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