224#define CHECK_FULL_PAGE 1L
3225
3226#ifdef CONFIG_RING_BUFFER_VALIDATE_TIME_DELTAS
3227
3228static const char *show_irq_str(int bits)
3229{
3230 const char *type[] = {
3231 ".", // 0
3232 "s", // 1
3233 "h", // 2
3234 "Hs", // 3
3235 "n", // 4
3236 "Ns", // 5
3237 "Nh", // 6
3238 "NHs", // 7
3239 };
3240
3241 return type[bits];
3242}
3243
3244/* Assume this is an trace event */
3245static const char *show_flags(struct ring_buffer_event *event)
3246{
3247 struct trace_entry *entry;
3248 int bits = 0;
3249
3250 if (rb_event_data_length(event) - RB_EVNT_HDR_SIZE < sizeof(*entry))
3251 return "X";
3252
3253 entry = ring_buffer_event_data(event);
3254
3255 if (entry->flags & TRACE_FLAG_SOFTIRQ)
3256 bits |= 1;
3257
3258 if (entry->flags & TRACE_FLAG_HARDIRQ)
3259 bits |= 2;
3260
3261 if (entry->flags & TRACE_FLAG_NMI)
3262 bits |= 4;
3263
3264 return show_irq_str(bits);
3265}
3266
3267static const char *show_irq(struct ring_buffer_event *event)
3268{
3269 struct trace_entry *entry;
3270
3271 if (rb_event_data_length(event) - RB_EVNT_HDR_SIZE < sizeof(*entry))
3272 return "";
3273
3274 entry = ring_buffer_event_data(event);
3275 if (entry->flags & TRACE_FLAG_IRQS_OFF)
3276 return "d";
3277 return "";
3278}
3279
3280static const char *show_interrupt_level(void)
3281{
3282 unsigned long pc = preempt_count();
3283 unsigned char level = 0;
3284
3285 if (pc & SOFTIRQ_OFFSET)
3286 level |= 1;
3287
3288 if (pc & HARDIRQ_MASK)
3289 level |= 2;
3290
3291 if (pc & NMI_MASK)
3292 level |= 4;
3293
3294 return show_irq_str(level);
3295}
3296
3297static void dump_buffer_page(struct buffer_data_page *bpage,
3298 struct rb_event_info *info,
3299 unsigned long tail)
3300{
3301 struct ring_buffer_event *event;
3302 u64 ts, delta;
3303 int e;
3304
3305 ts = bpage->time_stamp;
3306 pr_warn(" [%lld] PAGE TIME STAMP\n", ts);
3307
3308 for (e = 0; e < tail; e += rb_event_length(event)) {
3309
3310 event = (struct ring_buffer_event *)(bpage->data + e);
3311
3312 switch (event->type_len) {
3313
3314 case RINGBUF_TYPE_TIME_EXTEND:
3315 delta = rb_event_time_stamp(event);
3316 ts += delta;
3317 pr_warn(" 0x%x: [%lld] delta:%lld TIME EXTEND\n",
3318 e, ts, delta);
3319 break;
3320
3321 case RINGBUF_TYPE_TIME_STAMP:
3322 delta = rb_event_time_stamp(event);
3323 ts = rb_fix_abs_ts(delta, ts);
3324 pr_warn(" 0x%x: [%lld] absolute:%lld TIME STAMP\n",
3325 e, ts, delta);
3326 break;
3327
3328 case RINGBUF_TYPE_PADDING:
3329 ts += event->time_delta;
3330 pr_warn(" 0x%x: [%lld] delta:%d PADDING\n",
3331 e, ts, event->time_delta);
3332 break;
3333
3334 case RINGBUF_TYPE_DATA:
3335 ts += event->time_delta;
3336 pr_warn(" 0x%x: [%lld] delta:%d %s%s\n",
3337 e, ts, event->time_delta,
3338 show_flags(event), show_irq(event));
3339 break;
3340
3341 default:
3342 break;
3343 }
3344 }
3345 pr_warn("expected end:0x%lx last event actually ended at:0x%x\n", tail, e);
3346}
3347
3348static DEFINE_PER_CPU(atomic_t, checking);
3349static atomic_t ts_dump;
3350
3351#define buffer_warn_return(fmt, ...) \
3352 do { \
3353 /* If another report is happening, ignore this one */ \
3354 if (atomic_inc_return(&ts_dump) != 1) { \
3355 atomic_dec(&ts_dump); \
3356 goto out; \
3357 } \
3358 atomic_inc(&cpu_buffer->record_disabled); \
3359 pr_warn(fmt, ##__VA_ARGS__); \
3360 dump_buffer_page(bpage, info, tail); \
3361 atomic_dec(&ts_dump); \
3362 /* There's some cases in boot up that this can happen */ \
3363 if (WARN_ON_ONCE(system_state != SYSTEM_BOOTING)) \
3364 /* Do not re-enable checking */ \
3365 return; \
3366 } while (0)
3367
3368/*
3369 * Check if the current event time stamp matches the deltas on
3370 * the buffer page.
3371 */
3372static void check_buffer(struct ring_buffer_per_cpu *cpu_buffer,
3373 struct rb_event_info *info,
3374 unsigned long tail)
3375{
3376 struct ring_buffer_event *event;
3377 struct buffer_data_page *bpage;
3378 u64 ts, delta;
3379 bool full = false;
3380 int e;
3381
3382 bpage = info->tail_page->page;
3383
3384 if (tail == CHECK_FULL_PAGE) {
3385 full = true;
3386 tail = local_read(&bpage->commit);
3387 } else if (info->add_timestamp &
3388 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE)) {
3389 /* Ignore events with absolute time stamps */
3390 return;
3391 }
3392
3393 /*
3394 * Do not check the first event (skip possible extends too).
3395 * Also do not check if previous events have not been committed.
3396 */
3397 if (tail <= 8 || tail > local_read(&bpage->commit))
3398 return;
3399
3400 /*
3401 * If this interrupted another event,
3402 */
3403 if (atomic_inc_return(this_cpu_ptr(&checking)) != 1)
3404 goto out;
3405
3406 ts = bpage->time_stamp;
3407
3408 for (e = 0; e < tail; e += rb_event_length(event)) {
3409
3410 event = (struct ring_buffer_event *)(bpage->data + e);
3411
3412 switch (event->type_len) {
3413
3414 case RINGBUF_TYPE_TIME_EXTEND:
3415 delta = rb_event_time_stamp(event);
3416 ts += delta;
3417 break;
3418
3419 case RINGBUF_TYPE_TIME_STAMP:
3420 delta = rb_event_time_stamp(event);
3421 delta = rb_fix_abs_ts(delta, ts);
3422 if (delta < ts) {
3423 buffer_warn_return("[CPU: %d]ABSOLUTE TIME WENT BACKWARDS: last ts: %lld absolute ts: %lld\n",
3424 cpu_buffer->cpu, ts, delta);
3425 }
3426 ts = delta;
3427 break;
3428
3429 case RINGBUF_TYPE_PADDING:
3430 if (event->time_delta == 1)
3431 break;
3432 fallthrough;
3433 case RINGBUF_TYPE_DATA:
3434 ts += event->time_delta;
3435 break;
3436
3437 default:
3438 RB_WARN_ON(cpu_buffer, 1);
3439 }
3440 }
3441 if ((full && ts > info->ts) ||
3442 (!full && ts + info->delta != info->ts)) {
3443 buffer_warn_return("[CPU: %d]TIME DOES NOT MATCH expected:%lld actual:%lld delta:%lld before:%lld after:%lld%s context:%s\n",
3444 cpu_buffer->cpu,
3445 ts + info->delta, info->ts, info->delta,
3446 info->before, info->after,
3447 full ? " (full)" : "", show_interrupt_level());
3448 }
3449out:
3450 atomic_dec(this_cpu_ptr(&checking));
3451}
3452#else
3453static inline void check_buffer(struct ring_buffer_per_cpu *cpu_buffer,
3454 struct rb_event_info *info,
3455 unsigned long tail)
3456{
3457}
3458#endif /* CONFIG_RING_BUFFER_VALIDATE_TIME_DELTAS */
3459
3460static struct ring_buffer_event *
3461__rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
3462 struct rb_event_info *info)
3463{
3464 struct ring_buffer_event *event;
3465 struct buffer_page *tail_page;
3466 unsigned long tail, write, w;
3467
3468 /* Don't let the compiler play games with cpu_buffer->tail_page */
3469 tail_page = info->tail_page = READ_ONCE(cpu_buffer->tail_page);
3470
3471 /*A*/ w = local_read(&tail_page->write) & RB_WRITE_MASK;
3472 barrier();
3473 rb_time_read(&cpu_buffer->before_stamp, &info->before);
3474 rb_time_read(&cpu_buffer->write_stamp, &info->after);
3475 barrier();
3476 info->ts = rb_time_stamp(cpu_buffer->buffer);
3477
3478 if ((info->add_timestamp & RB_ADD_STAMP_ABSOLUTE)) {
3479 info->delta = info->ts;
3480 } else {
3481 /*
3482 * If interrupting an event time update, we may need an
3483 * absolute timestamp.
3484 * Don't bother if this is the start of a new page (w == 0).
3485 */
3486 if (!w) {
3487 /* Use the sub-buffer timestamp */
3488 info->delta = 0;
3489 } else if (unlikely(info->before != info->after)) {
3490 info->add_timestamp |= RB_ADD_STAMP_FORCE | RB_ADD_STAMP_EXTEND;
3491 info->length += RB_LEN_TIME_EXTEND;
3492 } else {
3493 info->delta = info->ts - info->after;
3494 if (unlikely(test_time_stamp(info->delta))) {
3495 info->add_timestamp |= RB_ADD_STAMP_EXTEND;
3496 info->length += RB_LEN_TIME_EXTEND;
3497 }
3498 }
3499 }
3500
3501 /*B*/ rb_time_set(&cpu_buffer->before_stamp, info->ts);
3502
3503 /*C*/ write = local_add_return(info->length, &tail_page->write);
3504
3505 /* set write to only the index of the write */
3506 write &= RB_WRITE_MASK;
3507
3508 tail = write - info->length;
3509
3510 /* See if we shot pass the end of this buffer page */
3511 if (unlikely(write > cpu_buffer->buffer->subbuf_size)) {
3512 check_buffer(cpu_buffer, info, CHECK_FULL_PAGE);
3513 return rb_move_tail(cpu_buffer, tail, info);
3514 }
3515
3516 if (likely(tail == w)) {
3517 /* Nothing interrupted us between A and C */
3518 /*D*/ rb_time_set(&cpu_buffer->write_stamp, info->ts);
3519 /*
3520 * If something came in between C and D, the write stamp
3521 * may now not be in sync. But that's fine as the before_stamp
3522 * will be different and then next event will just be forced
3523 * to use an absolute timestamp.
3524 */
3525 if (likely(!(info->add_timestamp &
3526 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE))))
3527 /* This did not interrupt any time update */
3528 info->delta = info->ts - info->after;
3529 else
3530 /* Just use full timestamp for interrupting event */
3531 info->delta = info->ts;
3532 check_buffer(cpu_buffer, info, tail);
3533 } else {
3534 u64 ts;
3535 /* SLOW PATH - Interrupted between A and C */
3536
3537 /* Save the old before_stamp */
3538 rb_time_read(&cpu_buffer->before_stamp, &info->before);
3539
3540 /*
3541 * Read a new timestamp and update the before_stamp to make
3542 * the next event after this one force using an absolute
3543 * timestamp. This is in case an interrupt were to come in
3544 * between E and F.
3545 */
3546 ts = rb_time_stamp(cpu_buffer->buffer);
3547 rb_time_set(&cpu_buffer->before_stamp, ts);
3548
3549 barrier();
3550 /*E*/ rb_time_read(&cpu_buffer->write_stamp, &info->after);
3551 barrier();
3552 /*F*/ if (write == (local_read(&tail_page->write) & RB_WRITE_MASK) &&
3553 info->after == info->before && info->after < ts) {
3554 /*
3555 * Nothing came after this event between C and F, it is
3556 * safe to use info->after for the delta as it
3557 * matched info->before and is still valid.
3558 */
3559 info->delta = ts - info->after;
3560 } else {
3561 /*
3562 * Interrupted between C and F:
3563 * Lost the previous events time stamp. Just set the
3564 * delta to zero, and this will be the same time as
3565 * the event this event interrupted. And the events that
3566 * came after this will still be correct (as they would
3567 * have built their delta on the previous event.
3568 */
3569 info->delta = 0;
3570 }
3571 info->ts = ts;
3572 info->add_timestamp &= ~RB_ADD_STAMP_FORCE;
3573 }
3574
3575 /*
3576 * If this is the first commit on the page, then it has the same
3577 * timestamp as the page itself.
3578 */
3579 if (unlikely(!tail && !(info->add_timestamp &
3580 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE))))
3581 info->delta = 0;
3582
3583 /* We reserved something on the buffer */
3584
3585 event = __rb_page_index(tail_page, tail);
3586 rb_update_event(cpu_buffer, event, info);
3587
3588 local_inc(&tail_page->entries);
3589
3590 /*
3591 * If this is the first commit on the page, then update
3592 * its timestamp.
3593 */
3594 if (unlikely(!tail))
3595 tail_page->page->time_stamp = info->ts;
3596
3597 /* account for these added bytes */
3598 local_add(info->length, &cpu_buffer->entries_bytes);
3599
3600 return event;
3601}
3602
3603static __always_inline struct ring_buffer_event *
3604rb_reserve_next_event(struct trace_buffer *buffer,
3605 struct ring_buffer_per_cpu *cpu_buffer,
3606 unsigned long length)
3607{
3608 struct ring_buffer_event *event;
3609 struct rb_event_info info;
3610 int nr_loops = 0;
3611 int add_ts_default;
3612
3613 /* ring buffer does cmpxchg, make sure it is safe in NMI context */
3614 if (!IS_ENABLED(CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG) &&
3615 (unlikely(in_nmi()))) {
3616 return NULL;
3617 }
3618
3619 rb_start_commit(cpu_buffer);
3620 /* The commit page can not change after this */
3621
3622#ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
3623 /*
3624 * Due to the ability to swap a cpu buffer from a buffer
3625 * it is possible it was swapped before we committed.
3626 * (committing stops a swap). We check for it here and
3627 * if it happened, we have to fail the write.
3628 */
3629 barrier();
3630 if (unlikely(READ_ONCE(cpu_buffer->buffer) != buffer)) {
3631 local_dec(&cpu_buffer->committing);
3632 local_dec(&cpu_buffer->commits);
3633 return NULL;
3634 }
3635#endif
3636
3637 info.length = rb_calculate_event_length(length);
3638
3639 if (ring_buffer_time_stamp_abs(cpu_buffer->buffer)) {
3640 add_ts_default = RB_ADD_STAMP_ABSOLUTE;
3641 info.length += RB_LEN_TIME_EXTEND;
3642 if (info.length > cpu_buffer->buffer->max_data_size)
3643 goto out_fail;
3644 } else {
3645 add_ts_default = RB_ADD_STAMP_NONE;
3646 }
3647
3648 again:
3649 info.add_timestamp = add_ts_default;
3650 info.delta = 0;
3651
3652 /*
3653 * We allow for interrupts to reenter here and do a trace.
3654 * If one does, it will cause this original code to loop
3655 * back here. Even with heavy interrupts happening, this
3656 * should only happen a few times in a row. If this happens
3657 * 1000 times in a row, there must be either an interrupt
3658 * storm or we have something buggy.
3659 * Bail!
3660 */
3661 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
3662 goto out_fail;
3663
3664 event = __rb_reserve_next(cpu_buffer, &info);
3665
3666 if (unlikely(PTR_ERR(event) == -EAGAIN)) {
3667 if (info.add_timestamp & (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_EXTEND))
3668 info.length -= RB_LEN_TIME_EXTEND;
3669 goto again;
3670 }
3671
3672 if (likely(event))
3673 return event;
3674 out_fail:
3675 rb_end_commit(cpu_buffer);
3676 return NULL;
3677}
3678
3679/**
3680 * ring_buffer_lock_reserve - reserve a part of the buffer
3681 * @buffer: the ring buffer to reserve from
3682 * @length: the length of the data to reserve (excluding event header)
3683 *
3684 * Returns a reserved event on the ring buffer to copy directly to.
3685 * The user of this interface will need to get the body to write into
3686 * and can use the ring_buffer_event_data() interface.
3687 *
3688 * The length is the length of the data needed, not the event length
3689 * which also includes the event header.
3690 *
3691 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
3692 * If NULL is returned, then nothing has been allocated or locked.
3693 */
3694struct ring_buffer_event *
3695ring_buffer_lock_reserve(struct trace_buffer *buffer, unsigned long length)
3696{
3697 struct ring_buffer_per_cpu *cpu_buffer;
3698 struct ring_buffer_event *event;
3699 int cpu;
3700
3701 /* If we are tracing schedule, we don't want to recurse */
3702 preempt_disable_notrace();
3703
3704 if (unlikely(atomic_read(&buffer->record_disabled)))
3705 goto out;
3706
3707 cpu = raw_smp_processor_id();
3708
3709 if (unlikely(!cpumask_test_cpu(cpu, buffer->cpumask)))
3710 goto out;
3711
3712 cpu_buffer = buffer->buffers[cpu];
3713
3714 if (unlikely(atomic_read(&cpu_buffer->record_disabled)))
3715 goto out;
3716
3717 if (unlikely(length > buffer->max_data_size))
3718 goto out;
3719
3720 if (unlikely(trace_recursive_lock(cpu_buffer)))
3721 goto out;
3722
3723 event = rb_reserve_next_event(buffer, cpu_buffer, length);
3724 if (!event)
3725 goto out_unlock;
3726
3727 return event;
3728
3729 out_unlock:
3730 trace_recursive_unlock(cpu_buffer);
3731 out:
3732 preempt_enable_notrace();
3733 return NULL;
3734}
3735EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
3736
3737/*
3738 * Decrement the entries to the page that an event is on.
3739 * The event does not even need to exist, only the pointer
3740 * to the page it is on. This may only be called before the commit
3741 * takes place.
3742 */
3743static inline void
3744rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
3745 struct ring_buffer_event *event)
3746{
3747 unsigned long addr = (unsigned long)event;
3748 struct buffer_page *bpage = cpu_buffer->commit_page;
3749 struct buffer_page *start;
3750
3751 addr &= ~((PAGE_SIZE << cpu_buffer->buffer->subbuf_order) - 1);
3752
3753 /* Do the likely case first */
3754 if (likely(bpage->page == (void *)addr)) {
3755 local_dec(&bpage->entries);
3756 return;
3757 }
3758
3759 /*
3760 * Because the commit page may be on the reader page we
3761 * start with the next page and check the end loop there.
3762 */
3763 rb_inc_page(&bpage);
3764 start = bpage;
3765 do {
3766 if (bpage->page == (void *)addr) {
3767 local_dec(&bpage->entries);
3768 return;
3769 }
3770 rb_inc_page(&bpage);
3771 } while (bpage != start);
3772
3773 /* commit not part of this buffer?? */
3774 RB_WARN_ON(cpu_buffer, 1);
3775}
3776
3777/**
3778 * ring_buffer_discard_commit - discard an event that has not been committed
3779 * @buffer: the ring buffer
3780 * @event: non committed event to discard
3781 *
3782 * Sometimes an event that is in the ring buffer needs to be ignored.
3783 * This function lets the user discard an event in the ring buffer
3784 * and then that event will not be read later.
3785 *
3786 * This function only works if it is called before the item has been
3787 * committed. It will try to free the event from the ring buffer
3788 * if another event has not been added behind it.
3789 *
3790 * If another event has been added behind it, it will set the event
3791 * up as discarded, and perform the commit.
3792 *
3793 * If this function is called, do not call ring_buffer_unlock_commit on
3794 * the event.
3795 */
3796void ring_buffer_discard_commit(struct trace_buffer *buffer,
3797 struct ring_buffer_event *event)
3798{
3799 struct ring_buffer_per_cpu *cpu_buffer;
3800 int cpu;
3801
3802 /* The event is discarded regardless */
3803 rb_event_discard(event);
3804
3805 cpu = smp_processor_id();
3806 cpu_buffer = buffer->buffers[cpu];
3807
3808 /*
3809 * This must only be called if the event has not been
3810 * committed yet. Thus we can assume that preemption
3811 * is still disabled.
3812 */
3813 RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
3814
3815 rb_decrement_entry(cpu_buffer, event);
3816 if (rb_try_to_discard(cpu_buffer, event))
3817 goto out;
3818
3819 out:
3820 rb_end_commit(cpu_buffer);
3821
3822 trace_recursive_unlock(cpu_buffer);
3823
3824 preempt_enable_notrace();
3825
3826}
3827EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
3828
3829/**
3830 * ring_buffer_write - write data to the buffer without reserving
3831 * @buffer: The ring buffer to write to.
3832 * @length: The length of the data being written (excluding the event header)
3833 * @data: The data to write to the buffer.
3834 *
3835 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
3836 * one function. If you already have the data to write to the buffer, it
3837 * may be easier to simply call this function.
3838 *
3839 * Note, like ring_buffer_lock_reserve, the length is the length of the data
3840 * and not the length of the event which would hold the header.
3841 */
3842int ring_buffer_write(struct trace_buffer *buffer,
3843 unsigned long length,
3844 void *data)
3845{
3846 struct ring_buffer_per_cpu *cpu_buffer;
3847 struct ring_buffer_event *event;
3848 void *body;
3849 int ret = -EBUSY;
3850 int cpu;
3851
3852 preempt_disable_notrace();
3853
3854 if (atomic_read(&buffer->record_disabled))
3855 goto out;
3856
3857 cpu = raw_smp_processor_id();
3858
3859 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3860 goto out;
3861
3862 cpu_buffer = buffer->buffers[cpu];
3863
3864 if (atomic_read(&cpu_buffer->record_disabled))
3865 goto out;
3866
3867 if (length > buffer->max_data_size)
3868 goto out;
3869
3870 if (unlikely(trace_recursive_lock(cpu_buffer)))
3871 goto out;
3872
3873 event = rb_reserve_next_event(buffer, cpu_buffer, length);
3874 if (!event)
3875 goto out_unlock;
3876
3877 body = rb_event_data(event);
3878
3879 memcpy(body, data, length);
3880
3881 rb_commit(cpu_buffer);
3882
3883 rb_wakeups(buffer, cpu_buffer);
3884
3885 ret = 0;
3886
3887 out_unlock:
3888 trace_recursive_unlock(cpu_buffer);
3889
3890 out:
3891 preempt_enable_notrace();
3892
3893 return ret;
3894}
3895EXPORT_SYMBOL_GPL(ring_buffer_write);
3896
3897static bool rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
3898{
3899 struct buffer_page *reader = cpu_buffer->reader_page;
3900 struct buffer_page *head = rb_set_head_page(cpu_buffer);
3901 struct buffer_page *commit = cpu_buffer->commit_page;
3902
3903 /* In case of error, head will be NULL */
3904 if (unlikely(!head))
3905 return true;
3906
3907 /* Reader should exhaust content in reader page */
3908 if (reader->read != rb_page_commit(reader))
3909 return false;
3910
3911 /*
3912 * If writers are committing on the reader page, knowing all
3913 * committed content has been read, the ring buffer is empty.
3914 */
3915 if (commit == reader)
3916 return true;
3917
3918 /*
3919 * If writers are committing on a page other than reader page
3920 * and head page, there should always be content to read.
3921 */
3922 if (commit != head)
3923 return false;
3924
3925 /*
3926 * Writers are committing on the head page, we just need
3927 * to care about there're committed data, and the reader will
3928 * swap reader page with head page when it is to read data.
3929 */
3930 return rb_page_commit(commit) == 0;
3931}
3932
3933/**
3934 * ring_buffer_record_disable - stop all writes into the buffer
3935 * @buffer: The ring buffer to stop writes to.
3936 *
3937 * This prevents all writes to the buffer. Any attempt to write
3938 * to the buffer after this will fail and return NULL.
3939 *
3940 * The caller should call synchronize_rcu() after this.
3941 */
3942void ring_buffer_record_disable(struct trace_buffer *buffer)
3943{
3944 atomic_inc(&buffer->record_disabled);
3945}
3946EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
3947
3948/**
3949 * ring_buffer_record_enable - enable writes to the buffer
3950 * @buffer: The ring buffer to enable writes
3951 *
3952 * Note, multiple disables will need the same number of enables
3953 * to truly enable the writing (much like preempt_disable).
3954 */
3955void ring_buffer_record_enable(struct trace_buffer *buffer)
3956{
3957 atomic_dec(&buffer->record_disabled);
3958}
3959EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
3960
3961/**
3962 * ring_buffer_record_off - stop all writes into the buffer
3963 * @buffer: The ring buffer to stop writes to.
3964 *
3965 * This prevents all writes to the buffer. Any attempt to write
3966 * to the buffer after this will fail and return NULL.
3967 *
3968 * This is different than ring_buffer_record_disable() as
3969 * it works like an on/off switch, where as the disable() version
3970 * must be paired with a enable().
3971 */
3972void ring_buffer_record_off(struct trace_buffer *buffer)
3973{
3974 unsigned int rd;
3975 unsigned int new_rd;
3976
3977 rd = atomic_read(&buffer->record_disabled);
3978 do {
3979 new_rd = rd | RB_BUFFER_OFF;
3980 } while (!atomic_try_cmpxchg(&buffer->record_disabled, &rd, new_rd));
3981}
3982EXPORT_SYMBOL_GPL(ring_buffer_record_off);
3983
3984/**
3985 * ring_buffer_record_on - restart writes into the buffer
3986 * @buffer: The ring buffer to start writes to.
3987 *
3988 * This enables all writes to the buffer that was disabled by
3989 * ring_buffer_record_off().
3990 *
3991 * This is different than ring_buffer_record_enable() as
3992 * it works like an on/off switch, where as the enable() version
3993 * must be paired with a disable().
3994 */
3995void ring_buffer_record_on(struct trace_buffer *buffer)
3996{
3997 unsigned int rd;
3998 unsigned int new_rd;
3999
4000 rd = atomic_read(&buffer->record_disabled);
4001 do {
4002 new_rd = rd & ~RB_BUFFER_OFF;
4003 } while (!atomic_try_cmpxchg(&buffer->record_disabled, &rd, new_rd));
4004}
4005EXPORT_SYMBOL_GPL(ring_buffer_record_on);
4006
4007/**
4008 * ring_buffer_record_is_on - return true if the ring buffer can write
4009 * @buffer: The ring buffer to see if write is enabled
4010 *
4011 * Returns true if the ring buffer is in a state that it accepts writes.
4012 */
4013bool ring_buffer_record_is_on(struct trace_buffer *buffer)
4014{
4015 return !atomic_read(&buffer->record_disabled);
4016}
4017
4018/**
4019 * ring_buffer_record_is_set_on - return true if the ring buffer is set writable
4020 * @buffer: The ring buffer to see if write is set enabled
4021 *
4022 * Returns true if the ring buffer is set writable by ring_buffer_record_on().
4023 * Note that this does NOT mean it is in a writable state.
4024 *
4025 * It may return true when the ring buffer has been disabled by
4026 * ring_buffer_record_disable(), as that is a temporary disabling of
4027 * the ring buffer.
4028 */
4029bool ring_buffer_record_is_set_on(struct trace_buffer *buffer)
4030{
4031 return !(atomic_read(&buffer->record_disabled) & RB_BUFFER_OFF);
4032}
4033
4034/**
4035 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
4036 * @buffer: The ring buffer to stop writes to.
4037 * @cpu: The CPU buffer to stop
4038 *
4039 * This prevents all writes to the buffer. Any attempt to write
4040 * to the buffer after this will fail and return NULL.
4041 *
4042 * The caller should call synchronize_rcu() after this.
4043 */
4044void ring_buffer_record_disable_cpu(struct trace_buffer *buffer, int cpu)
4045{
4046 struct ring_buffer_per_cpu *cpu_buffer;
4047
4048 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4049 return;
4050
4051 cpu_buffer = buffer->buffers[cpu];
4052 atomic_inc(&cpu_buffer->record_disabled);
4053}
4054EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
4055
4056/**
4057 * ring_buffer_record_enable_cpu - enable writes to the buffer
4058 * @buffer: The ring buffer to enable writes
4059 * @cpu: The CPU to enable.
4060 *
4061 * Note, multiple disables will need the same number of enables
4062 * to truly enable the writing (much like preempt_disable).
4063 */
4064void ring_buffer_record_enable_cpu(struct trace_buffer *buffer, int cpu)
4065{
4066 struct ring_buffer_per_cpu *cpu_buffer;
4067
4068 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4069 return;
4070
4071 cpu_buffer = buffer->buffers[cpu];
4072 atomic_dec(&cpu_buffer->record_disabled);
4073}
4074EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
4075
4076/*
4077 * The total entries in the ring buffer is the running counter
4078 * of entries entered into the ring buffer, minus the sum of
4079 * the entries read from the ring buffer and the number of
4080 * entries that were overwritten.
4081 */
4082static inline unsigned long
4083rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
4084{
4085 return local_read(&cpu_buffer->entries) -
4086 (local_read(&cpu_buffer->overrun) + cpu_buffer->read);
4087}
4088
4089/**
4090 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
4091 * @buffer: The ring buffer
4092 * @cpu: The per CPU buffer to read from.
4093 */
4094u64 ring_buffer_oldest_event_ts(struct trace_buffer *buffer, int cpu)
4095{
4096 unsigned long flags;
4097 struct ring_buffer_per_cpu *cpu_buffer;
4098 struct buffer_page *bpage;
4099 u64 ret = 0;
4100
4101 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4102 return 0;
4103
4104 cpu_buffer = buffer->buffers[cpu];
4105 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4106 /*
4107 * if the tail is on reader_page, oldest time stamp is on the reader
4108 * page
4109 */
4110 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
4111 bpage = cpu_buffer->reader_page;
4112 else
4113 bpage = rb_set_head_page(cpu_buffer);
4114 if (bpage)
4115 ret = bpage->page->time_stamp;
4116 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4117
4118 return ret;
4119}
4120EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
4121
4122/**
4123 * ring_buffer_bytes_cpu - get the number of bytes unconsumed in a cpu buffer
4124 * @buffer: The ring buffer
4125 * @cpu: The per CPU buffer to read from.
4126 */
4127unsigned long ring_buffer_bytes_cpu(struct trace_buffer *buffer, int cpu)
4128{
4129 struct ring_buffer_per_cpu *cpu_buffer;
4130 unsigned long ret;
4131
4132 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4133 return 0;
4134
4135 cpu_buffer = buffer->buffers[cpu];
4136 ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
4137
4138 return ret;
4139}
4140EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
4141
4142/**
4143 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
4144 * @buffer: The ring buffer
4145 * @cpu: The per CPU buffer to get the entries from.
4146 */
4147unsigned long ring_buffer_entries_cpu(struct trace_buffer *buffer, int cpu)
4148{
4149 struct ring_buffer_per_cpu *cpu_buffer;
4150
4151 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4152 return 0;
4153
4154 cpu_buffer = buffer->buffers[cpu];
4155
4156 return rb_num_of_entries(cpu_buffer);
4157}
4158EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
4159
4160/**
4161 * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
4162 * buffer wrapping around (only if RB_FL_OVERWRITE is on).
4163 * @buffer: The ring buffer
4164 * @cpu: The per CPU buffer to get the number of overruns from
4165 */
4166unsigned long ring_buffer_overrun_cpu(struct trace_buffer *buffer, int cpu)
4167{
4168 struct ring_buffer_per_cpu *cpu_buffer;
4169 unsigned long ret;
4170
4171 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4172 return 0;
4173
4174 cpu_buffer = buffer->buffers[cpu];
4175 ret = local_read(&cpu_buffer->overrun);
4176
4177 return ret;
4178}
4179EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
4180
4181/**
4182 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
4183 * commits failing due to the buffer wrapping around while there are uncommitted
4184 * events, such as during an interrupt storm.
4185 * @buffer: The ring buffer
4186 * @cpu: The per CPU buffer to get the number of overruns from
4187 */
4188unsigned long
4189ring_buffer_commit_overrun_cpu(struct trace_buffer *buffer, int cpu)
4190{
4191 struct ring_buffer_per_cpu *cpu_buffer;
4192 unsigned long ret;
4193
4194 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4195 return 0;
4196
4197 cpu_buffer = buffer->buffers[cpu];
4198 ret = local_read(&cpu_buffer->commit_overrun);
4199
4200 return ret;
4201}
4202EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
4203
4204/**
4205 * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
4206 * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
4207 * @buffer: The ring buffer
4208 * @cpu: The per CPU buffer to get the number of overruns from
4209 */
4210unsigned long
4211ring_buffer_dropped_events_cpu(struct trace_buffer *buffer, int cpu)
4212{
4213 struct ring_buffer_per_cpu *cpu_buffer;
4214 unsigned long ret;
4215
4216 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4217 return 0;
4218
4219 cpu_buffer = buffer->buffers[cpu];
4220 ret = local_read(&cpu_buffer->dropped_events);
4221
4222 return ret;
4223}
4224EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
4225
4226/**
4227 * ring_buffer_read_events_cpu - get the number of events successfully read
4228 * @buffer: The ring buffer
4229 * @cpu: The per CPU buffer to get the number of events read
4230 */
4231unsigned long
4232ring_buffer_read_events_cpu(struct trace_buffer *buffer, int cpu)
4233{
4234 struct ring_buffer_per_cpu *cpu_buffer;
4235
4236 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4237 return 0;
4238
4239 cpu_buffer = buffer->buffers[cpu];
4240 return cpu_buffer->read;
4241}
4242EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu);
4243
4244/**
4245 * ring_buffer_entries - get the number of entries in a buffer
4246 * @buffer: The ring buffer
4247 *
4248 * Returns the total number of entries in the ring buffer
4249 * (all CPU entries)
4250 */
4251unsigned long ring_buffer_entries(struct trace_buffer *buffer)
4252{
4253 struct ring_buffer_per_cpu *cpu_buffer;
4254 unsigned long entries = 0;
4255 int cpu;
4256
4257 /* if you care about this being correct, lock the buffer */
4258 for_each_buffer_cpu(buffer, cpu) {
4259 cpu_buffer = buffer->buffers[cpu];
4260 entries += rb_num_of_entries(cpu_buffer);
4261 }
4262
4263 return entries;
4264}
4265EXPORT_SYMBOL_GPL(ring_buffer_entries);
4266
4267/**
4268 * ring_buffer_overruns - get the number of overruns in buffer
4269 * @buffer: The ring buffer
4270 *
4271 * Returns the total number of overruns in the ring buffer
4272 * (all CPU entries)
4273 */
4274unsigned long ring_buffer_overruns(struct trace_buffer *buffer)
4275{
4276 struct ring_buffer_per_cpu *cpu_buffer;
4277 unsigned long overruns = 0;
4278 int cpu;
4279
4280 /* if you care about this being correct, lock the buffer */
4281 for_each_buffer_cpu(buffer, cpu) {
4282 cpu_buffer = buffer->buffers[cpu];
4283 overruns += local_read(&cpu_buffer->overrun);
4284 }
4285
4286 return overruns;
4287}
4288EXPORT_SYMBOL_GPL(ring_buffer_overruns);
4289
4290static void rb_iter_reset(struct ring_buffer_iter *iter)
4291{
4292 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4293
4294 /* Iterator usage is expected to have record disabled */
4295 iter->head_page = cpu_buffer->reader_page;
4296 iter->head = cpu_buffer->reader_page->read;
4297 iter->next_event = iter->head;
4298
4299 iter->cache_reader_page = iter->head_page;
4300 iter->cache_read = cpu_buffer->read;
4301 iter->cache_pages_removed = cpu_buffer->pages_removed;
4302
4303 if (iter->head) {
4304 iter->read_stamp = cpu_buffer->read_stamp;
4305 iter->page_stamp = cpu_buffer->reader_page->page->time_stamp;
4306 } else {
4307 iter->read_stamp = iter->head_page->page->time_stamp;
4308 iter->page_stamp = iter->read_stamp;
4309 }
4310}
4311
4312/**
4313 * ring_buffer_iter_reset - reset an iterator
4314 * @iter: The iterator to reset
4315 *
4316 * Resets the iterator, so that it will start from the beginning
4317 * again.
4318 */
4319void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
4320{
4321 struct ring_buffer_per_cpu *cpu_buffer;
4322 unsigned long flags;
4323
4324 if (!iter)
4325 return;
4326
4327 cpu_buffer = iter->cpu_buffer;
4328
4329 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4330 rb_iter_reset(iter);
4331 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4332}
4333EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
4334
4335/**
4336 * ring_buffer_iter_empty - check if an iterator has no more to read
4337 * @iter: The iterator to check
4338 */
4339int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
4340{
4341 struct ring_buffer_per_cpu *cpu_buffer;
4342 struct buffer_page *reader;
4343 struct buffer_page *head_page;
4344 struct buffer_page *commit_page;
4345 struct buffer_page *curr_commit_page;
4346 unsigned commit;
4347 u64 curr_commit_ts;
4348 u64 commit_ts;
4349
4350 cpu_buffer = iter->cpu_buffer;
4351 reader = cpu_buffer->reader_page;
4352 head_page = cpu_buffer->head_page;
4353 commit_page = cpu_buffer->commit_page;
4354 commit_ts = commit_page->page->time_stamp;
4355
4356 /*
4357 * When the writer goes across pages, it issues a cmpxchg which
4358 * is a mb(), which will synchronize with the rmb here.
4359 * (see rb_tail_page_update())
4360 */
4361 smp_rmb();
4362 commit = rb_page_commit(commit_page);
4363 /* We want to make sure that the commit page doesn't change */
4364 smp_rmb();
4365
4366 /* Make sure commit page didn't change */
4367 curr_commit_page = READ_ONCE(cpu_buffer->commit_page);
4368 curr_commit_ts = READ_ONCE(curr_commit_page->page->time_stamp);
4369
4370 /* If the commit page changed, then there's more data */
4371 if (curr_commit_page != commit_page ||
4372 curr_commit_ts != commit_ts)
4373 return 0;
4374
4375 /* Still racy, as it may return a false positive, but that's OK */
4376 return ((iter->head_page == commit_page && iter->head >= commit) ||
4377 (iter->head_page == reader && commit_page == head_page &&
4378 head_page->read == commit &&
4379 iter->head == rb_page_commit(cpu_buffer->reader_page)));
4380}
4381EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
4382
4383static void
4384rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
4385 struct ring_buffer_event *event)
4386{
4387 u64 delta;
4388
4389 switch (event->type_len) {
4390 case RINGBUF_TYPE_PADDING:
4391 return;
4392
4393 case RINGBUF_TYPE_TIME_EXTEND:
4394 delta = rb_event_time_stamp(event);
4395 cpu_buffer->read_stamp += delta;
4396 return;
4397
4398 case RINGBUF_TYPE_TIME_STAMP:
4399 delta = rb_event_time_stamp(event);
4400 delta = rb_fix_abs_ts(delta, cpu_buffer->read_stamp);
4401 cpu_buffer->read_stamp = delta;
4402 return;
4403
4404 case RINGBUF_TYPE_DATA:
4405 cpu_buffer->read_stamp += event->time_delta;
4406 return;
4407
4408 default:
4409 RB_WARN_ON(cpu_buffer, 1);
4410 }
4411}
4412
4413static void
4414rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
4415 struct ring_buffer_event *event)
4416{
4417 u64 delta;
4418
4419 switch (event->type_len) {
4420 case RINGBUF_TYPE_PADDING:
4421 return;
4422
4423 case RINGBUF_TYPE_TIME_EXTEND:
4424 delta = rb_event_time_stamp(event);
4425 iter->read_stamp += delta;
4426 return;
4427
4428 case RINGBUF_TYPE_TIME_STAMP:
4429 delta = rb_event_time_stamp(event);
4430 delta = rb_fix_abs_ts(delta, iter->read_stamp);
4431 iter->read_stamp = delta;
4432 return;
4433
4434 case RINGBUF_TYPE_DATA:
4435 iter->read_stamp += event->time_delta;
4436 return;
4437
4438 default:
4439 RB_WARN_ON(iter->cpu_buffer, 1);
4440 }
4441}
4442
4443static struct buffer_page *
4444rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
4445{
4446 struct buffer_page *reader = NULL;
4447 unsigned long bsize = READ_ONCE(cpu_buffer->buffer->subbuf_size);
4448 unsigned long overwrite;
4449 unsigned long flags;
4450 int nr_loops = 0;
4451 bool ret;
4452
4453 local_irq_save(flags);
4454 arch_spin_lock(&cpu_buffer->lock);
4455
4456 again:
4457 /*
4458 * This should normally only loop twice. But because the
4459 * start of the reader inserts an empty page, it causes
4460 * a case where we will loop three times. There should be no
4461 * reason to loop four times (that I know of).
4462 */
4463 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
4464 reader = NULL;
4465 goto out;
4466 }
4467
4468 reader = cpu_buffer->reader_page;
4469
4470 /* If there's more to read, return this page */
4471 if (cpu_buffer->reader_page->read < rb_page_size(reader))
4472 goto out;
4473
4474 /* Never should we have an index greater than the size */
4475 if (RB_WARN_ON(cpu_buffer,
4476 cpu_buffer->reader_page->read > rb_page_size(reader)))
4477 goto out;
4478
4479 /* check if we caught up to the tail */
4480 reader = NULL;
4481 if (cpu_buffer->commit_page == cpu_buffer->reader_page)
4482 goto out;
4483
4484 /* Don't bother swapping if the ring buffer is empty */
4485 if (rb_num_of_entries(cpu_buffer) == 0)
4486 goto out;
4487
4488 /*
4489 * Reset the reader page to size zero.
4490 */
4491 local_set(&cpu_buffer->reader_page->write, 0);
4492 local_set(&cpu_buffer->reader_page->entries, 0);
4493 local_set(&cpu_buffer->reader_page->page->commit, 0);
4494 cpu_buffer->reader_page->real_end = 0;
4495
4496 spin:
4497 /*
4498 * Splice the empty reader page into the list around the head.
4499 */
4500 reader = rb_set_head_page(cpu_buffer);
4501 if (!reader)
4502 goto out;
4503 cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
4504 cpu_buffer->reader_page->list.prev = reader->list.prev;
4505
4506 /*
4507 * cpu_buffer->pages just needs to point to the buffer, it
4508 * has no specific buffer page to point to. Lets move it out
4509 * of our way so we don't accidentally swap it.
4510 */
4511 cpu_buffer->pages = reader->list.prev;
4512
4513 /* The reader page will be pointing to the new head */
4514 rb_set_list_to_head(&cpu_buffer->reader_page->list);
4515
4516 /*
4517 * We want to make sure we read the overruns after we set up our
4518 * pointers to the next object. The writer side does a
4519 * cmpxchg to cross pages which acts as the mb on the writer
4520 * side. Note, the reader will constantly fail the swap
4521 * while the writer is updating the pointers, so this
4522 * guarantees that the overwrite recorded here is the one we
4523 * want to compare with the last_overrun.
4524 */
4525 smp_mb();
4526 overwrite = local_read(&(cpu_buffer->overrun));
4527
4528 /*
4529 * Here's the tricky part.
4530 *
4531 * We need to move the pointer past the header page.
4532 * But we can only do that if a writer is not currently
4533 * moving it. The page before the header page has the
4534 * flag bit '1' set if it is pointing to the page we want.
4535 * but if the writer is in the process of moving it
4536 * than it will be '2' or already moved '0'.
4537 */
4538
4539 ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
4540
4541 /*
4542 * If we did not convert it, then we must try again.
4543 */
4544 if (!ret)
4545 goto spin;
4546
4547 /*
4548 * Yay! We succeeded in replacing the page.
4549 *
4550 * Now make the new head point back to the reader page.
4551 */
4552 rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
4553 rb_inc_page(&cpu_buffer->head_page);
4554
4555 local_inc(&cpu_buffer->pages_read);
4556
4557 /* Finally update the reader page to the new head */
4558 cpu_buffer->reader_page = reader;
4559 cpu_buffer->reader_page->read = 0;
4560
4561 if (overwrite != cpu_buffer->last_overrun) {
4562 cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
4563 cpu_buffer->last_overrun = overwrite;
4564 }
4565
4566 goto again;
4567
4568 out:
4569 /* Update the read_stamp on the first event */
4570 if (reader && reader->read == 0)
4571 cpu_buffer->read_stamp = reader->page->time_stamp;
4572
4573 arch_spin_unlock(&cpu_buffer->lock);
4574 local_irq_restore(flags);
4575
4576 /*
4577 * The writer has preempt disable, wait for it. But not forever
4578 * Although, 1 second is pretty much "forever"
4579 */
4580#define USECS_WAIT 1000000
4581 for (nr_loops = 0; nr_loops < USECS_WAIT; nr_loops++) {
4582 /* If the write is past the end of page, a writer is still updating it */
4583 if (likely(!reader || rb_page_write(reader) <= bsize))
4584 break;
4585
4586 udelay(1);
4587
4588 /* Get the latest version of the reader write value */
4589 smp_rmb();
4590 }
4591
4592 /* The writer is not moving forward? Something is wrong */
4593 if (RB_WARN_ON(cpu_buffer, nr_loops == USECS_WAIT))
4594 reader = NULL;
4595
4596 /*
4597 * Make sure we see any padding after the write update
4598 * (see rb_reset_tail()).
4599 *
4600 * In addition, a writer may be writing on the reader page
4601 * if the page has not been fully filled, so the read barrier
4602 * is also needed to make sure we see the content of what is
4603 * committed by the writer (see rb_set_commit_to_write()).
4604 */
4605 smp_rmb();
4606
4607
4608 return reader;
4609}
4610
4611static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
4612{
4613 struct ring_buffer_event *event;
4614 struct buffer_page *reader;
4615 unsigned length;
4616
4617 reader = rb_get_reader_page(cpu_buffer);
4618
4619 /* This function should not be called when buffer is empty */
4620 if (RB_WARN_ON(cpu_buffer, !reader))
4621 return;
4622
4623 event = rb_reader_event(cpu_buffer);
4624
4625 if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
4626 cpu_buffer->read++;
4627
4628 rb_update_read_stamp(cpu_buffer, event);
4629
4630 length = rb_event_length(event);
4631 cpu_buffer->reader_page->read += length;
4632 cpu_buffer->read_bytes += length;
4633}
4634
4635static void rb_advance_iter(struct ring_buffer_iter *iter)
4636{
4637 struct ring_buffer_per_cpu *cpu_buffer;
4638
4639 cpu_buffer = iter->cpu_buffer;
4640
4641 /* If head == next_event then we need to jump to the next event */
4642 if (iter->head == iter->next_event) {
4643 /* If the event gets overwritten again, there's nothing to do */
4644 if (rb_iter_head_event(iter) == NULL)
4645 return;
4646 }
4647
4648 iter->head = iter->next_event;
4649
4650 /*
4651 * Check if we are at the end of the buffer.
4652 */
4653 if (iter->next_event >= rb_page_size(iter->head_page)) {
4654 /* discarded commits can make the page empty */
4655 if (iter->head_page == cpu_buffer->commit_page)
4656 return;
4657 rb_inc_iter(iter);
4658 return;
4659 }
4660
4661 rb_update_iter_read_stamp(iter, iter->event);
4662}
4663
4664static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
4665{
4666 return cpu_buffer->lost_events;
4667}
4668
4669static struct ring_buffer_event *
4670rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
4671 unsigned long *lost_events)
4672{
4673 struct ring_buffer_event *event;
4674 struct buffer_page *reader;
4675 int nr_loops = 0;
4676
4677 if (ts)
4678 *ts = 0;
4679 again:
4680 /*
4681 * We repeat when a time extend is encountered.
4682 * Since the time extend is always attached to a data event,
4683 * we should never loop more than once.
4684 * (We never hit the following condition more than twice).
4685 */
4686 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
4687 return NULL;
4688
4689 reader = rb_get_reader_page(cpu_buffer);
4690 if (!reader)
4691 return NULL;
4692
4693 event = rb_reader_event(cpu_buffer);
4694
4695 switch (event->type_len) {
4696 case RINGBUF_TYPE_PADDING:
4697 if (rb_null_event(event))
4698 RB_WARN_ON(cpu_buffer, 1);
4699 /*
4700 * Because the writer could be discarding every
4701 * event it creates (which would probably be bad)
4702 * if we were to go back to "again" then we may never
4703 * catch up, and will trigger the warn on, or lock
4704 * the box. Return the padding, and we will release
4705 * the current locks, and try again.
4706 */
4707 return event;
4708
4709 case RINGBUF_TYPE_TIME_EXTEND:
4710 /* Internal data, OK to advance */
4711 rb_advance_reader(cpu_buffer);
4712 goto again;
4713
4714 case RINGBUF_TYPE_TIME_STAMP:
4715 if (ts) {
4716 *ts = rb_event_time_stamp(event);
4717 *ts = rb_fix_abs_ts(*ts, reader->page->time_stamp);
4718 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
4719 cpu_buffer->cpu, ts);
4720 }
4721 /* Internal data, OK to advance */
4722 rb_advance_reader(cpu_buffer);
4723 goto again;
4724
4725 case RINGBUF_TYPE_DATA:
4726 if (ts && !(*ts)) {
4727 *ts = cpu_buffer->read_stamp + event->time_delta;
4728 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
4729 cpu_buffer->cpu, ts);
4730 }
4731 if (lost_events)
4732 *lost_events = rb_lost_events(cpu_buffer);
4733 return event;
4734
4735 default:
4736 RB_WARN_ON(cpu_buffer, 1);
4737 }
4738
4739 return NULL;
4740}
4741EXPORT_SYMBOL_GPL(ring_buffer_peek);
4742
4743static struct ring_buffer_event *
4744rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
4745{
4746 struct trace_buffer *buffer;
4747 struct ring_buffer_per_cpu *cpu_buffer;
4748 struct ring_buffer_event *event;
4749 int nr_loops = 0;
4750
4751 if (ts)
4752 *ts = 0;
4753
4754 cpu_buffer = iter->cpu_buffer;
4755 buffer = cpu_buffer->buffer;
4756
4757 /*
4758 * Check if someone performed a consuming read to the buffer
4759 * or removed some pages from the buffer. In these cases,
4760 * iterator was invalidated and we need to reset it.
4761 */
4762 if (unlikely(iter->cache_read != cpu_buffer->read ||
4763 iter->cache_reader_page != cpu_buffer->reader_page ||
4764 iter->cache_pages_removed != cpu_buffer->pages_removed))
4765 rb_iter_reset(iter);
4766
4767 again:
4768 if (ring_buffer_iter_empty(iter))
4769 return NULL;
4770
4771 /*
4772 * As the writer can mess with what the iterator is trying
4773 * to read, just give up if we fail to get an event after
4774 * three tries. The iterator is not as reliable when reading
4775 * the ring buffer with an active write as the consumer is.
4776 * Do not warn if the three failures is reached.
4777 */
4778 if (++nr_loops > 3)
4779 return NULL;
4780
4781 if (rb_per_cpu_empty(cpu_buffer))
4782 return NULL;
4783
4784 if (iter->head >= rb_page_size(iter->head_page)) {
4785 rb_inc_iter(iter);
4786 goto again;
4787 }
4788
4789 event = rb_iter_head_event(iter);
4790 if (!event)
4791 goto again;
4792
4793 switch (event->type_len) {
4794 case RINGBUF_TYPE_PADDING:
4795 if (rb_null_event(event)) {
4796 rb_inc_iter(iter);
4797 goto again;
4798 }
4799 rb_advance_iter(iter);
4800 return event;
4801
4802 case RINGBUF_TYPE_TIME_EXTEND:
4803 /* Internal data, OK to advance */
4804 rb_advance_iter(iter);
4805 goto again;
4806
4807 case RINGBUF_TYPE_TIME_STAMP:
4808 if (ts) {
4809 *ts = rb_event_time_stamp(event);
4810 *ts = rb_fix_abs_ts(*ts, iter->head_page->page->time_stamp);
4811 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
4812 cpu_buffer->cpu, ts);
4813 }
4814 /* Internal data, OK to advance */
4815 rb_advance_iter(iter);
4816 goto again;
4817
4818 case RINGBUF_TYPE_DATA:
4819 if (ts && !(*ts)) {
4820 *ts = iter->read_stamp + event->time_delta;
4821 ring_buffer_normalize_time_stamp(buffer,
4822 cpu_buffer->cpu, ts);
4823 }
4824 return event;
4825
4826 default:
4827 RB_WARN_ON(cpu_buffer, 1);
4828 }
4829
4830 return NULL;
4831}
4832EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
4833
4834static inline bool rb_reader_lock(struct ring_buffer_per_cpu *cpu_buffer)
4835{
4836 if (likely(!in_nmi())) {
4837 raw_spin_lock(&cpu_buffer->reader_lock);
4838 return true;
4839 }
4840
4841 /*
4842 * If an NMI die dumps out the content of the ring buffer
4843 * trylock must be used to prevent a deadlock if the NMI
4844 * preempted a task that holds the ring buffer locks. If
4845 * we get the lock then all is fine, if not, then continue
4846 * to do the read, but this can corrupt the ring buffer,
4847 * so it must be permanently disabled from future writes.
4848 * Reading from NMI is a oneshot deal.
4849 */
4850 if (raw_spin_trylock(&cpu_buffer->reader_lock))
4851 return true;
4852
4853 /* Continue without locking, but disable the ring buffer */
4854 atomic_inc(&cpu_buffer->record_disabled);
4855 return false;
4856}
4857
4858static inline void
4859rb_reader_unlock(struct ring_buffer_per_cpu *cpu_buffer, bool locked)
4860{
4861 if (likely(locked))
4862 raw_spin_unlock(&cpu_buffer->reader_lock);
4863}
4864
4865/**
4866 * ring_buffer_peek - peek at the next event to be read
4867 * @buffer: The ring buffer to read
4868 * @cpu: The cpu to peak at
4869 * @ts: The timestamp counter of this event.
4870 * @lost_events: a variable to store if events were lost (may be NULL)
4871 *
4872 * This will return the event that will be read next, but does
4873 * not consume the data.
4874 */
4875struct ring_buffer_event *
4876ring_buffer_peek(struct trace_buffer *buffer, int cpu, u64 *ts,
4877 unsigned long *lost_events)
4878{
4879 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4880 struct ring_buffer_event *event;
4881 unsigned long flags;
4882 bool dolock;
4883
4884 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4885 return NULL;
4886
4887 again:
4888 local_irq_save(flags);
4889 dolock = rb_reader_lock(cpu_buffer);
4890 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
4891 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4892 rb_advance_reader(cpu_buffer);
4893 rb_reader_unlock(cpu_buffer, dolock);
4894 local_irq_restore(flags);
4895
4896 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4897 goto again;
4898
4899 return event;
4900}
4901
4902/** ring_buffer_iter_dropped - report if there are dropped events
4903 * @iter: The ring buffer iterator
4904 *
4905 * Returns true if there was dropped events since the last peek.
4906 */
4907bool ring_buffer_iter_dropped(struct ring_buffer_iter *iter)
4908{
4909 bool ret = iter->missed_events != 0;
4910
4911 iter->missed_events = 0;
4912 return ret;
4913}
4914EXPORT_SYMBOL_GPL(ring_buffer_iter_dropped);
4915
4916/**
4917 * ring_buffer_iter_peek - peek at the next event to be read
4918 * @iter: The ring buffer iterator
4919 * @ts: The timestamp counter of this event.
4920 *
4921 * This will return the event that will be read next, but does
4922 * not increment the iterator.
4923 */
4924struct ring_buffer_event *
4925ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
4926{
4927 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4928 struct ring_buffer_event *event;
4929 unsigned long flags;
4930
4931 again:
4932 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4933 event = rb_iter_peek(iter, ts);
4934 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4935
4936 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4937 goto again;
4938
4939 return event;
4940}
4941
4942/**
4943 * ring_buffer_consume - return an event and consume it
4944 * @buffer: The ring buffer to get the next event from
4945 * @cpu: the cpu to read the buffer from
4946 * @ts: a variable to store the timestamp (may be NULL)
4947 * @lost_events: a variable to store if events were lost (may be NULL)
4948 *
4949 * Returns the next event in the ring buffer, and that event is consumed.
4950 * Meaning, that sequential reads will keep returning a different event,
4951 * and eventually empty the ring buffer if the producer is slower.
4952 */
4953struct ring_buffer_event *
4954ring_buffer_consume(struct trace_buffer *buffer, int cpu, u64 *ts,
4955 unsigned long *lost_events)
4956{
4957 struct ring_buffer_per_cpu *cpu_buffer;
4958 struct ring_buffer_event *event = NULL;
4959 unsigned long flags;
4960 bool dolock;
4961
4962 again:
4963 /* might be called in atomic */
4964 preempt_disable();
4965
4966 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4967 goto out;
4968
4969 cpu_buffer = buffer->buffers[cpu];
4970 local_irq_save(flags);
4971 dolock = rb_reader_lock(cpu_buffer);
4972
4973 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
4974 if (event) {
4975 cpu_buffer->lost_events = 0;
4976 rb_advance_reader(cpu_buffer);
4977 }
4978
4979 rb_reader_unlock(cpu_buffer, dolock);
4980 local_irq_restore(flags);
4981
4982 out:
4983 preempt_enable();
4984
4985 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4986 goto again;
4987
4988 return event;
4989}
4990EXPORT_SYMBOL_GPL(ring_buffer_consume);
4991
4992/**
4993 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
4994 * @buffer: The ring buffer to read from
4995 * @cpu: The cpu buffer to iterate over
4996 * @flags: gfp flags to use for memory allocation
4997 *
4998 * This performs the initial preparations necessary to iterate
4999 * through the buffer. Memory is allocated, buffer recording
5000 * is disabled, and the iterator pointer is returned to the caller.
5001 *
5002 * Disabling buffer recording prevents the reading from being
5003 * corrupted. This is not a consuming read, so a producer is not
5004 * expected.
5005 *
5006 * After a sequence of ring_buffer_read_prepare calls, the user is
5007 * expected to make at least one call to ring_buffer_read_prepare_sync.
5008 * Afterwards, ring_buffer_read_start is invoked to get things going
5009 * for real.
5010 *
5011 * This overall must be paired with ring_buffer_read_finish.
5012 */
5013struct ring_buffer_iter *
5014ring_buffer_read_prepare(struct trace_buffer *buffer, int cpu, gfp_t flags)
5015{
5016 struct ring_buffer_per_cpu *cpu_buffer;
5017 struct ring_buffer_iter *iter;
5018
5019 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5020 return NULL;
5021
5022 iter = kzalloc(sizeof(*iter), flags);
5023 if (!iter)
5024 return NULL;
5025
5026 /* Holds the entire event: data and meta data */
5027 iter->event_size = buffer->subbuf_size;
5028 iter->event = kmalloc(iter->event_size, flags);
5029 if (!iter->event) {
5030 kfree(iter);
5031 return NULL;
5032 }
5033
5034 cpu_buffer = buffer->buffers[cpu];
5035
5036 iter->cpu_buffer = cpu_buffer;
5037
5038 atomic_inc(&cpu_buffer->resize_disabled);
5039
5040 return iter;
5041}
5042EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
5043
5044/**
5045 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
5046 *
5047 * All previously invoked ring_buffer_read_prepare calls to prepare
5048 * iterators will be synchronized. Afterwards, read_buffer_read_start
5049 * calls on those iterators are allowed.
5050 */
5051void
5052ring_buffer_read_prepare_sync(void)
5053{
5054 synchronize_rcu();
5055}
5056EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
5057
5058/**
5059 * ring_buffer_read_start - start a non consuming read of the buffer
5060 * @iter: The iterator returned by ring_buffer_read_prepare
5061 *
5062 * This finalizes the startup of an iteration through the buffer.
5063 * The iterator comes from a call to ring_buffer_read_prepare and
5064 * an intervening ring_buffer_read_prepare_sync must have been
5065 * performed.
5066 *
5067 * Must be paired with ring_buffer_read_finish.
5068 */
5069void
5070ring_buffer_read_start(struct ring_buffer_iter *iter)
5071{
5072 struct ring_buffer_per_cpu *cpu_buffer;
5073 unsigned long flags;
5074
5075 if (!iter)
5076 return;
5077
5078 cpu_buffer = iter->cpu_buffer;
5079
5080 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5081 arch_spin_lock(&cpu_buffer->lock);
5082 rb_iter_reset(iter);
5083 arch_spin_unlock(&cpu_buffer->lock);
5084 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5085}
5086EXPORT_SYMBOL_GPL(ring_buffer_read_start);
5087
5088/**
5089 * ring_buffer_read_finish - finish reading the iterator of the buffer
5090 * @iter: The iterator retrieved by ring_buffer_start
5091 *
5092 * This re-enables the recording to the buffer, and frees the
5093 * iterator.
5094 */
5095void
5096ring_buffer_read_finish(struct ring_buffer_iter *iter)
5097{
5098 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
5099 unsigned long flags;
5100
5101 /*
5102 * Ring buffer is disabled from recording, here's a good place
5103 * to check the integrity of the ring buffer.
5104 * Must prevent readers from trying to read, as the check
5105 * clears the HEAD page and readers require it.
5106 */
5107 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5108 rb_check_pages(cpu_buffer);
5109 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5110
5111 atomic_dec(&cpu_buffer->resize_disabled);
5112 kfree(iter->event);
5113 kfree(iter);
5114}
5115EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
5116
5117/**
5118 * ring_buffer_iter_advance - advance the iterator to the next location
5119 * @iter: The ring buffer iterator
5120 *
5121 * Move the location of the iterator such that the next read will
5122 * be the next location of the iterator.
5123 */
5124void ring_buffer_iter_advance(struct ring_buffer_iter *iter)
5125{
5126 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
5127 unsigned long flags;
5128
5129 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5130
5131 rb_advance_iter(iter);
5132
5133 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5134}
5135EXPORT_SYMBOL_GPL(ring_buffer_iter_advance);
5136
5137/**
5138 * ring_buffer_size - return the size of the ring buffer (in bytes)
5139 * @buffer: The ring buffer.
5140 * @cpu: The CPU to get ring buffer size from.
5141 */
5142unsigned long ring_buffer_size(struct trace_buffer *buffer, int cpu)
5143{
5144 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5145 return 0;
5146
5147 return buffer->subbuf_size * buffer->buffers[cpu]->nr_pages;
5148}
5149EXPORT_SYMBOL_GPL(ring_buffer_size);
5150
5151/**
5152 * ring_buffer_max_event_size - return the max data size of an event
5153 * @buffer: The ring buffer.
5154 *
5155 * Returns the maximum size an event can be.
5156 */
5157unsigned long ring_buffer_max_event_size(struct trace_buffer *buffer)
5158{
5159 /* If abs timestamp is requested, events have a timestamp too */
5160 if (ring_buffer_time_stamp_abs(buffer))
5161 return buffer->max_data_size - RB_LEN_TIME_EXTEND;
5162 return buffer->max_data_size;
5163}
5164EXPORT_SYMBOL_GPL(ring_buffer_max_event_size);
5165
5166static void rb_clear_buffer_page(struct buffer_page *page)
5167{
5168 local_set(&page->write, 0);
5169 local_set(&page->entries, 0);
5170 rb_init_page(page->page);
5171 page->read = 0;
5172}
5173
5174static void
5175rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
5176{
5177 struct buffer_page *page;
5178
5179 rb_head_page_deactivate(cpu_buffer);
5180
5181 cpu_buffer->head_page
5182 = list_entry(cpu_buffer->pages, struct buffer_page, list);
5183 rb_clear_buffer_page(cpu_buffer->head_page);
5184 list_for_each_entry(page, cpu_buffer->pages, list) {
5185 rb_clear_buffer_page(page);
5186 }
5187
5188 cpu_buffer->tail_page = cpu_buffer->head_page;
5189 cpu_buffer->commit_page = cpu_buffer->head_page;
5190
5191 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
5192 INIT_LIST_HEAD(&cpu_buffer->new_pages);
5193 rb_clear_buffer_page(cpu_buffer->reader_page);
5194
5195 local_set(&cpu_buffer->entries_bytes, 0);
5196 local_set(&cpu_buffer->overrun, 0);
5197 local_set(&cpu_buffer->commit_overrun, 0);
5198 local_set(&cpu_buffer->dropped_events, 0);
5199 local_set(&cpu_buffer->entries, 0);
5200 local_set(&cpu_buffer->committing, 0);
5201 local_set(&cpu_buffer->commits, 0);
5202 local_set(&cpu_buffer->pages_touched, 0);
5203 local_set(&cpu_buffer->pages_lost, 0);
5204 local_set(&cpu_buffer->pages_read, 0);
5205 cpu_buffer->last_pages_touch = 0;
5206 cpu_buffer->shortest_full = 0;
5207 cpu_buffer->read = 0;
5208 cpu_buffer->read_bytes = 0;
5209
5210 rb_time_set(&cpu_buffer->write_stamp, 0);
5211 rb_time_set(&cpu_buffer->before_stamp, 0);
5212
5213 memset(cpu_buffer->event_stamp, 0, sizeof(cpu_buffer->event_stamp));
5214
5215 cpu_buffer->lost_events = 0;
5216 cpu_buffer->last_overrun = 0;
5217
5218 rb_head_page_activate(cpu_buffer);
5219 cpu_buffer->pages_removed = 0;
5220}
5221
5222/* Must have disabled the cpu buffer then done a synchronize_rcu */
5223static void reset_disabled_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
5224{
5225 unsigned long flags;
5226
5227 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5228
5229 if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
5230 goto out;
5231
5232 arch_spin_lock(&cpu_buffer->lock);
5233
5234 rb_reset_cpu(cpu_buffer);
5235
5236 arch_spin_unlock(&cpu_buffer->lock);
5237
5238 out:
5239 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5240}
5241
5242/**
5243 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
5244 * @buffer: The ring buffer to reset a per cpu buffer of
5245 * @cpu: The CPU buffer to be reset
5246 */
5247void ring_buffer_reset_cpu(struct trace_buffer *buffer, int cpu)
5248{
5249 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
5250
5251 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5252 return;
5253
5254 /* prevent another thread from changing buffer sizes */
5255 mutex_lock(&buffer->mutex);
5256
5257 atomic_inc(&cpu_buffer->resize_disabled);
5258 atomic_inc(&cpu_buffer->record_disabled);
5259
5260 /* Make sure all commits have finished */
5261 synchronize_rcu();
5262
5263 reset_disabled_cpu_buffer(cpu_buffer);
5264
5265 atomic_dec(&cpu_buffer->record_disabled);
5266 atomic_dec(&cpu_buffer->resize_disabled);
5267
5268 mutex_unlock(&buffer->mutex);
5269}
5270EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
5271
5272/* Flag to ensure proper resetting of atomic variables */
5273#define RESET_BIT (1 << 30)
5274
5275/**
5276 * ring_buffer_reset_online_cpus - reset a ring buffer per CPU buffer
5277 * @buffer: The ring buffer to reset a per cpu buffer of
5278 */
5279void ring_buffer_reset_online_cpus(struct trace_buffer *buffer)
5280{
5281 struct ring_buffer_per_cpu *cpu_buffer;
5282 int cpu;
5283
5284 /* prevent another thread from changing buffer sizes */
5285 mutex_lock(&buffer->mutex);
5286
5287 for_each_online_buffer_cpu(buffer, cpu) {
5288 cpu_buffer = buffer->buffers[cpu];
5289
5290 atomic_add(RESET_BIT, &cpu_buffer->resize_disabled);
5291 atomic_inc(&cpu_buffer->record_disabled);
5292 }
5293
5294 /* Make sure all commits have finished */
5295 synchronize_rcu();
5296
5297 for_each_buffer_cpu(buffer, cpu) {
5298 cpu_buffer = buffer->buffers[cpu];
5299
5300 /*
5301 * If a CPU came online during the synchronize_rcu(), then
5302 * ignore it.
5303 */
5304 if (!(atomic_read(&cpu_buffer->resize_disabled) & RESET_BIT))
5305 continue;
5306
5307 reset_disabled_cpu_buffer(cpu_buffer);
5308
5309 atomic_dec(&cpu_buffer->record_disabled);
5310 atomic_sub(RESET_BIT, &cpu_buffer->resize_disabled);
5311 }
5312
5313 mutex_unlock(&buffer->mutex);
5314}
5315
5316/**
5317 * ring_buffer_reset - reset a ring buffer
5318 * @buffer: The ring buffer to reset all cpu buffers
5319 */
5320void ring_buffer_reset(struct trace_buffer *buffer)
5321{
5322 struct ring_buffer_per_cpu *cpu_buffer;
5323 int cpu;
5324
5325 /* prevent another thread from changing buffer sizes */
5326 mutex_lock(&buffer->mutex);
5327
5328 for_each_buffer_cpu(buffer, cpu) {
5329 cpu_buffer = buffer->buffers[cpu];
5330
5331 atomic_inc(&cpu_buffer->resize_disabled);
5332 atomic_inc(&cpu_buffer->record_disabled);
5333 }
5334
5335 /* Make sure all commits have finished */
5336 synchronize_rcu();
5337
5338 for_each_buffer_cpu(buffer, cpu) {
5339 cpu_buffer = buffer->buffers[cpu];
5340
5341 reset_disabled_cpu_buffer(cpu_buffer);
5342
5343 atomic_dec(&cpu_buffer->record_disabled);
5344 atomic_dec(&cpu_buffer->resize_disabled);
5345 }
5346
5347 mutex_unlock(&buffer->mutex);
5348}
5349EXPORT_SYMBOL_GPL(ring_buffer_reset);
5350
5351/**
5352 * ring_buffer_empty - is the ring buffer empty?
5353 * @buffer: The ring buffer to test
5354 */
5355bool ring_buffer_empty(struct trace_buffer *buffer)
5356{
5357 struct ring_buffer_per_cpu *cpu_buffer;
5358 unsigned long flags;
5359 bool dolock;
5360 bool ret;
5361 int cpu;
5362
5363 /* yes this is racy, but if you don't like the race, lock the buffer */
5364 for_each_buffer_cpu(buffer, cpu) {
5365 cpu_buffer = buffer->buffers[cpu];
5366 local_irq_save(flags);
5367 dolock = rb_reader_lock(cpu_buffer);
5368 ret = rb_per_cpu_empty(cpu_buffer);
5369 rb_reader_unlock(cpu_buffer, dolock);
5370 local_irq_restore(flags);
5371
5372 if (!ret)
5373 return false;
5374 }
5375
5376 return true;
5377}
5378EXPORT_SYMBOL_GPL(ring_buffer_empty);
5379
5380/**
5381 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
5382 * @buffer: The ring buffer
5383 * @cpu: The CPU buffer to test
5384 */
5385bool ring_buffer_empty_cpu(struct trace_buffer *buffer, int cpu)
5386{
5387 struct ring_buffer_per_cpu *cpu_buffer;
5388 unsigned long flags;
5389 bool dolock;
5390 bool ret;
5391
5392 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5393 return true;
5394
5395 cpu_buffer = buffer->buffers[cpu];
5396 local_irq_save(flags);
5397 dolock = rb_reader_lock(cpu_buffer);
5398 ret = rb_per_cpu_empty(cpu_buffer);
5399 rb_reader_unlock(cpu_buffer, dolock);
5400 local_irq_restore(flags);
5401
5402 return ret;
5403}
5404EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
5405
5406#ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
5407/**
5408 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
5409 * @buffer_a: One buffer to swap with
5410 * @buffer_b: The other buffer to swap with
5411 * @cpu: the CPU of the buffers to swap
5412 *
5413 * This function is useful for tracers that want to take a "snapshot"
5414 * of a CPU buffer and has another back up buffer lying around.
5415 * it is expected that the tracer handles the cpu buffer not being
5416 * used at the moment.
5417 */
5418int ring_buffer_swap_cpu(struct trace_buffer *buffer_a,
5419 struct trace_buffer *buffer_b, int cpu)
5420{
5421 struct ring_buffer_per_cpu *cpu_buffer_a;
5422 struct ring_buffer_per_cpu *cpu_buffer_b;
5423 int ret = -EINVAL;
5424
5425 if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
5426 !cpumask_test_cpu(cpu, buffer_b->cpumask))
5427 goto out;
5428
5429 cpu_buffer_a = buffer_a->buffers[cpu];
5430 cpu_buffer_b = buffer_b->buffers[cpu];
5431
5432 /* At least make sure the two buffers are somewhat the same */
5433 if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
5434 goto out;
5435
5436 if (buffer_a->subbuf_order != buffer_b->subbuf_order)
5437 goto out;
5438
5439 ret = -EAGAIN;
5440
5441 if (atomic_read(&buffer_a->record_disabled))
5442 goto out;
5443
5444 if (atomic_read(&buffer_b->record_disabled))
5445 goto out;
5446
5447 if (atomic_read(&cpu_buffer_a->record_disabled))
5448 goto out;
5449
5450 if (atomic_read(&cpu_buffer_b->record_disabled))
5451 goto out;
5452
5453 /*
5454 * We can't do a synchronize_rcu here because this
5455 * function can be called in atomic context.
5456 * Normally this will be called from the same CPU as cpu.
5457 * If not it's up to the caller to protect this.
5458 */
5459 atomic_inc(&cpu_buffer_a->record_disabled);
5460 atomic_inc(&cpu_buffer_b->record_disabled);
5461
5462 ret = -EBUSY;
5463 if (local_read(&cpu_buffer_a->committing))
5464 goto out_dec;
5465 if (local_read(&cpu_buffer_b->committing))
5466 goto out_dec;
5467
5468 /*
5469 * When resize is in progress, we cannot swap it because
5470 * it will mess the state of the cpu buffer.
5471 */
5472 if (atomic_read(&buffer_a->resizing))
5473 goto out_dec;
5474 if (atomic_read(&buffer_b->resizing))
5475 goto out_dec;
5476
5477 buffer_a->buffers[cpu] = cpu_buffer_b;
5478 buffer_b->buffers[cpu] = cpu_buffer_a;
5479
5480 cpu_buffer_b->buffer = buffer_a;
5481 cpu_buffer_a->buffer = buffer_b;
5482
5483 ret = 0;
5484
5485out_dec:
5486 atomic_dec(&cpu_buffer_a->record_disabled);
5487 atomic_dec(&cpu_buffer_b->record_disabled);
5488out:
5489 return ret;
5490}
5491EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
5492#endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
5493
5494/**
5495 * ring_buffer_alloc_read_page - allocate a page to read from buffer
5496 * @buffer: the buffer to allocate for.
5497 * @cpu: the cpu buffer to allocate.
5498 *
5499 * This function is used in conjunction with ring_buffer_read_page.
5500 * When reading a full page from the ring buffer, these functions
5501 * can be used to speed up the process. The calling function should
5502 * allocate a few pages first with this function. Then when it
5503 * needs to get pages from the ring buffer, it passes the result
5504 * of this function into ring_buffer_read_page, which will swap
5505 * the page that was allocated, with the read page of the buffer.
5506 *
5507 * Returns:
5508 * The page allocated, or ERR_PTR
5509 */
5510struct buffer_data_read_page *
5511ring_buffer_alloc_read_page(struct trace_buffer *buffer, int cpu)
5512{
5513 struct ring_buffer_per_cpu *cpu_buffer;
5514 struct buffer_data_read_page *bpage = NULL;
5515 unsigned long flags;
5516 struct page *page;
5517
5518 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5519 return ERR_PTR(-ENODEV);
5520
5521 bpage = kzalloc(sizeof(*bpage), GFP_KERNEL);
5522 if (!bpage)
5523 return ERR_PTR(-ENOMEM);
5524
5525 bpage->order = buffer->subbuf_order;
5526 cpu_buffer = buffer->buffers[cpu];
5527 local_irq_save(flags);
5528 arch_spin_lock(&cpu_buffer->lock);
5529
5530 if (cpu_buffer->free_page) {
5531 bpage->data = cpu_buffer->free_page;
5532 cpu_buffer->free_page = NULL;
5533 }
5534
5535 arch_spin_unlock(&cpu_buffer->lock);
5536 local_irq_restore(flags);
5537
5538 if (bpage->data)
5539 goto out;
5540
5541 page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL | __GFP_NORETRY,
5542 cpu_buffer->buffer->subbuf_order);
5543 if (!page) {
5544 kfree(bpage);
5545 return ERR_PTR(-ENOMEM);
5546 }
5547
5548 bpage->data = page_address(page);
5549
5550 out:
5551 rb_init_page(bpage->data);
5552
5553 return bpage;
5554}
5555EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
5556
5557/**
5558 * ring_buffer_free_read_page - free an allocated read page
5559 * @buffer: the buffer the page was allocate for
5560 * @cpu: the cpu buffer the page came from
5561 * @data_page: the page to free
5562 *
5563 * Free a page allocated from ring_buffer_alloc_read_page.
5564 */
5565void ring_buffer_free_read_page(struct trace_buffer *buffer, int cpu,
5566 struct buffer_data_read_page *data_page)
5567{
5568 struct ring_buffer_per_cpu *cpu_buffer;
5569 struct buffer_data_page *bpage = data_page->data;
5570 struct page *page = virt_to_page(bpage);
5571 unsigned long flags;
5572
5573 if (!buffer || !buffer->buffers || !buffer->buffers[cpu])
5574 return;
5575
5576 cpu_buffer = buffer->buffers[cpu];
5577
5578 /*
5579 * If the page is still in use someplace else, or order of the page
5580 * is different from the subbuffer order of the buffer -
5581 * we can't reuse it
5582 */
5583 if (page_ref_count(page) > 1 || data_page->order != buffer->subbuf_order)
5584 goto out;
5585
5586 local_irq_save(flags);
5587 arch_spin_lock(&cpu_buffer->lock);
5588
5589 if (!cpu_buffer->free_page) {
5590 cpu_buffer->free_page = bpage;
5591 bpage = NULL;
5592 }
5593
5594 arch_spin_unlock(&cpu_buffer->lock);
5595 local_irq_restore(flags);
5596
5597 out:
5598 free_pages((unsigned long)bpage, data_page->order);
5599 kfree(data_page);
5600}
5601EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
5602
5603/**
5604 * ring_buffer_read_page - extract a page from the ring buffer
5605 * @buffer: buffer to extract from
5606 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
5607 * @len: amount to extract
5608 * @cpu: the cpu of the buffer to extract
5609 * @full: should the extraction only happen when the page is full.
5610 *
5611 * This function will pull out a page from the ring buffer and consume it.
5612 * @data_page must be the address of the variable that was returned
5613 * from ring_buffer_alloc_read_page. This is because the page might be used
5614 * to swap with a page in the ring buffer.
5615 *
5616 * for example:
5617 * rpage = ring_buffer_alloc_read_page(buffer, cpu);
5618 * if (IS_ERR(rpage))
5619 * return PTR_ERR(rpage);
5620 * ret = ring_buffer_read_page(buffer, rpage, len, cpu, 0);
5621 * if (ret >= 0)
5622 * process_page(ring_buffer_read_page_data(rpage), ret);
5623 * ring_buffer_free_read_page(buffer, cpu, rpage);
5624 *
5625 * When @full is set, the function will not return true unless
5626 * the writer is off the reader page.
5627 *
5628 * Note: it is up to the calling functions to handle sleeps and wakeups.
5629 * The ring buffer can be used anywhere in the kernel and can not
5630 * blindly call wake_up. The layer that uses the ring buffer must be
5631 * responsible for that.
5632 *
5633 * Returns:
5634 * >=0 if data has been transferred, returns the offset of consumed data.
5635 * <0 if no data has been transferred.
5636 */
5637int ring_buffer_read_page(struct trace_buffer *buffer,
5638 struct buffer_data_read_page *data_page,
5639 size_t len, int cpu, int full)
5640{
5641 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
5642 struct ring_buffer_event *event;
5643 struct buffer_data_page *bpage;
5644 struct buffer_page *reader;
5645 unsigned long missed_events;
5646 unsigned long flags;
5647 unsigned int commit;
5648 unsigned int read;
5649 u64 save_timestamp;
5650 int ret = -1;
5651
5652 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5653 goto out;
5654
5655 /*
5656 * If len is not big enough to hold the page header, then
5657 * we can not copy anything.
5658 */
5659 if (len <= BUF_PAGE_HDR_SIZE)
5660 goto out;
5661
5662 len -= BUF_PAGE_HDR_SIZE;
5663
5664 if (!data_page || !data_page->data)
5665 goto out;
5666 if (data_page->order != buffer->subbuf_order)
5667 goto out;
5668
5669 bpage = data_page->data;
5670 if (!bpage)
5671 goto out;
5672
5673 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5674
5675 reader = rb_get_reader_page(cpu_buffer);
5676 if (!reader)
5677 goto out_unlock;
5678
5679 event = rb_reader_event(cpu_buffer);
5680
5681 read = reader->read;
5682 commit = rb_page_commit(reader);
5683
5684 /* Check if any events were dropped */
5685 missed_events = cpu_buffer->lost_events;
5686
5687 /*
5688 * If this page has been partially read or
5689 * if len is not big enough to read the rest of the page or
5690 * a writer is still on the page, then
5691 * we must copy the data from the page to the buffer.
5692 * Otherwise, we can simply swap the page with the one passed in.
5693 */
5694 if (read || (len < (commit - read)) ||
5695 cpu_buffer->reader_page == cpu_buffer->commit_page) {
5696 struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
5697 unsigned int rpos = read;
5698 unsigned int pos = 0;
5699 unsigned int size;
5700
5701 /*
5702 * If a full page is expected, this can still be returned
5703 * if there's been a previous partial read and the
5704 * rest of the page can be read and the commit page is off
5705 * the reader page.
5706 */
5707 if (full &&
5708 (!read || (len < (commit - read)) ||
5709 cpu_buffer->reader_page == cpu_buffer->commit_page))
5710 goto out_unlock;
5711
5712 if (len > (commit - read))
5713 len = (commit - read);
5714
5715 /* Always keep the time extend and data together */
5716 size = rb_event_ts_length(event);
5717
5718 if (len < size)
5719 goto out_unlock;
5720
5721 /* save the current timestamp, since the user will need it */
5722 save_timestamp = cpu_buffer->read_stamp;
5723
5724 /* Need to copy one event at a time */
5725 do {
5726 /* We need the size of one event, because
5727 * rb_advance_reader only advances by one event,
5728 * whereas rb_event_ts_length may include the size of
5729 * one or two events.
5730 * We have already ensured there's enough space if this
5731 * is a time extend. */
5732 size = rb_event_length(event);
5733 memcpy(bpage->data + pos, rpage->data + rpos, size);
5734
5735 len -= size;
5736
5737 rb_advance_reader(cpu_buffer);
5738 rpos = reader->read;
5739 pos += size;
5740
5741 if (rpos >= commit)
5742 break;
5743
5744 event = rb_reader_event(cpu_buffer);
5745 /* Always keep the time extend and data together */
5746 size = rb_event_ts_length(event);
5747 } while (len >= size);
5748
5749 /* update bpage */
5750 local_set(&bpage->commit, pos);
5751 bpage->time_stamp = save_timestamp;
5752
5753 /* we copied everything to the beginning */
5754 read = 0;
5755 } else {
5756 /* update the entry counter */
5757 cpu_buffer->read += rb_page_entries(reader);
5758 cpu_buffer->read_bytes += rb_page_commit(reader);
5759
5760 /* swap the pages */
5761 rb_init_page(bpage);
5762 bpage = reader->page;
5763 reader->page = data_page->data;
5764 local_set(&reader->write, 0);
5765 local_set(&reader->entries, 0);
5766 reader->read = 0;
5767 data_page->data = bpage;
5768
5769 /*
5770 * Use the real_end for the data size,
5771 * This gives us a chance to store the lost events
5772 * on the page.
5773 */
5774 if (reader->real_end)
5775 local_set(&bpage->commit, reader->real_end);
5776 }
5777 ret = read;
5778
5779 cpu_buffer->lost_events = 0;
5780
5781 commit = local_read(&bpage->commit);
5782 /*
5783 * Set a flag in the commit field if we lost events
5784 */
5785 if (missed_events) {
5786 /* If there is room at the end of the page to save the
5787 * missed events, then record it there.
5788 */
5789 if (buffer->subbuf_size - commit >= sizeof(missed_events)) {
5790 memcpy(&bpage->data[commit], &missed_events,
5791 sizeof(missed_events));
5792 local_add(RB_MISSED_STORED, &bpage->commit);
5793 commit += sizeof(missed_events);
5794 }
5795 local_add(RB_MISSED_EVENTS, &bpage->commit);
5796 }
5797
5798 /*
5799 * This page may be off to user land. Zero it out here.
5800 */
5801 if (commit < buffer->subbuf_size)
5802 memset(&bpage->data[commit], 0, buffer->subbuf_size - commit);
5803
5804 out_unlock:
5805 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5806
5807 out:
5808 return ret;
5809}
5810EXPORT_SYMBOL_GPL(ring_buffer_read_page);
5811
5812/**
5813 * ring_buffer_read_page_data - get pointer to the data in the page.
5814 * @page: the page to get the data from
5815 *
5816 * Returns pointer to the actual data in this page.
5817 */
5818void *ring_buffer_read_page_data(struct buffer_data_read_page *page)
5819{
5820 return page->data;
5821}
5822EXPORT_SYMBOL_GPL(ring_buffer_read_page_data);
5823
5824/**
5825 * ring_buffer_subbuf_size_get - get size of the sub buffer.
5826 * @buffer: the buffer to get the sub buffer size from
5827 *
5828 * Returns size of the sub buffer, in bytes.
5829 */
5830int ring_buffer_subbuf_size_get(struct trace_buffer *buffer)
5831{
5832 return buffer->subbuf_size + BUF_PAGE_HDR_SIZE;
5833}
5834EXPORT_SYMBOL_GPL(ring_buffer_subbuf_size_get);
5835
5836/**
5837 * ring_buffer_subbuf_order_get - get order of system sub pages in one buffer page.
5838 * @buffer: The ring_buffer to get the system sub page order from
5839 *
5840 * By default, one ring buffer sub page equals to one system page. This parameter
5841 * is configurable, per ring buffer. The size of the ring buffer sub page can be
5842 * extended, but must be an order of system page size.
5843 *
5844 * Returns the order of buffer sub page size, in system pages:
5845 * 0 means the sub buffer size is 1 system page and so forth.
5846 * In case of an error < 0 is returned.
5847 */
5848int ring_buffer_subbuf_order_get(struct trace_buffer *buffer)
5849{
5850 if (!buffer)
5851 return -EINVAL;
5852
5853 return buffer->subbuf_order;
5854}
5855EXPORT_SYMBOL_GPL(ring_buffer_subbuf_order_get);
5856
5857/**
5858 * ring_buffer_subbuf_order_set - set the size of ring buffer sub page.
5859 * @buffer: The ring_buffer to set the new page size.
5860 * @order: Order of the system pages in one sub buffer page
5861 *
5862 * By default, one ring buffer pages equals to one system page. This API can be
5863 * used to set new size of the ring buffer page. The size must be order of
5864 * system page size, that's why the input parameter @order is the order of
5865 * system pages that are allocated for one ring buffer page:
5866 * 0 - 1 system page
5867 * 1 - 2 system pages
5868 * 3 - 4 system pages
5869 * ...
5870 *
5871 * Returns 0 on success or < 0 in case of an error.
5872 */
5873int ring_buffer_subbuf_order_set(struct trace_buffer *buffer, int order)
5874{
5875 struct ring_buffer_per_cpu *cpu_buffer;
5876 struct buffer_page *bpage, *tmp;
5877 int old_order, old_size;
5878 int nr_pages;
5879 int psize;
5880 int err;
5881 int cpu;
5882
5883 if (!buffer || order < 0)
5884 return -EINVAL;
5885
5886 if (buffer->subbuf_order == order)
5887 return 0;
5888
5889 psize = (1 << order) * PAGE_SIZE;
5890 if (psize <= BUF_PAGE_HDR_SIZE)
5891 return -EINVAL;
5892
5893 /* Size of a subbuf cannot be greater than the write counter */
5894 if (psize > RB_WRITE_MASK + 1)
5895 return -EINVAL;
5896
5897 old_order = buffer->subbuf_order;
5898 old_size = buffer->subbuf_size;
5899
5900 /* prevent another thread from changing buffer sizes */
5901 mutex_lock(&buffer->mutex);
5902 atomic_inc(&buffer->record_disabled);
5903
5904 /* Make sure all commits have finished */
5905 synchronize_rcu();
5906
5907 buffer->subbuf_order = order;
5908 buffer->subbuf_size = psize - BUF_PAGE_HDR_SIZE;
5909
5910 /* Make sure all new buffers are allocated, before deleting the old ones */
5911 for_each_buffer_cpu(buffer, cpu) {
5912
5913 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5914 continue;
5915
5916 cpu_buffer = buffer->buffers[cpu];
5917
5918 /* Update the number of pages to match the new size */
5919 nr_pages = old_size * buffer->buffers[cpu]->nr_pages;
5920 nr_pages = DIV_ROUND_UP(nr_pages, buffer->subbuf_size);
5921
5922 /* we need a minimum of two pages */
5923 if (nr_pages < 2)
5924 nr_pages = 2;
5925
5926 cpu_buffer->nr_pages_to_update = nr_pages;
5927
5928 /* Include the reader page */
5929 nr_pages++;
5930
5931 /* Allocate the new size buffer */
5932 INIT_LIST_HEAD(&cpu_buffer->new_pages);
5933 if (__rb_allocate_pages(cpu_buffer, nr_pages,
5934 &cpu_buffer->new_pages)) {
5935 /* not enough memory for new pages */
5936 err = -ENOMEM;
5937 goto error;
5938 }
5939 }
5940
5941 for_each_buffer_cpu(buffer, cpu) {
5942
5943 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5944 continue;
5945
5946 cpu_buffer = buffer->buffers[cpu];
5947
5948 /* Clear the head bit to make the link list normal to read */
5949 rb_head_page_deactivate(cpu_buffer);
5950
5951 /* Now walk the list and free all the old sub buffers */
5952 list_for_each_entry_safe(bpage, tmp, cpu_buffer->pages, list) {
5953 list_del_init(&bpage->list);
5954 free_buffer_page(bpage);
5955 }
5956 /* The above loop stopped an the last page needing to be freed */
5957 bpage = list_entry(cpu_buffer->pages, struct buffer_page, list);
5958 free_buffer_page(bpage);
5959
5960 /* Free the current reader page */
5961 free_buffer_page(cpu_buffer->reader_page);
5962
5963 /* One page was allocated for the reader page */
5964 cpu_buffer->reader_page = list_entry(cpu_buffer->new_pages.next,
5965 struct buffer_page, list);
5966 list_del_init(&cpu_buffer->reader_page->list);
5967
5968 /* The cpu_buffer pages are a link list with no head */
5969 cpu_buffer->pages = cpu_buffer->new_pages.next;
5970 cpu_buffer->new_pages.next->prev = cpu_buffer->new_pages.prev;
5971 cpu_buffer->new_pages.prev->next = cpu_buffer->new_pages.next;
5972
5973 /* Clear the new_pages list */
5974 INIT_LIST_HEAD(&cpu_buffer->new_pages);
5975
5976 cpu_buffer->head_page
5977 = list_entry(cpu_buffer->pages, struct buffer_page, list);
5978 cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
5979
5980 cpu_buffer->nr_pages = cpu_buffer->nr_pages_to_update;
5981 cpu_buffer->nr_pages_to_update = 0;
5982
5983 free_pages((unsigned long)cpu_buffer->free_page, old_order);
5984 cpu_buffer->free_page = NULL;
5985
5986 rb_head_page_activate(cpu_buffer);
5987
5988 rb_check_pages(cpu_buffer);
5989 }
5990
5991 atomic_dec(&buffer->record_disabled);
5992 mutex_unlock(&buffer->mutex);
5993
5994 return 0;
5995
5996error:
5997 buffer->subbuf_order = old_order;
5998 buffer->subbuf_size = old_size;
5999
6000 atomic_dec(&buffer->record_disabled);
6001 mutex_unlock(&buffer->mutex);
6002
6003 for_each_buffer_cpu(buffer, cpu) {
6004 cpu_buffer = buffer->buffers[cpu];
6005
6006 if (!cpu_buffer->nr_pages_to_update)
6007 continue;
6008
6009 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages, list) {
6010 list_del_init(&bpage->list);
6011 free_buffer_page(bpage);
6012 }
6013 }
6014
6015 return err;
6016}
6017EXPORT_SYMBOL_GPL(ring_buffer_subbuf_order_set);
6018
6019/*
6020 * We only allocate new buffers, never free them if the CPU goes down.
6021 * If we were to free the buffer, then the user would lose any trace that was in
6022 * the buffer.
6023 */
6024int trace_rb_cpu_prepare(unsigned int cpu, struct hlist_node *node)
6025{
6026 struct trace_buffer *buffer;
6027 long nr_pages_same;
6028 int cpu_i;
6029 unsigned long nr_pages;
6030
6031 buffer = container_of(node, struct trace_buffer, node);
6032 if (cpumask_test_cpu(cpu, buffer->cpumask))
6033 return 0;
6034
6035 nr_pages = 0;
6036 nr_pages_same = 1;
6037 /* check if all cpu sizes are same */
6038 for_each_buffer_cpu(buffer, cpu_i) {
6039 /* fill in the size from first enabled cpu */
6040 if (nr_pages == 0)
6041 nr_pages = buffer->buffers[cpu_i]->nr_pages;
6042 if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
6043 nr_pages_same = 0;
6044 break;
6045 }
6046 }
6047 /* allocate minimum pages, user can later expand it */
6048 if (!nr_pages_same)
6049 nr_pages = 2;
6050 buffer->buffers[cpu] =
6051 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
6052 if (!buffer->buffers[cpu]) {
6053 WARN(1, "failed to allocate ring buffer on CPU %u\n",
6054 cpu);
6055 return -ENOMEM;
6056 }
6057 smp_wmb();
6058 cpumask_set_cpu(cpu, buffer->cpumask);
6059 return 0;
6060}
6061
6062#ifdef CONFIG_RING_BUFFER_STARTUP_TEST
6063/*
6064 * This is a basic integrity check of the ring buffer.
6065 * Late in the boot cycle this test will run when configured in.
6066 * It will kick off a thread per CPU that will go into a loop
6067 * writing to the per cpu ring buffer various sizes of data.
6068 * Some of the data will be large items, some small.
6069 *
6070 * Another thread is created that goes into a spin, sending out
6071 * IPIs to the other CPUs to also write into the ring buffer.
6072 * this is to test the nesting ability of the buffer.
6073 *
6074 * Basic stats are recorded and reported. If something in the
6075 * ring buffer should happen that's not expected, a big warning
6076 * is displayed and all ring buffers are disabled.
6077 */
6078static struct task_struct *rb_threads[NR_CPUS] __initdata;
6079
6080struct rb_test_data {
6081 struct trace_buffer *buffer;
6082 unsigned long events;
6083 unsigned long bytes_written;
6084 unsigned long bytes_alloc;
6085 unsigned long bytes_dropped;
6086 unsigned long events_nested;
6087 unsigned long bytes_written_nested;
6088 unsigned long bytes_alloc_nested;
6089 unsigned long bytes_dropped_nested;
6090 int min_size_nested;
6091 int max_size_nested;
6092 int max_size;
6093 int min_size;
6094 int cpu;
6095 int cnt;
6096};
6097
6098static struct rb_test_data rb_data[NR_CPUS] __initdata;
6099
6100/* 1 meg per cpu */
6101#define RB_TEST_BUFFER_SIZE 1048576
6102
6103static char rb_string[] __initdata =
6104 "abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
6105 "?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
6106 "!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
6107
6108static bool rb_test_started __initdata;
6109
6110struct rb_item {
6111 int size;
6112 char str[];
6113};
6114
6115static __init int rb_write_something(struct rb_test_data *data, bool nested)
6116{
6117 struct ring_buffer_event *event;
6118 struct rb_item *item;
6119 bool started;
6120 int event_len;
6121 int size;
6122 int len;
6123 int cnt;
6124
6125 /* Have nested writes different that what is written */
6126 cnt = data->cnt + (nested ? 27 : 0);
6127
6128 /* Multiply cnt by ~e, to make some unique increment */
6129 size = (cnt * 68 / 25) % (sizeof(rb_string) - 1);
6130
6131 len = size + sizeof(struct rb_item);
6132
6133 started = rb_test_started;
6134 /* read rb_test_started before checking buffer enabled */
6135 smp_rmb();
6136
6137 event = ring_buffer_lock_reserve(data->buffer, len);
6138 if (!event) {
6139 /* Ignore dropped events before test starts. */
6140 if (started) {
6141 if (nested)
6142 data->bytes_dropped += len;
6143 else
6144 data->bytes_dropped_nested += len;
6145 }
6146 return len;
6147 }
6148
6149 event_len = ring_buffer_event_length(event);
6150
6151 if (RB_WARN_ON(data->buffer, event_len < len))
6152 goto out;
6153
6154 item = ring_buffer_event_data(event);
6155 item->size = size;
6156 memcpy(item->str, rb_string, size);
6157
6158 if (nested) {
6159 data->bytes_alloc_nested += event_len;
6160 data->bytes_written_nested += len;
6161 data->events_nested++;
6162 if (!data->min_size_nested || len < data->min_size_nested)
6163 data->min_size_nested = len;
6164 if (len > data->max_size_nested)
6165 data->max_size_nested = len;
6166 } else {
6167 data->bytes_alloc += event_len;
6168 data->bytes_written += len;
6169 data->events++;
6170 if (!data->min_size || len < data->min_size)
6171 data->max_size = len;
6172 if (len > data->max_size)
6173 data->max_size = len;
6174 }
6175
6176 out:
6177 ring_buffer_unlock_commit(data->buffer);
6178
6179 return 0;
6180}
6181
6182static __init int rb_test(void *arg)
6183{
6184 struct rb_test_data *data = arg;
6185
6186 while (!kthread_should_stop()) {
6187 rb_write_something(data, false);
6188 data->cnt++;
6189
6190 set_current_state(TASK_INTERRUPTIBLE);
6191 /* Now sleep between a min of 100-300us and a max of 1ms */
6192 usleep_range(((data->cnt % 3) + 1) * 100, 1000);
6193 }
6194
6195 return 0;
6196}
6197
6198static __init void rb_ipi(void *ignore)
6199{
6200 struct rb_test_data *data;
6201 int cpu = smp_processor_id();
6202
6203 data = &rb_data[cpu];
6204 rb_write_something(data, true);
6205}
6206
6207static __init int rb_hammer_test(void *arg)
6208{
6209 while (!kthread_should_stop()) {
6210
6211 /* Send an IPI to all cpus to write data! */
6212 smp_call_function(rb_ipi, NULL, 1);
6213 /* No sleep, but for non preempt, let others run */
6214 schedule();
6215 }
6216
6217 return 0;
6218}
6219
6220static __init int test_ringbuffer(void)
6221{
6222 struct task_struct *rb_hammer;
6223 struct trace_buffer *buffer;
6224 int cpu;
6225 int ret = 0;
6226
6227 if (security_locked_down(LOCKDOWN_TRACEFS)) {
6228 pr_warn("Lockdown is enabled, skipping ring buffer tests\n");
6229 return 0;
6230 }
6231
6232 pr_info("Running ring buffer tests...\n");
6233
6234 buffer = ring_buffer_alloc(RB_TEST_BUFFER_SIZE, RB_FL_OVERWRITE);
6235 if (WARN_ON(!buffer))
6236 return 0;
6237
6238 /* Disable buffer so that threads can't write to it yet */
6239 ring_buffer_record_off(buffer);
6240
6241 for_each_online_cpu(cpu) {
6242 rb_data[cpu].buffer = buffer;
6243 rb_data[cpu].cpu = cpu;
6244 rb_data[cpu].cnt = cpu;
6245 rb_threads[cpu] = kthread_run_on_cpu(rb_test, &rb_data[cpu],
6246 cpu, "rbtester/%u");
6247 if (WARN_ON(IS_ERR(rb_threads[cpu]))) {
6248 pr_cont("FAILED\n");
6249 ret = PTR_ERR(rb_threads[cpu]);
6250 goto out_free;
6251 }
6252 }
6253
6254 /* Now create the rb hammer! */
6255 rb_hammer = kthread_run(rb_hammer_test, NULL, "rbhammer");
6256 if (WARN_ON(IS_ERR(rb_hammer))) {
6257 pr_cont("FAILED\n");
6258 ret = PTR_ERR(rb_hammer);
6259 goto out_free;
6260 }
6261
6262 ring_buffer_record_on(buffer);
6263 /*
6264 * Show buffer is enabled before setting rb_test_started.
6265 * Yes there's a small race window where events could be
6266 * dropped and the thread wont catch it. But when a ring
6267 * buffer gets enabled, there will always be some kind of
6268 * delay before other CPUs see it. Thus, we don't care about
6269 * those dropped events. We care about events dropped after
6270 * the threads see that the buffer is active.
6271 */
6272 smp_wmb();
6273 rb_test_started = true;
6274
6275 set_current_state(TASK_INTERRUPTIBLE);
6276 /* Just run for 10 seconds */;
6277 schedule_timeout(10 * HZ);
6278
6279 kthread_stop(rb_hammer);
6280
6281 out_free:
6282 for_each_online_cpu(cpu) {
6283 if (!rb_threads[cpu])
6284 break;
6285 kthread_stop(rb_threads[cpu]);
6286 }
6287 if (ret) {
6288 ring_buffer_free(buffer);
6289 return ret;
6290 }
6291
6292 /* Report! */
6293 pr_info("finished\n");
6294 for_each_online_cpu(cpu) {
6295 struct ring_buffer_event *event;
6296 struct rb_test_data *data = &rb_data[cpu];
6297 struct rb_item *item;
6298 unsigned long total_events;
6299 unsigned long total_dropped;
6300 unsigned long total_written;
6301 unsigned long total_alloc;
6302 unsigned long total_read = 0;
6303 unsigned long total_size = 0;
6304 unsigned long total_len = 0;
6305 unsigned long total_lost = 0;
6306 unsigned long lost;
6307 int big_event_size;
6308 int small_event_size;
6309
6310 ret = -1;
6311
6312 total_events = data->events + data->events_nested;
6313 total_written = data->bytes_written + data->bytes_written_nested;
6314 total_alloc = data->bytes_alloc + data->bytes_alloc_nested;
6315 total_dropped = data->bytes_dropped + data->bytes_dropped_nested;
6316
6317 big_event_size = data->max_size + data->max_size_nested;
6318 small_event_size = data->min_size + data->min_size_nested;
6319
6320 pr_info("CPU %d:\n", cpu);
6321 pr_info(" events: %ld\n", total_events);
6322 pr_info(" dropped bytes: %ld\n", total_dropped);
6323 pr_info(" alloced bytes: %ld\n", total_alloc);
6324 pr_info(" written bytes: %ld\n", total_written);
6325 pr_info(" biggest event: %d\n", big_event_size);
6326 pr_info(" smallest event: %d\n", small_event_size);
6327
6328 if (RB_WARN_ON(buffer, total_dropped))
6329 break;
6330
6331 ret = 0;
6332
6333 while ((event = ring_buffer_consume(buffer, cpu, NULL, &lost))) {
6334 total_lost += lost;
6335 item = ring_buffer_event_data(event);
6336 total_len += ring_buffer_event_length(event);
6337 total_size += item->size + sizeof(struct rb_item);
6338 if (memcmp(&item->str[0], rb_string, item->size) != 0) {
6339 pr_info("FAILED!\n");
6340 pr_info("buffer had: %.*s\n", item->size, item->str);
6341 pr_info("expected: %.*s\n", item->size, rb_string);
6342 RB_WARN_ON(buffer, 1);
6343 ret = -1;
6344 break;
6345 }
6346 total_read++;
6347 }
6348 if (ret)
6349 break;
6350
6351 ret = -1;
6352
6353 pr_info(" read events: %ld\n", total_read);
6354 pr_info(" lost events: %ld\n", total_lost);
6355 pr_info(" total events: %ld\n", total_lost + total_read);
6356 pr_info(" recorded len bytes: %ld\n", total_len);
6357 pr_info(" recorded size bytes: %ld\n", total_size);
6358 if (total_lost) {
6359 pr_info(" With dropped events, record len and size may not match\n"
6360 " alloced and written from above\n");
6361 } else {
6362 if (RB_WARN_ON(buffer, total_len != total_alloc ||
6363 total_size != total_written))
6364 break;
6365 }
6366 if (RB_WARN_ON(buffer, total_lost + total_read != total_events))
6367 break;
6368
6369 ret = 0;
6370 }
6371 if (!ret)
6372 pr_info("Ring buffer PASSED!\n");
6373
6374 ring_buffer_free(buffer);
6375 return 0;
6376}
6377
6378late_initcall(test_ringbuffer);
6379#endif /* CONFIG_RING_BUFFER_STARTUP_TEST */