projects/linux·blob master · refs

8444 lines · 226.8 KiB

1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Generic ring buffer
4 *
5 * Copyright (C) 2008 Steven Rostedt <[email protected]>
6 */
7#include <linux/ring_buffer_types.h>
8#include <linux/sched/isolation.h>
9#include <linux/trace_recursion.h>
10#include <linux/panic_notifier.h>
11#include <linux/trace_events.h>
12#include <linux/ring_buffer.h>
13#include <linux/trace_clock.h>
14#include <linux/sched/clock.h>
15#include <linux/cacheflush.h>
16#include <linux/trace_seq.h>
17#include <linux/spinlock.h>
18#include <linux/irq_work.h>
19#include <linux/security.h>
20#include <linux/uaccess.h>
21#include <linux/hardirq.h>
22#include <linux/kthread.h> /* for self test */
23#include <linux/module.h>
24#include <linux/percpu.h>
25#include <linux/mutex.h>
26#include <linux/delay.h>
27#include <linux/slab.h>
28#include <linux/init.h>
29#include <linux/hash.h>
30#include <linux/list.h>
31#include <linux/cpu.h>
32#include <linux/oom.h>
33#include <linux/mm.h>
34
35#include <asm/ring_buffer.h>
36#include <asm/local64.h>
37#include <asm/local.h>
38#include <asm/setup.h>
39
40#include "trace.h"
41
42/*
43 * The "absolute" timestamp in the buffer is only 59 bits.
44 * If a clock has the 5 MSBs set, it needs to be saved and
45 * reinserted.
46 */
47#define TS_MSB (0xf8ULL << 56)
48#define ABS_TS_MASK (~TS_MSB)
49
50static void update_pages_handler(struct work_struct *work);
51
52#define RING_BUFFER_META_MAGIC 0xBADFEED
53
54struct ring_buffer_meta {
55 int magic;
56 int struct_sizes;
57 unsigned long total_size;
58 unsigned long buffers_offset;
59};
60
61struct ring_buffer_cpu_meta {
62 unsigned long first_buffer;
63 unsigned long head_buffer;
64 unsigned long commit_buffer;
65 __u32 subbuf_size;
66 __u32 nr_subbufs;
67#ifdef CONFIG_RING_BUFFER_PERSISTENT_INJECT
68 __u32 nr_invalid;
69 __u32 entry_bytes;
70#endif
71 int buffers[];
72};
73
74/*
75 * The ring buffer header is special. We must manually up keep it.
76 */
77int ring_buffer_print_entry_header(struct trace_seq *s)
78{
79 trace_seq_puts(s, "# compressed entry header\n");
80 trace_seq_puts(s, "\ttype_len : 5 bits\n");
81 trace_seq_puts(s, "\ttime_delta : 27 bits\n");
82 trace_seq_puts(s, "\tarray : 32 bits\n");
83 trace_seq_putc(s, '\n');
84 trace_seq_printf(s, "\tpadding : type == %d\n",
85 RINGBUF_TYPE_PADDING);
86 trace_seq_printf(s, "\ttime_extend : type == %d\n",
87 RINGBUF_TYPE_TIME_EXTEND);
88 trace_seq_printf(s, "\ttime_stamp : type == %d\n",
89 RINGBUF_TYPE_TIME_STAMP);
90 trace_seq_printf(s, "\tdata max type_len == %d\n",
91 RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
92
93 return !trace_seq_has_overflowed(s);
94}
95
96/*
97 * The ring buffer is made up of a list of pages. A separate list of pages is
98 * allocated for each CPU. A writer may only write to a buffer that is
99 * associated with the CPU it is currently executing on. A reader may read
100 * from any per cpu buffer.
101 *
102 * The reader is special. For each per cpu buffer, the reader has its own
103 * reader page. When a reader has read the entire reader page, this reader
104 * page is swapped with another page in the ring buffer.
105 *
106 * Now, as long as the writer is off the reader page, the reader can do what
107 * ever it wants with that page. The writer will never write to that page
108 * again (as long as it is out of the ring buffer).
109 *
110 * Here's some silly ASCII art.
111 *
112 * +------+
113 * |reader| RING BUFFER
114 * |page |
115 * +------+ +---+ +---+ +---+
116 * | |-->| |-->| |
117 * +---+ +---+ +---+
118 * ^ |
119 * | |
120 * +---------------+
121 *
122 *
123 * +------+
124 * |reader| RING BUFFER
125 * |page |------------------v
126 * +------+ +---+ +---+ +---+
127 * | |-->| |-->| |
128 * +---+ +---+ +---+
129 * ^ |
130 * | |
131 * +---------------+
132 *
133 *
134 * +------+
135 * |reader| RING BUFFER
136 * |page |------------------v
137 * +------+ +---+ +---+ +---+
138 * ^ | |-->| |-->| |
139 * | +---+ +---+ +---+
140 * | |
141 * | |
142 * +------------------------------+
143 *
144 *
145 * +------+
146 * |buffer| RING BUFFER
147 * |page |------------------v
148 * +------+ +---+ +---+ +---+
149 * ^ | | | |-->| |
150 * | New +---+ +---+ +---+
151 * | Reader------^ |
152 * | page |
153 * +------------------------------+
154 *
155 *
156 * After we make this swap, the reader can hand this page off to the splice
157 * code and be done with it. It can even allocate a new page if it needs to
158 * and swap that into the ring buffer.
159 *
160 * We will be using cmpxchg soon to make all this lockless.
161 *
162 */
163
164/* Used for individual buffers (after the counter) */
165#define RB_BUFFER_OFF (1 << 20)
166
167/* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
168#define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
169
170enum {
171 RB_LEN_TIME_EXTEND = 8,
172 RB_LEN_TIME_STAMP = 8,
173};
174
175#define skip_time_extend(event) \
176 ((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
177
178#define extended_time(event) \
179 (event->type_len >= RINGBUF_TYPE_TIME_EXTEND)
180
181static inline bool rb_null_event(struct ring_buffer_event *event)
182{
183 return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
184}
185
186static void rb_event_set_padding(struct ring_buffer_event *event)
187{
188 /* padding has a NULL time_delta */
189 event->type_len = RINGBUF_TYPE_PADDING;
190 event->time_delta = 0;
191}
192
193static unsigned
194rb_event_data_length(struct ring_buffer_event *event)
195{
196 unsigned length;
197
198 if (event->type_len)
199 length = event->type_len * RB_ALIGNMENT;
200 else
201 length = event->array[0];
202 return length + RB_EVNT_HDR_SIZE;
203}
204
205/*
206 * Return the length of the given event. Will return
207 * the length of the time extend if the event is a
208 * time extend.
209 */
210static inline unsigned
211rb_event_length(struct ring_buffer_event *event)
212{
213 switch (event->type_len) {
214 case RINGBUF_TYPE_PADDING:
215 if (rb_null_event(event))
216 /* undefined */
217 return -1;
218 return event->array[0] + RB_EVNT_HDR_SIZE;
219
220 case RINGBUF_TYPE_TIME_EXTEND:
221 return RB_LEN_TIME_EXTEND;
222
223 case RINGBUF_TYPE_TIME_STAMP:
224 return RB_LEN_TIME_STAMP;
225
226 case RINGBUF_TYPE_DATA:
227 return rb_event_data_length(event);
228 default:
229 WARN_ON_ONCE(1);
230 }
231 /* not hit */
232 return 0;
233}
234
235/*
236 * Return total length of time extend and data,
237 * or just the event length for all other events.
238 */
239static inline unsigned
240rb_event_ts_length(struct ring_buffer_event *event)
241{
242 unsigned len = 0;
243
244 if (extended_time(event)) {
245 /* time extends include the data event after it */
246 len = RB_LEN_TIME_EXTEND;
247 event = skip_time_extend(event);
248 }
249 return len + rb_event_length(event);
250}
251
252/**
253 * ring_buffer_event_length - return the length of the event
254 * @event: the event to get the length of
255 *
256 * Returns the size of the data load of a data event.
257 * If the event is something other than a data event, it
258 * returns the size of the event itself. With the exception
259 * of a TIME EXTEND, where it still returns the size of the
260 * data load of the data event after it.
261 */
262unsigned ring_buffer_event_length(struct ring_buffer_event *event)
263{
264 unsigned length;
265
266 if (extended_time(event))
267 event = skip_time_extend(event);
268
269 length = rb_event_length(event);
270 if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
271 return length;
272 length -= RB_EVNT_HDR_SIZE;
273 if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]) ||
274 RB_FORCE_8BYTE_ALIGNMENT)
275 length -= sizeof(event->array[0]);
276 return length;
277}
278EXPORT_SYMBOL_GPL(ring_buffer_event_length);
279
280/* inline for ring buffer fast paths */
281static __always_inline void *
282rb_event_data(struct ring_buffer_event *event)
283{
284 if (extended_time(event))
285 event = skip_time_extend(event);
286 WARN_ON_ONCE(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
287 /* If length is in len field, then array[0] has the data */
288 if (event->type_len)
289 return (void *)&event->array[0];
290 /* Otherwise length is in array[0] and array[1] has the data */
291 return (void *)&event->array[1];
292}
293
294/**
295 * ring_buffer_event_data - return the data of the event
296 * @event: the event to get the data from
297 */
298void *ring_buffer_event_data(struct ring_buffer_event *event)
299{
300 return rb_event_data(event);
301}
302EXPORT_SYMBOL_GPL(ring_buffer_event_data);
303
304#define for_each_buffer_cpu(buffer, cpu) \
305 for_each_cpu(cpu, buffer->cpumask)
306
307#define for_each_online_buffer_cpu(buffer, cpu) \
308 for_each_cpu_and(cpu, buffer->cpumask, cpu_online_mask)
309
310static u64 rb_event_time_stamp(struct ring_buffer_event *event)
311{
312 u64 ts;
313
314 ts = event->array[0];
315 ts <<= TS_SHIFT;
316 ts += event->time_delta;
317
318 return ts;
319}
320
321/* Flag when events were overwritten */
322#define RB_MISSED_EVENTS (1 << 31)
323/* Missed count stored at end */
324#define RB_MISSED_STORED (1 << 30)
325
326#define RB_MISSED_MASK (3 << 30)
327
328struct buffer_data_read_page {
329 unsigned order; /* order of the page */
330 struct buffer_data_page *data; /* actual data, stored in this page */
331};
332
333static __always_inline unsigned int rb_read_page_capacity(struct buffer_data_read_page *rpage)
334{
335 return (PAGE_SIZE << rpage->order) - BUF_PAGE_HDR_SIZE;
336}
337
338/* The number of bits for static buffer ids */
339#define RB_STATIC_BITS 30
340
341/*
342 * Note, the buffer_page list must be first. The buffer pages
343 * are allocated in cache lines, which means that each buffer
344 * page will be at the beginning of a cache line, and thus
345 * the least significant bits will be zero. We use this to
346 * add flags in the list struct pointers, to make the ring buffer
347 * lockless.
348 */
349struct buffer_page {
350 struct list_head list; /* list of buffer pages */
351 local_t write; /* index for next write */
352 unsigned read; /* index for next read */
353 local_t entries; /* entries on this page */
354 unsigned long real_end; /* real end of data */
355 unsigned order; /* order of the page */
356 u32 id:RB_STATIC_BITS; /* ID for external mapping */
357 u32 range:1; /* Mapped via a range */
358 struct buffer_data_page *page; /* Actual data page */
359};
360
361/*
362 * The buffer page counters, write and entries, must be reset
363 * atomically when crossing page boundaries. To synchronize this
364 * update, two counters are inserted into the number. One is
365 * the actual counter for the write position or count on the page.
366 *
367 * The other is a counter of updaters. Before an update happens
368 * the update partition of the counter is incremented. This will
369 * allow the updater to update the counter atomically.
370 *
371 * The counter is 20 bits, and the state data is 12.
372 */
373#define RB_WRITE_MASK 0xfffff
374#define RB_WRITE_INTCNT (1 << 20)
375
376static void rb_init_data_page(struct buffer_data_page *bpage)
377{
378 local_set(&bpage->commit, 0);
379 bpage->time_stamp = 0;
380}
381
382static __always_inline long rb_data_page_commit(struct buffer_data_page *dpage)
383{
384 return local_read(&dpage->commit);
385}
386
387static __always_inline long rb_data_page_size(struct buffer_data_page *dpage)
388{
389 return rb_data_page_commit(dpage) & ~RB_MISSED_MASK;
390}
391
392static __always_inline unsigned int rb_page_commit(struct buffer_page *bpage)
393{
394 return rb_data_page_commit(bpage->page);
395}
396
397static __always_inline unsigned int rb_page_size(struct buffer_page *bpage)
398{
399 return rb_data_page_size(bpage->page);
400}
401
402/**
403 * rb_page_capacity - Get the capacity of a buffer page
404 * @bpage: The buffer page
405 *
406 * Return: The maximum size available for events in the given buffer page.
407 */
408static __always_inline unsigned int rb_page_capacity(struct buffer_page *bpage)
409{
410 return (PAGE_SIZE << bpage->order) - BUF_PAGE_HDR_SIZE;
411}
412
413static void free_buffer_page(struct buffer_page *bpage)
414{
415 /* Range pages are not to be freed */
416 if (!bpage->range)
417 free_pages((unsigned long)bpage->page, bpage->order);
418 kfree(bpage);
419}
420
421/*
422 * For best performance, allocate cpu buffer data cache line sized
423 * and per CPU.
424 */
425#define alloc_cpu_buffer(cpu) (struct ring_buffer_per_cpu *) \
426 kzalloc_node(ALIGN(sizeof(struct ring_buffer_per_cpu), \
427 cache_line_size()), GFP_KERNEL, cpu_to_node(cpu))
428
429#define alloc_cpu_page(cpu) (struct buffer_page *) \
430 kzalloc_node(ALIGN(sizeof(struct buffer_page), \
431 cache_line_size()), GFP_KERNEL, cpu_to_node(cpu))
432
433static struct buffer_data_page *alloc_cpu_data(int cpu, int order)
434{
435 struct buffer_data_page *dpage;
436 struct page *page;
437 gfp_t mflags;
438
439 /*
440 * __GFP_RETRY_MAYFAIL flag makes sure that the allocation fails
441 * gracefully without invoking oom-killer and the system is not
442 * destabilized.
443 */
444 mflags = GFP_KERNEL | __GFP_RETRY_MAYFAIL | __GFP_COMP | __GFP_ZERO;
445
446 page = alloc_pages_node(cpu_to_node(cpu), mflags, order);
447 if (!page)
448 return NULL;
449
450 dpage = page_address(page);
451 rb_init_data_page(dpage);
452
453 return dpage;
454}
455
456struct rb_irq_work {
457 struct irq_work work;
458 wait_queue_head_t waiters;
459 wait_queue_head_t full_waiters;
460 atomic_t seq;
461 bool waiters_pending;
462 bool full_waiters_pending;
463 bool wakeup_full;
464};
465
466/*
467 * Structure to hold event state and handle nested events.
468 */
469struct rb_event_info {
470 u64 ts;
471 u64 delta;
472 u64 before;
473 u64 after;
474 unsigned long length;
475 struct buffer_page *tail_page;
476 int add_timestamp;
477};
478
479/*
480 * Used for the add_timestamp
481 * NONE
482 * EXTEND - wants a time extend
483 * ABSOLUTE - the buffer requests all events to have absolute time stamps
484 * FORCE - force a full time stamp.
485 */
486enum {
487 RB_ADD_STAMP_NONE = 0,
488 RB_ADD_STAMP_EXTEND = BIT(1),
489 RB_ADD_STAMP_ABSOLUTE = BIT(2),
490 RB_ADD_STAMP_FORCE = BIT(3)
491};
492/*
493 * Used for which event context the event is in.
494 * TRANSITION = 0
495 * NMI = 1
496 * IRQ = 2
497 * SOFTIRQ = 3
498 * NORMAL = 4
499 *
500 * See trace_recursive_lock() comment below for more details.
501 */
502enum {
503 RB_CTX_TRANSITION,
504 RB_CTX_NMI,
505 RB_CTX_IRQ,
506 RB_CTX_SOFTIRQ,
507 RB_CTX_NORMAL,
508 RB_CTX_MAX
509};
510
511struct rb_time_struct {
512 local64_t time;
513};
514typedef struct rb_time_struct rb_time_t;
515
516#define MAX_NEST 5
517
518/*
519 * head_page == tail_page && head == tail then buffer is empty.
520 */
521struct ring_buffer_per_cpu {
522 int cpu;
523 atomic_t record_disabled;
524 atomic_t resize_disabled;
525 struct trace_buffer *buffer;
526 raw_spinlock_t reader_lock; /* serialize readers */
527 arch_spinlock_t lock;
528 struct lock_class_key lock_key;
529 struct buffer_data_read_page free_page;
530 unsigned long nr_pages;
531 unsigned int current_context;
532 struct list_head *pages;
533 /* pages generation counter, incremented when the list changes */
534 unsigned long cnt;
535 struct buffer_page *head_page; /* read from head */
536 struct buffer_page *tail_page; /* write to tail */
537 struct buffer_page *commit_page; /* committed pages */
538 struct buffer_page *reader_page;
539 unsigned long lost_events;
540 unsigned long last_overrun;
541 unsigned long nest;
542 local_t entries_bytes;
543 local_t entries;
544 local_t overrun;
545 local_t commit_overrun;
546 local_t dropped_events;
547 local_t committing;
548 local_t commits;
549 local_t pages_touched;
550 local_t pages_lost;
551 local_t pages_read;
552 long last_pages_touch;
553 size_t shortest_full;
554 unsigned long read;
555 unsigned long read_bytes;
556 rb_time_t write_stamp;
557 rb_time_t before_stamp;
558 u64 event_stamp[MAX_NEST];
559 u64 read_stamp;
560 /* pages removed since last reset */
561 unsigned long pages_removed;
562
563 unsigned int user_mapped; /* user space mapping */
564 struct mutex mapping_lock;
565 struct buffer_page **subbuf_ids; /* ID to subbuf VA */
566 struct trace_buffer_meta *meta_page;
567 struct ring_buffer_cpu_meta *ring_meta;
568
569 struct ring_buffer_remote *remote;
570
571 /* ring buffer pages to update, > 0 to add, < 0 to remove */
572 long nr_pages_to_update;
573 struct list_head new_pages; /* new pages to add */
574 struct work_struct update_pages_work;
575 struct completion update_done;
576
577 struct rb_irq_work irq_work;
578};
579
580struct trace_buffer {
581 unsigned flags;
582 atomic_t record_disabled;
583 atomic_t resizing;
584 cpumask_var_t cpumask;
585
586 struct lock_class_key *reader_lock_key;
587
588 struct mutex mutex;
589
590 struct ring_buffer_per_cpu **buffers;
591
592 struct ring_buffer_remote *remote;
593
594 struct hlist_node node;
595 u64 (*clock)(void);
596
597 struct rb_irq_work irq_work;
598 bool time_stamp_abs;
599
600 unsigned long range_addr_start;
601 unsigned long range_addr_end;
602 struct notifier_block flush_nb;
603
604 struct ring_buffer_meta *meta;
605
606 unsigned int subbuf_order;
607};
608
609static __always_inline unsigned int rb_subbuf_size(struct trace_buffer *buffer)
610{
611 return PAGE_SIZE << buffer->subbuf_order;
612}
613
614/**
615 * rb_subbuf_capacity - Get the capacity of a subbuffer
616 * @buffer: A trace buffer
617 *
618 * Unsafe to use without holding trace_buffer::mutex or with resizing enabled.
619 * Consider rb_page_capacity() instead.
620 *
621 * Return: The maximum size available for events in a trace buffer subbuffer.
622 */
623static __always_inline unsigned int rb_subbuf_capacity(struct trace_buffer *buffer)
624{
625 return rb_subbuf_size(buffer) - BUF_PAGE_HDR_SIZE;
626}
627
628/**
629 * rb_subbuf_max_data_size - Get the maximum payload size of a single event
630 * @buffer: A trace buffer
631 *
632 * Return: The maximum data payload size that can be stored in a single event.
633 */
634static __always_inline unsigned int rb_subbuf_max_data_size(struct trace_buffer *buffer)
635{
636 struct ring_buffer_event *event;
637
638 /*
639 * surely rb_subbuf_capacity() is bigger than
640 * RINGBUF_TYPE_DATA_TYPE_LEN_MAX (see ring_buffer_event_length).
641 */
642 return rb_subbuf_capacity(buffer) - RB_EVNT_HDR_SIZE - sizeof(event->array[0]);
643}
644
645/**
646 * rb_subbuf_start - Get the start address of a subbuffer
647 * @buffer: A trace buffer
648 * @addr: An address of an event on a subbuffer
649 *
650 * Return: The start of the subbuffer for where @addr sits
651 */
652static __always_inline
653unsigned long rb_subbuf_start(struct trace_buffer *buffer, unsigned long addr)
654{
655 return addr & ~((unsigned long)(rb_subbuf_size(buffer) - 1));
656}
657
658static bool rb_is_static(struct ring_buffer_per_cpu *cpu_buffer)
659{
660 return cpu_buffer->user_mapped || cpu_buffer->remote || cpu_buffer->ring_meta;
661}
662
663static unsigned long rb_static_max_pages(void)
664{
665 /*
666 * Static ring buffers are using bpage::id and must account for the
667 * reader page.
668 */
669 return (1UL << RB_STATIC_BITS) - 1;
670}
671
672struct ring_buffer_iter {
673 struct ring_buffer_per_cpu *cpu_buffer;
674 unsigned long head;
675 unsigned long next_event;
676 struct buffer_page *head_page;
677 struct buffer_page *cache_reader_page;
678 unsigned long cache_read;
679 unsigned long cache_pages_removed;
680 u64 read_stamp;
681 u64 page_stamp;
682 struct ring_buffer_event *event;
683 size_t event_size;
684 int missed_events;
685};
686
687int ring_buffer_print_page_header(struct trace_buffer *buffer, struct trace_seq *s)
688{
689 struct buffer_data_page field;
690
691 trace_seq_printf(s, "\tfield: u64 timestamp;\t"
692 "offset:0;\tsize:%u;\tsigned:%u;\n",
693 (unsigned int)sizeof(field.time_stamp),
694 (unsigned int)is_signed_type(u64));
695
696 trace_seq_printf(s, "\tfield: local_t commit;\t"
697 "offset:%u;\tsize:%u;\tsigned:%u;\n",
698 (unsigned int)offsetof(typeof(field), commit),
699 (unsigned int)sizeof(field.commit),
700 (unsigned int)is_signed_type(long));
701
702 trace_seq_printf(s, "\tfield: char overwrite;\t"
703 "offset:%u;\tsize:%u;\tsigned:%u;\n",
704 (unsigned int)offsetof(typeof(field), commit),
705 1,
706 (unsigned int)is_signed_type(char));
707
708 trace_seq_printf(s, "\tfield: char data;\t"
709 "offset:%u;\tsize:%u;\tsigned:%u;\n",
710 (unsigned int)offsetof(typeof(field), data),
711 (unsigned int)(buffer ? rb_subbuf_capacity(buffer) :
712 PAGE_SIZE - BUF_PAGE_HDR_SIZE),
713 (unsigned int)is_signed_type(char));
714
715 return !trace_seq_has_overflowed(s);
716}
717
718static inline void rb_time_read(rb_time_t *t, u64 *ret)
719{
720 *ret = local64_read(&t->time);
721}
722static void rb_time_set(rb_time_t *t, u64 val)
723{
724 local64_set(&t->time, val);
725}
726
727/*
728 * Enable this to make sure that the event passed to
729 * ring_buffer_event_time_stamp() is not committed and also
730 * is on the buffer that it passed in.
731 */
732//#define RB_VERIFY_EVENT
733#ifdef RB_VERIFY_EVENT
734static struct list_head *rb_list_head(struct list_head *list);
735static void verify_event(struct ring_buffer_per_cpu *cpu_buffer,
736 void *event)
737{
738 struct buffer_page *page = cpu_buffer->commit_page;
739 struct buffer_page *tail_page = READ_ONCE(cpu_buffer->tail_page);
740 struct list_head *next;
741 long commit, write;
742 unsigned long addr = (unsigned long)event;
743 bool done = false;
744 int stop = 0;
745
746 /* Make sure the event exists and is not committed yet */
747 do {
748 if (page == tail_page || WARN_ON_ONCE(stop++ > 100))
749 done = true;
750 commit = rb_page_commit(page);
751 write = local_read(&page->write);
752 if (addr >= (unsigned long)&page->page->data[commit] &&
753 addr < (unsigned long)&page->page->data[write])
754 return;
755
756 next = rb_list_head(page->list.next);
757 page = list_entry(next, struct buffer_page, list);
758 } while (!done);
759 WARN_ON_ONCE(1);
760}
761#else
762static inline void verify_event(struct ring_buffer_per_cpu *cpu_buffer,
763 void *event)
764{
765}
766#endif
767
768/*
769 * The absolute time stamp drops the 5 MSBs and some clocks may
770 * require them. The rb_fix_abs_ts() will take a previous full
771 * time stamp, and add the 5 MSB of that time stamp on to the
772 * saved absolute time stamp. Then they are compared in case of
773 * the unlikely event that the latest time stamp incremented
774 * the 5 MSB.
775 */
776static inline u64 rb_fix_abs_ts(u64 abs, u64 save_ts)
777{
778 if (save_ts & TS_MSB) {
779 abs |= save_ts & TS_MSB;
780 /* Check for overflow */
781 if (unlikely(abs < save_ts))
782 abs += 1ULL << 59;
783 }
784 return abs;
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 ts = rb_event_time_stamp(event);
816 return rb_fix_abs_ts(ts, cpu_buffer->tail_page->page->time_stamp);
817 }
818
819 nest = local_read(&cpu_buffer->committing);
820 verify_event(cpu_buffer, event);
821 if (WARN_ON_ONCE(!nest))
822 goto fail;
823
824 /* Read the current saved nesting level time stamp */
825 if (likely(--nest < MAX_NEST))
826 return cpu_buffer->event_stamp[nest];
827
828 /* Shouldn't happen, warn if it does */
829 WARN_ONCE(1, "nest (%d) greater than max", nest);
830
831 fail:
832 rb_time_read(&cpu_buffer->write_stamp, &ts);
833
834 return ts;
835}
836
837/**
838 * ring_buffer_nr_dirty_pages - get the number of used pages in the ring buffer
839 * @buffer: The ring_buffer to get the number of pages from
840 * @cpu: The cpu of the ring_buffer to get the number of pages from
841 *
842 * Returns the number of pages that have content in the ring buffer.
843 */
844size_t ring_buffer_nr_dirty_pages(struct trace_buffer *buffer, int cpu)
845{
846 size_t read;
847 size_t lost;
848 size_t cnt;
849
850 read = local_read(&buffer->buffers[cpu]->pages_read);
851 lost = local_read(&buffer->buffers[cpu]->pages_lost);
852 cnt = local_read(&buffer->buffers[cpu]->pages_touched);
853
854 if (WARN_ON_ONCE(cnt < lost))
855 return 0;
856
857 cnt -= lost;
858
859 /* The reader can read an empty page, but not more than that */
860 if (cnt < read) {
861 WARN_ON_ONCE(read > cnt + 1);
862 return 0;
863 }
864
865 return cnt - read;
866}
867
868static __always_inline bool full_hit(struct trace_buffer *buffer, int cpu, int full)
869{
870 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
871 size_t nr_pages;
872 size_t dirty;
873
874 nr_pages = cpu_buffer->nr_pages;
875 if (!nr_pages || !full)
876 return true;
877
878 /*
879 * Add one as dirty will never equal nr_pages, as the sub-buffer
880 * that the writer is on is not counted as dirty.
881 * This is needed if "buffer_percent" is set to 100.
882 */
883 dirty = ring_buffer_nr_dirty_pages(buffer, cpu) + 1;
884
885 return (dirty * 100) >= (full * nr_pages);
886}
887
888/*
889 * rb_wake_up_waiters - wake up tasks waiting for ring buffer input
890 *
891 * Schedules a delayed work to wake up any task that is blocked on the
892 * ring buffer waiters queue.
893 */
894static void rb_wake_up_waiters(struct irq_work *work)
895{
896 struct rb_irq_work *rbwork = container_of(work, struct rb_irq_work, work);
897
898 /* For waiters waiting for the first wake up */
899 (void)atomic_fetch_inc_release(&rbwork->seq);
900
901 wake_up_all(&rbwork->waiters);
902 if (rbwork->full_waiters_pending || rbwork->wakeup_full) {
903 /* Only cpu_buffer sets the above flags */
904 struct ring_buffer_per_cpu *cpu_buffer =
905 container_of(rbwork, struct ring_buffer_per_cpu, irq_work);
906
907 scoped_guard(raw_spinlock_irqsave, &cpu_buffer->reader_lock) {
908 rbwork->wakeup_full = false;
909 rbwork->full_waiters_pending = false;
910
911 /* Waking up all waiters, they will reset the shortest full */
912 cpu_buffer->shortest_full = 0;
913 }
914
915 wake_up_all(&rbwork->full_waiters);
916 }
917}
918
919/**
920 * ring_buffer_wake_waiters - wake up any waiters on this ring buffer
921 * @buffer: The ring buffer to wake waiters on
922 * @cpu: The CPU buffer to wake waiters on
923 *
924 * In the case of a file that represents a ring buffer is closing,
925 * it is prudent to wake up any waiters that are on this.
926 */
927void ring_buffer_wake_waiters(struct trace_buffer *buffer, int cpu)
928{
929 struct ring_buffer_per_cpu *cpu_buffer;
930 struct rb_irq_work *rbwork;
931
932 if (!buffer)
933 return;
934
935 if (cpu == RING_BUFFER_ALL_CPUS) {
936
937 /* Wake up individual ones too. One level recursion */
938 for_each_buffer_cpu(buffer, cpu)
939 ring_buffer_wake_waiters(buffer, cpu);
940
941 rbwork = &buffer->irq_work;
942 } else {
943 if (WARN_ON_ONCE(!buffer->buffers))
944 return;
945 if (WARN_ON_ONCE(cpu >= nr_cpu_ids))
946 return;
947
948 cpu_buffer = buffer->buffers[cpu];
949 /* The CPU buffer may not have been initialized yet */
950 if (!cpu_buffer)
951 return;
952 rbwork = &cpu_buffer->irq_work;
953 }
954
955 /* This can be called in any context */
956 irq_work_queue(&rbwork->work);
957}
958
959static bool rb_watermark_hit(struct trace_buffer *buffer, int cpu, int full)
960{
961 struct ring_buffer_per_cpu *cpu_buffer;
962 bool ret = false;
963
964 /* Reads of all CPUs always waits for any data */
965 if (cpu == RING_BUFFER_ALL_CPUS)
966 return !ring_buffer_empty(buffer);
967
968 cpu_buffer = buffer->buffers[cpu];
969
970 if (!ring_buffer_empty_cpu(buffer, cpu)) {
971 unsigned long flags;
972 bool pagebusy;
973
974 if (!full)
975 return true;
976
977 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
978 pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
979 ret = !pagebusy && full_hit(buffer, cpu, full);
980
981 if (!ret && (!cpu_buffer->shortest_full ||
982 cpu_buffer->shortest_full > full)) {
983 cpu_buffer->shortest_full = full;
984 }
985 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
986 }
987 return ret;
988}
989
990static inline bool
991rb_wait_cond(struct rb_irq_work *rbwork, struct trace_buffer *buffer,
992 int cpu, int full, ring_buffer_cond_fn cond, void *data)
993{
994 if (rb_watermark_hit(buffer, cpu, full))
995 return true;
996
997 if (cond(data))
998 return true;
999
1000 /*
1001 * The events can happen in critical sections where
1002 * checking a work queue can cause deadlocks.
1003 * After adding a task to the queue, this flag is set
1004 * only to notify events to try to wake up the queue
1005 * using irq_work.
1006 *
1007 * We don't clear it even if the buffer is no longer
1008 * empty. The flag only causes the next event to run
1009 * irq_work to do the work queue wake up. The worse
1010 * that can happen if we race with !trace_empty() is that
1011 * an event will cause an irq_work to try to wake up
1012 * an empty queue.
1013 *
1014 * There's no reason to protect this flag either, as
1015 * the work queue and irq_work logic will do the necessary
1016 * synchronization for the wake ups. The only thing
1017 * that is necessary is that the wake up happens after
1018 * a task has been queued. It's OK for spurious wake ups.
1019 */
1020 if (full)
1021 rbwork->full_waiters_pending = true;
1022 else
1023 rbwork->waiters_pending = true;
1024
1025 return false;
1026}
1027
1028struct rb_wait_data {
1029 struct rb_irq_work *irq_work;
1030 int seq;
1031};
1032
1033/*
1034 * The default wait condition for ring_buffer_wait() is to just to exit the
1035 * wait loop the first time it is woken up.
1036 */
1037static bool rb_wait_once(void *data)
1038{
1039 struct rb_wait_data *rdata = data;
1040 struct rb_irq_work *rbwork = rdata->irq_work;
1041
1042 return atomic_read_acquire(&rbwork->seq) != rdata->seq;
1043}
1044
1045/**
1046 * ring_buffer_wait - wait for input to the ring buffer
1047 * @buffer: buffer to wait on
1048 * @cpu: the cpu buffer to wait on
1049 * @full: wait until the percentage of pages are available, if @cpu != RING_BUFFER_ALL_CPUS
1050 * @cond: condition function to break out of wait (NULL to run once)
1051 * @data: the data to pass to @cond.
1052 *
1053 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
1054 * as data is added to any of the @buffer's cpu buffers. Otherwise
1055 * it will wait for data to be added to a specific cpu buffer.
1056 */
1057int ring_buffer_wait(struct trace_buffer *buffer, int cpu, int full,
1058 ring_buffer_cond_fn cond, void *data)
1059{
1060 struct ring_buffer_per_cpu *cpu_buffer;
1061 struct wait_queue_head *waitq;
1062 struct rb_irq_work *rbwork;
1063 struct rb_wait_data rdata;
1064 int ret = 0;
1065
1066 /*
1067 * Depending on what the caller is waiting for, either any
1068 * data in any cpu buffer, or a specific buffer, put the
1069 * caller on the appropriate wait queue.
1070 */
1071 if (cpu == RING_BUFFER_ALL_CPUS) {
1072 rbwork = &buffer->irq_work;
1073 /* Full only makes sense on per cpu reads */
1074 full = 0;
1075 } else {
1076 if (!cpumask_test_cpu(cpu, buffer->cpumask))
1077 return -ENODEV;
1078 cpu_buffer = buffer->buffers[cpu];
1079 rbwork = &cpu_buffer->irq_work;
1080 }
1081
1082 if (full)
1083 waitq = &rbwork->full_waiters;
1084 else
1085 waitq = &rbwork->waiters;
1086
1087 /* Set up to exit loop as soon as it is woken */
1088 if (!cond) {
1089 cond = rb_wait_once;
1090 rdata.irq_work = rbwork;
1091 rdata.seq = atomic_read_acquire(&rbwork->seq);
1092 data = &rdata;
1093 }
1094
1095 ret = wait_event_interruptible((*waitq),
1096 rb_wait_cond(rbwork, buffer, cpu, full, cond, data));
1097
1098 return ret;
1099}
1100
1101/**
1102 * ring_buffer_poll_wait - poll on buffer input
1103 * @buffer: buffer to wait on
1104 * @cpu: the cpu buffer to wait on
1105 * @filp: the file descriptor
1106 * @poll_table: The poll descriptor
1107 * @full: wait until the percentage of pages are available, if @cpu != RING_BUFFER_ALL_CPUS
1108 *
1109 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
1110 * as data is added to any of the @buffer's cpu buffers. Otherwise
1111 * it will wait for data to be added to a specific cpu buffer.
1112 *
1113 * Returns EPOLLIN | EPOLLRDNORM if data exists in the buffers,
1114 * zero otherwise.
1115 */
1116__poll_t ring_buffer_poll_wait(struct trace_buffer *buffer, int cpu,
1117 struct file *filp, poll_table *poll_table, int full)
1118{
1119 struct ring_buffer_per_cpu *cpu_buffer;
1120 struct rb_irq_work *rbwork;
1121
1122 if (cpu == RING_BUFFER_ALL_CPUS) {
1123 rbwork = &buffer->irq_work;
1124 full = 0;
1125 } else {
1126 if (!cpumask_test_cpu(cpu, buffer->cpumask))
1127 return EPOLLERR;
1128
1129 cpu_buffer = buffer->buffers[cpu];
1130 rbwork = &cpu_buffer->irq_work;
1131 }
1132
1133 if (full) {
1134 poll_wait(filp, &rbwork->full_waiters, poll_table);
1135
1136 if (rb_watermark_hit(buffer, cpu, full))
1137 return EPOLLIN | EPOLLRDNORM;
1138 /*
1139 * Only allow full_waiters_pending update to be seen after
1140 * the shortest_full is set (in rb_watermark_hit). If the
1141 * writer sees the full_waiters_pending flag set, it will
1142 * compare the amount in the ring buffer to shortest_full.
1143 * If the amount in the ring buffer is greater than the
1144 * shortest_full percent, it will call the irq_work handler
1145 * to wake up this list. The irq_handler will reset shortest_full
1146 * back to zero. That's done under the reader_lock, but
1147 * the below smp_mb() makes sure that the update to
1148 * full_waiters_pending doesn't leak up into the above.
1149 */
1150 smp_mb();
1151 rbwork->full_waiters_pending = true;
1152 return 0;
1153 }
1154
1155 poll_wait(filp, &rbwork->waiters, poll_table);
1156 rbwork->waiters_pending = true;
1157
1158 /*
1159 * There's a tight race between setting the waiters_pending and
1160 * checking if the ring buffer is empty. Once the waiters_pending bit
1161 * is set, the next event will wake the task up, but we can get stuck
1162 * if there's only a single event in.
1163 *
1164 * FIXME: Ideally, we need a memory barrier on the writer side as well,
1165 * but adding a memory barrier to all events will cause too much of a
1166 * performance hit in the fast path. We only need a memory barrier when
1167 * the buffer goes from empty to having content. But as this race is
1168 * extremely small, and it's not a problem if another event comes in, we
1169 * will fix it later.
1170 */
1171 smp_mb();
1172
1173 if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
1174 (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
1175 return EPOLLIN | EPOLLRDNORM;
1176 return 0;
1177}
1178
1179/* buffer may be either ring_buffer or ring_buffer_per_cpu */
1180#define RB_WARN_ON(b, cond) \
1181 ({ \
1182 int _____ret = unlikely(cond); \
1183 if (_____ret) { \
1184 if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
1185 struct ring_buffer_per_cpu *__b = \
1186 (void *)b; \
1187 atomic_inc(&__b->buffer->record_disabled); \
1188 } else \
1189 atomic_inc(&b->record_disabled); \
1190 WARN_ON(1); \
1191 } \
1192 _____ret; \
1193 })
1194
1195/* Up this if you want to test the TIME_EXTENTS and normalization */
1196#define DEBUG_SHIFT 0
1197
1198static inline u64 rb_time_stamp(struct trace_buffer *buffer)
1199{
1200 u64 ts;
1201
1202 /* Skip retpolines :-( */
1203 if (IS_ENABLED(CONFIG_MITIGATION_RETPOLINE) && likely(buffer->clock == trace_clock_local))
1204 ts = trace_clock_local();
1205 else
1206 ts = buffer->clock();
1207
1208 /* shift to debug/test normalization and TIME_EXTENTS */
1209 return ts << DEBUG_SHIFT;
1210}
1211
1212u64 ring_buffer_time_stamp(struct trace_buffer *buffer)
1213{
1214 u64 time;
1215
1216 preempt_disable_notrace();
1217 time = rb_time_stamp(buffer);
1218 preempt_enable_notrace();
1219
1220 return time;
1221}
1222EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
1223
1224void ring_buffer_normalize_time_stamp(struct trace_buffer *buffer,
1225 int cpu, u64 *ts)
1226{
1227 /* Just stupid testing the normalize function and deltas */
1228 *ts >>= DEBUG_SHIFT;
1229}
1230EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
1231
1232/*
1233 * Making the ring buffer lockless makes things tricky.
1234 * Although writes only happen on the CPU that they are on,
1235 * and they only need to worry about interrupts. Reads can
1236 * happen on any CPU.
1237 *
1238 * The reader page is always off the ring buffer, but when the
1239 * reader finishes with a page, it needs to swap its page with
1240 * a new one from the buffer. The reader needs to take from
1241 * the head (writes go to the tail). But if a writer is in overwrite
1242 * mode and wraps, it must push the head page forward.
1243 *
1244 * Here lies the problem.
1245 *
1246 * The reader must be careful to replace only the head page, and
1247 * not another one. As described at the top of the file in the
1248 * ASCII art, the reader sets its old page to point to the next
1249 * page after head. It then sets the page after head to point to
1250 * the old reader page. But if the writer moves the head page
1251 * during this operation, the reader could end up with the tail.
1252 *
1253 * We use cmpxchg to help prevent this race. We also do something
1254 * special with the page before head. We set the LSB to 1.
1255 *
1256 * When the writer must push the page forward, it will clear the
1257 * bit that points to the head page, move the head, and then set
1258 * the bit that points to the new head page.
1259 *
1260 * We also don't want an interrupt coming in and moving the head
1261 * page on another writer. Thus we use the second LSB to catch
1262 * that too. Thus:
1263 *
1264 * head->list->prev->next bit 1 bit 0
1265 * ------- -------
1266 * Normal page 0 0
1267 * Points to head page 0 1
1268 * New head page 1 0
1269 *
1270 * Note we can not trust the prev pointer of the head page, because:
1271 *
1272 * +----+ +-----+ +-----+
1273 * | |------>| T |---X--->| N |
1274 * | |<------| | | |
1275 * +----+ +-----+ +-----+
1276 * ^ ^ |
1277 * | +-----+ | |
1278 * +----------| R |----------+ |
1279 * | |<-----------+
1280 * +-----+
1281 *
1282 * Key: ---X--> HEAD flag set in pointer
1283 * T Tail page
1284 * R Reader page
1285 * N Next page
1286 *
1287 * (see __rb_reserve_next() to see where this happens)
1288 *
1289 * What the above shows is that the reader just swapped out
1290 * the reader page with a page in the buffer, but before it
1291 * could make the new header point back to the new page added
1292 * it was preempted by a writer. The writer moved forward onto
1293 * the new page added by the reader and is about to move forward
1294 * again.
1295 *
1296 * You can see, it is legitimate for the previous pointer of
1297 * the head (or any page) not to point back to itself. But only
1298 * temporarily.
1299 */
1300
1301#define RB_PAGE_NORMAL 0UL
1302#define RB_PAGE_HEAD 1UL
1303#define RB_PAGE_UPDATE 2UL
1304
1305
1306#define RB_FLAG_MASK 3UL
1307
1308/* PAGE_MOVED is not part of the mask */
1309#define RB_PAGE_MOVED 4UL
1310
1311/*
1312 * rb_list_head - remove any bit
1313 */
1314static struct list_head *rb_list_head(struct list_head *list)
1315{
1316 unsigned long val = (unsigned long)list;
1317
1318 return (struct list_head *)(val & ~RB_FLAG_MASK);
1319}
1320
1321/*
1322 * rb_is_head_page - test if the given page is the head page
1323 *
1324 * Because the reader may move the head_page pointer, we can
1325 * not trust what the head page is (it may be pointing to
1326 * the reader page). But if the next page is a header page,
1327 * its flags will be non zero.
1328 */
1329static inline int
1330rb_is_head_page(struct buffer_page *page, struct list_head *list)
1331{
1332 unsigned long val;
1333
1334 val = (unsigned long)list->next;
1335
1336 if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
1337 return RB_PAGE_MOVED;
1338
1339 return val & RB_FLAG_MASK;
1340}
1341
1342/*
1343 * rb_is_reader_page
1344 *
1345 * The unique thing about the reader page, is that, if the
1346 * writer is ever on it, the previous pointer never points
1347 * back to the reader page.
1348 */
1349static bool rb_is_reader_page(struct buffer_page *page)
1350{
1351 struct list_head *list = page->list.prev;
1352
1353 return rb_list_head(list->next) != &page->list;
1354}
1355
1356/*
1357 * rb_set_list_to_head - set a list_head to be pointing to head.
1358 */
1359static void rb_set_list_to_head(struct list_head *list)
1360{
1361 unsigned long *ptr;
1362
1363 ptr = (unsigned long *)&list->next;
1364 *ptr |= RB_PAGE_HEAD;
1365 *ptr &= ~RB_PAGE_UPDATE;
1366}
1367
1368/*
1369 * rb_head_page_activate - sets up head page
1370 */
1371static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
1372{
1373 struct buffer_page *head;
1374
1375 head = cpu_buffer->head_page;
1376 if (!head)
1377 return;
1378
1379 /*
1380 * Set the previous list pointer to have the HEAD flag.
1381 */
1382 rb_set_list_to_head(head->list.prev);
1383
1384 if (cpu_buffer->ring_meta) {
1385 struct ring_buffer_cpu_meta *meta = cpu_buffer->ring_meta;
1386 meta->head_buffer = (unsigned long)head->page;
1387 }
1388}
1389
1390static void rb_list_head_clear(struct list_head *list)
1391{
1392 unsigned long *ptr = (unsigned long *)&list->next;
1393
1394 *ptr &= ~RB_FLAG_MASK;
1395}
1396
1397/*
1398 * rb_head_page_deactivate - clears head page ptr (for free list)
1399 */
1400static void
1401rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
1402{
1403 struct list_head *hd;
1404
1405 /* Go through the whole list and clear any pointers found. */
1406 rb_list_head_clear(cpu_buffer->pages);
1407
1408 list_for_each(hd, cpu_buffer->pages)
1409 rb_list_head_clear(hd);
1410}
1411
1412static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
1413 struct buffer_page *head,
1414 struct buffer_page *prev,
1415 int old_flag, int new_flag)
1416{
1417 struct list_head *list;
1418 unsigned long val = (unsigned long)&head->list;
1419 unsigned long ret;
1420
1421 list = &prev->list;
1422
1423 val &= ~RB_FLAG_MASK;
1424
1425 ret = cmpxchg((unsigned long *)&list->next,
1426 val | old_flag, val | new_flag);
1427
1428 /* check if the reader took the page */
1429 if ((ret & ~RB_FLAG_MASK) != val)
1430 return RB_PAGE_MOVED;
1431
1432 return ret & RB_FLAG_MASK;
1433}
1434
1435static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
1436 struct buffer_page *head,
1437 struct buffer_page *prev,
1438 int old_flag)
1439{
1440 return rb_head_page_set(cpu_buffer, head, prev,
1441 old_flag, RB_PAGE_UPDATE);
1442}
1443
1444static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
1445 struct buffer_page *head,
1446 struct buffer_page *prev,
1447 int old_flag)
1448{
1449 return rb_head_page_set(cpu_buffer, head, prev,
1450 old_flag, RB_PAGE_HEAD);
1451}
1452
1453static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
1454 struct buffer_page *head,
1455 struct buffer_page *prev,
1456 int old_flag)
1457{
1458 return rb_head_page_set(cpu_buffer, head, prev,
1459 old_flag, RB_PAGE_NORMAL);
1460}
1461
1462static inline void rb_inc_page(struct buffer_page **bpage)
1463{
1464 struct list_head *p = rb_list_head((*bpage)->list.next);
1465
1466 *bpage = list_entry(p, struct buffer_page, list);
1467}
1468
1469static inline void rb_dec_page(struct buffer_page **bpage)
1470{
1471 struct list_head *p = rb_list_head((*bpage)->list.prev);
1472
1473 *bpage = list_entry(p, struct buffer_page, list);
1474}
1475
1476static struct buffer_page *
1477rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
1478{
1479 struct buffer_page *head;
1480 struct buffer_page *page;
1481 struct list_head *list;
1482 int i;
1483
1484 if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
1485 return NULL;
1486
1487 /* sanity check */
1488 list = cpu_buffer->pages;
1489 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
1490 return NULL;
1491
1492 page = head = cpu_buffer->head_page;
1493 /*
1494 * It is possible that the writer moves the header behind
1495 * where we started, and we miss in one loop.
1496 * A second loop should grab the header, but we'll do
1497 * three loops just because I'm paranoid.
1498 */
1499 for (i = 0; i < 3; i++) {
1500 do {
1501 if (rb_is_head_page(page, page->list.prev)) {
1502 cpu_buffer->head_page = page;
1503 return page;
1504 }
1505 rb_inc_page(&page);
1506 } while (page != head);
1507 }
1508
1509 RB_WARN_ON(cpu_buffer, 1);
1510
1511 return NULL;
1512}
1513
1514static bool rb_head_page_replace(struct buffer_page *old,
1515 struct buffer_page *new)
1516{
1517 unsigned long *ptr = (unsigned long *)&old->list.prev->next;
1518 unsigned long val;
1519
1520 val = *ptr & ~RB_FLAG_MASK;
1521 val |= RB_PAGE_HEAD;
1522
1523 return try_cmpxchg(ptr, &val, (unsigned long)&new->list);
1524}
1525
1526/*
1527 * rb_tail_page_update - move the tail page forward
1528 */
1529static void rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
1530 struct buffer_page *tail_page,
1531 struct buffer_page *next_page)
1532{
1533 unsigned long old_entries;
1534 unsigned long old_write;
1535
1536 /*
1537 * The tail page now needs to be moved forward.
1538 *
1539 * We need to reset the tail page, but without messing
1540 * with possible erasing of data brought in by interrupts
1541 * that have moved the tail page and are currently on it.
1542 *
1543 * We add a counter to the write field to denote this.
1544 */
1545 old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
1546 old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
1547
1548 /*
1549 * Just make sure we have seen our old_write and synchronize
1550 * with any interrupts that come in.
1551 */
1552 barrier();
1553
1554 /*
1555 * If the tail page is still the same as what we think
1556 * it is, then it is up to us to update the tail
1557 * pointer.
1558 */
1559 if (tail_page == READ_ONCE(cpu_buffer->tail_page)) {
1560 /* Zero the write counter */
1561 unsigned long val = old_write & ~RB_WRITE_MASK;
1562 unsigned long eval = old_entries & ~RB_WRITE_MASK;
1563
1564 /*
1565 * This will only succeed if an interrupt did
1566 * not come in and change it. In which case, we
1567 * do not want to modify it.
1568 *
1569 * We add (void) to let the compiler know that we do not care
1570 * about the return value of these functions. We use the
1571 * cmpxchg to only update if an interrupt did not already
1572 * do it for us. If the cmpxchg fails, we don't care.
1573 */
1574 (void)local_cmpxchg(&next_page->write, old_write, val);
1575 (void)local_cmpxchg(&next_page->entries, old_entries, eval);
1576
1577 /*
1578 * No need to worry about races with clearing out the commit.
1579 * it only can increment when a commit takes place. But that
1580 * only happens in the outer most nested commit.
1581 */
1582 local_set(&next_page->page->commit, 0);
1583
1584 /* Either we update tail_page or an interrupt does */
1585 if (try_cmpxchg(&cpu_buffer->tail_page, &tail_page, next_page))
1586 local_inc(&cpu_buffer->pages_touched);
1587 }
1588}
1589
1590static void rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
1591 struct buffer_page *bpage)
1592{
1593 unsigned long val = (unsigned long)bpage;
1594
1595 RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK);
1596}
1597
1598static bool rb_check_links(struct ring_buffer_per_cpu *cpu_buffer,
1599 struct list_head *list)
1600{
1601 if (RB_WARN_ON(cpu_buffer,
1602 rb_list_head(rb_list_head(list->next)->prev) != list))
1603 return false;
1604
1605 if (RB_WARN_ON(cpu_buffer,
1606 rb_list_head(rb_list_head(list->prev)->next) != list))
1607 return false;
1608
1609 return true;
1610}
1611
1612/**
1613 * rb_check_pages - integrity check of buffer pages
1614 * @cpu_buffer: CPU buffer with pages to test
1615 *
1616 * As a safety measure we check to make sure the data pages have not
1617 * been corrupted.
1618 */
1619static void rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
1620{
1621 struct list_head *head, *tmp;
1622 unsigned long buffer_cnt;
1623 unsigned long flags;
1624 int nr_loops = 0;
1625
1626 /*
1627 * Walk the linked list underpinning the ring buffer and validate all
1628 * its next and prev links.
1629 *
1630 * The check acquires the reader_lock to avoid concurrent processing
1631 * with code that could be modifying the list. However, the lock cannot
1632 * be held for the entire duration of the walk, as this would make the
1633 * time when interrupts are disabled non-deterministic, dependent on the
1634 * ring buffer size. Therefore, the code releases and re-acquires the
1635 * lock after checking each page. The ring_buffer_per_cpu.cnt variable
1636 * is then used to detect if the list was modified while the lock was
1637 * not held, in which case the check needs to be restarted.
1638 *
1639 * The code attempts to perform the check at most three times before
1640 * giving up. This is acceptable because this is only a self-validation
1641 * to detect problems early on. In practice, the list modification
1642 * operations are fairly spaced, and so this check typically succeeds at
1643 * most on the second try.
1644 */
1645again:
1646 if (++nr_loops > 3)
1647 return;
1648
1649 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
1650 head = rb_list_head(cpu_buffer->pages);
1651 if (!rb_check_links(cpu_buffer, head))
1652 goto out_locked;
1653 buffer_cnt = cpu_buffer->cnt;
1654 tmp = head;
1655 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
1656
1657 while (true) {
1658 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
1659
1660 if (buffer_cnt != cpu_buffer->cnt) {
1661 /* The list was updated, try again. */
1662 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
1663 goto again;
1664 }
1665
1666 tmp = rb_list_head(tmp->next);
1667 if (tmp == head)
1668 /* The iteration circled back, all is done. */
1669 goto out_locked;
1670
1671 if (!rb_check_links(cpu_buffer, tmp))
1672 goto out_locked;
1673
1674 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
1675 }
1676
1677out_locked:
1678 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
1679}
1680
1681/*
1682 * Take an address, add the meta data size as well as the array of
1683 * array subbuffer indexes, then align it to a subbuffer size.
1684 *
1685 * This is used to help find the next per cpu subbuffer within a mapped range.
1686 */
1687static unsigned long
1688rb_range_align_subbuf(unsigned long addr, unsigned int subbuf_size, unsigned long nr_subbufs)
1689{
1690 addr += sizeof(struct ring_buffer_cpu_meta) +
1691 sizeof(int) * nr_subbufs;
1692 return ALIGN(addr, subbuf_size);
1693}
1694
1695/*
1696 * Return the ring_buffer_meta for a given @cpu.
1697 */
1698static void *rb_range_meta(struct trace_buffer *buffer, unsigned long nr_pages, int cpu)
1699{
1700 unsigned int subbuf_size = rb_subbuf_size(buffer);
1701 struct ring_buffer_cpu_meta *meta;
1702 struct ring_buffer_meta *bmeta;
1703 unsigned long ptr, nr_subbufs;
1704
1705 bmeta = buffer->meta;
1706 if (!bmeta)
1707 return NULL;
1708
1709 ptr = (unsigned long)bmeta + bmeta->buffers_offset;
1710 meta = (struct ring_buffer_cpu_meta *)ptr;
1711
1712 /* When nr_pages passed in is zero, the first meta has already been initialized */
1713 if (!nr_pages) {
1714 nr_subbufs = meta->nr_subbufs;
1715 } else {
1716 /* Include the reader page */
1717 nr_subbufs = nr_pages + 1;
1718 }
1719
1720 /*
1721 * The first chunk may not be subbuffer aligned, where as
1722 * the rest of the chunks are.
1723 */
1724 if (cpu) {
1725 ptr = rb_range_align_subbuf(ptr, subbuf_size, nr_subbufs);
1726 ptr += subbuf_size * nr_subbufs;
1727
1728 /* We can use multiplication to find chunks greater than 1 */
1729 if (cpu > 1) {
1730 unsigned long size;
1731 unsigned long p;
1732
1733 /* Save the beginning of this CPU chunk */
1734 p = ptr;
1735 ptr = rb_range_align_subbuf(ptr, subbuf_size, nr_subbufs);
1736 ptr += subbuf_size * nr_subbufs;
1737
1738 /* Now all chunks after this are the same size */
1739 size = ptr - p;
1740 ptr += size * (cpu - 2);
1741 }
1742 }
1743 return (void *)ptr;
1744}
1745
1746/* Return the start of subbufs given the meta pointer */
1747static void *rb_subbufs_from_meta(struct ring_buffer_cpu_meta *meta)
1748{
1749 unsigned int subbuf_size = meta->subbuf_size;
1750 unsigned long ptr;
1751
1752 ptr = (unsigned long)meta;
1753 ptr = rb_range_align_subbuf(ptr, subbuf_size, meta->nr_subbufs);
1754
1755 return (void *)ptr;
1756}
1757
1758/*
1759 * Return a specific sub-buffer for a given @cpu defined by @idx.
1760 */
1761static void *rb_range_buffer(struct ring_buffer_per_cpu *cpu_buffer, unsigned long idx)
1762{
1763 struct ring_buffer_cpu_meta *meta;
1764 unsigned int subbuf_size;
1765 unsigned long ptr;
1766
1767 meta = rb_range_meta(cpu_buffer->buffer, 0, cpu_buffer->cpu);
1768 if (!meta)
1769 return NULL;
1770
1771 if (WARN_ON_ONCE(idx >= meta->nr_subbufs))
1772 return NULL;
1773
1774 subbuf_size = meta->subbuf_size;
1775
1776 /* Map this buffer to the order that's in meta->buffers[] */
1777 idx = meta->buffers[idx];
1778
1779 ptr = (unsigned long)rb_subbufs_from_meta(meta);
1780
1781 ptr += (unsigned long)subbuf_size * idx;
1782 if (ptr + subbuf_size > cpu_buffer->buffer->range_addr_end)
1783 return NULL;
1784
1785 return (void *)ptr;
1786}
1787
1788/*
1789 * See if the existing memory contains a valid meta section.
1790 * if so, use that, otherwise initialize it.
1791 */
1792static bool rb_meta_init(struct trace_buffer *buffer, int scratch_size)
1793{
1794 unsigned long ptr = buffer->range_addr_start;
1795 struct ring_buffer_meta *bmeta;
1796 unsigned long total_size;
1797 int struct_sizes;
1798
1799 bmeta = (struct ring_buffer_meta *)ptr;
1800 buffer->meta = bmeta;
1801
1802 total_size = buffer->range_addr_end - buffer->range_addr_start;
1803
1804 struct_sizes = sizeof(struct ring_buffer_cpu_meta);
1805 struct_sizes |= sizeof(*bmeta) << 16;
1806
1807 /* The first buffer will start word size after the meta page */
1808 ptr += sizeof(*bmeta);
1809 ptr = ALIGN(ptr, sizeof(long));
1810 ptr += scratch_size;
1811
1812 if (bmeta->magic != RING_BUFFER_META_MAGIC) {
1813 pr_info("Ring buffer boot meta mismatch of magic\n");
1814 goto init;
1815 }
1816
1817 if (bmeta->struct_sizes != struct_sizes) {
1818 pr_info("Ring buffer boot meta mismatch of struct size\n");
1819 goto init;
1820 }
1821
1822 if (bmeta->total_size != total_size) {
1823 pr_info("Ring buffer boot meta mismatch of total size\n");
1824 goto init;
1825 }
1826
1827 if (bmeta->buffers_offset > bmeta->total_size) {
1828 pr_info("Ring buffer boot meta mismatch of offset outside of total size\n");
1829 goto init;
1830 }
1831
1832 if (bmeta->buffers_offset != (void *)ptr - (void *)bmeta) {
1833 pr_info("Ring buffer boot meta mismatch of first buffer offset\n");
1834 goto init;
1835 }
1836
1837 return true;
1838
1839 init:
1840 bmeta->magic = RING_BUFFER_META_MAGIC;
1841 bmeta->struct_sizes = struct_sizes;
1842 bmeta->total_size = total_size;
1843 bmeta->buffers_offset = (void *)ptr - (void *)bmeta;
1844
1845 /* Zero out the scratch pad */
1846 memset((void *)bmeta + sizeof(*bmeta), 0, bmeta->buffers_offset - sizeof(*bmeta));
1847
1848 return false;
1849}
1850
1851/*
1852 * See if the existing memory contains valid ring buffer data.
1853 * As the previous kernel must be the same as this kernel, all
1854 * the calculations (size of buffers and number of buffers)
1855 * must be the same.
1856 */
1857static bool rb_cpu_meta_valid(struct ring_buffer_cpu_meta *meta, int cpu,
1858 struct trace_buffer *buffer, unsigned long nr_pages,
1859 unsigned long *subbuf_mask)
1860{
1861 unsigned long buffers_start;
1862 unsigned long buffers_end;
1863 unsigned long i;
1864
1865 if (!subbuf_mask)
1866 return false;
1867
1868 if (meta->subbuf_size != PAGE_SIZE) {
1869 pr_info("Ring buffer boot meta [%d] invalid subbuf_size\n", cpu);
1870 return false;
1871 }
1872
1873 if (meta->nr_subbufs != nr_pages + 1) {
1874 pr_info("Ring buffer boot meta [%d] invalid nr_subbufs\n", cpu);
1875 return false;
1876 }
1877
1878 buffers_start = meta->first_buffer;
1879 buffers_end = meta->first_buffer + (meta->nr_subbufs * PAGE_SIZE);
1880
1881 /* Is the head and commit buffers within the range of buffers? */
1882 if (meta->head_buffer < buffers_start ||
1883 meta->head_buffer >= buffers_end) {
1884 pr_info("Ring buffer boot meta [%d] head buffer out of range\n", cpu);
1885 return false;
1886 }
1887
1888 if (meta->commit_buffer < buffers_start ||
1889 meta->commit_buffer >= buffers_end) {
1890 pr_info("Ring buffer boot meta [%d] commit buffer out of range\n", cpu);
1891 return false;
1892 }
1893
1894 bitmap_clear(subbuf_mask, 0, meta->nr_subbufs);
1895
1896 /*
1897 * Ensure the meta::buffers array has correct data. The data in each subbufs
1898 * are checked later in rb_meta_validate_events().
1899 */
1900 for (i = 0; i < meta->nr_subbufs; i++) {
1901 if (meta->buffers[i] < 0 ||
1902 meta->buffers[i] >= meta->nr_subbufs) {
1903 pr_info("Ring buffer boot meta [%d] array out of range\n", cpu);
1904 return false;
1905 }
1906
1907 if (test_bit(meta->buffers[i], subbuf_mask)) {
1908 pr_info("Ring buffer boot meta [%d] array has duplicates\n", cpu);
1909 return false;
1910 }
1911
1912 set_bit(meta->buffers[i], subbuf_mask);
1913 }
1914
1915 return true;
1916}
1917
1918static int rb_meta_subbuf_idx(struct ring_buffer_cpu_meta *meta, void *subbuf);
1919
1920static int rb_read_data_buffer(struct buffer_data_page *dpage, int tail, int cpu,
1921 unsigned long long *timestamp, u64 *delta_ptr)
1922{
1923 struct ring_buffer_event *event;
1924 u64 ts, delta;
1925 int events = 0;
1926 int len;
1927 int e;
1928
1929 *delta_ptr = 0;
1930 *timestamp = 0;
1931
1932 ts = dpage->time_stamp;
1933
1934 for (e = 0; e < tail; e += len) {
1935
1936 event = (struct ring_buffer_event *)(dpage->data + e);
1937 len = rb_event_length(event);
1938 if (len <= 0 || len > tail - e)
1939 return -1;
1940
1941 switch (event->type_len) {
1942
1943 case RINGBUF_TYPE_TIME_EXTEND:
1944 delta = rb_event_time_stamp(event);
1945 ts += delta;
1946 break;
1947
1948 case RINGBUF_TYPE_TIME_STAMP:
1949 delta = rb_event_time_stamp(event);
1950 delta = rb_fix_abs_ts(delta, ts);
1951 if (delta < ts) {
1952 *delta_ptr = delta;
1953 *timestamp = ts;
1954 return -1;
1955 }
1956 ts = delta;
1957 break;
1958
1959 case RINGBUF_TYPE_PADDING:
1960 if (event->time_delta == 1)
1961 break;
1962 fallthrough;
1963 case RINGBUF_TYPE_DATA:
1964 events++;
1965 ts += event->time_delta;
1966 break;
1967
1968 default:
1969 return -1;
1970 }
1971 }
1972 *timestamp = ts;
1973 return events;
1974}
1975
1976struct rb_validation_state {
1977 unsigned long entries;
1978 unsigned long entry_bytes;
1979 int discarded;
1980 u64 ts;
1981};
1982
1983static int __rb_validate_buffer(struct buffer_page *bpage, int cpu,
1984 struct ring_buffer_cpu_meta *meta,
1985 u64 prev_ts, u64 next_ts)
1986{
1987 struct buffer_data_page *dpage = bpage->page;
1988 unsigned long long ts;
1989 unsigned long tail;
1990 u64 delta;
1991 int ret;
1992
1993 /*
1994 * When a sub-buffer is recovered from a read, the commit value may
1995 * have RB_MISSED_* bits set, as these bits are reset on reuse.
1996 * Even after clearing these bits, a commit value greater than the
1997 * subbuf_size is considered invalid.
1998 */
1999 tail = rb_data_page_commit(dpage);
2000 if (tail <= meta->subbuf_size - BUF_PAGE_HDR_SIZE)
2001 ret = rb_read_data_buffer(dpage, tail, cpu, &ts, &delta);
2002 else
2003 ret = -1;
2004
2005 /*
2006 * The timestamp must be greater than @prev_ts and smaller than @next_ts.
2007 * Since this function works in both forward (verify) and reverse (unwind)
2008 * loop, we don't know both @prev_ts and @next_ts at the same time.
2009 * So use the known boundary as the boundary.
2010 */
2011 if (ret < 0 || (prev_ts && prev_ts > ts) || (next_ts && ts > next_ts)) {
2012 local_set(&bpage->entries, 0);
2013 /*
2014 * Note, the RB_MISSED_EVENTS is only set inside the main write
2015 * buffer by this verification logic. The normal ring buffer
2016 * has this bit set when the page is read and passed to the
2017 * consumers.
2018 */
2019 local_set(&dpage->commit, RB_MISSED_EVENTS);
2020 dpage->time_stamp = prev_ts ? prev_ts : next_ts;
2021 ret = -1;
2022 } else {
2023 local_set(&bpage->entries, ret);
2024 }
2025
2026 return ret;
2027}
2028
2029/**
2030 * rb_validate_buffer - validates a single buffer page and updates the state.
2031 * @bpage: buffer page to validate
2032 * @cpu_buffer: cpu_buffer this page belongs to
2033 * @meta: meta of the cpu_buffer
2034 * @state: validation state
2035 * @prev_ts: previous buffer's timestamp (optional)
2036 * @next_ts: next buffer's timestamp (optional)
2037 *
2038 * If the page is invalid (wrong event length or timestamp), it increments the
2039 * discarded counter and warns it. Otherwise, it updates the validation state.
2040 */
2041static void rb_validate_buffer(struct buffer_page *bpage,
2042 struct ring_buffer_per_cpu *cpu_buffer,
2043 struct ring_buffer_cpu_meta *meta,
2044 struct rb_validation_state *state,
2045 u64 prev_ts, u64 next_ts)
2046{
2047 int ret;
2048
2049 ret = __rb_validate_buffer(bpage, cpu_buffer->cpu, meta, prev_ts, next_ts);
2050 if (ret < 0) {
2051 if (!state->discarded)
2052 pr_info("Ring buffer meta [%d] invalid buffer page detected\n",
2053 cpu_buffer->cpu);
2054 state->discarded++;
2055 } else {
2056 /* If the buffer has content, update pages_touched */
2057 if (ret)
2058 local_inc(&cpu_buffer->pages_touched);
2059
2060 state->entries += ret;
2061 state->entry_bytes += rb_page_size(bpage);
2062 state->ts = bpage->page->time_stamp;
2063 }
2064}
2065
2066static void rb_meta_inject_reader_page(struct ring_buffer_per_cpu *cpu_buffer,
2067 struct ring_buffer_cpu_meta *meta,
2068 struct buffer_page *orig_head,
2069 struct buffer_page *head_page)
2070{
2071 struct buffer_page *bpage = orig_head;
2072 int i;
2073
2074 rb_dec_page(&bpage);
2075 /*
2076 * Insert the reader_page before the original head page.
2077 * Since the list encode RB_PAGE flags, general list
2078 * operations should be avoided.
2079 */
2080 cpu_buffer->reader_page->list.next = &orig_head->list;
2081 cpu_buffer->reader_page->list.prev = orig_head->list.prev;
2082 orig_head->list.prev = &cpu_buffer->reader_page->list;
2083 bpage->list.next = &cpu_buffer->reader_page->list;
2084
2085 /* Make the head_page the reader page */
2086 cpu_buffer->reader_page = head_page;
2087 bpage = head_page;
2088 rb_inc_page(&head_page);
2089 head_page->list.prev = bpage->list.prev;
2090 rb_dec_page(&bpage);
2091 bpage->list.next = &head_page->list;
2092 rb_set_list_to_head(&bpage->list);
2093 cpu_buffer->pages = &head_page->list;
2094
2095 cpu_buffer->head_page = head_page;
2096 meta->head_buffer = (unsigned long)head_page->page;
2097
2098 /* Reset all the indexes */
2099 bpage = cpu_buffer->reader_page;
2100 meta->buffers[0] = rb_meta_subbuf_idx(meta, bpage->page);
2101 bpage->id = 0;
2102
2103 for (i = 1, bpage = head_page; i < meta->nr_subbufs;
2104 i++, rb_inc_page(&bpage)) {
2105 meta->buffers[i] = rb_meta_subbuf_idx(meta, bpage->page);
2106 bpage->id = i;
2107 }
2108}
2109
2110/* If the meta data has been validated, now validate the events */
2111static void rb_meta_validate_events(struct ring_buffer_per_cpu *cpu_buffer)
2112{
2113 struct ring_buffer_cpu_meta *meta = cpu_buffer->ring_meta;
2114 struct buffer_page *head_page, *orig_head, *orig_reader;
2115 struct rb_validation_state state = { 0 };
2116 bool skip = false;
2117 unsigned long i;
2118 int ret;
2119
2120 if (!meta || !meta->head_buffer)
2121 return;
2122
2123 orig_head = head_page = cpu_buffer->head_page;
2124 orig_reader = cpu_buffer->reader_page;
2125
2126 /* Do the head page first */
2127 ret = __rb_validate_buffer(head_page, cpu_buffer->cpu, meta, 0, 0);
2128 if (ret < 0) {
2129 pr_info("Ring buffer meta [%d] invalid head page detected\n",
2130 cpu_buffer->cpu);
2131 /* Don't bother rewinding */
2132 skip = true;
2133 state.ts = 0;
2134 } else {
2135 state.ts = head_page->page->time_stamp;
2136 }
2137
2138 /* Do the reader page - reader must be previous to head. */
2139 rb_validate_buffer(orig_reader, cpu_buffer, meta, &state, 0, state.ts);
2140
2141 if (skip)
2142 goto skip_rewind;
2143
2144 /*
2145 * Try to rewind the head so that we can read the pages which are already
2146 * read in the previous boot.
2147 */
2148 if (head_page == cpu_buffer->tail_page)
2149 goto skip_rewind;
2150
2151 rb_dec_page(&head_page);
2152 for (i = 0; i < meta->nr_subbufs + 1; i++, rb_dec_page(&head_page)) {
2153
2154 /* Rewind until tail (writer) page. */
2155 if (head_page == cpu_buffer->tail_page)
2156 break;
2157
2158 /* Rewind until unused page (no timestamp, no commit). */
2159 if (!head_page->page->time_stamp && rb_page_commit(head_page) == 0)
2160 break;
2161
2162 /*
2163 * Skip if the page is invalid, or its timestamp is newer than the
2164 * previous valid page.
2165 */
2166 rb_validate_buffer(head_page, cpu_buffer, meta, &state, 0, state.ts);
2167 }
2168 if (i)
2169 pr_info("Ring buffer [%d] rewound %lu pages\n", cpu_buffer->cpu, i);
2170
2171 /* The last rewound page must be skipped. */
2172 if (head_page != orig_head)
2173 rb_inc_page(&head_page);
2174
2175 /*
2176 * If the ring buffer was rewound, then inject the reader page
2177 * into the location just before the original head page.
2178 */
2179 if (head_page != orig_head) {
2180 rb_meta_inject_reader_page(cpu_buffer, meta, orig_head, head_page);
2181 /* We'll restart verifying from orig_head */
2182 head_page = orig_head;
2183 }
2184
2185 skip_rewind:
2186 /* If the commit_buffer is the reader page, update the commit page */
2187 if (meta->commit_buffer == (unsigned long)cpu_buffer->reader_page->page) {
2188 cpu_buffer->commit_page = cpu_buffer->reader_page;
2189 /* Nothing more to do, the only page is the reader page */
2190 goto done;
2191 }
2192 state.ts = head_page->page->time_stamp;
2193
2194 /* Iterate until finding the commit page */
2195 for (i = 0; i < meta->nr_subbufs + 1; i++, rb_inc_page(&head_page)) {
2196
2197 /* The original reader page has already been checked/counted. */
2198 if (head_page == orig_reader)
2199 continue;
2200
2201 rb_validate_buffer(head_page, cpu_buffer, meta, &state, state.ts, 0);
2202
2203 if (head_page == cpu_buffer->commit_page)
2204 break;
2205 }
2206
2207 if (head_page != cpu_buffer->commit_page) {
2208 pr_info("Ring buffer meta [%d] commit page not found\n",
2209 cpu_buffer->cpu);
2210 goto invalid;
2211 }
2212 done:
2213 local_set(&cpu_buffer->entries, state.entries);
2214 local_set(&cpu_buffer->entries_bytes, state.entry_bytes);
2215
2216 pr_info("Ring buffer meta [%d] is from previous boot!", cpu_buffer->cpu);
2217 if (state.discarded)
2218 pr_cont(" (%d pages discarded)", state.discarded);
2219 pr_cont("\n");
2220
2221#ifdef CONFIG_RING_BUFFER_PERSISTENT_INJECT
2222 if (meta->nr_invalid)
2223 pr_warn("Ring buffer testing [%d] invalid pages: %s (%d/%d)\n",
2224 cpu_buffer->cpu,
2225 (state.discarded == meta->nr_invalid) ? "PASSED" : "FAILED",
2226 state.discarded, meta->nr_invalid);
2227 if (meta->entry_bytes)
2228 pr_warn("Ring buffer testing [%d] entry_bytes: %s (%ld/%ld)\n",
2229 cpu_buffer->cpu,
2230 (state.entry_bytes == meta->entry_bytes) ? "PASSED" : "FAILED",
2231 (long)state.entry_bytes, (long)meta->entry_bytes);
2232 meta->nr_invalid = 0;
2233 meta->entry_bytes = 0;
2234#endif
2235 return;
2236
2237 invalid:
2238 /* The content of the buffers are invalid, reset the meta data */
2239 meta->head_buffer = 0;
2240 meta->commit_buffer = 0;
2241
2242 /* Reset the reader page */
2243 local_set(&cpu_buffer->reader_page->entries, 0);
2244 rb_init_data_page(cpu_buffer->reader_page->page);
2245
2246 /* Reset all the subbuffers */
2247 for (i = 0; i < meta->nr_subbufs - 1; i++, rb_inc_page(&head_page)) {
2248 local_set(&head_page->entries, 0);
2249 rb_init_data_page(head_page->page);
2250 }
2251}
2252
2253static void rb_range_meta_init(struct trace_buffer *buffer, unsigned long nr_pages,
2254 int scratch_size)
2255{
2256 struct ring_buffer_cpu_meta *meta;
2257 unsigned long *subbuf_mask;
2258 unsigned long delta;
2259 void *subbuf;
2260 bool valid = false;
2261 int cpu;
2262 int i;
2263
2264 /* Create a mask to test the subbuf array */
2265 subbuf_mask = bitmap_alloc(nr_pages + 1, GFP_KERNEL);
2266 /* If subbuf_mask fails to allocate, then rb_meta_valid() will return false */
2267
2268 if (rb_meta_init(buffer, scratch_size))
2269 valid = true;
2270
2271 for (cpu = 0; cpu < nr_cpu_ids; cpu++) {
2272 void *next_meta;
2273
2274 meta = rb_range_meta(buffer, nr_pages, cpu);
2275
2276 if (valid && rb_cpu_meta_valid(meta, cpu, buffer, nr_pages, subbuf_mask)) {
2277 /* Make the mappings match the current address */
2278 subbuf = rb_subbufs_from_meta(meta);
2279 delta = (unsigned long)subbuf - meta->first_buffer;
2280 meta->first_buffer += delta;
2281 meta->head_buffer += delta;
2282 meta->commit_buffer += delta;
2283 continue;
2284 }
2285
2286 if (cpu < nr_cpu_ids - 1)
2287 next_meta = rb_range_meta(buffer, nr_pages, cpu + 1);
2288 else
2289 next_meta = (void *)buffer->range_addr_end;
2290
2291 memset(meta, 0, next_meta - (void *)meta);
2292
2293 meta->nr_subbufs = nr_pages + 1;
2294 meta->subbuf_size = PAGE_SIZE;
2295
2296 subbuf = rb_subbufs_from_meta(meta);
2297
2298 meta->first_buffer = (unsigned long)subbuf;
2299
2300 /*
2301 * The buffers[] array holds the order of the sub-buffers
2302 * that are after the meta data. The sub-buffers may
2303 * be swapped out when read and inserted into a different
2304 * location of the ring buffer. Although their addresses
2305 * remain the same, the buffers[] array contains the
2306 * index into the sub-buffers holding their actual order.
2307 */
2308 for (i = 0; i < meta->nr_subbufs; i++) {
2309 meta->buffers[i] = i;
2310 rb_init_data_page(subbuf);
2311 subbuf += meta->subbuf_size;
2312 }
2313 }
2314 bitmap_free(subbuf_mask);
2315}
2316
2317static void *rbm_start(struct seq_file *m, loff_t *pos)
2318{
2319 struct ring_buffer_per_cpu *cpu_buffer = m->private;
2320 struct ring_buffer_cpu_meta *meta = cpu_buffer->ring_meta;
2321 unsigned long val;
2322
2323 if (!meta)
2324 return NULL;
2325
2326 if (*pos > meta->nr_subbufs)
2327 return NULL;
2328
2329 val = *pos;
2330 val++;
2331
2332 return (void *)val;
2333}
2334
2335static void *rbm_next(struct seq_file *m, void *v, loff_t *pos)
2336{
2337 (*pos)++;
2338
2339 return rbm_start(m, pos);
2340}
2341
2342static int rbm_show(struct seq_file *m, void *v)
2343{
2344 struct ring_buffer_per_cpu *cpu_buffer = m->private;
2345 struct ring_buffer_cpu_meta *meta = cpu_buffer->ring_meta;
2346 unsigned long val = (unsigned long)v;
2347 struct buffer_data_page *dpage;
2348
2349 if (val == 1) {
2350 seq_printf(m, "head_buffer: %d\n",
2351 rb_meta_subbuf_idx(meta, (void *)meta->head_buffer));
2352 seq_printf(m, "commit_buffer: %d\n",
2353 rb_meta_subbuf_idx(meta, (void *)meta->commit_buffer));
2354 seq_printf(m, "subbuf_size: %u\n", meta->subbuf_size);
2355 seq_printf(m, "nr_subbufs: %u\n", meta->nr_subbufs);
2356 return 0;
2357 }
2358
2359 val -= 2;
2360 dpage = rb_range_buffer(cpu_buffer, val);
2361 seq_printf(m, "buffer[%ld]: %d (commit: %ld)\n",
2362 val, meta->buffers[val], dpage ? rb_data_page_commit(dpage) : -1);
2363
2364 return 0;
2365}
2366
2367static void rbm_stop(struct seq_file *m, void *p)
2368{
2369}
2370
2371static const struct seq_operations rb_meta_seq_ops = {
2372 .start = rbm_start,
2373 .next = rbm_next,
2374 .show = rbm_show,
2375 .stop = rbm_stop,
2376};
2377
2378int ring_buffer_meta_seq_init(struct file *file, struct trace_buffer *buffer, int cpu)
2379{
2380 struct seq_file *m;
2381 int ret;
2382
2383 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2384 return -ENODEV;
2385
2386 ret = seq_open(file, &rb_meta_seq_ops);
2387 if (ret)
2388 return ret;
2389
2390 m = file->private_data;
2391 m->private = buffer->buffers[cpu];
2392
2393 return 0;
2394}
2395
2396/* Map the buffer_pages to the previous head and commit pages */
2397static void rb_meta_buffer_update(struct ring_buffer_per_cpu *cpu_buffer,
2398 struct buffer_page *bpage)
2399{
2400 struct ring_buffer_cpu_meta *meta = cpu_buffer->ring_meta;
2401
2402 if (meta->head_buffer == (unsigned long)bpage->page)
2403 cpu_buffer->head_page = bpage;
2404
2405 if (meta->commit_buffer == (unsigned long)bpage->page) {
2406 cpu_buffer->commit_page = bpage;
2407 cpu_buffer->tail_page = bpage;
2408 }
2409}
2410
2411static struct ring_buffer_desc *ring_buffer_desc(struct trace_buffer_desc *trace_desc, int cpu)
2412{
2413 struct ring_buffer_desc *desc, *end;
2414 size_t len;
2415 int i;
2416
2417 if (!trace_desc || !trace_desc->nr_cpus)
2418 return NULL;
2419
2420 end = (struct ring_buffer_desc *)((void *)trace_desc + trace_desc->struct_len);
2421 desc = __first_ring_buffer_desc(trace_desc);
2422 len = struct_size(desc, page_va, desc->nr_page_va);
2423 desc = (struct ring_buffer_desc *)((void *)desc + (len * cpu));
2424
2425 if (desc < end && desc->cpu == cpu)
2426 return desc;
2427
2428 /* Missing CPUs, need to linear search */
2429 for_each_ring_buffer_desc(desc, i, trace_desc) {
2430 if (desc->cpu == cpu)
2431 return desc;
2432 }
2433
2434 return NULL;
2435}
2436
2437static void *ring_buffer_desc_page(struct ring_buffer_desc *desc, unsigned int page_id)
2438{
2439 return page_id >= desc->nr_page_va ? NULL : (void *)desc->page_va[page_id];
2440}
2441
2442static int __rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
2443 unsigned long nr_pages, struct list_head *pages)
2444{
2445 struct trace_buffer *buffer = cpu_buffer->buffer;
2446 unsigned int subbuf_order = buffer->subbuf_order;
2447 struct ring_buffer_cpu_meta *meta = NULL;
2448 struct buffer_page *bpage, *tmp;
2449 struct ring_buffer_desc *desc = NULL;
2450 bool user_thread = false;
2451 long i;
2452
2453 if (buffer->range_addr_start) {
2454 meta = rb_range_meta(buffer, nr_pages, cpu_buffer->cpu);
2455 } else if (buffer->remote) {
2456 desc = ring_buffer_desc(buffer->remote->desc, cpu_buffer->cpu);
2457 if (!desc || WARN_ON(desc->nr_page_va != (nr_pages + 1)))
2458 return -EINVAL;
2459 } else {
2460 /*
2461 * Check if the available memory is there first.
2462 * Note, si_mem_available() only gives us a rough estimate of available
2463 * memory. It may not be accurate. But we don't care, we just want
2464 * to prevent doing any allocation when it is obvious that it is
2465 * not going to succeed.
2466 */
2467 i = si_mem_available();
2468 if (i < (nr_pages << subbuf_order))
2469 return -ENOMEM;
2470
2471 user_thread = current->mm != NULL;
2472 }
2473
2474 /*
2475 * If a user thread allocates too much, and si_mem_available()
2476 * reports there's enough memory, even though there is not.
2477 * Make sure the OOM killer kills this thread. This can happen
2478 * even with RETRY_MAYFAIL because another task may be doing
2479 * an allocation after this task has taken all memory.
2480 * This is the task the OOM killer needs to take out during this
2481 * loop, even if it was triggered by an allocation somewhere else.
2482 */
2483 if (user_thread)
2484 set_current_oom_origin();
2485
2486 for (i = 0; i < nr_pages; i++) {
2487
2488 bpage = alloc_cpu_page(cpu_buffer->cpu);
2489 if (!bpage)
2490 goto free_pages;
2491
2492 rb_check_bpage(cpu_buffer, bpage);
2493
2494 /*
2495 * Append the pages as for mapped buffers we want to keep
2496 * the order
2497 */
2498 list_add_tail(&bpage->list, pages);
2499
2500 if (meta) {
2501 /* A range was given. Use that for the buffer page */
2502 bpage->page = rb_range_buffer(cpu_buffer, i + 1);
2503 if (!bpage->page)
2504 goto free_pages;
2505 /* If this is valid from a previous boot */
2506 if (meta->head_buffer)
2507 rb_meta_buffer_update(cpu_buffer, bpage);
2508 bpage->range = 1;
2509 bpage->id = i + 1;
2510 } else if (desc) {
2511 void *p = ring_buffer_desc_page(desc, i + 1);
2512
2513 if (WARN_ON(!p))
2514 goto free_pages;
2515
2516 bpage->page = p;
2517 bpage->range = 1; /* bpage->page can't be freed */
2518 bpage->id = i + 1;
2519 cpu_buffer->subbuf_ids[i + 1] = bpage;
2520 } else {
2521 bpage->page = alloc_cpu_data(cpu_buffer->cpu, subbuf_order);
2522 if (!bpage->page)
2523 goto free_pages;
2524 }
2525 bpage->order = subbuf_order;
2526
2527 if (user_thread && fatal_signal_pending(current))
2528 goto free_pages;
2529 }
2530 if (user_thread)
2531 clear_current_oom_origin();
2532
2533 return 0;
2534
2535free_pages:
2536 list_for_each_entry_safe(bpage, tmp, pages, list) {
2537 list_del_init(&bpage->list);
2538 free_buffer_page(bpage);
2539 }
2540 if (user_thread)
2541 clear_current_oom_origin();
2542
2543 return -ENOMEM;
2544}
2545
2546static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
2547 unsigned long nr_pages)
2548{
2549 LIST_HEAD(pages);
2550
2551 WARN_ON(!nr_pages);
2552
2553 if (__rb_allocate_pages(cpu_buffer, nr_pages, &pages))
2554 return -ENOMEM;
2555
2556 /*
2557 * The ring buffer page list is a circular list that does not
2558 * start and end with a list head. All page list items point to
2559 * other pages.
2560 */
2561 cpu_buffer->pages = pages.next;
2562 list_del(&pages);
2563
2564 cpu_buffer->nr_pages = nr_pages;
2565
2566 rb_check_pages(cpu_buffer);
2567
2568 return 0;
2569}
2570
2571static struct ring_buffer_per_cpu *
2572rb_allocate_cpu_buffer(struct trace_buffer *buffer, unsigned long nr_pages, int cpu)
2573{
2574 struct ring_buffer_per_cpu *cpu_buffer __free(kfree) =
2575 alloc_cpu_buffer(cpu);
2576 struct ring_buffer_cpu_meta *meta;
2577 struct buffer_page *bpage;
2578 int ret;
2579
2580 if (!cpu_buffer)
2581 return NULL;
2582
2583 cpu_buffer->cpu = cpu;
2584 cpu_buffer->buffer = buffer;
2585 raw_spin_lock_init(&cpu_buffer->reader_lock);
2586 lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
2587 cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
2588 INIT_WORK(&cpu_buffer->update_pages_work, update_pages_handler);
2589 init_completion(&cpu_buffer->update_done);
2590 init_irq_work(&cpu_buffer->irq_work.work, rb_wake_up_waiters);
2591 init_waitqueue_head(&cpu_buffer->irq_work.waiters);
2592 init_waitqueue_head(&cpu_buffer->irq_work.full_waiters);
2593 mutex_init(&cpu_buffer->mapping_lock);
2594
2595 bpage = alloc_cpu_page(cpu);
2596 if (!bpage)
2597 return NULL;
2598 bpage->order = cpu_buffer->buffer->subbuf_order;
2599
2600 rb_check_bpage(cpu_buffer, bpage);
2601
2602 cpu_buffer->reader_page = bpage;
2603
2604 if (buffer->range_addr_start) {
2605 /*
2606 * Range mapped buffers have the same restrictions as memory
2607 * mapped ones do.
2608 */
2609 cpu_buffer->ring_meta = rb_range_meta(buffer, nr_pages, cpu);
2610 bpage->page = rb_range_buffer(cpu_buffer, 0);
2611 if (!bpage->page)
2612 goto fail_free_reader;
2613 if (cpu_buffer->ring_meta->head_buffer)
2614 rb_meta_buffer_update(cpu_buffer, bpage);
2615 bpage->range = 1;
2616
2617 atomic_inc(&cpu_buffer->resize_disabled);
2618 } else if (buffer->remote) {
2619 struct ring_buffer_desc *desc = ring_buffer_desc(buffer->remote->desc, cpu);
2620
2621 if (!desc)
2622 goto fail_free_reader;
2623
2624 cpu_buffer->remote = buffer->remote;
2625 cpu_buffer->meta_page = (struct trace_buffer_meta *)(void *)desc->meta_va;
2626 cpu_buffer->nr_pages = nr_pages;
2627 cpu_buffer->subbuf_ids = kzalloc_objs(*cpu_buffer->subbuf_ids,
2628 cpu_buffer->nr_pages + 1);
2629 if (!cpu_buffer->subbuf_ids)
2630 goto fail_free_reader;
2631
2632 /* Remote buffers are read-only and immutable */
2633 atomic_inc(&cpu_buffer->record_disabled);
2634 atomic_inc(&cpu_buffer->resize_disabled);
2635
2636 bpage->page = ring_buffer_desc_page(desc, cpu_buffer->meta_page->reader.id);
2637 if (!bpage->page)
2638 goto fail_free_reader;
2639
2640 bpage->range = 1;
2641 cpu_buffer->subbuf_ids[0] = bpage;
2642 } else {
2643 bpage->page = alloc_cpu_data(cpu, bpage->order);
2644 if (!bpage->page)
2645 goto fail_free_reader;
2646 }
2647
2648 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
2649 INIT_LIST_HEAD(&cpu_buffer->new_pages);
2650
2651 ret = rb_allocate_pages(cpu_buffer, nr_pages);
2652 if (ret < 0)
2653 goto fail_free_reader;
2654
2655 rb_meta_validate_events(cpu_buffer);
2656
2657 /* If the boot meta was valid then this has already been updated */
2658 meta = cpu_buffer->ring_meta;
2659 if (!meta || !meta->head_buffer ||
2660 !cpu_buffer->head_page || !cpu_buffer->commit_page || !cpu_buffer->tail_page) {
2661 if (meta && meta->head_buffer &&
2662 (cpu_buffer->head_page || cpu_buffer->commit_page || cpu_buffer->tail_page)) {
2663 pr_warn("Ring buffer meta buffers not all mapped\n");
2664 if (!cpu_buffer->head_page)
2665 pr_warn(" Missing head_page\n");
2666 if (!cpu_buffer->commit_page)
2667 pr_warn(" Missing commit_page\n");
2668 if (!cpu_buffer->tail_page)
2669 pr_warn(" Missing tail_page\n");
2670 }
2671
2672 cpu_buffer->head_page
2673 = list_entry(cpu_buffer->pages, struct buffer_page, list);
2674 cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
2675
2676 rb_head_page_activate(cpu_buffer);
2677
2678 if (cpu_buffer->ring_meta)
2679 meta->commit_buffer = meta->head_buffer;
2680 } else {
2681 /* The valid meta buffer still needs to activate the head page */
2682 rb_head_page_activate(cpu_buffer);
2683 }
2684
2685 return_ptr(cpu_buffer);
2686
2687 fail_free_reader:
2688 kfree(cpu_buffer->subbuf_ids);
2689 free_buffer_page(cpu_buffer->reader_page);
2690
2691 return NULL;
2692}
2693
2694static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
2695{
2696 struct list_head *head = cpu_buffer->pages;
2697 struct buffer_page *bpage, *tmp;
2698
2699 irq_work_sync(&cpu_buffer->irq_work.work);
2700
2701 if (cpu_buffer->remote)
2702 kfree(cpu_buffer->subbuf_ids);
2703
2704 free_buffer_page(cpu_buffer->reader_page);
2705
2706 if (head) {
2707 rb_head_page_deactivate(cpu_buffer);
2708
2709 list_for_each_entry_safe(bpage, tmp, head, list) {
2710 list_del_init(&bpage->list);
2711 free_buffer_page(bpage);
2712 }
2713 bpage = list_entry(head, struct buffer_page, list);
2714 free_buffer_page(bpage);
2715 }
2716
2717 free_pages((unsigned long)cpu_buffer->free_page.data, cpu_buffer->free_page.order);
2718
2719 kfree(cpu_buffer);
2720}
2721
2722#ifdef CONFIG_RING_BUFFER_PERSISTENT_INJECT
2723static void rb_test_inject_invalid_pages(struct trace_buffer *buffer)
2724{
2725 struct ring_buffer_per_cpu *cpu_buffer;
2726 struct ring_buffer_cpu_meta *meta;
2727 struct buffer_data_page *dpage;
2728 unsigned long entry_bytes = 0;
2729 unsigned int subbuf_size;
2730 unsigned long ptr;
2731 int invalid = 0;
2732 int cpu;
2733 int i;
2734
2735 if (!(buffer->flags & RB_FL_TESTING))
2736 return;
2737
2738 guard(preempt)();
2739 cpu = smp_processor_id();
2740
2741 cpu_buffer = buffer->buffers[cpu];
2742 if (!cpu_buffer)
2743 return;
2744 meta = cpu_buffer->ring_meta;
2745 if (!meta)
2746 return;
2747
2748 ptr = (unsigned long)rb_subbufs_from_meta(meta);
2749 subbuf_size = meta->subbuf_size;
2750
2751 for (i = 0; i < meta->nr_subbufs; i++) {
2752 unsigned long idx = meta->buffers[i];
2753
2754 dpage = (void *)(ptr + idx * subbuf_size);
2755 /* Skip unused pages */
2756 if (!rb_data_page_commit(dpage))
2757 continue;
2758
2759 /*
2760 * Invalidate even pages or multiples of 5. This will cause 3
2761 * contiguous invalidated(empty) pages.
2762 */
2763 if (!(i & 0x1) || !(i % 5)) {
2764 local_add(subbuf_size + 1, &dpage->commit);
2765 invalid++;
2766 } else {
2767 /* Count total commit bytes. */
2768 entry_bytes += rb_data_page_size(dpage);
2769 }
2770 }
2771
2772 pr_info("Inject invalidated %d pages on CPU%d, total size: %ld\n",
2773 invalid, cpu, (long)entry_bytes);
2774 meta->nr_invalid = invalid;
2775 meta->entry_bytes = entry_bytes;
2776}
2777#else /* !CONFIG_RING_BUFFER_PERSISTENT_INJECT */
2778#define rb_test_inject_invalid_pages(buffer) do { } while (0)
2779#endif
2780
2781/* Stop recording on a persistent buffer and flush cache if needed. */
2782static int rb_flush_buffer_cb(struct notifier_block *nb, unsigned long event, void *data)
2783{
2784 struct trace_buffer *buffer = container_of(nb, struct trace_buffer, flush_nb);
2785
2786 ring_buffer_record_off(buffer);
2787 rb_test_inject_invalid_pages(buffer);
2788 arch_ring_buffer_flush_range(buffer->range_addr_start, buffer->range_addr_end);
2789 return NOTIFY_DONE;
2790}
2791
2792static struct trace_buffer *alloc_buffer(unsigned long size, unsigned flags,
2793 int order, unsigned long start,
2794 unsigned long end,
2795 unsigned long scratch_size,
2796 struct lock_class_key *key,
2797 struct ring_buffer_remote *remote)
2798{
2799 struct trace_buffer *buffer __free(kfree) = NULL;
2800 unsigned int subbuf_size;
2801 unsigned long nr_pages;
2802 int bsize;
2803 int cpu;
2804 int ret;
2805
2806 /* keep it in its own cache line */
2807 buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
2808 GFP_KERNEL);
2809 if (!buffer)
2810 return NULL;
2811
2812 if (!zalloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
2813 return NULL;
2814
2815 buffer->subbuf_order = order;
2816 subbuf_size = (PAGE_SIZE << order);
2817
2818 buffer->flags = flags;
2819 buffer->clock = trace_clock_local;
2820 buffer->reader_lock_key = key;
2821
2822 init_irq_work(&buffer->irq_work.work, rb_wake_up_waiters);
2823 init_waitqueue_head(&buffer->irq_work.waiters);
2824
2825
2826 bsize = sizeof(void *) * nr_cpu_ids;
2827 buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
2828 GFP_KERNEL);
2829 if (!buffer->buffers)
2830 goto fail_free_cpumask;
2831
2832 cpu = raw_smp_processor_id();
2833
2834 /* If start/end are specified, then that overrides size */
2835 if (start && end) {
2836 unsigned long buffers_start;
2837 unsigned long ptr;
2838 int n;
2839
2840 /* Make sure that start is word aligned */
2841 start = ALIGN(start, sizeof(long));
2842
2843 /* scratch_size needs to be aligned too */
2844 scratch_size = ALIGN(scratch_size, sizeof(long));
2845
2846 /* Subtract the buffer meta data and word aligned */
2847 buffers_start = start + sizeof(struct ring_buffer_cpu_meta);
2848 buffers_start = ALIGN(buffers_start, sizeof(long));
2849 buffers_start += scratch_size;
2850
2851 /* Calculate the size for the per CPU data */
2852 size = end - buffers_start;
2853 size = size / nr_cpu_ids;
2854
2855 if (size < sizeof(struct ring_buffer_cpu_meta))
2856 goto fail_free_buffers;
2857 /*
2858 * The number of sub-buffers (nr_pages) is determined by the
2859 * total size allocated minus the meta data size.
2860 * Then that is divided by the number of per CPU buffers
2861 * needed, plus account for the integer array index that
2862 * will be appended to the meta data.
2863 */
2864 nr_pages = (size - sizeof(struct ring_buffer_cpu_meta)) /
2865 (subbuf_size + sizeof(int));
2866
2867 if (nr_pages > rb_static_max_pages())
2868 goto fail_free_buffers;
2869
2870 /* Need at least two pages plus the reader page */
2871 if (nr_pages < 3)
2872 goto fail_free_buffers;
2873
2874 again:
2875 /* Make sure that the size fits aligned */
2876 for (n = 0, ptr = buffers_start; n < nr_cpu_ids; n++) {
2877 ptr += sizeof(struct ring_buffer_cpu_meta) +
2878 sizeof(int) * nr_pages;
2879 ptr = ALIGN(ptr, subbuf_size);
2880 ptr += subbuf_size * nr_pages;
2881 }
2882 if (ptr > end) {
2883 if (nr_pages <= 3)
2884 goto fail_free_buffers;
2885 nr_pages--;
2886 goto again;
2887 }
2888
2889 /* nr_pages should not count the reader page */
2890 nr_pages--;
2891 buffer->range_addr_start = start;
2892 buffer->range_addr_end = end;
2893
2894 rb_range_meta_init(buffer, nr_pages, scratch_size);
2895 } else if (remote) {
2896 struct ring_buffer_desc *desc = ring_buffer_desc(remote->desc, cpu);
2897
2898 buffer->remote = remote;
2899 /* The writer is remote. This ring-buffer is read-only */
2900 atomic_inc(&buffer->record_disabled);
2901 nr_pages = desc->nr_page_va - 1;
2902
2903 if (nr_pages > rb_static_max_pages())
2904 goto fail_free_buffers;
2905
2906 if (nr_pages < 2)
2907 goto fail_free_buffers;
2908 } else {
2909 /* need at least two pages */
2910 nr_pages = DIV_ROUND_UP(size, rb_subbuf_capacity(buffer));
2911 if (nr_pages < 2)
2912 nr_pages = 2;
2913 }
2914
2915 cpumask_set_cpu(cpu, buffer->cpumask);
2916 buffer->buffers[cpu] = rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
2917 if (!buffer->buffers[cpu])
2918 goto fail_free_buffers;
2919
2920 ret = cpuhp_state_add_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
2921 if (ret < 0)
2922 goto fail_free_buffers;
2923
2924 mutex_init(&buffer->mutex);
2925
2926 /* Persistent ring buffer needs to flush cache before reboot. */
2927 if (start && end) {
2928 buffer->flush_nb.notifier_call = rb_flush_buffer_cb;
2929 atomic_notifier_chain_register(&panic_notifier_list, &buffer->flush_nb);
2930 }
2931
2932 return_ptr(buffer);
2933
2934 fail_free_buffers:
2935 for_each_buffer_cpu(buffer, cpu) {
2936 if (buffer->buffers[cpu])
2937 rb_free_cpu_buffer(buffer->buffers[cpu]);
2938 }
2939 kfree(buffer->buffers);
2940
2941 fail_free_cpumask:
2942 free_cpumask_var(buffer->cpumask);
2943
2944 return NULL;
2945}
2946
2947/**
2948 * __ring_buffer_alloc - allocate a new ring_buffer
2949 * @size: the size in bytes per cpu that is needed.
2950 * @flags: attributes to set for the ring buffer.
2951 * @key: ring buffer reader_lock_key.
2952 *
2953 * Currently the only flag that is available is the RB_FL_OVERWRITE
2954 * flag. This flag means that the buffer will overwrite old data
2955 * when the buffer wraps. If this flag is not set, the buffer will
2956 * drop data when the tail hits the head.
2957 */
2958struct trace_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
2959 struct lock_class_key *key)
2960{
2961 /* Default buffer page size - one system page */
2962 return alloc_buffer(size, flags, 0, 0, 0, 0, key, NULL);
2963
2964}
2965EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
2966
2967/**
2968 * __ring_buffer_alloc_range - allocate a new ring_buffer from existing memory
2969 * @size: the size in bytes per cpu that is needed.
2970 * @flags: attributes to set for the ring buffer.
2971 * @order: sub-buffer order
2972 * @start: start of allocated range
2973 * @range_size: size of allocated range
2974 * @scratch_size: size of scratch area (for preallocated memory buffers)
2975 * @key: ring buffer reader_lock_key.
2976 *
2977 * Currently the only flag that is available is the RB_FL_OVERWRITE
2978 * flag. This flag means that the buffer will overwrite old data
2979 * when the buffer wraps. If this flag is not set, the buffer will
2980 * drop data when the tail hits the head.
2981 */
2982struct trace_buffer *__ring_buffer_alloc_range(unsigned long size, unsigned flags,
2983 int order, unsigned long start,
2984 unsigned long range_size,
2985 unsigned long scratch_size,
2986 struct lock_class_key *key)
2987{
2988 return alloc_buffer(size, flags, order, start, start + range_size,
2989 scratch_size, key, NULL);
2990}
2991
2992/**
2993 * __ring_buffer_alloc_remote - allocate a new ring_buffer from a remote
2994 * @remote: Contains a description of the ring-buffer pages and remote callbacks.
2995 * @key: ring buffer reader_lock_key.
2996 */
2997struct trace_buffer *__ring_buffer_alloc_remote(struct ring_buffer_remote *remote,
2998 struct lock_class_key *key)
2999{
3000 return alloc_buffer(0, 0, 0, 0, 0, 0, key, remote);
3001}
3002
3003void *ring_buffer_meta_scratch(struct trace_buffer *buffer, unsigned int *size)
3004{
3005 struct ring_buffer_meta *meta;
3006 void *ptr;
3007
3008 if (!buffer || !buffer->meta)
3009 return NULL;
3010
3011 meta = buffer->meta;
3012
3013 ptr = (void *)ALIGN((unsigned long)meta + sizeof(*meta), sizeof(long));
3014
3015 if (size)
3016 *size = (void *)meta + meta->buffers_offset - ptr;
3017
3018 return ptr;
3019}
3020
3021/**
3022 * ring_buffer_free - free a ring buffer.
3023 * @buffer: the buffer to free.
3024 */
3025void
3026ring_buffer_free(struct trace_buffer *buffer)
3027{
3028 int cpu;
3029
3030 if (buffer->range_addr_start && buffer->range_addr_end)
3031 atomic_notifier_chain_unregister(&panic_notifier_list, &buffer->flush_nb);
3032
3033 cpuhp_state_remove_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
3034
3035 irq_work_sync(&buffer->irq_work.work);
3036
3037 for_each_buffer_cpu(buffer, cpu)
3038 rb_free_cpu_buffer(buffer->buffers[cpu]);
3039
3040 kfree(buffer->buffers);
3041 free_cpumask_var(buffer->cpumask);
3042
3043 kfree(buffer);
3044}
3045EXPORT_SYMBOL_GPL(ring_buffer_free);
3046
3047void ring_buffer_set_clock(struct trace_buffer *buffer,
3048 u64 (*clock)(void))
3049{
3050 buffer->clock = clock;
3051}
3052
3053void ring_buffer_set_time_stamp_abs(struct trace_buffer *buffer, bool abs)
3054{
3055 buffer->time_stamp_abs = abs;
3056}
3057
3058bool ring_buffer_time_stamp_abs(struct trace_buffer *buffer)
3059{
3060 return buffer->time_stamp_abs;
3061}
3062
3063static inline unsigned long rb_page_entries(struct buffer_page *bpage)
3064{
3065 return local_read(&bpage->entries) & RB_WRITE_MASK;
3066}
3067
3068static inline unsigned long rb_page_write(struct buffer_page *bpage)
3069{
3070 return local_read(&bpage->write) & RB_WRITE_MASK;
3071}
3072
3073static bool
3074rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned long nr_pages)
3075{
3076 struct list_head *tail_page, *to_remove, *next_page;
3077 struct buffer_page *to_remove_page, *tmp_iter_page;
3078 struct buffer_page *last_page, *first_page;
3079 unsigned long nr_removed;
3080 unsigned long head_bit;
3081 int page_entries;
3082
3083 head_bit = 0;
3084
3085 raw_spin_lock_irq(&cpu_buffer->reader_lock);
3086 atomic_inc(&cpu_buffer->record_disabled);
3087 /*
3088 * We don't race with the readers since we have acquired the reader
3089 * lock. We also don't race with writers after disabling recording.
3090 * This makes it easy to figure out the first and the last page to be
3091 * removed from the list. We unlink all the pages in between including
3092 * the first and last pages. This is done in a busy loop so that we
3093 * lose the least number of traces.
3094 * The pages are freed after we restart recording and unlock readers.
3095 */
3096 tail_page = &cpu_buffer->tail_page->list;
3097
3098 /*
3099 * tail page might be on reader page, we remove the next page
3100 * from the ring buffer
3101 */
3102 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
3103 tail_page = rb_list_head(tail_page->next);
3104 to_remove = tail_page;
3105
3106 /* start of pages to remove */
3107 first_page = list_entry(rb_list_head(to_remove->next),
3108 struct buffer_page, list);
3109
3110 for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
3111 to_remove = rb_list_head(to_remove)->next;
3112 head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
3113 }
3114 /* Read iterators need to reset themselves when some pages removed */
3115 cpu_buffer->pages_removed += nr_removed;
3116
3117 next_page = rb_list_head(to_remove)->next;
3118
3119 /*
3120 * Now we remove all pages between tail_page and next_page.
3121 * Make sure that we have head_bit value preserved for the
3122 * next page
3123 */
3124 tail_page->next = (struct list_head *)((unsigned long)next_page |
3125 head_bit);
3126 next_page = rb_list_head(next_page);
3127 next_page->prev = tail_page;
3128
3129 /* make sure pages points to a valid page in the ring buffer */
3130 cpu_buffer->pages = next_page;
3131 cpu_buffer->cnt++;
3132
3133 /* update head page */
3134 if (head_bit)
3135 cpu_buffer->head_page = list_entry(next_page,
3136 struct buffer_page, list);
3137
3138 /* pages are removed, resume tracing and then free the pages */
3139 atomic_dec(&cpu_buffer->record_disabled);
3140 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
3141
3142 RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
3143
3144 /* last buffer page to remove */
3145 last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
3146 list);
3147 tmp_iter_page = first_page;
3148
3149 do {
3150 cond_resched();
3151
3152 to_remove_page = tmp_iter_page;
3153 rb_inc_page(&tmp_iter_page);
3154
3155 /* update the counters */
3156 page_entries = rb_page_entries(to_remove_page);
3157 if (page_entries) {
3158 /*
3159 * If something was added to this page, it was full
3160 * since it is not the tail page. So we deduct the
3161 * bytes consumed in ring buffer from here.
3162 * Increment overrun to account for the lost events.
3163 */
3164 local_add(page_entries, &cpu_buffer->overrun);
3165 local_sub(rb_page_commit(to_remove_page), &cpu_buffer->entries_bytes);
3166 local_inc(&cpu_buffer->pages_lost);
3167 }
3168
3169 /*
3170 * We have already removed references to this list item, just
3171 * free up the buffer_page and its page
3172 */
3173 free_buffer_page(to_remove_page);
3174 nr_removed--;
3175
3176 } while (to_remove_page != last_page);
3177
3178 RB_WARN_ON(cpu_buffer, nr_removed);
3179
3180 return nr_removed == 0;
3181}
3182
3183static bool
3184rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
3185{
3186 struct list_head *pages = &cpu_buffer->new_pages;
3187 unsigned long flags;
3188 bool success;
3189 int retries;
3190
3191 /* Can be called at early boot up, where interrupts must not been enabled */
3192 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3193 /*
3194 * We are holding the reader lock, so the reader page won't be swapped
3195 * in the ring buffer. Now we are racing with the writer trying to
3196 * move head page and the tail page.
3197 * We are going to adapt the reader page update process where:
3198 * 1. We first splice the start and end of list of new pages between
3199 * the head page and its previous page.
3200 * 2. We cmpxchg the prev_page->next to point from head page to the
3201 * start of new pages list.
3202 * 3. Finally, we update the head->prev to the end of new list.
3203 *
3204 * We will try this process 10 times, to make sure that we don't keep
3205 * spinning.
3206 */
3207 retries = 10;
3208 success = false;
3209 while (retries--) {
3210 struct list_head *head_page, *prev_page;
3211 struct list_head *last_page, *first_page;
3212 struct list_head *head_page_with_bit;
3213 struct buffer_page *hpage = rb_set_head_page(cpu_buffer);
3214
3215 if (!hpage)
3216 break;
3217 head_page = &hpage->list;
3218 prev_page = head_page->prev;
3219
3220 first_page = pages->next;
3221 last_page = pages->prev;
3222
3223 head_page_with_bit = (struct list_head *)
3224 ((unsigned long)head_page | RB_PAGE_HEAD);
3225
3226 last_page->next = head_page_with_bit;
3227 first_page->prev = prev_page;
3228
3229 /* caution: head_page_with_bit gets updated on cmpxchg failure */
3230 if (try_cmpxchg(&prev_page->next,
3231 &head_page_with_bit, first_page)) {
3232 /*
3233 * yay, we replaced the page pointer to our new list,
3234 * now, we just have to update to head page's prev
3235 * pointer to point to end of list
3236 */
3237 head_page->prev = last_page;
3238 cpu_buffer->cnt++;
3239 success = true;
3240 break;
3241 }
3242 }
3243
3244 if (success)
3245 INIT_LIST_HEAD(pages);
3246 /*
3247 * If we weren't successful in adding in new pages, warn and stop
3248 * tracing
3249 */
3250 RB_WARN_ON(cpu_buffer, !success);
3251 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3252
3253 /* free pages if they weren't inserted */
3254 if (!success) {
3255 struct buffer_page *bpage, *tmp;
3256 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
3257 list) {
3258 list_del_init(&bpage->list);
3259 free_buffer_page(bpage);
3260 }
3261 }
3262 return success;
3263}
3264
3265static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
3266{
3267 bool success;
3268
3269 if (cpu_buffer->nr_pages_to_update > 0)
3270 success = rb_insert_pages(cpu_buffer);
3271 else
3272 success = rb_remove_pages(cpu_buffer,
3273 -cpu_buffer->nr_pages_to_update);
3274
3275 if (success)
3276 cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
3277}
3278
3279static void update_pages_handler(struct work_struct *work)
3280{
3281 struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
3282 struct ring_buffer_per_cpu, update_pages_work);
3283 rb_update_pages(cpu_buffer);
3284 complete(&cpu_buffer->update_done);
3285}
3286
3287/**
3288 * ring_buffer_resize - resize the ring buffer
3289 * @buffer: the buffer to resize.
3290 * @size: the new size.
3291 * @cpu_id: the cpu buffer to resize
3292 *
3293 * Minimum size is 2 * rb_subbuf_capacity(buffer).
3294 *
3295 * Returns 0 on success and < 0 on failure.
3296 */
3297int ring_buffer_resize(struct trace_buffer *buffer, unsigned long size,
3298 int cpu_id)
3299{
3300 struct ring_buffer_per_cpu *cpu_buffer;
3301 unsigned long nr_pages;
3302 int cpu, err;
3303
3304 /*
3305 * Always succeed at resizing a non-existent buffer:
3306 */
3307 if (!buffer)
3308 return 0;
3309
3310 /* Make sure the requested buffer exists */
3311 if (cpu_id != RING_BUFFER_ALL_CPUS &&
3312 !cpumask_test_cpu(cpu_id, buffer->cpumask))
3313 return 0;
3314
3315 /*
3316 * Keep CPUs from coming online while resizing to synchronize
3317 * with new per CPU buffers being created.
3318 */
3319 guard(cpus_read_lock)();
3320
3321 /* prevent another thread from changing buffer sizes */
3322 mutex_lock(&buffer->mutex);
3323 atomic_inc(&buffer->resizing);
3324
3325 nr_pages = DIV_ROUND_UP(size, rb_subbuf_capacity(buffer));
3326
3327 /* we need a minimum of two pages */
3328 if (nr_pages < 2)
3329 nr_pages = 2;
3330
3331 if (cpu_id == RING_BUFFER_ALL_CPUS) {
3332 /*
3333 * Don't succeed if resizing is disabled, as a reader might be
3334 * manipulating the ring buffer and is expecting a sane state while
3335 * this is true.
3336 */
3337 for_each_buffer_cpu(buffer, cpu) {
3338 cpu_buffer = buffer->buffers[cpu];
3339 if (atomic_read(&cpu_buffer->resize_disabled)) {
3340 err = -EBUSY;
3341 goto out_err_unlock;
3342 }
3343 }
3344
3345 /* calculate the pages to update */
3346 for_each_buffer_cpu(buffer, cpu) {
3347 cpu_buffer = buffer->buffers[cpu];
3348
3349 cpu_buffer->nr_pages_to_update = nr_pages -
3350 cpu_buffer->nr_pages;
3351 /*
3352 * nothing more to do for removing pages or no update
3353 */
3354 if (cpu_buffer->nr_pages_to_update <= 0)
3355 continue;
3356 /*
3357 * to add pages, make sure all new pages can be
3358 * allocated without receiving ENOMEM
3359 */
3360 INIT_LIST_HEAD(&cpu_buffer->new_pages);
3361 if (__rb_allocate_pages(cpu_buffer, cpu_buffer->nr_pages_to_update,
3362 &cpu_buffer->new_pages)) {
3363 /* not enough memory for new pages */
3364 err = -ENOMEM;
3365 goto out_err;
3366 }
3367
3368 cond_resched();
3369 }
3370
3371 /*
3372 * Fire off all the required work handlers
3373 * We can't schedule on offline CPUs, but it's not necessary
3374 * since we can change their buffer sizes without any race.
3375 */
3376 for_each_buffer_cpu(buffer, cpu) {
3377 cpu_buffer = buffer->buffers[cpu];
3378 if (!cpu_buffer->nr_pages_to_update)
3379 continue;
3380
3381 /* Can't run something on an offline CPU. */
3382 if (!cpu_online(cpu)) {
3383 rb_update_pages(cpu_buffer);
3384 cpu_buffer->nr_pages_to_update = 0;
3385 } else {
3386 /* Run directly if possible. */
3387 migrate_disable();
3388 if (cpu != smp_processor_id()) {
3389 migrate_enable();
3390 schedule_work_on(cpu,
3391 &cpu_buffer->update_pages_work);
3392 } else {
3393 update_pages_handler(&cpu_buffer->update_pages_work);
3394 migrate_enable();
3395 }
3396 }
3397 }
3398
3399 /* wait for all the updates to complete */
3400 for_each_buffer_cpu(buffer, cpu) {
3401 cpu_buffer = buffer->buffers[cpu];
3402 if (!cpu_buffer->nr_pages_to_update)
3403 continue;
3404
3405 if (cpu_online(cpu))
3406 wait_for_completion(&cpu_buffer->update_done);
3407 cpu_buffer->nr_pages_to_update = 0;
3408 }
3409
3410 } else {
3411 cpu_buffer = buffer->buffers[cpu_id];
3412
3413 if (nr_pages == cpu_buffer->nr_pages)
3414 goto out;
3415
3416 /*
3417 * Don't succeed if resizing is disabled, as a reader might be
3418 * manipulating the ring buffer and is expecting a sane state while
3419 * this is true.
3420 */
3421 if (atomic_read(&cpu_buffer->resize_disabled)) {
3422 err = -EBUSY;
3423 goto out_err_unlock;
3424 }
3425
3426 cpu_buffer->nr_pages_to_update = nr_pages -
3427 cpu_buffer->nr_pages;
3428
3429 INIT_LIST_HEAD(&cpu_buffer->new_pages);
3430 if (cpu_buffer->nr_pages_to_update > 0 &&
3431 __rb_allocate_pages(cpu_buffer, cpu_buffer->nr_pages_to_update,
3432 &cpu_buffer->new_pages)) {
3433 err = -ENOMEM;
3434 goto out_err;
3435 }
3436
3437 /* Can't run something on an offline CPU. */
3438 if (!cpu_online(cpu_id))
3439 rb_update_pages(cpu_buffer);
3440 else {
3441 /* Run directly if possible. */
3442 migrate_disable();
3443 if (cpu_id == smp_processor_id()) {
3444 rb_update_pages(cpu_buffer);
3445 migrate_enable();
3446 } else {
3447 migrate_enable();
3448 schedule_work_on(cpu_id,
3449 &cpu_buffer->update_pages_work);
3450 wait_for_completion(&cpu_buffer->update_done);
3451 }
3452 }
3453
3454 cpu_buffer->nr_pages_to_update = 0;
3455 }
3456
3457 out:
3458 /*
3459 * The ring buffer resize can happen with the ring buffer
3460 * enabled, so that the update disturbs the tracing as little
3461 * as possible. But if the buffer is disabled, we do not need
3462 * to worry about that, and we can take the time to verify
3463 * that the buffer is not corrupt.
3464 */
3465 if (atomic_read(&buffer->record_disabled)) {
3466 atomic_inc(&buffer->record_disabled);
3467 /*
3468 * Even though the buffer was disabled, we must make sure
3469 * that it is truly disabled before calling rb_check_pages.
3470 * There could have been a race between checking
3471 * record_disable and incrementing it.
3472 */
3473 synchronize_rcu();
3474 for_each_buffer_cpu(buffer, cpu) {
3475 cpu_buffer = buffer->buffers[cpu];
3476 rb_check_pages(cpu_buffer);
3477 }
3478 atomic_dec(&buffer->record_disabled);
3479 }
3480
3481 atomic_dec(&buffer->resizing);
3482 mutex_unlock(&buffer->mutex);
3483 return 0;
3484
3485 out_err:
3486 for_each_buffer_cpu(buffer, cpu) {
3487 struct buffer_page *bpage, *tmp;
3488
3489 cpu_buffer = buffer->buffers[cpu];
3490 cpu_buffer->nr_pages_to_update = 0;
3491
3492 if (list_empty(&cpu_buffer->new_pages))
3493 continue;
3494
3495 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
3496 list) {
3497 list_del_init(&bpage->list);
3498 free_buffer_page(bpage);
3499
3500 cond_resched();
3501 }
3502 }
3503 out_err_unlock:
3504 atomic_dec(&buffer->resizing);
3505 mutex_unlock(&buffer->mutex);
3506 return err;
3507}
3508EXPORT_SYMBOL_GPL(ring_buffer_resize);
3509
3510void ring_buffer_change_overwrite(struct trace_buffer *buffer, int val)
3511{
3512 mutex_lock(&buffer->mutex);
3513 if (val)
3514 buffer->flags |= RB_FL_OVERWRITE;
3515 else
3516 buffer->flags &= ~RB_FL_OVERWRITE;
3517 mutex_unlock(&buffer->mutex);
3518}
3519EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
3520
3521static __always_inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
3522{
3523 return bpage->page->data + index;
3524}
3525
3526static __always_inline struct ring_buffer_event *
3527rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
3528{
3529 return __rb_page_index(cpu_buffer->reader_page,
3530 cpu_buffer->reader_page->read);
3531}
3532
3533static struct ring_buffer_event *
3534rb_iter_head_event(struct ring_buffer_iter *iter)
3535{
3536 struct ring_buffer_event *event;
3537 struct buffer_page *iter_head_page = iter->head_page;
3538 unsigned long commit;
3539 unsigned length;
3540
3541 if (iter->head != iter->next_event)
3542 return iter->event;
3543
3544 /*
3545 * When the writer goes across pages, it issues a cmpxchg which
3546 * is a mb(), which will synchronize with the rmb here.
3547 * (see rb_tail_page_update() and __rb_reserve_next())
3548 */
3549 commit = rb_page_size(iter_head_page);
3550 smp_rmb();
3551
3552 /* An event needs to be at least 8 bytes in size */
3553 if (iter->head > commit - 8)
3554 goto reset;
3555
3556 event = __rb_page_index(iter_head_page, iter->head);
3557 length = rb_event_length(event);
3558
3559 /*
3560 * READ_ONCE() doesn't work on functions and we don't want the
3561 * compiler doing any crazy optimizations with length.
3562 */
3563 barrier();
3564
3565 if ((iter->head + length) > commit || length > iter->event_size)
3566 /* Writer corrupted the read? */
3567 goto reset;
3568
3569 memcpy(iter->event, event, length);
3570 /*
3571 * If the page stamp is still the same after this rmb() then the
3572 * event was safely copied without the writer entering the page.
3573 */
3574 smp_rmb();
3575
3576 /* Make sure the page didn't change since we read this */
3577 if (iter->page_stamp != iter_head_page->page->time_stamp ||
3578 commit > rb_page_size(iter_head_page))
3579 goto reset;
3580
3581 iter->next_event = iter->head + length;
3582 return iter->event;
3583 reset:
3584 /* Reset to the beginning */
3585 iter->page_stamp = iter->read_stamp = iter->head_page->page->time_stamp;
3586 iter->head = 0;
3587 iter->next_event = 0;
3588 iter->missed_events = 1;
3589 return NULL;
3590}
3591
3592static __always_inline unsigned
3593rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
3594{
3595 return rb_page_commit(cpu_buffer->commit_page);
3596}
3597
3598static __always_inline unsigned
3599rb_event_index(struct ring_buffer_per_cpu *cpu_buffer, struct ring_buffer_event *event)
3600{
3601 unsigned long addr = (unsigned long)event;
3602
3603 addr &= (unsigned long)rb_subbuf_size(cpu_buffer->buffer) - 1;
3604
3605 return addr - BUF_PAGE_HDR_SIZE;
3606}
3607
3608static void rb_inc_iter(struct ring_buffer_iter *iter)
3609{
3610 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3611
3612 /*
3613 * The iterator could be on the reader page (it starts there).
3614 * But the head could have moved, since the reader was
3615 * found. Check for this case and assign the iterator
3616 * to the head page instead of next.
3617 */
3618 if (iter->head_page == cpu_buffer->reader_page)
3619 iter->head_page = rb_set_head_page(cpu_buffer);
3620 else
3621 rb_inc_page(&iter->head_page);
3622
3623 if (rb_page_commit(iter->head_page) & RB_MISSED_EVENTS)
3624 iter->missed_events = -1;
3625
3626 iter->page_stamp = iter->read_stamp = iter->head_page->page->time_stamp;
3627 iter->head = 0;
3628 iter->next_event = 0;
3629}
3630
3631/* Return the index into the sub-buffers for a given sub-buffer */
3632static int rb_meta_subbuf_idx(struct ring_buffer_cpu_meta *meta, void *subbuf)
3633{
3634 void *subbuf_array = rb_subbufs_from_meta(meta);
3635
3636 return (subbuf - subbuf_array) / meta->subbuf_size;
3637}
3638
3639static void rb_update_meta_head(struct ring_buffer_per_cpu *cpu_buffer,
3640 struct buffer_page *next_page)
3641{
3642 struct ring_buffer_cpu_meta *meta = cpu_buffer->ring_meta;
3643 unsigned long old_head = (unsigned long)next_page->page;
3644 unsigned long new_head;
3645
3646 rb_inc_page(&next_page);
3647 new_head = (unsigned long)next_page->page;
3648
3649 /*
3650 * Only move it forward once, if something else came in and
3651 * moved it forward, then we don't want to touch it.
3652 */
3653 (void)cmpxchg(&meta->head_buffer, old_head, new_head);
3654}
3655
3656static void rb_update_meta_reader(struct ring_buffer_per_cpu *cpu_buffer,
3657 struct buffer_page *reader)
3658{
3659 struct ring_buffer_cpu_meta *meta = cpu_buffer->ring_meta;
3660 void *old_reader = cpu_buffer->reader_page->page;
3661 void *new_reader = reader->page;
3662 int id;
3663
3664 id = reader->id;
3665 cpu_buffer->reader_page->id = id;
3666 reader->id = 0;
3667
3668 meta->buffers[0] = rb_meta_subbuf_idx(meta, new_reader);
3669 meta->buffers[id] = rb_meta_subbuf_idx(meta, old_reader);
3670
3671 /* The head pointer is the one after the reader */
3672 rb_update_meta_head(cpu_buffer, reader);
3673}
3674
3675/*
3676 * rb_handle_head_page - writer hit the head page
3677 *
3678 * Returns: +1 to retry page
3679 * 0 to continue
3680 * -1 on error
3681 */
3682static int
3683rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
3684 struct buffer_page *tail_page,
3685 struct buffer_page *next_page)
3686{
3687 struct buffer_page *new_head;
3688 int entries;
3689 int type;
3690 int ret;
3691
3692 entries = rb_page_entries(next_page);
3693
3694 /*
3695 * The hard part is here. We need to move the head
3696 * forward, and protect against both readers on
3697 * other CPUs and writers coming in via interrupts.
3698 */
3699 type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
3700 RB_PAGE_HEAD);
3701
3702 /*
3703 * type can be one of four:
3704 * NORMAL - an interrupt already moved it for us
3705 * HEAD - we are the first to get here.
3706 * UPDATE - we are the interrupt interrupting
3707 * a current move.
3708 * MOVED - a reader on another CPU moved the next
3709 * pointer to its reader page. Give up
3710 * and try again.
3711 */
3712
3713 switch (type) {
3714 case RB_PAGE_HEAD:
3715 /*
3716 * We changed the head to UPDATE, thus
3717 * it is our responsibility to update
3718 * the counters.
3719 */
3720 local_add(entries, &cpu_buffer->overrun);
3721 local_sub(rb_page_commit(next_page), &cpu_buffer->entries_bytes);
3722 local_inc(&cpu_buffer->pages_lost);
3723
3724 if (cpu_buffer->ring_meta)
3725 rb_update_meta_head(cpu_buffer, next_page);
3726 /*
3727 * The entries will be zeroed out when we move the
3728 * tail page.
3729 */
3730
3731 /* still more to do */
3732 break;
3733
3734 case RB_PAGE_UPDATE:
3735 /*
3736 * This is an interrupt that interrupt the
3737 * previous update. Still more to do.
3738 */
3739 break;
3740 case RB_PAGE_NORMAL:
3741 /*
3742 * An interrupt came in before the update
3743 * and processed this for us.
3744 * Nothing left to do.
3745 */
3746 return 1;
3747 case RB_PAGE_MOVED:
3748 /*
3749 * The reader is on another CPU and just did
3750 * a swap with our next_page.
3751 * Try again.
3752 */
3753 return 1;
3754 default:
3755 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
3756 return -1;
3757 }
3758
3759 /*
3760 * Now that we are here, the old head pointer is
3761 * set to UPDATE. This will keep the reader from
3762 * swapping the head page with the reader page.
3763 * The reader (on another CPU) will spin till
3764 * we are finished.
3765 *
3766 * We just need to protect against interrupts
3767 * doing the job. We will set the next pointer
3768 * to HEAD. After that, we set the old pointer
3769 * to NORMAL, but only if it was HEAD before.
3770 * otherwise we are an interrupt, and only
3771 * want the outer most commit to reset it.
3772 */
3773 new_head = next_page;
3774 rb_inc_page(&new_head);
3775
3776 ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
3777 RB_PAGE_NORMAL);
3778
3779 /*
3780 * Valid returns are:
3781 * HEAD - an interrupt came in and already set it.
3782 * NORMAL - One of two things:
3783 * 1) We really set it.
3784 * 2) A bunch of interrupts came in and moved
3785 * the page forward again.
3786 */
3787 switch (ret) {
3788 case RB_PAGE_HEAD:
3789 case RB_PAGE_NORMAL:
3790 /* OK */
3791 break;
3792 default:
3793 RB_WARN_ON(cpu_buffer, 1);
3794 return -1;
3795 }
3796
3797 /*
3798 * It is possible that an interrupt came in,
3799 * set the head up, then more interrupts came in
3800 * and moved it again. When we get back here,
3801 * the page would have been set to NORMAL but we
3802 * just set it back to HEAD.
3803 *
3804 * How do you detect this? Well, if that happened
3805 * the tail page would have moved.
3806 */
3807 if (ret == RB_PAGE_NORMAL) {
3808 struct buffer_page *buffer_tail_page;
3809
3810 buffer_tail_page = READ_ONCE(cpu_buffer->tail_page);
3811 /*
3812 * If the tail had moved passed next, then we need
3813 * to reset the pointer.
3814 */
3815 if (buffer_tail_page != tail_page &&
3816 buffer_tail_page != next_page)
3817 rb_head_page_set_normal(cpu_buffer, new_head,
3818 next_page,
3819 RB_PAGE_HEAD);
3820 }
3821
3822 /*
3823 * If this was the outer most commit (the one that
3824 * changed the original pointer from HEAD to UPDATE),
3825 * then it is up to us to reset it to NORMAL.
3826 */
3827 if (type == RB_PAGE_HEAD) {
3828 ret = rb_head_page_set_normal(cpu_buffer, next_page,
3829 tail_page,
3830 RB_PAGE_UPDATE);
3831 if (RB_WARN_ON(cpu_buffer,
3832 ret != RB_PAGE_UPDATE))
3833 return -1;
3834 }
3835
3836 return 0;
3837}
3838
3839static inline void
3840rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
3841 unsigned long tail, struct rb_event_info *info)
3842{
3843 struct buffer_page *tail_page = info->tail_page;
3844 unsigned long bsize = rb_page_capacity(tail_page);
3845 struct ring_buffer_event *event;
3846 unsigned long length = info->length;
3847
3848 /*
3849 * Only the event that crossed the page boundary
3850 * must fill the old tail_page with padding.
3851 */
3852 if (tail >= bsize) {
3853 /*
3854 * If the page was filled, then we still need
3855 * to update the real_end. Reset it to zero
3856 * and the reader will ignore it.
3857 */
3858 if (tail == bsize)
3859 tail_page->real_end = 0;
3860
3861 local_sub(length, &tail_page->write);
3862 return;
3863 }
3864
3865 event = __rb_page_index(tail_page, tail);
3866
3867 /*
3868 * Save the original length to the meta data.
3869 * This will be used by the reader to add lost event
3870 * counter.
3871 */
3872 tail_page->real_end = tail;
3873
3874 /*
3875 * If this event is bigger than the minimum size, then
3876 * we need to be careful that we don't subtract the
3877 * write counter enough to allow another writer to slip
3878 * in on this page.
3879 * We put in a discarded commit instead, to make sure
3880 * that this space is not used again, and this space will
3881 * not be accounted into 'entries_bytes'.
3882 *
3883 * If we are less than the minimum size, we don't need to
3884 * worry about it.
3885 */
3886 if (tail > (bsize - RB_EVNT_MIN_SIZE)) {
3887 /* No room for any events */
3888
3889 /* Mark the rest of the page with padding */
3890 rb_event_set_padding(event);
3891
3892 /* Make sure the padding is visible before the write update */
3893 smp_wmb();
3894
3895 /* Set the write back to the previous setting */
3896 local_sub(length, &tail_page->write);
3897 return;
3898 }
3899
3900 /* Put in a discarded event */
3901 event->array[0] = (bsize - tail) - RB_EVNT_HDR_SIZE;
3902 event->type_len = RINGBUF_TYPE_PADDING;
3903 /* time delta must be non zero */
3904 event->time_delta = 1;
3905
3906 /* account for padding bytes */
3907 local_add(bsize - tail, &cpu_buffer->entries_bytes);
3908
3909 /* Make sure the padding is visible before the tail_page->write update */
3910 smp_wmb();
3911
3912 /* Set write to end of buffer */
3913 length = (tail + length) - bsize;
3914 local_sub(length, &tail_page->write);
3915}
3916
3917static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer);
3918
3919/*
3920 * This is the slow path, force gcc not to inline it.
3921 */
3922static noinline struct ring_buffer_event *
3923rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
3924 unsigned long tail, struct rb_event_info *info)
3925{
3926 struct buffer_page *tail_page = info->tail_page;
3927 struct buffer_page *commit_page = cpu_buffer->commit_page;
3928 struct trace_buffer *buffer = cpu_buffer->buffer;
3929 struct buffer_page *next_page;
3930 int ret;
3931
3932 next_page = tail_page;
3933
3934 rb_inc_page(&next_page);
3935
3936 /*
3937 * If for some reason, we had an interrupt storm that made
3938 * it all the way around the buffer, bail, and warn
3939 * about it.
3940 */
3941 if (unlikely(next_page == commit_page)) {
3942 local_inc(&cpu_buffer->commit_overrun);
3943 goto out_reset;
3944 }
3945
3946 /*
3947 * This is where the fun begins!
3948 *
3949 * We are fighting against races between a reader that
3950 * could be on another CPU trying to swap its reader
3951 * page with the buffer head.
3952 *
3953 * We are also fighting against interrupts coming in and
3954 * moving the head or tail on us as well.
3955 *
3956 * If the next page is the head page then we have filled
3957 * the buffer, unless the commit page is still on the
3958 * reader page.
3959 */
3960 if (rb_is_head_page(next_page, &tail_page->list)) {
3961
3962 /*
3963 * If the commit is not on the reader page, then
3964 * move the header page.
3965 */
3966 if (!rb_is_reader_page(cpu_buffer->commit_page)) {
3967 /*
3968 * If we are not in overwrite mode,
3969 * this is easy, just stop here.
3970 */
3971 if (!(buffer->flags & RB_FL_OVERWRITE)) {
3972 local_inc(&cpu_buffer->dropped_events);
3973 goto out_reset;
3974 }
3975
3976 ret = rb_handle_head_page(cpu_buffer,
3977 tail_page,
3978 next_page);
3979 if (ret < 0)
3980 goto out_reset;
3981 if (ret)
3982 goto out_again;
3983 } else {
3984 /*
3985 * We need to be careful here too. The
3986 * commit page could still be on the reader
3987 * page. We could have a small buffer, and
3988 * have filled up the buffer with events
3989 * from interrupts and such, and wrapped.
3990 *
3991 * Note, if the tail page is also on the
3992 * reader_page, we let it move out.
3993 */
3994 if (unlikely((cpu_buffer->commit_page !=
3995 cpu_buffer->tail_page) &&
3996 (cpu_buffer->commit_page ==
3997 cpu_buffer->reader_page))) {
3998 local_inc(&cpu_buffer->commit_overrun);
3999 goto out_reset;
4000 }
4001 }
4002 }
4003
4004 rb_tail_page_update(cpu_buffer, tail_page, next_page);
4005
4006 out_again:
4007
4008 rb_reset_tail(cpu_buffer, tail, info);
4009
4010 /* Commit what we have for now. */
4011 rb_end_commit(cpu_buffer);
4012 /* rb_end_commit() decs committing */
4013 local_inc(&cpu_buffer->committing);
4014
4015 /* fail and let the caller try again */
4016 return ERR_PTR(-EAGAIN);
4017
4018 out_reset:
4019 /* reset write */
4020 rb_reset_tail(cpu_buffer, tail, info);
4021
4022 return NULL;
4023}
4024
4025/* Slow path */
4026static struct ring_buffer_event *
4027rb_add_time_stamp(struct ring_buffer_per_cpu *cpu_buffer,
4028 struct ring_buffer_event *event, u64 delta, bool abs)
4029{
4030 if (abs)
4031 event->type_len = RINGBUF_TYPE_TIME_STAMP;
4032 else
4033 event->type_len = RINGBUF_TYPE_TIME_EXTEND;
4034
4035 /* Not the first event on the page, or not delta? */
4036 if (abs || rb_event_index(cpu_buffer, event)) {
4037 event->time_delta = delta & TS_MASK;
4038 event->array[0] = delta >> TS_SHIFT;
4039 } else {
4040 /* nope, just zero it */
4041 event->time_delta = 0;
4042 event->array[0] = 0;
4043 }
4044
4045 return skip_time_extend(event);
4046}
4047
4048static void
4049rb_check_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
4050 struct rb_event_info *info)
4051{
4052 u64 write_stamp;
4053
4054 WARN_ONCE(1, "Delta way too big! %llu ts=%llu before=%llu after=%llu write stamp=%llu\n%s",
4055 (unsigned long long)info->delta,
4056 (unsigned long long)info->ts,
4057 (unsigned long long)info->before,
4058 (unsigned long long)info->after,
4059 (unsigned long long)({rb_time_read(&cpu_buffer->write_stamp, &write_stamp); write_stamp;}),
4060 sched_clock_stable() ? "" :
4061 "If you just came from a suspend/resume,\n"
4062 "please switch to the trace global clock:\n"
4063 " echo global > /sys/kernel/tracing/trace_clock\n"
4064 "or add trace_clock=global to the kernel command line\n");
4065}
4066
4067static void rb_add_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
4068 struct ring_buffer_event **event,
4069 struct rb_event_info *info,
4070 u64 *delta,
4071 unsigned int *length)
4072{
4073 bool abs = info->add_timestamp &
4074 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE);
4075
4076 if (unlikely(info->delta > (1ULL << 59))) {
4077 /*
4078 * Some timers can use more than 59 bits, and when a timestamp
4079 * is added to the buffer, it will lose those bits.
4080 */
4081 if (abs && (info->ts & TS_MSB)) {
4082 info->delta &= ABS_TS_MASK;
4083
4084 /* did the clock go backwards */
4085 } else if (info->before == info->after && info->before > info->ts) {
4086 /* not interrupted */
4087 static int once;
4088
4089 /*
4090 * This is possible with a recalibrating of the TSC.
4091 * Do not produce a call stack, but just report it.
4092 */
4093 if (!once) {
4094 once++;
4095 pr_warn("Ring buffer clock went backwards: %llu -> %llu\n",
4096 info->before, info->ts);
4097 }
4098 } else
4099 rb_check_timestamp(cpu_buffer, info);
4100 if (!abs)
4101 info->delta = 0;
4102 }
4103 *event = rb_add_time_stamp(cpu_buffer, *event, info->delta, abs);
4104 *length -= RB_LEN_TIME_EXTEND;
4105 *delta = 0;
4106}
4107
4108/**
4109 * rb_update_event - update event type and data
4110 * @cpu_buffer: The per cpu buffer of the @event
4111 * @event: the event to update
4112 * @info: The info to update the @event with (contains length and delta)
4113 *
4114 * Update the type and data fields of the @event. The length
4115 * is the actual size that is written to the ring buffer,
4116 * and with this, we can determine what to place into the
4117 * data field.
4118 */
4119static void
4120rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
4121 struct ring_buffer_event *event,
4122 struct rb_event_info *info)
4123{
4124 unsigned length = info->length;
4125 u64 delta = info->delta;
4126 unsigned int nest = local_read(&cpu_buffer->committing) - 1;
4127
4128 if (!WARN_ON_ONCE(nest >= MAX_NEST))
4129 cpu_buffer->event_stamp[nest] = info->ts;
4130
4131 /*
4132 * If we need to add a timestamp, then we
4133 * add it to the start of the reserved space.
4134 */
4135 if (unlikely(info->add_timestamp))
4136 rb_add_timestamp(cpu_buffer, &event, info, &delta, &length);
4137
4138 event->time_delta = delta;
4139 length -= RB_EVNT_HDR_SIZE;
4140 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
4141 event->type_len = 0;
4142 event->array[0] = length;
4143 } else
4144 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
4145}
4146
4147static unsigned rb_calculate_event_length(unsigned length)
4148{
4149 struct ring_buffer_event event; /* Used only for sizeof array */
4150
4151 /* zero length can cause confusions */
4152 if (!length)
4153 length++;
4154
4155 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
4156 length += sizeof(event.array[0]);
4157
4158 length += RB_EVNT_HDR_SIZE;
4159 length = ALIGN(length, RB_ARCH_ALIGNMENT);
4160
4161 /*
4162 * In case the time delta is larger than the 27 bits for it
4163 * in the header, we need to add a timestamp. If another
4164 * event comes in when trying to discard this one to increase
4165 * the length, then the timestamp will be added in the allocated
4166 * space of this event. If length is bigger than the size needed
4167 * for the TIME_EXTEND, then padding has to be used. The events
4168 * length must be either RB_LEN_TIME_EXTEND, or greater than or equal
4169 * to RB_LEN_TIME_EXTEND + 8, as 8 is the minimum size for padding.
4170 * As length is a multiple of 4, we only need to worry if it
4171 * is 12 (RB_LEN_TIME_EXTEND + 4).
4172 */
4173 if (length == RB_LEN_TIME_EXTEND + RB_ALIGNMENT)
4174 length += RB_ALIGNMENT;
4175
4176 return length;
4177}
4178
4179static inline bool
4180rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
4181 struct ring_buffer_event *event)
4182{
4183 unsigned long new_index, old_index;
4184 struct buffer_page *bpage;
4185 unsigned long addr;
4186
4187 new_index = rb_event_index(cpu_buffer, event);
4188 old_index = new_index + rb_event_ts_length(event);
4189 addr = rb_subbuf_start(cpu_buffer->buffer, (unsigned long)event);
4190
4191 bpage = READ_ONCE(cpu_buffer->tail_page);
4192
4193 /*
4194 * Make sure the tail_page is still the same and
4195 * the next write location is the end of this event
4196 */
4197 if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
4198 unsigned long write_mask =
4199 local_read(&bpage->write) & ~RB_WRITE_MASK;
4200 unsigned long event_length = rb_event_length(event);
4201
4202 /*
4203 * For the before_stamp to be different than the write_stamp
4204 * to make sure that the next event adds an absolute
4205 * value and does not rely on the saved write stamp, which
4206 * is now going to be bogus.
4207 *
4208 * By setting the before_stamp to zero, the next event
4209 * is not going to use the write_stamp and will instead
4210 * create an absolute timestamp. This means there's no
4211 * reason to update the wirte_stamp!
4212 */
4213 rb_time_set(&cpu_buffer->before_stamp, 0);
4214
4215 /*
4216 * If an event were to come in now, it would see that the
4217 * write_stamp and the before_stamp are different, and assume
4218 * that this event just added itself before updating
4219 * the write stamp. The interrupting event will fix the
4220 * write stamp for us, and use an absolute timestamp.
4221 */
4222
4223 /*
4224 * This is on the tail page. It is possible that
4225 * a write could come in and move the tail page
4226 * and write to the next page. That is fine
4227 * because we just shorten what is on this page.
4228 */
4229 old_index += write_mask;
4230 new_index += write_mask;
4231
4232 /* caution: old_index gets updated on cmpxchg failure */
4233 if (local_try_cmpxchg(&bpage->write, &old_index, new_index)) {
4234 /* update counters */
4235 local_sub(event_length, &cpu_buffer->entries_bytes);
4236 return true;
4237 }
4238 }
4239
4240 /* could not discard */
4241 return false;
4242}
4243
4244static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
4245{
4246 local_inc(&cpu_buffer->committing);
4247 local_inc(&cpu_buffer->commits);
4248}
4249
4250static __always_inline void
4251rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
4252{
4253 unsigned long max_count;
4254
4255 /*
4256 * We only race with interrupts and NMIs on this CPU.
4257 * If we own the commit event, then we can commit
4258 * all others that interrupted us, since the interruptions
4259 * are in stack format (they finish before they come
4260 * back to us). This allows us to do a simple loop to
4261 * assign the commit to the tail.
4262 */
4263 again:
4264 max_count = cpu_buffer->nr_pages * 100;
4265
4266 while (cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)) {
4267 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
4268 return;
4269 if (RB_WARN_ON(cpu_buffer,
4270 rb_is_reader_page(cpu_buffer->tail_page)))
4271 return;
4272 /*
4273 * No need for a memory barrier here, as the update
4274 * of the tail_page did it for this page.
4275 */
4276 local_set(&cpu_buffer->commit_page->page->commit,
4277 rb_page_write(cpu_buffer->commit_page));
4278 rb_inc_page(&cpu_buffer->commit_page);
4279 if (cpu_buffer->ring_meta) {
4280 struct ring_buffer_cpu_meta *meta = cpu_buffer->ring_meta;
4281 meta->commit_buffer = (unsigned long)cpu_buffer->commit_page->page;
4282 }
4283 /* add barrier to keep gcc from optimizing too much */
4284 barrier();
4285 }
4286 while (rb_commit_index(cpu_buffer) !=
4287 rb_page_write(cpu_buffer->commit_page)) {
4288
4289 /* Make sure the readers see the content of what is committed. */
4290 smp_wmb();
4291 local_set(&cpu_buffer->commit_page->page->commit,
4292 rb_page_write(cpu_buffer->commit_page));
4293 RB_WARN_ON(cpu_buffer,
4294 rb_page_commit(cpu_buffer->commit_page) & ~RB_WRITE_MASK);
4295 barrier();
4296 }
4297
4298 /* again, keep gcc from optimizing */
4299 barrier();
4300
4301 /*
4302 * If an interrupt came in just after the first while loop
4303 * and pushed the tail page forward, we will be left with
4304 * a dangling commit that will never go forward.
4305 */
4306 if (unlikely(cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)))
4307 goto again;
4308}
4309
4310static __always_inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
4311{
4312 unsigned long commits;
4313
4314 if (RB_WARN_ON(cpu_buffer,
4315 !local_read(&cpu_buffer->committing)))
4316 return;
4317
4318 again:
4319 commits = local_read(&cpu_buffer->commits);
4320 /* synchronize with interrupts */
4321 barrier();
4322 if (local_read(&cpu_buffer->committing) == 1)
4323 rb_set_commit_to_write(cpu_buffer);
4324
4325 local_dec(&cpu_buffer->committing);
4326
4327 /* synchronize with interrupts */
4328 barrier();
4329
4330 /*
4331 * Need to account for interrupts coming in between the
4332 * updating of the commit page and the clearing of the
4333 * committing counter.
4334 */
4335 if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
4336 !local_read(&cpu_buffer->committing)) {
4337 local_inc(&cpu_buffer->committing);
4338 goto again;
4339 }
4340}
4341
4342static inline void rb_event_discard(struct ring_buffer_event *event)
4343{
4344 if (extended_time(event))
4345 event = skip_time_extend(event);
4346
4347 /* array[0] holds the actual length for the discarded event */
4348 event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
4349 event->type_len = RINGBUF_TYPE_PADDING;
4350 /* time delta must be non zero */
4351 if (!event->time_delta)
4352 event->time_delta = 1;
4353}
4354
4355static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer)
4356{
4357 local_inc(&cpu_buffer->entries);
4358 rb_end_commit(cpu_buffer);
4359}
4360
4361static bool
4362rb_irq_work_queue(struct rb_irq_work *irq_work)
4363{
4364 int cpu;
4365
4366 /* irq_work_queue_on() is not NMI-safe */
4367 if (unlikely(in_nmi()))
4368 return irq_work_queue(&irq_work->work);
4369
4370 /*
4371 * If CPU isolation is not active, cpu is always the current
4372 * CPU, and the following is equivallent to irq_work_queue().
4373 */
4374 cpu = housekeeping_any_cpu(HK_TYPE_KERNEL_NOISE);
4375 return irq_work_queue_on(&irq_work->work, cpu);
4376}
4377
4378static __always_inline void
4379rb_wakeups(struct trace_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer)
4380{
4381 if (buffer->irq_work.waiters_pending) {
4382 buffer->irq_work.waiters_pending = false;
4383 /* irq_work_queue() supplies it's own memory barriers */
4384 rb_irq_work_queue(&buffer->irq_work);
4385 }
4386
4387 if (cpu_buffer->irq_work.waiters_pending) {
4388 cpu_buffer->irq_work.waiters_pending = false;
4389 /* irq_work_queue() supplies it's own memory barriers */
4390 rb_irq_work_queue(&cpu_buffer->irq_work);
4391 }
4392
4393 if (cpu_buffer->last_pages_touch == local_read(&cpu_buffer->pages_touched))
4394 return;
4395
4396 if (cpu_buffer->reader_page == cpu_buffer->commit_page)
4397 return;
4398
4399 if (!cpu_buffer->irq_work.full_waiters_pending)
4400 return;
4401
4402 cpu_buffer->last_pages_touch = local_read(&cpu_buffer->pages_touched);
4403
4404 if (!full_hit(buffer, cpu_buffer->cpu, cpu_buffer->shortest_full))
4405 return;
4406
4407 cpu_buffer->irq_work.wakeup_full = true;
4408 cpu_buffer->irq_work.full_waiters_pending = false;
4409 /* irq_work_queue() supplies it's own memory barriers */
4410 rb_irq_work_queue(&cpu_buffer->irq_work);
4411}
4412
4413#ifdef CONFIG_RING_BUFFER_RECORD_RECURSION
4414# define do_ring_buffer_record_recursion() \
4415 do_ftrace_record_recursion(_THIS_IP_, _RET_IP_)
4416#else
4417# define do_ring_buffer_record_recursion() do { } while (0)
4418#endif
4419
4420/*
4421 * The lock and unlock are done within a preempt disable section.
4422 * The current_context per_cpu variable can only be modified
4423 * by the current task between lock and unlock. But it can
4424 * be modified more than once via an interrupt. To pass this
4425 * information from the lock to the unlock without having to
4426 * access the 'in_interrupt()' functions again (which do show
4427 * a bit of overhead in something as critical as function tracing,
4428 * we use a bitmask trick.
4429 *
4430 * bit 1 = NMI context
4431 * bit 2 = IRQ context
4432 * bit 3 = SoftIRQ context
4433 * bit 4 = normal context.
4434 *
4435 * This works because this is the order of contexts that can
4436 * preempt other contexts. A SoftIRQ never preempts an IRQ
4437 * context.
4438 *
4439 * When the context is determined, the corresponding bit is
4440 * checked and set (if it was set, then a recursion of that context
4441 * happened).
4442 *
4443 * On unlock, we need to clear this bit. To do so, just subtract
4444 * 1 from the current_context and AND it to itself.
4445 *
4446 * (binary)
4447 * 101 - 1 = 100
4448 * 101 & 100 = 100 (clearing bit zero)
4449 *
4450 * 1010 - 1 = 1001
4451 * 1010 & 1001 = 1000 (clearing bit 1)
4452 *
4453 * The least significant bit can be cleared this way, and it
4454 * just so happens that it is the same bit corresponding to
4455 * the current context.
4456 *
4457 * Now the TRANSITION bit breaks the above slightly. The TRANSITION bit
4458 * is set when a recursion is detected at the current context, and if
4459 * the TRANSITION bit is already set, it will fail the recursion.
4460 * This is needed because there's a lag between the changing of
4461 * interrupt context and updating the preempt count. In this case,
4462 * a false positive will be found. To handle this, one extra recursion
4463 * is allowed, and this is done by the TRANSITION bit. If the TRANSITION
4464 * bit is already set, then it is considered a recursion and the function
4465 * ends. Otherwise, the TRANSITION bit is set, and that bit is returned.
4466 *
4467 * On the trace_recursive_unlock(), the TRANSITION bit will be the first
4468 * to be cleared. Even if it wasn't the context that set it. That is,
4469 * if an interrupt comes in while NORMAL bit is set and the ring buffer
4470 * is called before preempt_count() is updated, since the check will
4471 * be on the NORMAL bit, the TRANSITION bit will then be set. If an
4472 * NMI then comes in, it will set the NMI bit, but when the NMI code
4473 * does the trace_recursive_unlock() it will clear the TRANSITION bit
4474 * and leave the NMI bit set. But this is fine, because the interrupt
4475 * code that set the TRANSITION bit will then clear the NMI bit when it
4476 * calls trace_recursive_unlock(). If another NMI comes in, it will
4477 * set the TRANSITION bit and continue.
4478 *
4479 * Note: The TRANSITION bit only handles a single transition between context.
4480 */
4481
4482static __always_inline bool
4483trace_recursive_lock(struct ring_buffer_per_cpu *cpu_buffer)
4484{
4485 unsigned int val = cpu_buffer->current_context;
4486 int bit = interrupt_context_level();
4487
4488 bit = RB_CTX_NORMAL - bit;
4489
4490 if (unlikely(val & (1 << (bit + cpu_buffer->nest)))) {
4491 /*
4492 * It is possible that this was called by transitioning
4493 * between interrupt context, and preempt_count() has not
4494 * been updated yet. In this case, use the TRANSITION bit.
4495 */
4496 bit = RB_CTX_TRANSITION;
4497 if (val & (1 << (bit + cpu_buffer->nest))) {
4498 do_ring_buffer_record_recursion();
4499 return true;
4500 }
4501 }
4502
4503 val |= (1 << (bit + cpu_buffer->nest));
4504 cpu_buffer->current_context = val;
4505
4506 return false;
4507}
4508
4509static __always_inline void
4510trace_recursive_unlock(struct ring_buffer_per_cpu *cpu_buffer)
4511{
4512 cpu_buffer->current_context &=
4513 cpu_buffer->current_context - (1 << cpu_buffer->nest);
4514}
4515
4516/* The recursive locking above uses 5 bits */
4517#define NESTED_BITS 5
4518
4519/**
4520 * ring_buffer_nest_start - Allow to trace while nested
4521 * @buffer: The ring buffer to modify
4522 *
4523 * The ring buffer has a safety mechanism to prevent recursion.
4524 * But there may be a case where a trace needs to be done while
4525 * tracing something else. In this case, calling this function
4526 * will allow this function to nest within a currently active
4527 * ring_buffer_lock_reserve().
4528 *
4529 * Call this function before calling another ring_buffer_lock_reserve() and
4530 * call ring_buffer_nest_end() after the nested ring_buffer_unlock_commit().
4531 */
4532void ring_buffer_nest_start(struct trace_buffer *buffer)
4533{
4534 struct ring_buffer_per_cpu *cpu_buffer;
4535 int cpu;
4536
4537 /* Enabled by ring_buffer_nest_end() */
4538 preempt_disable_notrace();
4539 cpu = raw_smp_processor_id();
4540 cpu_buffer = buffer->buffers[cpu];
4541 /* This is the shift value for the above recursive locking */
4542 cpu_buffer->nest += NESTED_BITS;
4543}
4544
4545/**
4546 * ring_buffer_nest_end - Allow to trace while nested
4547 * @buffer: The ring buffer to modify
4548 *
4549 * Must be called after ring_buffer_nest_start() and after the
4550 * ring_buffer_unlock_commit().
4551 */
4552void ring_buffer_nest_end(struct trace_buffer *buffer)
4553{
4554 struct ring_buffer_per_cpu *cpu_buffer;
4555 int cpu;
4556
4557 /* disabled by ring_buffer_nest_start() */
4558 cpu = raw_smp_processor_id();
4559 cpu_buffer = buffer->buffers[cpu];
4560 /* This is the shift value for the above recursive locking */
4561 cpu_buffer->nest -= NESTED_BITS;
4562 preempt_enable_notrace();
4563}
4564
4565/**
4566 * ring_buffer_unlock_commit - commit a reserved
4567 * @buffer: The buffer to commit to
4568 *
4569 * This commits the data to the ring buffer, and releases any locks held.
4570 *
4571 * Must be paired with ring_buffer_lock_reserve.
4572 */
4573int ring_buffer_unlock_commit(struct trace_buffer *buffer)
4574{
4575 struct ring_buffer_per_cpu *cpu_buffer;
4576 int cpu = raw_smp_processor_id();
4577
4578 cpu_buffer = buffer->buffers[cpu];
4579
4580 rb_commit(cpu_buffer);
4581
4582 rb_wakeups(buffer, cpu_buffer);
4583
4584 trace_recursive_unlock(cpu_buffer);
4585
4586 preempt_enable_notrace();
4587
4588 return 0;
4589}
4590EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
4591
4592/* Special value to validate all deltas on a page. */
4593#define CHECK_FULL_PAGE 1L
4594
4595#ifdef CONFIG_RING_BUFFER_VALIDATE_TIME_DELTAS
4596
4597static const char *show_irq_str(int bits)
4598{
4599 static const char * type[] = {
4600 ".", // 0
4601 "s", // 1
4602 "h", // 2
4603 "Hs", // 3
4604 "n", // 4
4605 "Ns", // 5
4606 "Nh", // 6
4607 "NHs", // 7
4608 };
4609
4610 return type[bits];
4611}
4612
4613/* Assume this is a trace event */
4614static const char *show_flags(struct ring_buffer_event *event)
4615{
4616 struct trace_entry *entry;
4617 int bits = 0;
4618
4619 if (rb_event_data_length(event) - RB_EVNT_HDR_SIZE < sizeof(*entry))
4620 return "X";
4621
4622 entry = ring_buffer_event_data(event);
4623
4624 if (entry->flags & TRACE_FLAG_SOFTIRQ)
4625 bits |= 1;
4626
4627 if (entry->flags & TRACE_FLAG_HARDIRQ)
4628 bits |= 2;
4629
4630 if (entry->flags & TRACE_FLAG_NMI)
4631 bits |= 4;
4632
4633 return show_irq_str(bits);
4634}
4635
4636static const char *show_irq(struct ring_buffer_event *event)
4637{
4638 struct trace_entry *entry;
4639
4640 if (rb_event_data_length(event) - RB_EVNT_HDR_SIZE < sizeof(*entry))
4641 return "";
4642
4643 entry = ring_buffer_event_data(event);
4644 if (entry->flags & TRACE_FLAG_IRQS_OFF)
4645 return "d";
4646 return "";
4647}
4648
4649static const char *show_interrupt_level(void)
4650{
4651 unsigned long pc = preempt_count();
4652 unsigned char level = 0;
4653
4654 if (pc & SOFTIRQ_OFFSET)
4655 level |= 1;
4656
4657 if (pc & HARDIRQ_MASK)
4658 level |= 2;
4659
4660 if (pc & NMI_MASK)
4661 level |= 4;
4662
4663 return show_irq_str(level);
4664}
4665
4666static void dump_buffer_page(struct buffer_data_page *dpage,
4667 struct rb_event_info *info,
4668 unsigned long tail)
4669{
4670 struct ring_buffer_event *event;
4671 u64 ts, delta;
4672 int e;
4673
4674 ts = dpage->time_stamp;
4675 pr_warn(" [%lld] PAGE TIME STAMP\n", ts);
4676
4677 for (e = 0; e < tail; e += rb_event_length(event)) {
4678
4679 event = (struct ring_buffer_event *)(dpage->data + e);
4680
4681 switch (event->type_len) {
4682
4683 case RINGBUF_TYPE_TIME_EXTEND:
4684 delta = rb_event_time_stamp(event);
4685 ts += delta;
4686 pr_warn(" 0x%x: [%lld] delta:%lld TIME EXTEND\n",
4687 e, ts, delta);
4688 break;
4689
4690 case RINGBUF_TYPE_TIME_STAMP:
4691 delta = rb_event_time_stamp(event);
4692 ts = rb_fix_abs_ts(delta, ts);
4693 pr_warn(" 0x%x: [%lld] absolute:%lld TIME STAMP\n",
4694 e, ts, delta);
4695 break;
4696
4697 case RINGBUF_TYPE_PADDING:
4698 ts += event->time_delta;
4699 pr_warn(" 0x%x: [%lld] delta:%d PADDING\n",
4700 e, ts, event->time_delta);
4701 break;
4702
4703 case RINGBUF_TYPE_DATA:
4704 ts += event->time_delta;
4705 pr_warn(" 0x%x: [%lld] delta:%d %s%s\n",
4706 e, ts, event->time_delta,
4707 show_flags(event), show_irq(event));
4708 break;
4709
4710 default:
4711 break;
4712 }
4713 }
4714 pr_warn("expected end:0x%lx last event actually ended at:0x%x\n", tail, e);
4715}
4716
4717static DEFINE_PER_CPU(atomic_t, checking);
4718static atomic_t ts_dump;
4719
4720#define buffer_warn_return(fmt, ...) \
4721 do { \
4722 /* If another report is happening, ignore this one */ \
4723 if (atomic_inc_return(&ts_dump) != 1) { \
4724 atomic_dec(&ts_dump); \
4725 goto out; \
4726 } \
4727 atomic_inc(&cpu_buffer->record_disabled); \
4728 pr_warn(fmt, ##__VA_ARGS__); \
4729 dump_buffer_page(dpage, info, tail); \
4730 atomic_dec(&ts_dump); \
4731 /* There's some cases in boot up that this can happen */ \
4732 if (WARN_ON_ONCE(system_state != SYSTEM_BOOTING)) \
4733 /* Do not re-enable checking */ \
4734 return; \
4735 } while (0)
4736
4737/*
4738 * Check if the current event time stamp matches the deltas on
4739 * the buffer page.
4740 */
4741static void check_buffer(struct ring_buffer_per_cpu *cpu_buffer,
4742 struct rb_event_info *info,
4743 unsigned long tail)
4744{
4745 struct buffer_data_page *dpage;
4746 u64 ts, delta;
4747 bool full = false;
4748 int ret;
4749
4750 dpage = info->tail_page->page;
4751
4752 if (tail == CHECK_FULL_PAGE) {
4753 full = true;
4754 tail = rb_data_page_commit(dpage);
4755 } else if (info->add_timestamp &
4756 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE)) {
4757 /* Ignore events with absolute time stamps */
4758 return;
4759 }
4760
4761 /*
4762 * Do not check the first event (skip possible extends too).
4763 * Also do not check if previous events have not been committed.
4764 */
4765 if (tail <= 8 || tail > rb_data_page_commit(dpage))
4766 return;
4767
4768 /*
4769 * If this interrupted another event,
4770 */
4771 if (atomic_inc_return(this_cpu_ptr(&checking)) != 1)
4772 goto out;
4773
4774 ret = rb_read_data_buffer(dpage, tail, cpu_buffer->cpu, &ts, &delta);
4775 if (ret < 0) {
4776 if (delta < ts) {
4777 buffer_warn_return("[CPU: %d]ABSOLUTE TIME WENT BACKWARDS: last ts: %lld absolute ts: %lld clock:%pS\n",
4778 cpu_buffer->cpu, ts, delta,
4779 cpu_buffer->buffer->clock);
4780 goto out;
4781 }
4782 }
4783 if ((full && ts > info->ts) ||
4784 (!full && ts + info->delta != info->ts)) {
4785 buffer_warn_return("[CPU: %d]TIME DOES NOT MATCH expected:%lld actual:%lld delta:%lld before:%lld after:%lld%s context:%s\ntrace clock:%pS",
4786 cpu_buffer->cpu,
4787 ts + info->delta, info->ts, info->delta,
4788 info->before, info->after,
4789 full ? " (full)" : "", show_interrupt_level(),
4790 cpu_buffer->buffer->clock);
4791 }
4792out:
4793 atomic_dec(this_cpu_ptr(&checking));
4794}
4795#else
4796static inline void check_buffer(struct ring_buffer_per_cpu *cpu_buffer,
4797 struct rb_event_info *info,
4798 unsigned long tail)
4799{
4800}
4801#endif /* CONFIG_RING_BUFFER_VALIDATE_TIME_DELTAS */
4802
4803static struct ring_buffer_event *
4804__rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
4805 struct rb_event_info *info)
4806{
4807 struct ring_buffer_event *event;
4808 struct buffer_page *tail_page;
4809 unsigned long tail, write, w;
4810
4811 /* Don't let the compiler play games with cpu_buffer->tail_page */
4812 tail_page = info->tail_page = READ_ONCE(cpu_buffer->tail_page);
4813
4814 /*A*/ w = local_read(&tail_page->write) & RB_WRITE_MASK;
4815 barrier();
4816 rb_time_read(&cpu_buffer->before_stamp, &info->before);
4817 rb_time_read(&cpu_buffer->write_stamp, &info->after);
4818 barrier();
4819 info->ts = rb_time_stamp(cpu_buffer->buffer);
4820
4821 if ((info->add_timestamp & RB_ADD_STAMP_ABSOLUTE)) {
4822 info->delta = info->ts;
4823 } else {
4824 /*
4825 * If interrupting an event time update, we may need an
4826 * absolute timestamp.
4827 * Don't bother if this is the start of a new page (w == 0).
4828 */
4829 if (!w) {
4830 /* Use the sub-buffer timestamp */
4831 info->delta = 0;
4832 } else if (unlikely(info->before != info->after)) {
4833 info->add_timestamp |= RB_ADD_STAMP_FORCE | RB_ADD_STAMP_EXTEND;
4834 info->length += RB_LEN_TIME_EXTEND;
4835 } else {
4836 info->delta = info->ts - info->after;
4837 if (unlikely(test_time_stamp(info->delta))) {
4838 info->add_timestamp |= RB_ADD_STAMP_EXTEND;
4839 info->length += RB_LEN_TIME_EXTEND;
4840 }
4841 }
4842 }
4843
4844 /*B*/ rb_time_set(&cpu_buffer->before_stamp, info->ts);
4845
4846 /*C*/ write = local_add_return(info->length, &tail_page->write);
4847
4848 /* set write to only the index of the write */
4849 write &= RB_WRITE_MASK;
4850
4851 tail = write - info->length;
4852
4853 /* See if we shot pass the end of this buffer page */
4854 if (unlikely(write > rb_page_capacity(tail_page))) {
4855 check_buffer(cpu_buffer, info, CHECK_FULL_PAGE);
4856 return rb_move_tail(cpu_buffer, tail, info);
4857 }
4858
4859 if (likely(tail == w)) {
4860 /* Nothing interrupted us between A and C */
4861 /*D*/ rb_time_set(&cpu_buffer->write_stamp, info->ts);
4862 /*
4863 * If something came in between C and D, the write stamp
4864 * may now not be in sync. But that's fine as the before_stamp
4865 * will be different and then next event will just be forced
4866 * to use an absolute timestamp.
4867 */
4868 if (likely(!(info->add_timestamp &
4869 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE))))
4870 /* This did not interrupt any time update */
4871 info->delta = info->ts - info->after;
4872 else
4873 /* Just use full timestamp for interrupting event */
4874 info->delta = info->ts;
4875 check_buffer(cpu_buffer, info, tail);
4876 } else {
4877 u64 ts;
4878 /* SLOW PATH - Interrupted between A and C */
4879
4880 /* Save the old before_stamp */
4881 rb_time_read(&cpu_buffer->before_stamp, &info->before);
4882
4883 /*
4884 * Read a new timestamp and update the before_stamp to make
4885 * the next event after this one force using an absolute
4886 * timestamp. This is in case an interrupt were to come in
4887 * between E and F.
4888 */
4889 ts = rb_time_stamp(cpu_buffer->buffer);
4890 rb_time_set(&cpu_buffer->before_stamp, ts);
4891
4892 barrier();
4893 /*E*/ rb_time_read(&cpu_buffer->write_stamp, &info->after);
4894 barrier();
4895 /*F*/ if (write == (local_read(&tail_page->write) & RB_WRITE_MASK) &&
4896 info->after == info->before && info->after < ts) {
4897 /*
4898 * Nothing came after this event between C and F, it is
4899 * safe to use info->after for the delta as it
4900 * matched info->before and is still valid.
4901 */
4902 info->delta = ts - info->after;
4903 } else {
4904 /*
4905 * Interrupted between C and F:
4906 * Lost the previous events time stamp. Just set the
4907 * delta to zero, and this will be the same time as
4908 * the event this event interrupted. And the events that
4909 * came after this will still be correct (as they would
4910 * have built their delta on the previous event.
4911 */
4912 info->delta = 0;
4913 }
4914 info->ts = ts;
4915 info->add_timestamp &= ~RB_ADD_STAMP_FORCE;
4916 }
4917
4918 /*
4919 * If this is the first commit on the page, then it has the same
4920 * timestamp as the page itself.
4921 */
4922 if (unlikely(!tail && !(info->add_timestamp &
4923 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE))))
4924 info->delta = 0;
4925
4926 /* We reserved something on the buffer */
4927
4928 event = __rb_page_index(tail_page, tail);
4929 rb_update_event(cpu_buffer, event, info);
4930
4931 local_inc(&tail_page->entries);
4932
4933 /*
4934 * If this is the first commit on the page, then update
4935 * its timestamp.
4936 */
4937 if (unlikely(!tail))
4938 tail_page->page->time_stamp = info->ts;
4939
4940 /* account for these added bytes */
4941 local_add(info->length, &cpu_buffer->entries_bytes);
4942
4943 return event;
4944}
4945
4946static __always_inline struct ring_buffer_event *
4947rb_reserve_next_event(struct trace_buffer *buffer,
4948 struct ring_buffer_per_cpu *cpu_buffer,
4949 unsigned long length)
4950{
4951 struct ring_buffer_event *event;
4952 struct rb_event_info info;
4953 int nr_loops = 0;
4954 int add_ts_default;
4955
4956 /*
4957 * ring buffer does cmpxchg as well as atomic64 operations
4958 * (which some archs use locking for atomic64), make sure this
4959 * is safe in NMI context
4960 */
4961 if ((!IS_ENABLED(CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG) ||
4962 IS_ENABLED(CONFIG_GENERIC_ATOMIC64)) &&
4963 (unlikely(in_nmi()))) {
4964 return NULL;
4965 }
4966
4967 rb_start_commit(cpu_buffer);
4968 /* The commit page can not change after this */
4969
4970#ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
4971 /*
4972 * Due to the ability to swap a cpu buffer from a buffer
4973 * it is possible it was swapped before we committed.
4974 * (committing stops a swap). We check for it here and
4975 * if it happened, we have to fail the write.
4976 */
4977 barrier();
4978 if (unlikely(READ_ONCE(cpu_buffer->buffer) != buffer)) {
4979 local_dec(&cpu_buffer->committing);
4980 local_dec(&cpu_buffer->commits);
4981 return NULL;
4982 }
4983#endif
4984
4985 info.length = rb_calculate_event_length(length);
4986
4987 if (ring_buffer_time_stamp_abs(cpu_buffer->buffer)) {
4988 add_ts_default = RB_ADD_STAMP_ABSOLUTE;
4989 info.length += RB_LEN_TIME_EXTEND;
4990 if (info.length > rb_subbuf_max_data_size(cpu_buffer->buffer))
4991 goto out_fail;
4992 } else {
4993 add_ts_default = RB_ADD_STAMP_NONE;
4994 }
4995
4996 again:
4997 info.add_timestamp = add_ts_default;
4998 info.delta = 0;
4999
5000 /*
5001 * We allow for interrupts to reenter here and do a trace.
5002 * If one does, it will cause this original code to loop
5003 * back here. Even with heavy interrupts happening, this
5004 * should only happen a few times in a row. If this happens
5005 * 1000 times in a row, there must be either an interrupt
5006 * storm or we have something buggy.
5007 * Bail!
5008 */
5009 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
5010 goto out_fail;
5011
5012 event = __rb_reserve_next(cpu_buffer, &info);
5013
5014 if (unlikely(PTR_ERR(event) == -EAGAIN)) {
5015 if (info.add_timestamp & (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_EXTEND))
5016 info.length -= RB_LEN_TIME_EXTEND;
5017 goto again;
5018 }
5019
5020 if (likely(event))
5021 return event;
5022 out_fail:
5023 rb_end_commit(cpu_buffer);
5024 return NULL;
5025}
5026
5027/**
5028 * ring_buffer_lock_reserve - reserve a part of the buffer
5029 * @buffer: the ring buffer to reserve from
5030 * @length: the length of the data to reserve (excluding event header)
5031 *
5032 * Returns a reserved event on the ring buffer to copy directly to.
5033 * The user of this interface will need to get the body to write into
5034 * and can use the ring_buffer_event_data() interface.
5035 *
5036 * The length is the length of the data needed, not the event length
5037 * which also includes the event header.
5038 *
5039 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
5040 * If NULL is returned, then nothing has been allocated or locked.
5041 */
5042struct ring_buffer_event *
5043ring_buffer_lock_reserve(struct trace_buffer *buffer, unsigned long length)
5044{
5045 struct ring_buffer_per_cpu *cpu_buffer;
5046 struct ring_buffer_event *event;
5047 int cpu;
5048
5049 /* If we are tracing schedule, we don't want to recurse */
5050 preempt_disable_notrace();
5051
5052 if (unlikely(atomic_read(&buffer->record_disabled)))
5053 goto out;
5054
5055 cpu = raw_smp_processor_id();
5056
5057 if (unlikely(!cpumask_test_cpu(cpu, buffer->cpumask)))
5058 goto out;
5059
5060 cpu_buffer = buffer->buffers[cpu];
5061
5062 if (unlikely(atomic_read(&cpu_buffer->record_disabled)))
5063 goto out;
5064
5065 if (unlikely(length > rb_subbuf_max_data_size(buffer)))
5066 goto out;
5067
5068 if (unlikely(trace_recursive_lock(cpu_buffer)))
5069 goto out;
5070
5071 event = rb_reserve_next_event(buffer, cpu_buffer, length);
5072 if (!event)
5073 goto out_unlock;
5074
5075 return event;
5076
5077 out_unlock:
5078 trace_recursive_unlock(cpu_buffer);
5079 out:
5080 preempt_enable_notrace();
5081 return NULL;
5082}
5083EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
5084
5085/*
5086 * Decrement the entries to the page that an event is on.
5087 * The event does not even need to exist, only the pointer
5088 * to the page it is on. This may only be called before the commit
5089 * takes place.
5090 */
5091static inline void
5092rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
5093 struct ring_buffer_event *event)
5094{
5095 unsigned long addr = (unsigned long)event;
5096 struct buffer_page *bpage = cpu_buffer->commit_page;
5097 struct buffer_page *start;
5098
5099 addr = rb_subbuf_start(cpu_buffer->buffer, addr);
5100
5101 /* Do the likely case first */
5102 if (likely(bpage->page == (void *)addr)) {
5103 local_dec(&bpage->entries);
5104 return;
5105 }
5106
5107 /*
5108 * Because the commit page may be on the reader page we
5109 * start with the next page and check the end loop there.
5110 */
5111 rb_inc_page(&bpage);
5112 start = bpage;
5113 do {
5114 if (bpage->page == (void *)addr) {
5115 local_dec(&bpage->entries);
5116 return;
5117 }
5118 rb_inc_page(&bpage);
5119 } while (bpage != start);
5120
5121 /* commit not part of this buffer?? */
5122 RB_WARN_ON(cpu_buffer, 1);
5123}
5124
5125/**
5126 * ring_buffer_discard_commit - discard an event that has not been committed
5127 * @buffer: the ring buffer
5128 * @event: non committed event to discard
5129 *
5130 * Sometimes an event that is in the ring buffer needs to be ignored.
5131 * This function lets the user discard an event in the ring buffer
5132 * and then that event will not be read later.
5133 *
5134 * This function only works if it is called before the item has been
5135 * committed. It will try to free the event from the ring buffer
5136 * if another event has not been added behind it.
5137 *
5138 * If another event has been added behind it, it will set the event
5139 * up as discarded, and perform the commit.
5140 *
5141 * If this function is called, do not call ring_buffer_unlock_commit on
5142 * the event.
5143 */
5144void ring_buffer_discard_commit(struct trace_buffer *buffer,
5145 struct ring_buffer_event *event)
5146{
5147 struct ring_buffer_per_cpu *cpu_buffer;
5148 int cpu;
5149
5150 /* The event is discarded regardless */
5151 rb_event_discard(event);
5152
5153 cpu = smp_processor_id();
5154 cpu_buffer = buffer->buffers[cpu];
5155
5156 /*
5157 * This must only be called if the event has not been
5158 * committed yet. Thus we can assume that preemption
5159 * is still disabled.
5160 */
5161 RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
5162
5163 rb_decrement_entry(cpu_buffer, event);
5164 rb_try_to_discard(cpu_buffer, event);
5165 rb_end_commit(cpu_buffer);
5166
5167 trace_recursive_unlock(cpu_buffer);
5168
5169 preempt_enable_notrace();
5170
5171}
5172EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
5173
5174/**
5175 * ring_buffer_write - write data to the buffer without reserving
5176 * @buffer: The ring buffer to write to.
5177 * @length: The length of the data being written (excluding the event header)
5178 * @data: The data to write to the buffer.
5179 *
5180 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
5181 * one function. If you already have the data to write to the buffer, it
5182 * may be easier to simply call this function.
5183 *
5184 * Note, like ring_buffer_lock_reserve, the length is the length of the data
5185 * and not the length of the event which would hold the header.
5186 */
5187int ring_buffer_write(struct trace_buffer *buffer,
5188 unsigned long length,
5189 void *data)
5190{
5191 struct ring_buffer_per_cpu *cpu_buffer;
5192 struct ring_buffer_event *event;
5193 void *body;
5194 int ret = -EBUSY;
5195 int cpu;
5196
5197 guard(preempt_notrace)();
5198
5199 if (atomic_read(&buffer->record_disabled))
5200 return -EBUSY;
5201
5202 cpu = raw_smp_processor_id();
5203
5204 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5205 return -EBUSY;
5206
5207 cpu_buffer = buffer->buffers[cpu];
5208
5209 if (atomic_read(&cpu_buffer->record_disabled))
5210 return -EBUSY;
5211
5212 if (length > rb_subbuf_max_data_size(buffer))
5213 return -EBUSY;
5214
5215 if (unlikely(trace_recursive_lock(cpu_buffer)))
5216 return -EBUSY;
5217
5218 event = rb_reserve_next_event(buffer, cpu_buffer, length);
5219 if (!event)
5220 goto out_unlock;
5221
5222 body = rb_event_data(event);
5223
5224 memcpy(body, data, length);
5225
5226 rb_commit(cpu_buffer);
5227
5228 rb_wakeups(buffer, cpu_buffer);
5229
5230 ret = 0;
5231
5232 out_unlock:
5233 trace_recursive_unlock(cpu_buffer);
5234 return ret;
5235}
5236EXPORT_SYMBOL_GPL(ring_buffer_write);
5237
5238/*
5239 * The total entries in the ring buffer is the running counter
5240 * of entries entered into the ring buffer, minus the sum of
5241 * the entries read from the ring buffer and the number of
5242 * entries that were overwritten.
5243 */
5244static inline unsigned long
5245rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
5246{
5247 return local_read(&cpu_buffer->entries) -
5248 (local_read(&cpu_buffer->overrun) + cpu_buffer->read);
5249}
5250
5251static bool rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
5252{
5253 return !rb_num_of_entries(cpu_buffer);
5254}
5255
5256/**
5257 * ring_buffer_record_disable - stop all writes into the buffer
5258 * @buffer: The ring buffer to stop writes to.
5259 *
5260 * This prevents all writes to the buffer. Any attempt to write
5261 * to the buffer after this will fail and return NULL.
5262 *
5263 * The caller should call synchronize_rcu() after this.
5264 */
5265void ring_buffer_record_disable(struct trace_buffer *buffer)
5266{
5267 atomic_inc(&buffer->record_disabled);
5268}
5269EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
5270
5271/**
5272 * ring_buffer_record_enable - enable writes to the buffer
5273 * @buffer: The ring buffer to enable writes
5274 *
5275 * Note, multiple disables will need the same number of enables
5276 * to truly enable the writing (much like preempt_disable).
5277 */
5278void ring_buffer_record_enable(struct trace_buffer *buffer)
5279{
5280 atomic_dec(&buffer->record_disabled);
5281}
5282EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
5283
5284/**
5285 * ring_buffer_record_off - stop all writes into the buffer
5286 * @buffer: The ring buffer to stop writes to.
5287 *
5288 * This prevents all writes to the buffer. Any attempt to write
5289 * to the buffer after this will fail and return NULL.
5290 *
5291 * This is different than ring_buffer_record_disable() as
5292 * it works like an on/off switch, where as the disable() version
5293 * must be paired with a enable().
5294 */
5295void ring_buffer_record_off(struct trace_buffer *buffer)
5296{
5297 unsigned int rd;
5298 unsigned int new_rd;
5299
5300 rd = atomic_read(&buffer->record_disabled);
5301 do {
5302 new_rd = rd | RB_BUFFER_OFF;
5303 } while (!atomic_try_cmpxchg(&buffer->record_disabled, &rd, new_rd));
5304}
5305EXPORT_SYMBOL_GPL(ring_buffer_record_off);
5306
5307/**
5308 * ring_buffer_record_on - restart writes into the buffer
5309 * @buffer: The ring buffer to start writes to.
5310 *
5311 * This enables all writes to the buffer that was disabled by
5312 * ring_buffer_record_off().
5313 *
5314 * This is different than ring_buffer_record_enable() as
5315 * it works like an on/off switch, where as the enable() version
5316 * must be paired with a disable().
5317 */
5318void ring_buffer_record_on(struct trace_buffer *buffer)
5319{
5320 unsigned int rd;
5321 unsigned int new_rd;
5322
5323 rd = atomic_read(&buffer->record_disabled);
5324 do {
5325 new_rd = rd & ~RB_BUFFER_OFF;
5326 } while (!atomic_try_cmpxchg(&buffer->record_disabled, &rd, new_rd));
5327}
5328EXPORT_SYMBOL_GPL(ring_buffer_record_on);
5329
5330/**
5331 * ring_buffer_record_is_on - return true if the ring buffer can write
5332 * @buffer: The ring buffer to see if write is enabled
5333 *
5334 * Returns true if the ring buffer is in a state that it accepts writes.
5335 */
5336bool ring_buffer_record_is_on(struct trace_buffer *buffer)
5337{
5338 return !atomic_read(&buffer->record_disabled);
5339}
5340
5341/**
5342 * ring_buffer_record_is_set_on - return true if the ring buffer is set writable
5343 * @buffer: The ring buffer to see if write is set enabled
5344 *
5345 * Returns true if the ring buffer is set writable by ring_buffer_record_on().
5346 * Note that this does NOT mean it is in a writable state.
5347 *
5348 * It may return true when the ring buffer has been disabled by
5349 * ring_buffer_record_disable(), as that is a temporary disabling of
5350 * the ring buffer.
5351 */
5352bool ring_buffer_record_is_set_on(struct trace_buffer *buffer)
5353{
5354 return !(atomic_read(&buffer->record_disabled) & RB_BUFFER_OFF);
5355}
5356
5357/**
5358 * ring_buffer_record_is_on_cpu - return true if the ring buffer can write
5359 * @buffer: The ring buffer to see if write is enabled
5360 * @cpu: The CPU to test if the ring buffer can write too
5361 *
5362 * Returns true if the ring buffer is in a state that it accepts writes
5363 * for a particular CPU.
5364 */
5365bool ring_buffer_record_is_on_cpu(struct trace_buffer *buffer, int cpu)
5366{
5367 struct ring_buffer_per_cpu *cpu_buffer;
5368
5369 cpu_buffer = buffer->buffers[cpu];
5370
5371 return ring_buffer_record_is_set_on(buffer) &&
5372 !atomic_read(&cpu_buffer->record_disabled);
5373}
5374
5375/**
5376 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
5377 * @buffer: The ring buffer to stop writes to.
5378 * @cpu: The CPU buffer to stop
5379 *
5380 * This prevents all writes to the buffer. Any attempt to write
5381 * to the buffer after this will fail and return NULL.
5382 *
5383 * The caller should call synchronize_rcu() after this.
5384 */
5385void ring_buffer_record_disable_cpu(struct trace_buffer *buffer, int cpu)
5386{
5387 struct ring_buffer_per_cpu *cpu_buffer;
5388
5389 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5390 return;
5391
5392 cpu_buffer = buffer->buffers[cpu];
5393 atomic_inc(&cpu_buffer->record_disabled);
5394}
5395EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
5396
5397/**
5398 * ring_buffer_record_enable_cpu - enable writes to the buffer
5399 * @buffer: The ring buffer to enable writes
5400 * @cpu: The CPU to enable.
5401 *
5402 * Note, multiple disables will need the same number of enables
5403 * to truly enable the writing (much like preempt_disable).
5404 */
5405void ring_buffer_record_enable_cpu(struct trace_buffer *buffer, int cpu)
5406{
5407 struct ring_buffer_per_cpu *cpu_buffer;
5408
5409 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5410 return;
5411
5412 cpu_buffer = buffer->buffers[cpu];
5413 atomic_dec(&cpu_buffer->record_disabled);
5414}
5415EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
5416
5417/**
5418 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
5419 * @buffer: The ring buffer
5420 * @cpu: The per CPU buffer to read from.
5421 */
5422u64 ring_buffer_oldest_event_ts(struct trace_buffer *buffer, int cpu)
5423{
5424 unsigned long flags;
5425 struct ring_buffer_per_cpu *cpu_buffer;
5426 struct buffer_page *bpage;
5427 u64 ret = 0;
5428
5429 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5430 return 0;
5431
5432 cpu_buffer = buffer->buffers[cpu];
5433 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5434 /*
5435 * if the tail is on reader_page, oldest time stamp is on the reader
5436 * page
5437 */
5438 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
5439 bpage = cpu_buffer->reader_page;
5440 else
5441 bpage = rb_set_head_page(cpu_buffer);
5442 if (bpage)
5443 ret = bpage->page->time_stamp;
5444 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5445
5446 return ret;
5447}
5448EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
5449
5450/**
5451 * ring_buffer_bytes_cpu - get the number of bytes unconsumed in a cpu buffer
5452 * @buffer: The ring buffer
5453 * @cpu: The per CPU buffer to read from.
5454 */
5455unsigned long ring_buffer_bytes_cpu(struct trace_buffer *buffer, int cpu)
5456{
5457 struct ring_buffer_per_cpu *cpu_buffer;
5458 unsigned long ret;
5459
5460 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5461 return 0;
5462
5463 cpu_buffer = buffer->buffers[cpu];
5464 ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
5465
5466 return ret;
5467}
5468EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
5469
5470/**
5471 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
5472 * @buffer: The ring buffer
5473 * @cpu: The per CPU buffer to get the entries from.
5474 */
5475unsigned long ring_buffer_entries_cpu(struct trace_buffer *buffer, int cpu)
5476{
5477 struct ring_buffer_per_cpu *cpu_buffer;
5478
5479 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5480 return 0;
5481
5482 cpu_buffer = buffer->buffers[cpu];
5483
5484 return rb_num_of_entries(cpu_buffer);
5485}
5486EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
5487
5488/**
5489 * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
5490 * buffer wrapping around (only if RB_FL_OVERWRITE is on).
5491 * @buffer: The ring buffer
5492 * @cpu: The per CPU buffer to get the number of overruns from
5493 */
5494unsigned long ring_buffer_overrun_cpu(struct trace_buffer *buffer, int cpu)
5495{
5496 struct ring_buffer_per_cpu *cpu_buffer;
5497 unsigned long ret;
5498
5499 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5500 return 0;
5501
5502 cpu_buffer = buffer->buffers[cpu];
5503 ret = local_read(&cpu_buffer->overrun);
5504
5505 return ret;
5506}
5507EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
5508
5509/**
5510 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
5511 * commits failing due to the buffer wrapping around while there are uncommitted
5512 * events, such as during an interrupt storm.
5513 * @buffer: The ring buffer
5514 * @cpu: The per CPU buffer to get the number of overruns from
5515 */
5516unsigned long
5517ring_buffer_commit_overrun_cpu(struct trace_buffer *buffer, int cpu)
5518{
5519 struct ring_buffer_per_cpu *cpu_buffer;
5520 unsigned long ret;
5521
5522 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5523 return 0;
5524
5525 cpu_buffer = buffer->buffers[cpu];
5526 ret = local_read(&cpu_buffer->commit_overrun);
5527
5528 return ret;
5529}
5530EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
5531
5532/**
5533 * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
5534 * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
5535 * @buffer: The ring buffer
5536 * @cpu: The per CPU buffer to get the number of overruns from
5537 */
5538unsigned long
5539ring_buffer_dropped_events_cpu(struct trace_buffer *buffer, int cpu)
5540{
5541 struct ring_buffer_per_cpu *cpu_buffer;
5542 unsigned long ret;
5543
5544 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5545 return 0;
5546
5547 cpu_buffer = buffer->buffers[cpu];
5548 ret = local_read(&cpu_buffer->dropped_events);
5549
5550 return ret;
5551}
5552EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
5553
5554/**
5555 * ring_buffer_read_events_cpu - get the number of events successfully read
5556 * @buffer: The ring buffer
5557 * @cpu: The per CPU buffer to get the number of events read
5558 */
5559unsigned long
5560ring_buffer_read_events_cpu(struct trace_buffer *buffer, int cpu)
5561{
5562 struct ring_buffer_per_cpu *cpu_buffer;
5563
5564 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5565 return 0;
5566
5567 cpu_buffer = buffer->buffers[cpu];
5568 return cpu_buffer->read;
5569}
5570EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu);
5571
5572/**
5573 * ring_buffer_entries - get the number of entries in a buffer
5574 * @buffer: The ring buffer
5575 *
5576 * Returns the total number of entries in the ring buffer
5577 * (all CPU entries)
5578 */
5579unsigned long ring_buffer_entries(struct trace_buffer *buffer)
5580{
5581 struct ring_buffer_per_cpu *cpu_buffer;
5582 unsigned long entries = 0;
5583 int cpu;
5584
5585 /* if you care about this being correct, lock the buffer */
5586 for_each_buffer_cpu(buffer, cpu) {
5587 cpu_buffer = buffer->buffers[cpu];
5588 entries += rb_num_of_entries(cpu_buffer);
5589 }
5590
5591 return entries;
5592}
5593EXPORT_SYMBOL_GPL(ring_buffer_entries);
5594
5595/**
5596 * ring_buffer_overruns - get the number of overruns in buffer
5597 * @buffer: The ring buffer
5598 *
5599 * Returns the total number of overruns in the ring buffer
5600 * (all CPU entries)
5601 */
5602unsigned long ring_buffer_overruns(struct trace_buffer *buffer)
5603{
5604 struct ring_buffer_per_cpu *cpu_buffer;
5605 unsigned long overruns = 0;
5606 int cpu;
5607
5608 /* if you care about this being correct, lock the buffer */
5609 for_each_buffer_cpu(buffer, cpu) {
5610 cpu_buffer = buffer->buffers[cpu];
5611 overruns += local_read(&cpu_buffer->overrun);
5612 }
5613
5614 return overruns;
5615}
5616EXPORT_SYMBOL_GPL(ring_buffer_overruns);
5617
5618static bool rb_read_remote_meta_page(struct ring_buffer_per_cpu *cpu_buffer)
5619{
5620 local_set(&cpu_buffer->entries, READ_ONCE(cpu_buffer->meta_page->entries));
5621 local_set(&cpu_buffer->overrun, READ_ONCE(cpu_buffer->meta_page->overrun));
5622 local_set(&cpu_buffer->pages_touched, READ_ONCE(cpu_buffer->meta_page->pages_touched));
5623 local_set(&cpu_buffer->pages_lost, READ_ONCE(cpu_buffer->meta_page->pages_lost));
5624
5625 return rb_num_of_entries(cpu_buffer);
5626}
5627
5628static void rb_update_remote_head(struct ring_buffer_per_cpu *cpu_buffer)
5629{
5630 struct buffer_page *next, *orig;
5631 int retry = 3;
5632
5633 orig = next = cpu_buffer->head_page;
5634 rb_inc_page(&next);
5635
5636 /* Run after the writer */
5637 while (cpu_buffer->head_page->page->time_stamp > next->page->time_stamp) {
5638 rb_inc_page(&next);
5639
5640 rb_list_head_clear(cpu_buffer->head_page->list.prev);
5641 rb_inc_page(&cpu_buffer->head_page);
5642 rb_set_list_to_head(cpu_buffer->head_page->list.prev);
5643
5644 if (cpu_buffer->head_page == orig) {
5645 if (WARN_ON_ONCE(!(--retry)))
5646 return;
5647 }
5648 }
5649
5650 orig = cpu_buffer->commit_page = cpu_buffer->head_page;
5651 retry = 3;
5652
5653 while (cpu_buffer->commit_page->page->time_stamp < next->page->time_stamp) {
5654 rb_inc_page(&next);
5655 rb_inc_page(&cpu_buffer->commit_page);
5656
5657 if (cpu_buffer->commit_page == orig) {
5658 if (WARN_ON_ONCE(!(--retry)))
5659 return;
5660 }
5661 }
5662}
5663
5664static void rb_iter_reset(struct ring_buffer_iter *iter)
5665{
5666 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
5667
5668 if (cpu_buffer->remote) {
5669 rb_read_remote_meta_page(cpu_buffer);
5670 rb_update_remote_head(cpu_buffer);
5671 }
5672
5673 /* Iterator usage is expected to have record disabled */
5674 iter->head_page = cpu_buffer->reader_page;
5675 iter->head = cpu_buffer->reader_page->read;
5676 iter->next_event = iter->head;
5677 iter->missed_events = 0;
5678
5679 iter->cache_reader_page = iter->head_page;
5680 iter->cache_read = cpu_buffer->read;
5681 iter->cache_pages_removed = cpu_buffer->pages_removed;
5682
5683 if (iter->head) {
5684 iter->read_stamp = cpu_buffer->read_stamp;
5685 iter->page_stamp = cpu_buffer->reader_page->page->time_stamp;
5686 } else {
5687 iter->read_stamp = iter->head_page->page->time_stamp;
5688 iter->page_stamp = iter->read_stamp;
5689 }
5690}
5691
5692/**
5693 * ring_buffer_iter_reset - reset an iterator
5694 * @iter: The iterator to reset
5695 *
5696 * Resets the iterator, so that it will start from the beginning
5697 * again.
5698 */
5699void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
5700{
5701 struct ring_buffer_per_cpu *cpu_buffer;
5702 unsigned long flags;
5703
5704 if (!iter)
5705 return;
5706
5707 cpu_buffer = iter->cpu_buffer;
5708
5709 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5710 rb_iter_reset(iter);
5711 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5712}
5713EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
5714
5715/**
5716 * ring_buffer_iter_empty - check if an iterator has no more to read
5717 * @iter: The iterator to check
5718 */
5719int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
5720{
5721 struct ring_buffer_per_cpu *cpu_buffer;
5722 struct buffer_page *reader;
5723 struct buffer_page *head_page;
5724 struct buffer_page *commit_page;
5725 struct buffer_page *curr_commit_page;
5726 unsigned commit;
5727 u64 curr_commit_ts;
5728 u64 commit_ts;
5729
5730 cpu_buffer = iter->cpu_buffer;
5731 reader = cpu_buffer->reader_page;
5732 head_page = cpu_buffer->head_page;
5733 commit_page = READ_ONCE(cpu_buffer->commit_page);
5734 commit_ts = commit_page->page->time_stamp;
5735
5736 /*
5737 * When the writer goes across pages, it issues a cmpxchg which
5738 * is a mb(), which will synchronize with the rmb here.
5739 * (see rb_tail_page_update())
5740 */
5741 smp_rmb();
5742 commit = rb_page_size(commit_page);
5743 /* We want to make sure that the commit page doesn't change */
5744 smp_rmb();
5745
5746 /* Make sure commit page didn't change */
5747 curr_commit_page = READ_ONCE(cpu_buffer->commit_page);
5748 curr_commit_ts = READ_ONCE(curr_commit_page->page->time_stamp);
5749
5750 /* If the commit page changed, then there's more data */
5751 if (curr_commit_page != commit_page ||
5752 curr_commit_ts != commit_ts)
5753 return 0;
5754
5755 /* Still racy, as it may return a false positive, but that's OK */
5756 return ((iter->head_page == commit_page && iter->head >= commit) ||
5757 (iter->head_page == reader && commit_page == head_page &&
5758 head_page->read == commit &&
5759 iter->head == rb_page_size(cpu_buffer->reader_page)));
5760}
5761EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
5762
5763static void
5764rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
5765 struct ring_buffer_event *event)
5766{
5767 u64 delta;
5768
5769 switch (event->type_len) {
5770 case RINGBUF_TYPE_PADDING:
5771 return;
5772
5773 case RINGBUF_TYPE_TIME_EXTEND:
5774 delta = rb_event_time_stamp(event);
5775 cpu_buffer->read_stamp += delta;
5776 return;
5777
5778 case RINGBUF_TYPE_TIME_STAMP:
5779 delta = rb_event_time_stamp(event);
5780 delta = rb_fix_abs_ts(delta, cpu_buffer->read_stamp);
5781 cpu_buffer->read_stamp = delta;
5782 return;
5783
5784 case RINGBUF_TYPE_DATA:
5785 cpu_buffer->read_stamp += event->time_delta;
5786 return;
5787
5788 default:
5789 RB_WARN_ON(cpu_buffer, 1);
5790 }
5791}
5792
5793static void
5794rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
5795 struct ring_buffer_event *event)
5796{
5797 u64 delta;
5798
5799 switch (event->type_len) {
5800 case RINGBUF_TYPE_PADDING:
5801 return;
5802
5803 case RINGBUF_TYPE_TIME_EXTEND:
5804 delta = rb_event_time_stamp(event);
5805 iter->read_stamp += delta;
5806 return;
5807
5808 case RINGBUF_TYPE_TIME_STAMP:
5809 delta = rb_event_time_stamp(event);
5810 delta = rb_fix_abs_ts(delta, iter->read_stamp);
5811 iter->read_stamp = delta;
5812 return;
5813
5814 case RINGBUF_TYPE_DATA:
5815 iter->read_stamp += event->time_delta;
5816 return;
5817
5818 default:
5819 RB_WARN_ON(iter->cpu_buffer, 1);
5820 }
5821}
5822
5823static struct buffer_page *
5824__rb_get_reader_page_from_remote(struct ring_buffer_per_cpu *cpu_buffer)
5825{
5826 struct buffer_page *new_reader, *prev_reader, *prev_head, *new_head, *last;
5827
5828 if (!rb_read_remote_meta_page(cpu_buffer))
5829 return NULL;
5830
5831 /* More to read on the reader page */
5832 if (cpu_buffer->reader_page->read < rb_page_size(cpu_buffer->reader_page)) {
5833 if (!cpu_buffer->reader_page->read)
5834 cpu_buffer->read_stamp = cpu_buffer->reader_page->page->time_stamp;
5835 return cpu_buffer->reader_page;
5836 }
5837
5838 prev_reader = cpu_buffer->subbuf_ids[cpu_buffer->meta_page->reader.id];
5839
5840 if (cpu_buffer->remote->swap_reader_page(cpu_buffer->cpu,
5841 cpu_buffer->remote->priv)) {
5842 pr_warn_ratelimited("Remote reader page swap failed\n");
5843 return NULL;
5844 }
5845 /* nr_pages doesn't include the reader page */
5846 if (WARN_ON_ONCE(cpu_buffer->meta_page->reader.id > cpu_buffer->nr_pages))
5847 return NULL;
5848
5849 new_reader = cpu_buffer->subbuf_ids[cpu_buffer->meta_page->reader.id];
5850
5851 WARN_ON_ONCE(prev_reader == new_reader);
5852
5853 prev_head = new_reader; /* New reader was also the previous head */
5854 new_head = prev_head;
5855 rb_inc_page(&new_head);
5856 last = prev_head;
5857 rb_dec_page(&last);
5858
5859 /* Clear the old HEAD flag */
5860 rb_list_head_clear(cpu_buffer->head_page->list.prev);
5861
5862 prev_reader->list.next = prev_head->list.next;
5863 prev_reader->list.prev = prev_head->list.prev;
5864
5865 /* Swap prev_reader with new_reader */
5866 last->list.next = &prev_reader->list;
5867 new_head->list.prev = &prev_reader->list;
5868
5869 new_reader->list.prev = &new_reader->list;
5870 new_reader->list.next = &new_head->list;
5871
5872 /* Reactivate the HEAD flag */
5873 rb_set_list_to_head(&last->list);
5874
5875 cpu_buffer->head_page = new_head;
5876 cpu_buffer->reader_page = new_reader;
5877 cpu_buffer->reader_page->read = 0;
5878 cpu_buffer->pages = &new_head->list;
5879 cpu_buffer->read_stamp = new_reader->page->time_stamp;
5880 cpu_buffer->lost_events = cpu_buffer->meta_page->reader.lost_events;
5881
5882 return rb_page_size(cpu_buffer->reader_page) ? cpu_buffer->reader_page : NULL;
5883}
5884
5885static struct buffer_page *
5886__rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
5887{
5888 unsigned long max_loops = cpu_buffer->ring_meta ? cpu_buffer->nr_pages : 3;
5889 struct buffer_page *reader = NULL;
5890 unsigned long nr_loops = 0;
5891 unsigned long overwrite;
5892 unsigned long flags;
5893 int missed_events = 0;
5894 bool ret;
5895
5896 local_irq_save(flags);
5897 arch_spin_lock(&cpu_buffer->lock);
5898
5899 again:
5900 /*
5901 * This should normally only loop twice. But because the
5902 * start of the reader inserts an empty page, it causes a
5903 * case where we will loop three times. There should be no
5904 * reason to loop four times unless the ring buffer is a
5905 * recovered persistent ring buffer. For persistent ring buffers,
5906 * invalid pages are reset during recovery, so there may be more
5907 * than 3 contiguous pages can be empty, but less than nr_pages.
5908 */
5909 if (RB_WARN_ON(cpu_buffer, ++nr_loops > max_loops)) {
5910 reader = NULL;
5911 goto out;
5912 }
5913
5914 reader = cpu_buffer->reader_page;
5915
5916 /* If there's more to read, return this page */
5917 if (cpu_buffer->reader_page->read < rb_page_size(reader))
5918 goto out;
5919
5920 /* Never should we have an index greater than the size */
5921 if (RB_WARN_ON(cpu_buffer,
5922 cpu_buffer->reader_page->read > rb_page_size(reader)))
5923 goto out;
5924
5925 /* check if we caught up to the tail */
5926 reader = NULL;
5927 if (cpu_buffer->commit_page == cpu_buffer->reader_page)
5928 goto out;
5929
5930 /* Don't bother swapping if the ring buffer is empty */
5931 if (rb_num_of_entries(cpu_buffer) == 0)
5932 goto out;
5933
5934 /*
5935 * Reset the reader page to size zero.
5936 */
5937 local_set(&cpu_buffer->reader_page->write, 0);
5938 local_set(&cpu_buffer->reader_page->entries, 0);
5939 rb_init_data_page(cpu_buffer->reader_page->page);
5940 cpu_buffer->reader_page->real_end = 0;
5941
5942 spin:
5943 /*
5944 * Splice the empty reader page into the list around the head.
5945 */
5946 reader = rb_set_head_page(cpu_buffer);
5947 if (!reader)
5948 goto out;
5949 cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
5950 cpu_buffer->reader_page->list.prev = reader->list.prev;
5951
5952 /*
5953 * cpu_buffer->pages just needs to point to the buffer, it
5954 * has no specific buffer page to point to. Lets move it out
5955 * of our way so we don't accidentally swap it.
5956 */
5957 cpu_buffer->pages = reader->list.prev;
5958
5959 /* The reader page will be pointing to the new head */
5960 rb_set_list_to_head(&cpu_buffer->reader_page->list);
5961
5962 /*
5963 * We want to make sure we read the overruns after we set up our
5964 * pointers to the next object. The writer side does a
5965 * cmpxchg to cross pages which acts as the mb on the writer
5966 * side. Note, the reader will constantly fail the swap
5967 * while the writer is updating the pointers, so this
5968 * guarantees that the overwrite recorded here is the one we
5969 * want to compare with the last_overrun.
5970 */
5971 smp_mb();
5972 overwrite = local_read(&(cpu_buffer->overrun));
5973
5974 /*
5975 * Here's the tricky part.
5976 *
5977 * We need to move the pointer past the header page.
5978 * But we can only do that if a writer is not currently
5979 * moving it. The page before the header page has the
5980 * flag bit '1' set if it is pointing to the page we want.
5981 * but if the writer is in the process of moving it
5982 * then it will be '2' or already moved '0'.
5983 */
5984
5985 ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
5986
5987 /*
5988 * If we did not convert it, then we must try again.
5989 */
5990 if (!ret)
5991 goto spin;
5992
5993 if (rb_page_commit(reader) & RB_MISSED_EVENTS)
5994 missed_events = -1;
5995
5996 if (cpu_buffer->ring_meta)
5997 rb_update_meta_reader(cpu_buffer, reader);
5998
5999 /*
6000 * Yay! We succeeded in replacing the page.
6001 *
6002 * Now make the new head point back to the reader page.
6003 */
6004 rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
6005 rb_inc_page(&cpu_buffer->head_page);
6006
6007 cpu_buffer->cnt++;
6008 local_inc(&cpu_buffer->pages_read);
6009
6010 /* Finally update the reader page to the new head */
6011 cpu_buffer->reader_page = reader;
6012 cpu_buffer->reader_page->read = 0;
6013
6014 if (overwrite != cpu_buffer->last_overrun) {
6015 cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
6016 cpu_buffer->last_overrun = overwrite;
6017 }
6018
6019 goto again;
6020
6021 out:
6022 /* Update the read_stamp on the first event */
6023 if (reader && reader->read == 0)
6024 cpu_buffer->read_stamp = reader->page->time_stamp;
6025
6026 arch_spin_unlock(&cpu_buffer->lock);
6027 local_irq_restore(flags);
6028
6029 /*
6030 * The writer has preempt disable, wait for it. But not forever
6031 * Although, 1 second is pretty much "forever"
6032 */
6033#define USECS_WAIT 1000000
6034 for (nr_loops = 0; nr_loops < USECS_WAIT; nr_loops++) {
6035 /* If the write is past the end of page, a writer is still updating it */
6036 if (likely(!reader || rb_page_write(reader) <= rb_page_capacity(reader)))
6037 break;
6038
6039 udelay(1);
6040
6041 /* Get the latest version of the reader write value */
6042 smp_rmb();
6043 }
6044
6045 /* The writer is not moving forward? Something is wrong */
6046 if (RB_WARN_ON(cpu_buffer, nr_loops == USECS_WAIT))
6047 reader = NULL;
6048
6049 /*
6050 * Make sure we see any padding after the write update
6051 * (see rb_reset_tail()).
6052 *
6053 * In addition, a writer may be writing on the reader page
6054 * if the page has not been fully filled, so the read barrier
6055 * is also needed to make sure we see the content of what is
6056 * committed by the writer (see rb_set_commit_to_write()).
6057 */
6058 smp_rmb();
6059
6060 if (!cpu_buffer->lost_events)
6061 cpu_buffer->lost_events = missed_events;
6062
6063 return reader;
6064}
6065
6066static struct buffer_page *
6067rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
6068{
6069 return cpu_buffer->remote ? __rb_get_reader_page_from_remote(cpu_buffer) :
6070 __rb_get_reader_page(cpu_buffer);
6071}
6072
6073static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
6074{
6075 struct ring_buffer_event *event;
6076 struct buffer_page *reader;
6077 unsigned length;
6078
6079 reader = rb_get_reader_page(cpu_buffer);
6080
6081 /* This function should not be called when buffer is empty */
6082 if (RB_WARN_ON(cpu_buffer, !reader))
6083 return;
6084
6085 event = rb_reader_event(cpu_buffer);
6086
6087 if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
6088 cpu_buffer->read++;
6089
6090 rb_update_read_stamp(cpu_buffer, event);
6091
6092 length = rb_event_length(event);
6093 cpu_buffer->reader_page->read += length;
6094 cpu_buffer->read_bytes += length;
6095}
6096
6097static void rb_advance_iter(struct ring_buffer_iter *iter)
6098{
6099 struct ring_buffer_per_cpu *cpu_buffer;
6100
6101 cpu_buffer = iter->cpu_buffer;
6102
6103 /* If head == next_event then we need to jump to the next event */
6104 if (iter->head == iter->next_event) {
6105 /* If the event gets overwritten again, there's nothing to do */
6106 if (rb_iter_head_event(iter) == NULL)
6107 return;
6108 }
6109
6110 iter->head = iter->next_event;
6111
6112 /*
6113 * Check if we are at the end of the buffer.
6114 */
6115 if (iter->next_event >= rb_page_size(iter->head_page)) {
6116 /* discarded commits can make the page empty */
6117 if (iter->head_page == cpu_buffer->commit_page)
6118 return;
6119 rb_inc_iter(iter);
6120 return;
6121 }
6122
6123 rb_update_iter_read_stamp(iter, iter->event);
6124}
6125
6126static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
6127{
6128 return cpu_buffer->lost_events;
6129}
6130
6131static struct ring_buffer_event *
6132rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
6133 unsigned long *lost_events)
6134{
6135 struct ring_buffer_event *event;
6136 struct buffer_page *reader;
6137 int nr_loops = 0;
6138
6139 if (ts)
6140 *ts = 0;
6141 again:
6142 /*
6143 * We repeat when a time extend is encountered.
6144 * Since the time extend is always attached to a data event,
6145 * we should never loop more than once.
6146 * (We never hit the following condition more than twice).
6147 */
6148 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
6149 return NULL;
6150
6151 reader = rb_get_reader_page(cpu_buffer);
6152 if (!reader)
6153 return NULL;
6154
6155 event = rb_reader_event(cpu_buffer);
6156
6157 switch (event->type_len) {
6158 case RINGBUF_TYPE_PADDING:
6159 if (rb_null_event(event))
6160 RB_WARN_ON(cpu_buffer, 1);
6161 /*
6162 * Because the writer could be discarding every
6163 * event it creates (which would probably be bad)
6164 * if we were to go back to "again" then we may never
6165 * catch up, and will trigger the warn on, or lock
6166 * the box. Return the padding, and we will release
6167 * the current locks, and try again.
6168 */
6169 return event;
6170
6171 case RINGBUF_TYPE_TIME_EXTEND:
6172 /* Internal data, OK to advance */
6173 rb_advance_reader(cpu_buffer);
6174 goto again;
6175
6176 case RINGBUF_TYPE_TIME_STAMP:
6177 if (ts) {
6178 *ts = rb_event_time_stamp(event);
6179 *ts = rb_fix_abs_ts(*ts, reader->page->time_stamp);
6180 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
6181 cpu_buffer->cpu, ts);
6182 }
6183 /* Internal data, OK to advance */
6184 rb_advance_reader(cpu_buffer);
6185 goto again;
6186
6187 case RINGBUF_TYPE_DATA:
6188 if (ts && !(*ts)) {
6189 *ts = cpu_buffer->read_stamp + event->time_delta;
6190 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
6191 cpu_buffer->cpu, ts);
6192 }
6193 if (lost_events)
6194 *lost_events = rb_lost_events(cpu_buffer);
6195 return event;
6196
6197 default:
6198 RB_WARN_ON(cpu_buffer, 1);
6199 }
6200
6201 return NULL;
6202}
6203EXPORT_SYMBOL_GPL(ring_buffer_peek);
6204
6205static struct ring_buffer_event *
6206rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
6207{
6208 struct trace_buffer *buffer;
6209 struct ring_buffer_per_cpu *cpu_buffer;
6210 struct ring_buffer_event *event;
6211 unsigned long nr_loops = 0;
6212 unsigned long max_loops;
6213
6214 if (ts)
6215 *ts = 0;
6216
6217 cpu_buffer = iter->cpu_buffer;
6218 buffer = cpu_buffer->buffer;
6219 max_loops = cpu_buffer->ring_meta ? cpu_buffer->nr_pages : 3;
6220
6221 /*
6222 * Check if someone performed a consuming read to the buffer
6223 * or removed some pages from the buffer. In these cases,
6224 * iterator was invalidated and we need to reset it.
6225 */
6226 if (unlikely(iter->cache_read != cpu_buffer->read ||
6227 iter->cache_reader_page != cpu_buffer->reader_page ||
6228 iter->cache_pages_removed != cpu_buffer->pages_removed))
6229 rb_iter_reset(iter);
6230
6231 again:
6232 if (ring_buffer_iter_empty(iter))
6233 return NULL;
6234
6235 /*
6236 * As the writer can mess with what the iterator is trying
6237 * to read, just give up if we fail to get an event after
6238 * three tries. The iterator is not as reliable when reading
6239 * the ring buffer with an active write as the consumer is.
6240 * Do not warn if the three failures is reached.
6241 */
6242 if (++nr_loops > max_loops)
6243 return NULL;
6244
6245 if (rb_per_cpu_empty(cpu_buffer))
6246 return NULL;
6247
6248 if (iter->head >= rb_page_size(iter->head_page)) {
6249 rb_inc_iter(iter);
6250 goto again;
6251 }
6252
6253 event = rb_iter_head_event(iter);
6254 if (!event)
6255 goto again;
6256
6257 switch (event->type_len) {
6258 case RINGBUF_TYPE_PADDING:
6259 if (rb_null_event(event)) {
6260 rb_inc_iter(iter);
6261 goto again;
6262 }
6263 rb_advance_iter(iter);
6264 return event;
6265
6266 case RINGBUF_TYPE_TIME_EXTEND:
6267 /* Internal data, OK to advance */
6268 rb_advance_iter(iter);
6269 goto again;
6270
6271 case RINGBUF_TYPE_TIME_STAMP:
6272 if (ts) {
6273 *ts = rb_event_time_stamp(event);
6274 *ts = rb_fix_abs_ts(*ts, iter->head_page->page->time_stamp);
6275 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
6276 cpu_buffer->cpu, ts);
6277 }
6278 /* Internal data, OK to advance */
6279 rb_advance_iter(iter);
6280 goto again;
6281
6282 case RINGBUF_TYPE_DATA:
6283 if (ts && !(*ts)) {
6284 *ts = iter->read_stamp + event->time_delta;
6285 ring_buffer_normalize_time_stamp(buffer,
6286 cpu_buffer->cpu, ts);
6287 }
6288 return event;
6289
6290 default:
6291 RB_WARN_ON(cpu_buffer, 1);
6292 }
6293
6294 return NULL;
6295}
6296EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
6297
6298static inline bool rb_reader_lock(struct ring_buffer_per_cpu *cpu_buffer)
6299{
6300 if (likely(!in_nmi())) {
6301 raw_spin_lock(&cpu_buffer->reader_lock);
6302 return true;
6303 }
6304
6305 /*
6306 * If an NMI die dumps out the content of the ring buffer
6307 * trylock must be used to prevent a deadlock if the NMI
6308 * preempted a task that holds the ring buffer locks. If
6309 * we get the lock then all is fine, if not, then continue
6310 * to do the read, but this can corrupt the ring buffer,
6311 * so it must be permanently disabled from future writes.
6312 * Reading from NMI is a oneshot deal.
6313 */
6314 if (raw_spin_trylock(&cpu_buffer->reader_lock))
6315 return true;
6316
6317 /* Continue without locking, but disable the ring buffer */
6318 atomic_inc(&cpu_buffer->record_disabled);
6319 return false;
6320}
6321
6322static inline void
6323rb_reader_unlock(struct ring_buffer_per_cpu *cpu_buffer, bool locked)
6324{
6325 if (likely(locked))
6326 raw_spin_unlock(&cpu_buffer->reader_lock);
6327}
6328
6329/**
6330 * ring_buffer_peek - peek at the next event to be read
6331 * @buffer: The ring buffer to read
6332 * @cpu: The cpu to peak at
6333 * @ts: The timestamp counter of this event.
6334 * @lost_events: a variable to store if events were lost (may be NULL)
6335 *
6336 * This will return the event that will be read next, but does
6337 * not consume the data.
6338 */
6339struct ring_buffer_event *
6340ring_buffer_peek(struct trace_buffer *buffer, int cpu, u64 *ts,
6341 unsigned long *lost_events)
6342{
6343 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
6344 struct ring_buffer_event *event;
6345 unsigned long flags;
6346 bool dolock;
6347
6348 if (!cpumask_test_cpu(cpu, buffer->cpumask))
6349 return NULL;
6350
6351 again:
6352 local_irq_save(flags);
6353 dolock = rb_reader_lock(cpu_buffer);
6354 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
6355 if (event && event->type_len == RINGBUF_TYPE_PADDING)
6356 rb_advance_reader(cpu_buffer);
6357 rb_reader_unlock(cpu_buffer, dolock);
6358 local_irq_restore(flags);
6359
6360 if (event && event->type_len == RINGBUF_TYPE_PADDING)
6361 goto again;
6362
6363 return event;
6364}
6365
6366/** ring_buffer_iter_dropped - report if there are dropped events
6367 * @iter: The ring buffer iterator
6368 *
6369 * Returns true if there was dropped events since the last peek.
6370 */
6371bool ring_buffer_iter_dropped(struct ring_buffer_iter *iter)
6372{
6373 return iter->missed_events != 0;
6374}
6375EXPORT_SYMBOL_GPL(ring_buffer_iter_dropped);
6376
6377/**
6378 * ring_buffer_iter_peek - peek at the next event to be read
6379 * @iter: The ring buffer iterator
6380 * @ts: The timestamp counter of this event.
6381 *
6382 * This will return the event that will be read next, but does
6383 * not increment the iterator.
6384 */
6385struct ring_buffer_event *
6386ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
6387{
6388 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
6389 struct ring_buffer_event *event;
6390 unsigned long flags;
6391
6392 again:
6393 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
6394 event = rb_iter_peek(iter, ts);
6395 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
6396
6397 if (event && event->type_len == RINGBUF_TYPE_PADDING)
6398 goto again;
6399
6400 return event;
6401}
6402
6403/**
6404 * ring_buffer_consume - return an event and consume it
6405 * @buffer: The ring buffer to get the next event from
6406 * @cpu: the cpu to read the buffer from
6407 * @ts: a variable to store the timestamp (may be NULL)
6408 * @lost_events: a variable to store if events were lost (may be NULL)
6409 *
6410 * Returns the next event in the ring buffer, and that event is consumed.
6411 * Meaning, that sequential reads will keep returning a different event,
6412 * and eventually empty the ring buffer if the producer is slower.
6413 */
6414struct ring_buffer_event *
6415ring_buffer_consume(struct trace_buffer *buffer, int cpu, u64 *ts,
6416 unsigned long *lost_events)
6417{
6418 struct ring_buffer_per_cpu *cpu_buffer;
6419 struct ring_buffer_event *event = NULL;
6420 unsigned long flags;
6421 bool dolock;
6422
6423 again:
6424 /* might be called in atomic */
6425 preempt_disable();
6426
6427 if (!cpumask_test_cpu(cpu, buffer->cpumask))
6428 goto out;
6429
6430 cpu_buffer = buffer->buffers[cpu];
6431 local_irq_save(flags);
6432 dolock = rb_reader_lock(cpu_buffer);
6433
6434 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
6435 if (event) {
6436 cpu_buffer->lost_events = 0;
6437 rb_advance_reader(cpu_buffer);
6438 }
6439
6440 rb_reader_unlock(cpu_buffer, dolock);
6441 local_irq_restore(flags);
6442
6443 out:
6444 preempt_enable();
6445
6446 if (event && event->type_len == RINGBUF_TYPE_PADDING)
6447 goto again;
6448
6449 return event;
6450}
6451EXPORT_SYMBOL_GPL(ring_buffer_consume);
6452
6453/**
6454 * ring_buffer_read_start - start a non consuming read of the buffer
6455 * @buffer: The ring buffer to read from
6456 * @cpu: The cpu buffer to iterate over
6457 * @flags: gfp flags to use for memory allocation
6458 *
6459 * This creates an iterator to allow non-consuming iteration through
6460 * the buffer. If the buffer is disabled for writing, it will produce
6461 * the same information each time, but if the buffer is still writing
6462 * then the first hit of a write will cause the iteration to stop.
6463 *
6464 * Must be paired with ring_buffer_read_finish.
6465 */
6466struct ring_buffer_iter *
6467ring_buffer_read_start(struct trace_buffer *buffer, int cpu, gfp_t flags)
6468{
6469 struct ring_buffer_iter *iter __free(kfree) = kzalloc_obj(*iter, flags);
6470 struct ring_buffer_per_cpu *cpu_buffer;
6471
6472 if (!iter)
6473 return NULL;
6474
6475 if (!cpumask_test_cpu(cpu, buffer->cpumask))
6476 return NULL;
6477
6478 cpu_buffer = buffer->buffers[cpu];
6479
6480 /*
6481 * Only KDB is using GFP_ATOMIC, for the others, lock the buffer to
6482 * prevent concurrent resizing.
6483 */
6484 if (gfpflags_allow_blocking(flags))
6485 mutex_lock(&buffer->mutex);
6486
6487 atomic_inc(&cpu_buffer->resize_disabled);
6488
6489 if (gfpflags_allow_blocking(flags))
6490 mutex_unlock(&buffer->mutex);
6491
6492 /* Holds the entire event: data and meta data. */
6493 iter->event_size = rb_page_capacity(READ_ONCE(cpu_buffer->reader_page));
6494 iter->event = kmalloc(iter->event_size, flags);
6495 if (!iter->event) {
6496 atomic_dec(&cpu_buffer->resize_disabled);
6497 return NULL;
6498 }
6499 iter->cpu_buffer = cpu_buffer;
6500
6501 guard(raw_spinlock_irqsave)(&cpu_buffer->reader_lock);
6502 arch_spin_lock(&cpu_buffer->lock);
6503 rb_iter_reset(iter);
6504 arch_spin_unlock(&cpu_buffer->lock);
6505
6506 return_ptr(iter);
6507}
6508EXPORT_SYMBOL_GPL(ring_buffer_read_start);
6509
6510/**
6511 * ring_buffer_read_finish - finish reading the iterator of the buffer
6512 * @iter: The iterator retrieved by ring_buffer_start
6513 *
6514 * This re-enables resizing of the buffer, and frees the iterator.
6515 */
6516void
6517ring_buffer_read_finish(struct ring_buffer_iter *iter)
6518{
6519 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
6520
6521 /* Use this opportunity to check the integrity of the ring buffer. */
6522 rb_check_pages(cpu_buffer);
6523
6524 atomic_dec(&cpu_buffer->resize_disabled);
6525 kfree(iter->event);
6526 kfree(iter);
6527}
6528EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
6529
6530/**
6531 * ring_buffer_iter_advance - advance the iterator to the next location
6532 * @iter: The ring buffer iterator
6533 *
6534 * Move the location of the iterator such that the next read will
6535 * be the next location of the iterator.
6536 */
6537void ring_buffer_iter_advance(struct ring_buffer_iter *iter)
6538{
6539 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
6540 unsigned long flags;
6541
6542 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
6543 iter->missed_events = 0;
6544 rb_advance_iter(iter);
6545
6546 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
6547}
6548EXPORT_SYMBOL_GPL(ring_buffer_iter_advance);
6549
6550/**
6551 * ring_buffer_size - return the size of the ring buffer (in bytes)
6552 * @buffer: The ring buffer.
6553 * @cpu: The CPU to get ring buffer size from.
6554 */
6555unsigned long ring_buffer_size(struct trace_buffer *buffer, int cpu)
6556{
6557 if (!cpumask_test_cpu(cpu, buffer->cpumask))
6558 return 0;
6559
6560 return rb_subbuf_capacity(buffer) * buffer->buffers[cpu]->nr_pages;
6561}
6562EXPORT_SYMBOL_GPL(ring_buffer_size);
6563
6564/**
6565 * ring_buffer_max_event_size - return the max data size of an event
6566 * @buffer: The ring buffer.
6567 *
6568 * Returns the maximum size an event can be.
6569 */
6570unsigned long ring_buffer_max_event_size(struct trace_buffer *buffer)
6571{
6572 /* If abs timestamp is requested, events have a timestamp too */
6573 if (ring_buffer_time_stamp_abs(buffer))
6574 return rb_subbuf_max_data_size(buffer) - RB_LEN_TIME_EXTEND;
6575
6576 return rb_subbuf_max_data_size(buffer);
6577}
6578EXPORT_SYMBOL_GPL(ring_buffer_max_event_size);
6579
6580static void rb_clear_buffer_page(struct buffer_page *page)
6581{
6582 local_set(&page->write, 0);
6583 local_set(&page->entries, 0);
6584 rb_init_data_page(page->page);
6585 page->read = 0;
6586}
6587
6588/*
6589 * When the buffer is memory mapped to user space, each sub buffer
6590 * has a unique id that is used by the meta data to tell the user
6591 * where the current reader page is.
6592 *
6593 * For a normal allocated ring buffer, the id is saved in the buffer page
6594 * id field, and updated via this function.
6595 *
6596 * But for a fixed memory mapped buffer, the id is already assigned for
6597 * fixed memory ordering in the memory layout and can not be used. Instead
6598 * the index of where the page lies in the memory layout is used.
6599 *
6600 * For the normal pages, set the buffer page id with the passed in @id
6601 * value and return that.
6602 *
6603 * For fixed memory mapped pages, get the page index in the memory layout
6604 * and return that as the id.
6605 */
6606static int rb_page_id(struct ring_buffer_per_cpu *cpu_buffer,
6607 struct buffer_page *bpage, int id)
6608{
6609 /*
6610 * For boot buffers, the id is the index,
6611 * otherwise, set the buffer page with this id
6612 */
6613 if (cpu_buffer->ring_meta)
6614 id = rb_meta_subbuf_idx(cpu_buffer->ring_meta, bpage->page);
6615 else
6616 bpage->id = id;
6617
6618 return id;
6619}
6620
6621static void rb_update_meta_page(struct ring_buffer_per_cpu *cpu_buffer)
6622{
6623 struct trace_buffer_meta *meta = cpu_buffer->meta_page;
6624
6625 if (!meta)
6626 return;
6627
6628 meta->reader.read = cpu_buffer->reader_page->read;
6629 meta->reader.id = rb_page_id(cpu_buffer, cpu_buffer->reader_page,
6630 cpu_buffer->reader_page->id);
6631
6632 meta->reader.lost_events = cpu_buffer->lost_events;
6633
6634 meta->entries = local_read(&cpu_buffer->entries);
6635 meta->overrun = local_read(&cpu_buffer->overrun);
6636 meta->read = cpu_buffer->read;
6637 meta->pages_lost = local_read(&cpu_buffer->pages_lost);
6638 meta->pages_touched = local_read(&cpu_buffer->pages_touched);
6639
6640 /* Some archs do not have data cache coherency between kernel and user-space */
6641 flush_kernel_vmap_range(cpu_buffer->meta_page, PAGE_SIZE);
6642}
6643
6644static void
6645rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
6646{
6647 struct buffer_page *page;
6648
6649 if (cpu_buffer->remote) {
6650 if (!cpu_buffer->remote->reset)
6651 return;
6652
6653 cpu_buffer->remote->reset(cpu_buffer->cpu, cpu_buffer->remote->priv);
6654 rb_read_remote_meta_page(cpu_buffer);
6655
6656 /* Read related values, not covered by the meta-page */
6657 local_set(&cpu_buffer->pages_read, 0);
6658 cpu_buffer->read = 0;
6659 cpu_buffer->read_bytes = 0;
6660 cpu_buffer->last_overrun = 0;
6661 cpu_buffer->reader_page->read = 0;
6662
6663 return;
6664 }
6665
6666 rb_head_page_deactivate(cpu_buffer);
6667
6668 cpu_buffer->head_page
6669 = list_entry(cpu_buffer->pages, struct buffer_page, list);
6670 rb_clear_buffer_page(cpu_buffer->head_page);
6671 list_for_each_entry(page, cpu_buffer->pages, list) {
6672 rb_clear_buffer_page(page);
6673 }
6674
6675 cpu_buffer->tail_page = cpu_buffer->head_page;
6676 cpu_buffer->commit_page = cpu_buffer->head_page;
6677
6678 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
6679 INIT_LIST_HEAD(&cpu_buffer->new_pages);
6680 rb_clear_buffer_page(cpu_buffer->reader_page);
6681
6682 local_set(&cpu_buffer->entries_bytes, 0);
6683 local_set(&cpu_buffer->overrun, 0);
6684 local_set(&cpu_buffer->commit_overrun, 0);
6685 local_set(&cpu_buffer->dropped_events, 0);
6686 local_set(&cpu_buffer->entries, 0);
6687 local_set(&cpu_buffer->committing, 0);
6688 local_set(&cpu_buffer->commits, 0);
6689 local_set(&cpu_buffer->pages_touched, 0);
6690 local_set(&cpu_buffer->pages_lost, 0);
6691 local_set(&cpu_buffer->pages_read, 0);
6692 cpu_buffer->last_pages_touch = 0;
6693 cpu_buffer->shortest_full = 0;
6694 cpu_buffer->read = 0;
6695 cpu_buffer->read_bytes = 0;
6696
6697 rb_time_set(&cpu_buffer->write_stamp, 0);
6698 rb_time_set(&cpu_buffer->before_stamp, 0);
6699
6700 memset(cpu_buffer->event_stamp, 0, sizeof(cpu_buffer->event_stamp));
6701
6702 cpu_buffer->lost_events = 0;
6703 cpu_buffer->last_overrun = 0;
6704
6705 rb_head_page_activate(cpu_buffer);
6706 cpu_buffer->pages_removed = 0;
6707
6708 rb_update_meta_page(cpu_buffer);
6709 if (cpu_buffer->ring_meta) {
6710 struct ring_buffer_cpu_meta *meta = cpu_buffer->ring_meta;
6711
6712 meta->commit_buffer = meta->head_buffer;
6713 }
6714}
6715
6716/* Must have disabled the cpu buffer then done a synchronize_rcu */
6717static void reset_disabled_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
6718{
6719 guard(raw_spinlock_irqsave)(&cpu_buffer->reader_lock);
6720
6721 if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
6722 return;
6723
6724 arch_spin_lock(&cpu_buffer->lock);
6725
6726 rb_reset_cpu(cpu_buffer);
6727
6728 arch_spin_unlock(&cpu_buffer->lock);
6729}
6730
6731/**
6732 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
6733 * @buffer: The ring buffer to reset a per cpu buffer of
6734 * @cpu: The CPU buffer to be reset
6735 */
6736void ring_buffer_reset_cpu(struct trace_buffer *buffer, int cpu)
6737{
6738 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
6739
6740 if (!cpumask_test_cpu(cpu, buffer->cpumask))
6741 return;
6742
6743 /* prevent another thread from changing buffer sizes */
6744 mutex_lock(&buffer->mutex);
6745
6746 atomic_inc(&cpu_buffer->resize_disabled);
6747 atomic_inc(&cpu_buffer->record_disabled);
6748
6749 /* Make sure all commits have finished */
6750 synchronize_rcu();
6751
6752 reset_disabled_cpu_buffer(cpu_buffer);
6753
6754 atomic_dec(&cpu_buffer->record_disabled);
6755 atomic_dec(&cpu_buffer->resize_disabled);
6756
6757 mutex_unlock(&buffer->mutex);
6758}
6759EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
6760
6761/* Flag to ensure proper resetting of atomic variables */
6762#define RESET_BIT (1 << 30)
6763
6764/**
6765 * ring_buffer_reset_online_cpus - reset a ring buffer per CPU buffer
6766 * @buffer: The ring buffer to reset a per cpu buffer of
6767 */
6768void ring_buffer_reset_online_cpus(struct trace_buffer *buffer)
6769{
6770 struct ring_buffer_per_cpu *cpu_buffer;
6771 int cpu;
6772
6773 /* prevent another thread from changing buffer sizes */
6774 mutex_lock(&buffer->mutex);
6775
6776 for_each_online_buffer_cpu(buffer, cpu) {
6777 cpu_buffer = buffer->buffers[cpu];
6778
6779 atomic_add(RESET_BIT, &cpu_buffer->resize_disabled);
6780 atomic_inc(&cpu_buffer->record_disabled);
6781 }
6782
6783 /* Make sure all commits have finished */
6784 synchronize_rcu();
6785
6786 for_each_buffer_cpu(buffer, cpu) {
6787 cpu_buffer = buffer->buffers[cpu];
6788
6789 /*
6790 * If a CPU came online during the synchronize_rcu(), then
6791 * ignore it.
6792 */
6793 if (!(atomic_read(&cpu_buffer->resize_disabled) & RESET_BIT))
6794 continue;
6795
6796 reset_disabled_cpu_buffer(cpu_buffer);
6797
6798 atomic_dec(&cpu_buffer->record_disabled);
6799 atomic_sub(RESET_BIT, &cpu_buffer->resize_disabled);
6800 }
6801
6802 mutex_unlock(&buffer->mutex);
6803}
6804
6805/**
6806 * ring_buffer_reset - reset a ring buffer
6807 * @buffer: The ring buffer to reset all cpu buffers
6808 */
6809void ring_buffer_reset(struct trace_buffer *buffer)
6810{
6811 struct ring_buffer_per_cpu *cpu_buffer;
6812 int cpu;
6813
6814 /* prevent another thread from changing buffer sizes */
6815 mutex_lock(&buffer->mutex);
6816
6817 for_each_buffer_cpu(buffer, cpu) {
6818 cpu_buffer = buffer->buffers[cpu];
6819
6820 atomic_inc(&cpu_buffer->resize_disabled);
6821 atomic_inc(&cpu_buffer->record_disabled);
6822 }
6823
6824 /* Make sure all commits have finished */
6825 synchronize_rcu();
6826
6827 for_each_buffer_cpu(buffer, cpu) {
6828 cpu_buffer = buffer->buffers[cpu];
6829
6830 reset_disabled_cpu_buffer(cpu_buffer);
6831
6832 atomic_dec(&cpu_buffer->record_disabled);
6833 atomic_dec(&cpu_buffer->resize_disabled);
6834 }
6835
6836 mutex_unlock(&buffer->mutex);
6837}
6838EXPORT_SYMBOL_GPL(ring_buffer_reset);
6839
6840/**
6841 * ring_buffer_empty - is the ring buffer empty?
6842 * @buffer: The ring buffer to test
6843 */
6844bool ring_buffer_empty(struct trace_buffer *buffer)
6845{
6846 struct ring_buffer_per_cpu *cpu_buffer;
6847 unsigned long flags;
6848 bool dolock;
6849 bool ret;
6850 int cpu;
6851
6852 /* yes this is racy, but if you don't like the race, lock the buffer */
6853 for_each_buffer_cpu(buffer, cpu) {
6854 cpu_buffer = buffer->buffers[cpu];
6855 local_irq_save(flags);
6856 dolock = rb_reader_lock(cpu_buffer);
6857 ret = rb_per_cpu_empty(cpu_buffer);
6858 rb_reader_unlock(cpu_buffer, dolock);
6859 local_irq_restore(flags);
6860
6861 if (!ret)
6862 return false;
6863 }
6864
6865 return true;
6866}
6867EXPORT_SYMBOL_GPL(ring_buffer_empty);
6868
6869/**
6870 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
6871 * @buffer: The ring buffer
6872 * @cpu: The CPU buffer to test
6873 */
6874bool ring_buffer_empty_cpu(struct trace_buffer *buffer, int cpu)
6875{
6876 struct ring_buffer_per_cpu *cpu_buffer;
6877 unsigned long flags;
6878 bool dolock;
6879 bool ret;
6880
6881 if (!cpumask_test_cpu(cpu, buffer->cpumask))
6882 return true;
6883
6884 cpu_buffer = buffer->buffers[cpu];
6885 local_irq_save(flags);
6886 dolock = rb_reader_lock(cpu_buffer);
6887 ret = rb_per_cpu_empty(cpu_buffer);
6888 rb_reader_unlock(cpu_buffer, dolock);
6889 local_irq_restore(flags);
6890
6891 return ret;
6892}
6893EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
6894
6895int ring_buffer_poll_remote(struct trace_buffer *buffer, int cpu)
6896{
6897 struct ring_buffer_per_cpu *cpu_buffer;
6898
6899 if (cpu != RING_BUFFER_ALL_CPUS) {
6900 if (!cpumask_test_cpu(cpu, buffer->cpumask))
6901 return -EINVAL;
6902
6903 cpu_buffer = buffer->buffers[cpu];
6904
6905 guard(raw_spinlock)(&cpu_buffer->reader_lock);
6906 if (rb_read_remote_meta_page(cpu_buffer))
6907 rb_wakeups(buffer, cpu_buffer);
6908
6909 return 0;
6910 }
6911
6912 guard(cpus_read_lock)();
6913
6914 /*
6915 * Make sure all the ring buffers are up to date before we start reading
6916 * them.
6917 */
6918 for_each_buffer_cpu(buffer, cpu) {
6919 cpu_buffer = buffer->buffers[cpu];
6920
6921 guard(raw_spinlock)(&cpu_buffer->reader_lock);
6922 rb_read_remote_meta_page(cpu_buffer);
6923 }
6924
6925 for_each_buffer_cpu(buffer, cpu) {
6926 cpu_buffer = buffer->buffers[cpu];
6927
6928 if (rb_num_of_entries(cpu_buffer))
6929 rb_wakeups(buffer, cpu_buffer);
6930 }
6931
6932 return 0;
6933}
6934
6935#ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
6936/**
6937 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
6938 * @buffer_a: One buffer to swap with
6939 * @buffer_b: The other buffer to swap with
6940 * @cpu: the CPU of the buffers to swap
6941 *
6942 * This function is useful for tracers that want to take a "snapshot"
6943 * of a CPU buffer and has another back up buffer lying around.
6944 * it is expected that the tracer handles the cpu buffer not being
6945 * used at the moment.
6946 */
6947int ring_buffer_swap_cpu(struct trace_buffer *buffer_a,
6948 struct trace_buffer *buffer_b, int cpu)
6949{
6950 struct ring_buffer_per_cpu *cpu_buffer_a;
6951 struct ring_buffer_per_cpu *cpu_buffer_b;
6952 int ret = -EBUSY;
6953
6954 if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
6955 !cpumask_test_cpu(cpu, buffer_b->cpumask))
6956 return -EINVAL;
6957
6958 cpu_buffer_a = buffer_a->buffers[cpu];
6959 cpu_buffer_b = buffer_b->buffers[cpu];
6960
6961 /* It's up to the callers to not try to swap static buffers */
6962 if (WARN_ON_ONCE(rb_is_static(cpu_buffer_a) || rb_is_static(cpu_buffer_b)))
6963 return -EBUSY;
6964
6965 /* At least make sure the two buffers are somewhat the same */
6966 if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
6967 return -EINVAL;
6968
6969 if (buffer_a->subbuf_order != buffer_b->subbuf_order)
6970 return -EINVAL;
6971
6972 if (atomic_read(&buffer_a->record_disabled))
6973 return -EAGAIN;
6974
6975 if (atomic_read(&buffer_b->record_disabled))
6976 return -EAGAIN;
6977
6978 if (atomic_read(&cpu_buffer_a->record_disabled))
6979 return -EAGAIN;
6980
6981 if (atomic_read(&cpu_buffer_b->record_disabled))
6982 return -EAGAIN;
6983
6984 /*
6985 * We can't do a synchronize_rcu here because this
6986 * function can be called in atomic context.
6987 * Normally this will be called from the same CPU as cpu.
6988 * If not it's up to the caller to protect this.
6989 */
6990 atomic_inc(&cpu_buffer_a->record_disabled);
6991 atomic_inc(&cpu_buffer_b->record_disabled);
6992
6993 /* Do not swap if either buffer is in the process of writing */
6994 if (cpu_buffer_a->current_context)
6995 goto out_dec;
6996 if (cpu_buffer_b->current_context)
6997 goto out_dec;
6998
6999 /*
7000 * When resize is in progress, we cannot swap it because
7001 * it will mess the state of the cpu buffer.
7002 */
7003 if (atomic_read(&buffer_a->resizing))
7004 goto out_dec;
7005 if (atomic_read(&buffer_b->resizing))
7006 goto out_dec;
7007
7008 buffer_a->buffers[cpu] = cpu_buffer_b;
7009 buffer_b->buffers[cpu] = cpu_buffer_a;
7010
7011 cpu_buffer_b->buffer = buffer_a;
7012 cpu_buffer_a->buffer = buffer_b;
7013
7014 ret = 0;
7015
7016out_dec:
7017 atomic_dec(&cpu_buffer_a->record_disabled);
7018 atomic_dec(&cpu_buffer_b->record_disabled);
7019 return ret;
7020}
7021EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
7022#endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
7023
7024/**
7025 * ring_buffer_alloc_read_page - allocate a page to read from buffer
7026 * @buffer: the buffer to allocate for.
7027 * @cpu: the cpu buffer to allocate.
7028 * @rpage: pointer to pass in an already allocated page (can be NULL)
7029 * and returns the allocated page.
7030 *
7031 * This function is used in conjunction with ring_buffer_read_page().
7032 * When reading a full page from the ring buffer, these functions
7033 * can be used to speed up the process. The calling function should
7034 * allocate a few pages first with this function. Then when it
7035 * needs to get pages from the ring buffer, it passes the result
7036 * of this function into ring_buffer_read_page(), which will swap
7037 * the page that was allocated, with the read page of the buffer.
7038 *
7039 * If @rpage is provided, and it has a different order than the current
7040 * subbuffer order, its payload will be freed and re-allocated. If it
7041 * already matches the order, it is simply returned.
7042 *
7043 * Returns:
7044 * 0 on success, < 0 on error
7045 */
7046int ring_buffer_alloc_read_page(struct trace_buffer *buffer, int cpu,
7047 struct buffer_data_read_page **rpage)
7048{
7049 struct ring_buffer_per_cpu *cpu_buffer;
7050 unsigned long flags;
7051 unsigned int order;
7052
7053 if (!cpumask_test_cpu(cpu, buffer->cpumask))
7054 return -ENODEV;
7055
7056 if (!rpage)
7057 return -EINVAL;
7058
7059 order = READ_ONCE(buffer->subbuf_order);
7060
7061 if (*rpage) {
7062 if ((*rpage)->order == order)
7063 return 0;
7064
7065 /* We can reuse rpage, but we discard the payload */
7066 free_pages((unsigned long)(*rpage)->data, (*rpage)->order);
7067 (*rpage)->data = NULL;
7068 } else {
7069 *rpage = kzalloc_obj(**rpage);
7070 if (!*rpage)
7071 return -ENOMEM;
7072 }
7073
7074 (*rpage)->order = order;
7075 cpu_buffer = buffer->buffers[cpu];
7076
7077 local_irq_save(flags);
7078 arch_spin_lock(&cpu_buffer->lock);
7079
7080 if (cpu_buffer->free_page.data) {
7081 **rpage = cpu_buffer->free_page;
7082 cpu_buffer->free_page.data = NULL;
7083 }
7084
7085 arch_spin_unlock(&cpu_buffer->lock);
7086 local_irq_restore(flags);
7087
7088 if ((*rpage)->data) {
7089 rb_init_data_page((*rpage)->data);
7090 } else {
7091 (*rpage)->data = alloc_cpu_data(cpu, (*rpage)->order);
7092 if (!(*rpage)->data) {
7093 kfree(*rpage);
7094 *rpage = NULL;
7095 return -ENOMEM;
7096 }
7097 }
7098
7099 return 0;
7100}
7101EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
7102
7103/**
7104 * ring_buffer_free_read_page - free an allocated read page
7105 * @buffer: the buffer the page was allocate for
7106 * @cpu: the cpu buffer the page came from
7107 * @rpage: the buffer_data_read_page to free
7108 *
7109 * Free a page allocated from ring_buffer_alloc_read_page.
7110 */
7111void ring_buffer_free_read_page(struct trace_buffer *buffer, int cpu,
7112 struct buffer_data_read_page *rpage)
7113{
7114 struct ring_buffer_per_cpu *cpu_buffer;
7115 struct buffer_data_page *dpage;
7116 unsigned long flags;
7117 struct page *page;
7118
7119 if (!buffer || !buffer->buffers || !buffer->buffers[cpu])
7120 return;
7121
7122 if (!rpage)
7123 return;
7124
7125 dpage = rpage->data;
7126 if (!dpage)
7127 goto out;
7128
7129 page = virt_to_page(dpage);
7130
7131 cpu_buffer = buffer->buffers[cpu];
7132
7133 /*
7134 * If the page is still in use someplace else, or order of the page
7135 * is different from the subbuffer order of the buffer -
7136 * we can't reuse it
7137 */
7138 if (page_ref_count(page) > 1 || rpage->order != READ_ONCE(buffer->subbuf_order))
7139 goto out;
7140
7141 local_irq_save(flags);
7142 arch_spin_lock(&cpu_buffer->lock);
7143
7144 if (!cpu_buffer->free_page.data) {
7145 cpu_buffer->free_page = *rpage;
7146 dpage = NULL;
7147 }
7148
7149 arch_spin_unlock(&cpu_buffer->lock);
7150 local_irq_restore(flags);
7151
7152 out:
7153 free_pages((unsigned long)dpage, rpage->order);
7154 kfree(rpage);
7155}
7156EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
7157
7158/**
7159 * ring_buffer_read_page - extract a page from the ring buffer
7160 * @buffer: buffer to extract from
7161 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
7162 * @len: amount to extract
7163 * @cpu: the cpu of the buffer to extract
7164 * @full: should the extraction only happen when the page is full.
7165 *
7166 * This function will pull out a page from the ring buffer and consume it.
7167 * @data_page must be the address of the variable that was returned
7168 * from ring_buffer_alloc_read_page. This is because the page might be used
7169 * to swap with a page in the ring buffer.
7170 *
7171 * for example:
7172 * rpage = ring_buffer_alloc_read_page(buffer, cpu);
7173 * if (IS_ERR(rpage))
7174 * return PTR_ERR(rpage);
7175 * ret = ring_buffer_read_page(buffer, rpage, len, cpu, 0);
7176 * if (ret >= 0)
7177 * process_page(ring_buffer_read_page_data(rpage), ret);
7178 * ring_buffer_free_read_page(buffer, cpu, rpage);
7179 *
7180 * When @full is set, the function will not return true unless
7181 * the writer is off the reader page.
7182 *
7183 * Note: it is up to the calling functions to handle sleeps and wakeups.
7184 * The ring buffer can be used anywhere in the kernel and can not
7185 * blindly call wake_up. The layer that uses the ring buffer must be
7186 * responsible for that.
7187 *
7188 * Returns:
7189 * >=0 if data has been transferred, returns the offset of consumed data.
7190 * <0 if no data has been transferred.
7191 */
7192int ring_buffer_read_page(struct trace_buffer *buffer,
7193 struct buffer_data_read_page *data_page,
7194 size_t len, int cpu, int full)
7195{
7196 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
7197 struct ring_buffer_event *event;
7198 struct buffer_data_page *dpage;
7199 struct buffer_page *reader;
7200 long missed_events;
7201 unsigned int commit;
7202 unsigned int size;
7203 unsigned int read;
7204 u64 save_timestamp;
7205
7206 if (!cpumask_test_cpu(cpu, buffer->cpumask))
7207 return -1;
7208
7209 /*
7210 * If len is not big enough to hold the page header, then
7211 * we can not copy anything.
7212 */
7213 if (len <= BUF_PAGE_HDR_SIZE)
7214 return -1;
7215
7216 len -= BUF_PAGE_HDR_SIZE;
7217
7218 if (!data_page || !data_page->data)
7219 return -1;
7220
7221 dpage = data_page->data;
7222 if (!dpage)
7223 return -1;
7224
7225 len = min_t(size_t, len, rb_read_page_capacity(data_page));
7226
7227 guard(raw_spinlock_irqsave)(&cpu_buffer->reader_lock);
7228
7229 reader = rb_get_reader_page(cpu_buffer);
7230 if (!reader)
7231 return -1;
7232
7233 event = rb_reader_event(cpu_buffer);
7234
7235 read = reader->read;
7236 commit = rb_page_commit(reader);
7237 size = rb_page_size(reader);
7238
7239 /* Check if any events were dropped */
7240 missed_events = cpu_buffer->lost_events;
7241
7242 /*
7243 * It is not possible to swap the reader page if:
7244 * - It has been partially read
7245 * - len is not big enough to read it entirely
7246 * - A writer is still on it
7247 * - The ring buffer is static
7248 * - The order doesn't match
7249 */
7250 if (read || (len < (size - read)) ||
7251 cpu_buffer->reader_page == cpu_buffer->commit_page ||
7252 rb_is_static(cpu_buffer) ||
7253 data_page->order != reader->order) {
7254 struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
7255 unsigned int rpos = read;
7256 unsigned int pos = 0;
7257 unsigned int event_size;
7258 unsigned int flags = 0;
7259
7260 /* If a full page is requested, it cannot be the commit page */
7261 if (full && cpu_buffer->reader_page == cpu_buffer->commit_page)
7262 return -1;
7263
7264 if (len > (size - read))
7265 len = (size - read);
7266
7267 /* Always keep the time extend and data together */
7268 event_size = rb_event_ts_length(event);
7269
7270 if (len < event_size)
7271 return -1;
7272
7273 if (commit & RB_MISSED_EVENTS)
7274 flags = RB_MISSED_EVENTS;
7275
7276 /* save the current timestamp, since the user will need it */
7277 save_timestamp = cpu_buffer->read_stamp;
7278
7279 /* Need to copy one event at a time */
7280 do {
7281 /* We need the size of one event, because
7282 * rb_advance_reader only advances by one event,
7283 * whereas rb_event_ts_length may include the size of
7284 * one or two events.
7285 * We have already ensured there's enough space if this
7286 * is a time extend. */
7287 event_size = rb_event_length(event);
7288 memcpy(dpage->data + pos, rpage->data + rpos, event_size);
7289
7290 len -= event_size;
7291
7292 rb_advance_reader(cpu_buffer);
7293 rpos = reader->read;
7294 pos += event_size;
7295
7296 if (rpos >= size)
7297 break;
7298
7299 event = rb_reader_event(cpu_buffer);
7300 /* Always keep the time extend and data together */
7301 event_size = rb_event_ts_length(event);
7302 } while (len >= event_size);
7303
7304 /* update dpage */
7305 local_set(&dpage->commit, pos | flags);
7306 dpage->time_stamp = save_timestamp;
7307
7308 /* we copied everything to the beginning */
7309 read = 0;
7310 } else {
7311 /* update the entry counter */
7312 cpu_buffer->read += rb_page_entries(reader);
7313 cpu_buffer->read_bytes += rb_page_size(reader);
7314
7315 /* swap the pages */
7316 rb_init_data_page(dpage);
7317 dpage = reader->page;
7318 reader->page = data_page->data;
7319 local_set(&reader->write, 0);
7320 local_set(&reader->entries, 0);
7321 reader->read = 0;
7322 data_page->data = dpage;
7323 if (!missed_events && rb_data_page_commit(dpage) & RB_MISSED_EVENTS)
7324 missed_events = -1;
7325
7326 /*
7327 * Use the real_end for the data size,
7328 * This gives us a chance to store the lost events
7329 * on the page.
7330 */
7331 if (reader->real_end)
7332 local_set(&dpage->commit, reader->real_end);
7333 }
7334
7335 cpu_buffer->lost_events = 0;
7336
7337 size = rb_data_page_size(dpage);
7338 /*
7339 * Set a flag in the commit field if we lost events
7340 */
7341 if (missed_events) {
7342 /*
7343 * If there is room at the end of the page to save the
7344 * missed events, then record it there.
7345 */
7346 if (missed_events > 0 &&
7347 rb_read_page_capacity(data_page) - size >= sizeof(missed_events)) {
7348 memcpy(&dpage->data[size], &missed_events,
7349 sizeof(missed_events));
7350 local_add(RB_MISSED_STORED, &dpage->commit);
7351 size += sizeof(missed_events);
7352 }
7353 /*
7354 * Note, for the persistent ring buffer, the RB_MISSED_EVENTS
7355 * may have been set in the main buffer via the verification code.
7356 * But here, dpage is a copy of that page and has not yet had
7357 * the RB_MISSED_EVENTS set. As for the normal buffers,
7358 * the main write buffer does not set these bits and it needs
7359 * to be set here.
7360 */
7361 local_add(RB_MISSED_EVENTS, &dpage->commit);
7362 }
7363
7364 /*
7365 * This page may be off to user land. Zero it out here.
7366 */
7367 if (size < rb_read_page_capacity(data_page))
7368 memset(&dpage->data[size], 0, rb_read_page_capacity(data_page) - size);
7369
7370 return read;
7371}
7372EXPORT_SYMBOL_GPL(ring_buffer_read_page);
7373
7374/**
7375 * ring_buffer_read_page_data - get pointer to the data in the page.
7376 * @page: the page to get the data from
7377 *
7378 * Returns pointer to the actual data in this page.
7379 */
7380void *ring_buffer_read_page_data(struct buffer_data_read_page *page)
7381{
7382 return page->data;
7383}
7384EXPORT_SYMBOL_GPL(ring_buffer_read_page_data);
7385
7386/**
7387 * ring_buffer_read_page_size - get size of the read page.
7388 * @rpage: the page to get the size from
7389 *
7390 * Returns size of the page in bytes.
7391 */
7392unsigned int ring_buffer_read_page_size(struct buffer_data_read_page *rpage)
7393{
7394 return rpage ? PAGE_SIZE << rpage->order : 0;
7395}
7396EXPORT_SYMBOL_GPL(ring_buffer_read_page_size);
7397
7398/**
7399 * ring_buffer_subbuf_size_get - get size of the sub buffer.
7400 * @buffer: the buffer to get the sub buffer size from
7401 *
7402 * Returns size of the sub buffer, in bytes.
7403 */
7404int ring_buffer_subbuf_size_get(struct trace_buffer *buffer)
7405{
7406 return rb_subbuf_size(buffer);
7407}
7408EXPORT_SYMBOL_GPL(ring_buffer_subbuf_size_get);
7409
7410/**
7411 * ring_buffer_subbuf_order_get - get order of system sub pages in one buffer page.
7412 * @buffer: The ring_buffer to get the system sub page order from
7413 *
7414 * By default, one ring buffer sub page equals to one system page. This parameter
7415 * is configurable, per ring buffer. The size of the ring buffer sub page can be
7416 * extended, but must be an order of system page size.
7417 *
7418 * Returns the order of buffer sub page size, in system pages:
7419 * 0 means the sub buffer size is 1 system page and so forth.
7420 * In case of an error < 0 is returned.
7421 */
7422int ring_buffer_subbuf_order_get(struct trace_buffer *buffer)
7423{
7424 if (!buffer)
7425 return -EINVAL;
7426
7427 return buffer->subbuf_order;
7428}
7429EXPORT_SYMBOL_GPL(ring_buffer_subbuf_order_get);
7430
7431/**
7432 * ring_buffer_subbuf_order_set - set the size of ring buffer sub page.
7433 * @buffer: The ring_buffer to set the new page size.
7434 * @order: Order of the system pages in one sub buffer page
7435 *
7436 * By default, one ring buffer pages equals to one system page. This API can be
7437 * used to set new size of the ring buffer page. The size must be order of
7438 * system page size, that's why the input parameter @order is the order of
7439 * system pages that are allocated for one ring buffer page:
7440 * 0 - 1 system page
7441 * 1 - 2 system pages
7442 * 3 - 4 system pages
7443 * ...
7444 *
7445 * Returns 0 on success or < 0 in case of an error.
7446 */
7447int ring_buffer_subbuf_order_set(struct trace_buffer *buffer, int order)
7448{
7449 struct ring_buffer_per_cpu *cpu_buffer;
7450 struct buffer_page *bpage, *tmp;
7451 unsigned int old_capacity;
7452 unsigned long nr_pages;
7453 int old_order;
7454 int psize;
7455 int err;
7456 int cpu;
7457
7458 if (!buffer || order < 0)
7459 return -EINVAL;
7460
7461 psize = (1 << order) * PAGE_SIZE;
7462 if (psize <= BUF_PAGE_HDR_SIZE)
7463 return -EINVAL;
7464
7465 /* Size of a subbuf cannot be greater than the write counter */
7466 if (psize > RB_WRITE_MASK + 1)
7467 return -EINVAL;
7468
7469 /* prevent another thread from changing buffer sizes */
7470 guard(mutex)(&buffer->mutex);
7471
7472 old_order = buffer->subbuf_order;
7473 if (old_order == order)
7474 return 0;
7475
7476 old_capacity = rb_subbuf_capacity(buffer);
7477
7478 /* The mmap fast path reads subbuf_order without buffer->mutex. */
7479 for_each_buffer_cpu(buffer, cpu) {
7480 if (!cpumask_test_cpu(cpu, buffer->cpumask))
7481 continue;
7482 if (atomic_read(&buffer->buffers[cpu]->resize_disabled))
7483 return -EBUSY;
7484 }
7485
7486 atomic_inc(&buffer->record_disabled);
7487
7488 /* Make sure all commits have finished */
7489 synchronize_rcu();
7490
7491 WRITE_ONCE(buffer->subbuf_order, order);
7492
7493 /* Make sure all new buffers are allocated, before deleting the old ones */
7494 for_each_buffer_cpu(buffer, cpu) {
7495
7496 if (!cpumask_test_cpu(cpu, buffer->cpumask))
7497 continue;
7498
7499 cpu_buffer = buffer->buffers[cpu];
7500
7501 if (atomic_read(&cpu_buffer->resize_disabled)) {
7502 err = -EBUSY;
7503 goto error;
7504 }
7505
7506 /* Update the number of pages to match the new size */
7507 nr_pages = old_capacity * buffer->buffers[cpu]->nr_pages;
7508 nr_pages = DIV_ROUND_UP(nr_pages, rb_subbuf_capacity(buffer));
7509
7510 /* we need a minimum of two pages */
7511 if (nr_pages < 2)
7512 nr_pages = 2;
7513
7514 cpu_buffer->nr_pages_to_update = nr_pages;
7515
7516 /* Include the reader page */
7517 nr_pages++;
7518
7519 /* Allocate the new size buffer */
7520 INIT_LIST_HEAD(&cpu_buffer->new_pages);
7521 if (__rb_allocate_pages(cpu_buffer, nr_pages,
7522 &cpu_buffer->new_pages)) {
7523 /* not enough memory for new pages */
7524 err = -ENOMEM;
7525 goto error;
7526 }
7527 }
7528
7529 for_each_buffer_cpu(buffer, cpu) {
7530 struct buffer_data_read_page old_free_data_page;
7531 struct list_head old_pages;
7532 unsigned long flags;
7533
7534 if (!cpumask_test_cpu(cpu, buffer->cpumask))
7535 continue;
7536
7537 cpu_buffer = buffer->buffers[cpu];
7538
7539 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
7540
7541 /* Clear the head bit to make the link list normal to read */
7542 rb_head_page_deactivate(cpu_buffer);
7543
7544 /*
7545 * Collect buffers from the cpu_buffer pages list and the
7546 * reader_page on old_pages, so they can be freed later when not
7547 * under a spinlock. The pages list is a linked list with no
7548 * head, adding old_pages turns it into a regular list with
7549 * old_pages being the head.
7550 */
7551 list_add(&old_pages, cpu_buffer->pages);
7552 list_add(&cpu_buffer->reader_page->list, &old_pages);
7553
7554 /* One page was allocated for the reader page */
7555 cpu_buffer->reader_page = list_entry(cpu_buffer->new_pages.next,
7556 struct buffer_page, list);
7557 list_del_init(&cpu_buffer->reader_page->list);
7558
7559 /* Install the new pages, remove the head from the list */
7560 cpu_buffer->pages = cpu_buffer->new_pages.next;
7561 list_del_init(&cpu_buffer->new_pages);
7562 cpu_buffer->cnt++;
7563
7564 cpu_buffer->head_page
7565 = list_entry(cpu_buffer->pages, struct buffer_page, list);
7566 cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
7567
7568 cpu_buffer->nr_pages = cpu_buffer->nr_pages_to_update;
7569 cpu_buffer->nr_pages_to_update = 0;
7570
7571 arch_spin_lock(&cpu_buffer->lock);
7572 old_free_data_page = cpu_buffer->free_page;
7573 cpu_buffer->free_page.data = NULL;
7574 arch_spin_unlock(&cpu_buffer->lock);
7575
7576 rb_head_page_activate(cpu_buffer);
7577
7578 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
7579
7580 /* Free old sub buffers */
7581 list_for_each_entry_safe(bpage, tmp, &old_pages, list) {
7582 list_del_init(&bpage->list);
7583 free_buffer_page(bpage);
7584 }
7585 free_pages((unsigned long)old_free_data_page.data, old_free_data_page.order);
7586
7587 rb_check_pages(cpu_buffer);
7588 }
7589
7590 atomic_dec(&buffer->record_disabled);
7591
7592 return 0;
7593
7594error:
7595 WRITE_ONCE(buffer->subbuf_order, old_order);
7596
7597 atomic_dec(&buffer->record_disabled);
7598
7599 for_each_buffer_cpu(buffer, cpu) {
7600 cpu_buffer = buffer->buffers[cpu];
7601
7602 if (!cpu_buffer->nr_pages_to_update)
7603 continue;
7604
7605 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages, list) {
7606 list_del_init(&bpage->list);
7607 free_buffer_page(bpage);
7608 }
7609 }
7610
7611 return err;
7612}
7613EXPORT_SYMBOL_GPL(ring_buffer_subbuf_order_set);
7614
7615static int rb_alloc_meta_page(struct ring_buffer_per_cpu *cpu_buffer)
7616{
7617 struct page *page;
7618
7619 if (cpu_buffer->meta_page)
7620 return 0;
7621
7622 page = alloc_page(GFP_USER | __GFP_ZERO);
7623 if (!page)
7624 return -ENOMEM;
7625
7626 cpu_buffer->meta_page = page_to_virt(page);
7627
7628 return 0;
7629}
7630
7631static void rb_free_meta_page(struct ring_buffer_per_cpu *cpu_buffer)
7632{
7633 unsigned long addr = (unsigned long)cpu_buffer->meta_page;
7634
7635 free_page(addr);
7636 cpu_buffer->meta_page = NULL;
7637}
7638
7639static void rb_setup_ids_meta_page(struct ring_buffer_per_cpu *cpu_buffer,
7640 struct buffer_page **subbuf_ids)
7641{
7642 struct trace_buffer_meta *meta = cpu_buffer->meta_page;
7643 unsigned long nr_subbufs = cpu_buffer->nr_pages + 1;
7644 struct buffer_page *first_subbuf, *subbuf;
7645 unsigned int cnt = 0;
7646 unsigned int id = 0;
7647
7648 id = rb_page_id(cpu_buffer, cpu_buffer->reader_page, id);
7649 subbuf_ids[id++] = cpu_buffer->reader_page;
7650 cnt++;
7651
7652 first_subbuf = subbuf = rb_set_head_page(cpu_buffer);
7653 do {
7654 id = rb_page_id(cpu_buffer, subbuf, id);
7655
7656 if (WARN_ON(id >= nr_subbufs))
7657 break;
7658
7659 subbuf_ids[id] = subbuf;
7660
7661 rb_inc_page(&subbuf);
7662 id++;
7663 cnt++;
7664 } while (subbuf != first_subbuf);
7665
7666 WARN_ON(cnt != nr_subbufs);
7667
7668 /* install subbuf ID to bpage translation */
7669 cpu_buffer->subbuf_ids = subbuf_ids;
7670
7671 meta->meta_struct_len = sizeof(*meta);
7672 meta->nr_subbufs = nr_subbufs;
7673 meta->subbuf_size = rb_subbuf_size(cpu_buffer->buffer);
7674 meta->meta_page_size = meta->subbuf_size;
7675
7676 rb_update_meta_page(cpu_buffer);
7677}
7678
7679static struct ring_buffer_per_cpu *
7680rb_get_mapped_buffer(struct trace_buffer *buffer, int cpu)
7681{
7682 struct ring_buffer_per_cpu *cpu_buffer;
7683
7684 if (!cpumask_test_cpu(cpu, buffer->cpumask))
7685 return ERR_PTR(-EINVAL);
7686
7687 cpu_buffer = buffer->buffers[cpu];
7688
7689 mutex_lock(&cpu_buffer->mapping_lock);
7690
7691 if (!cpu_buffer->user_mapped) {
7692 mutex_unlock(&cpu_buffer->mapping_lock);
7693 return ERR_PTR(-ENODEV);
7694 }
7695
7696 return cpu_buffer;
7697}
7698
7699static void rb_put_mapped_buffer(struct ring_buffer_per_cpu *cpu_buffer)
7700{
7701 mutex_unlock(&cpu_buffer->mapping_lock);
7702}
7703
7704/*
7705 * Fast-path for rb_buffer_(un)map(). Called whenever the meta-page doesn't need
7706 * to be set-up or torn-down.
7707 */
7708static int __rb_inc_dec_mapped(struct ring_buffer_per_cpu *cpu_buffer,
7709 bool inc)
7710{
7711 unsigned long flags;
7712
7713 lockdep_assert_held(&cpu_buffer->mapping_lock);
7714
7715 if (inc && cpu_buffer->user_mapped == UINT_MAX)
7716 return -EBUSY;
7717
7718 if (WARN_ON(!inc && cpu_buffer->user_mapped == 0))
7719 return -EINVAL;
7720
7721 mutex_lock(&cpu_buffer->buffer->mutex);
7722 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
7723
7724 if (inc)
7725 cpu_buffer->user_mapped++;
7726 else
7727 cpu_buffer->user_mapped--;
7728
7729 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
7730 mutex_unlock(&cpu_buffer->buffer->mutex);
7731
7732 return 0;
7733}
7734
7735/*
7736 * +--------------+ pgoff == 0
7737 * | meta page |
7738 * +--------------+ pgoff == 1
7739 * | subbuffer 0 |
7740 * | |
7741 * +--------------+ pgoff == (1 + (1 << subbuf_order))
7742 * | subbuffer 1 |
7743 * | |
7744 * ...
7745 */
7746#ifdef CONFIG_MMU
7747static int __rb_map_vma(struct ring_buffer_per_cpu *cpu_buffer,
7748 struct vm_area_struct *vma)
7749{
7750 unsigned long nr_subbufs, nr_pages, nr_vma_pages;
7751 pgoff_t pgoff = vma_start_pgoff(vma);
7752 unsigned int subbuf_pages, subbuf_order;
7753 struct page **pages __free(kfree) = NULL;
7754 int p = 0, s = 0;
7755 int err;
7756
7757 /* Refuse MP_PRIVATE or writable mappings */
7758 if (vma->vm_flags & VM_WRITE || vma->vm_flags & VM_EXEC ||
7759 !(vma->vm_flags & VM_MAYSHARE))
7760 return -EPERM;
7761
7762 subbuf_order = cpu_buffer->buffer->subbuf_order;
7763 subbuf_pages = 1 << subbuf_order;
7764
7765 if (subbuf_order && pgoff % subbuf_pages)
7766 return -EINVAL;
7767
7768 /*
7769 * Make sure the mapping cannot become writable later. Also tell the VM
7770 * to not touch these pages (VM_DONTCOPY | VM_DONTEXPAND).
7771 */
7772 vm_flags_mod(vma, VM_DONTCOPY | VM_DONTEXPAND | VM_DONTDUMP,
7773 VM_MAYWRITE);
7774
7775 lockdep_assert_held(&cpu_buffer->mapping_lock);
7776
7777 nr_subbufs = cpu_buffer->nr_pages + 1; /* + reader-subbuf */
7778 nr_pages = ((nr_subbufs + 1) << subbuf_order); /* + meta-page */
7779 if (nr_pages <= pgoff)
7780 return -EINVAL;
7781
7782 nr_pages -= pgoff;
7783
7784 nr_vma_pages = vma_pages(vma);
7785 if (!nr_vma_pages || nr_vma_pages > nr_pages)
7786 return -EINVAL;
7787
7788 nr_pages = nr_vma_pages;
7789
7790 pages = kzalloc_objs(*pages, nr_pages);
7791 if (!pages)
7792 return -ENOMEM;
7793
7794 if (!pgoff) {
7795 unsigned long meta_page_padding;
7796
7797 pages[p++] = virt_to_page(cpu_buffer->meta_page);
7798
7799 /*
7800 * Pad with the zero-page to align the meta-page with the
7801 * sub-buffers.
7802 */
7803 meta_page_padding = subbuf_pages - 1;
7804 while (meta_page_padding-- && p < nr_pages) {
7805 unsigned long __maybe_unused zero_addr =
7806 vma->vm_start + (PAGE_SIZE * p);
7807
7808 pages[p++] = ZERO_PAGE(zero_addr);
7809 }
7810 } else {
7811 /* Skip the meta-page */
7812 pgoff -= subbuf_pages;
7813
7814 s += pgoff / subbuf_pages;
7815 }
7816
7817 while (p < nr_pages) {
7818 struct buffer_page *subbuf;
7819 struct page *page;
7820 int off = 0;
7821
7822 if (WARN_ON_ONCE(s >= nr_subbufs))
7823 return -EINVAL;
7824
7825 subbuf = cpu_buffer->subbuf_ids[s];
7826 page = virt_to_page((void *)subbuf->page);
7827
7828 for (; off < (1 << (subbuf_order)); off++, page++) {
7829 if (p >= nr_pages)
7830 break;
7831
7832 pages[p++] = page;
7833 }
7834 s++;
7835 }
7836
7837 err = vm_insert_pages(vma, vma->vm_start, pages, &nr_pages);
7838
7839 return err;
7840}
7841#else
7842static int __rb_map_vma(struct ring_buffer_per_cpu *cpu_buffer,
7843 struct vm_area_struct *vma)
7844{
7845 return -EOPNOTSUPP;
7846}
7847#endif
7848
7849int ring_buffer_map(struct trace_buffer *buffer, int cpu,
7850 struct vm_area_struct *vma)
7851{
7852 struct ring_buffer_per_cpu *cpu_buffer;
7853 struct buffer_page **subbuf_ids;
7854 unsigned long flags;
7855 int err;
7856
7857 if (!cpumask_test_cpu(cpu, buffer->cpumask) || buffer->remote)
7858 return -EINVAL;
7859
7860 cpu_buffer = buffer->buffers[cpu];
7861
7862 guard(mutex)(&cpu_buffer->mapping_lock);
7863
7864 if (cpu_buffer->user_mapped) {
7865 err = __rb_map_vma(cpu_buffer, vma);
7866 if (!err)
7867 err = __rb_inc_dec_mapped(cpu_buffer, true);
7868 return err;
7869 }
7870
7871 /* prevent another thread from changing buffer/sub-buffer sizes */
7872 guard(mutex)(&buffer->mutex);
7873
7874 if (cpu_buffer->nr_pages > rb_static_max_pages())
7875 return -E2BIG;
7876
7877 err = rb_alloc_meta_page(cpu_buffer);
7878 if (err)
7879 return err;
7880
7881 /* subbuf_ids includes the reader while nr_pages does not */
7882 subbuf_ids = kcalloc(cpu_buffer->nr_pages + 1, sizeof(*subbuf_ids), GFP_KERNEL);
7883 if (!subbuf_ids) {
7884 rb_free_meta_page(cpu_buffer);
7885 return -ENOMEM;
7886 }
7887
7888 atomic_inc(&cpu_buffer->resize_disabled);
7889
7890 /*
7891 * Lock all readers to block any subbuf swap until the subbuf IDs are
7892 * assigned.
7893 */
7894 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
7895 rb_setup_ids_meta_page(cpu_buffer, subbuf_ids);
7896
7897 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
7898
7899 err = __rb_map_vma(cpu_buffer, vma);
7900 if (!err) {
7901 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
7902 /* This is the first time it is mapped by user */
7903 cpu_buffer->user_mapped = 1;
7904 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
7905 } else {
7906 kfree(cpu_buffer->subbuf_ids);
7907 cpu_buffer->subbuf_ids = NULL;
7908 rb_free_meta_page(cpu_buffer);
7909 atomic_dec(&cpu_buffer->resize_disabled);
7910 }
7911
7912 return err;
7913}
7914
7915/*
7916 * This is called when a VMA is duplicated (e.g., on fork()) to increment
7917 * the user_mapped counter without remapping pages.
7918 */
7919void ring_buffer_map_dup(struct trace_buffer *buffer, int cpu)
7920{
7921 struct ring_buffer_per_cpu *cpu_buffer;
7922
7923 if (WARN_ON(!cpumask_test_cpu(cpu, buffer->cpumask)))
7924 return;
7925
7926 cpu_buffer = buffer->buffers[cpu];
7927
7928 guard(mutex)(&cpu_buffer->mapping_lock);
7929
7930 if (cpu_buffer->user_mapped)
7931 __rb_inc_dec_mapped(cpu_buffer, true);
7932 else
7933 WARN(1, "Unexpected buffer stat, it should be mapped");
7934}
7935
7936int ring_buffer_unmap(struct trace_buffer *buffer, int cpu)
7937{
7938 struct ring_buffer_per_cpu *cpu_buffer;
7939 unsigned long flags;
7940
7941 if (!cpumask_test_cpu(cpu, buffer->cpumask))
7942 return -EINVAL;
7943
7944 cpu_buffer = buffer->buffers[cpu];
7945
7946 guard(mutex)(&cpu_buffer->mapping_lock);
7947
7948 if (!cpu_buffer->user_mapped) {
7949 return -ENODEV;
7950 } else if (cpu_buffer->user_mapped > 1) {
7951 __rb_inc_dec_mapped(cpu_buffer, false);
7952 return 0;
7953 }
7954
7955 guard(mutex)(&buffer->mutex);
7956 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
7957
7958 /* This is the last user space mapping */
7959 cpu_buffer->user_mapped = 0;
7960
7961 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
7962
7963 kfree(cpu_buffer->subbuf_ids);
7964 cpu_buffer->subbuf_ids = NULL;
7965 rb_free_meta_page(cpu_buffer);
7966 atomic_dec(&cpu_buffer->resize_disabled);
7967
7968 return 0;
7969}
7970
7971int ring_buffer_map_get_reader(struct trace_buffer *buffer, int cpu)
7972{
7973 struct ring_buffer_per_cpu *cpu_buffer;
7974 struct buffer_page *reader;
7975 unsigned long missed_events;
7976 unsigned long reader_size;
7977 unsigned long flags;
7978
7979 cpu_buffer = rb_get_mapped_buffer(buffer, cpu);
7980 if (IS_ERR(cpu_buffer))
7981 return (int)PTR_ERR(cpu_buffer);
7982
7983 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
7984
7985consume:
7986 if (rb_per_cpu_empty(cpu_buffer))
7987 goto out;
7988
7989 reader_size = rb_page_size(cpu_buffer->reader_page);
7990
7991 /*
7992 * There are data to be read on the current reader page, we can
7993 * return to the caller. But before that, we assume the latter will read
7994 * everything. Let's update the kernel reader accordingly.
7995 */
7996 if (cpu_buffer->reader_page->read < reader_size) {
7997 while (cpu_buffer->reader_page->read < reader_size)
7998 rb_advance_reader(cpu_buffer);
7999 goto out;
8000 }
8001
8002 reader = rb_get_reader_page(cpu_buffer);
8003 if (!reader)
8004 goto out;
8005
8006 /* Check if any events were dropped */
8007 missed_events = cpu_buffer->lost_events;
8008
8009 if (missed_events) {
8010 if (cpu_buffer->reader_page != cpu_buffer->commit_page) {
8011 struct buffer_data_page *dpage = reader->page;
8012 unsigned int commit;
8013 /*
8014 * Use the real_end for the data size,
8015 * This gives us a chance to store the lost events
8016 * on the page.
8017 */
8018 if (reader->real_end)
8019 local_set(&dpage->commit, reader->real_end);
8020 /*
8021 * If there is room at the end of the page to save the
8022 * missed events, then record it there.
8023 */
8024 commit = rb_page_size(reader);
8025 if (rb_page_capacity(reader) - commit >= sizeof(missed_events)) {
8026 memcpy(&dpage->data[commit], &missed_events,
8027 sizeof(missed_events));
8028 local_add(RB_MISSED_STORED, &dpage->commit);
8029 }
8030 local_add(RB_MISSED_EVENTS, &dpage->commit);
8031 } else if (!WARN_ONCE(cpu_buffer->reader_page == cpu_buffer->tail_page,
8032 "Reader on commit with %ld missed events",
8033 missed_events)) {
8034 /*
8035 * There shouldn't be any missed events if the tail_page
8036 * is on the reader page. But if the tail page is not on the
8037 * reader page and the commit_page is, that would mean that
8038 * there's a commit_overrun (an interrupt preempted an
8039 * addition of an event and then filled the buffer
8040 * with new events). In this case it's not an
8041 * error, but it should still be reported.
8042 *
8043 * TODO: Add missed events to the page for user space to know.
8044 */
8045 pr_info("Ring buffer [%d] commit overrun lost %ld events at timestamp:%lld\n",
8046 cpu, missed_events, cpu_buffer->reader_page->page->time_stamp);
8047 }
8048 }
8049
8050 cpu_buffer->lost_events = 0;
8051
8052 goto consume;
8053
8054out:
8055 /* Some archs do not have data cache coherency between kernel and user-space */
8056 flush_kernel_vmap_range(cpu_buffer->reader_page->page,
8057 rb_subbuf_size(buffer));
8058
8059 rb_update_meta_page(cpu_buffer);
8060
8061 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
8062 rb_put_mapped_buffer(cpu_buffer);
8063
8064 return 0;
8065}
8066
8067static void rb_cpu_sync(void *data)
8068{
8069 /* Not really needed, but documents what is happening */
8070 smp_rmb();
8071}
8072
8073/*
8074 * We only allocate new buffers, never free them if the CPU goes down.
8075 * If we were to free the buffer, then the user would lose any trace that was in
8076 * the buffer.
8077 */
8078int trace_rb_cpu_prepare(unsigned int cpu, struct hlist_node *node)
8079{
8080 struct trace_buffer *buffer;
8081 long nr_pages_same;
8082 int cpu_i;
8083 unsigned long nr_pages;
8084
8085 buffer = container_of(node, struct trace_buffer, node);
8086 if (cpumask_test_cpu(cpu, buffer->cpumask))
8087 return 0;
8088
8089 nr_pages = 0;
8090 nr_pages_same = 1;
8091 /* check if all cpu sizes are same */
8092 for_each_buffer_cpu(buffer, cpu_i) {
8093 /* fill in the size from first enabled cpu */
8094 if (nr_pages == 0)
8095 nr_pages = buffer->buffers[cpu_i]->nr_pages;
8096 if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
8097 nr_pages_same = 0;
8098 break;
8099 }
8100 }
8101 /* allocate minimum pages, user can later expand it */
8102 if (!nr_pages_same)
8103 nr_pages = 2;
8104 buffer->buffers[cpu] =
8105 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
8106 if (!buffer->buffers[cpu]) {
8107 WARN(1, "failed to allocate ring buffer on CPU %u\n",
8108 cpu);
8109 return -ENOMEM;
8110 }
8111
8112 /*
8113 * Ensure trace_buffer readers observe the newly allocated
8114 * ring_buffer_per_cpu before they check the cpumask. Instead of using a
8115 * read barrier for all readers, send an IPI.
8116 */
8117 if (unlikely(system_state == SYSTEM_RUNNING)) {
8118 on_each_cpu(rb_cpu_sync, NULL, 1);
8119 /* Not really needed, but documents what is happening */
8120 smp_wmb();
8121 }
8122
8123 cpumask_set_cpu(cpu, buffer->cpumask);
8124 return 0;
8125}
8126
8127#ifdef CONFIG_RING_BUFFER_STARTUP_TEST
8128/*
8129 * This is a basic integrity check of the ring buffer.
8130 * Late in the boot cycle this test will run when configured in.
8131 * It will kick off a thread per CPU that will go into a loop
8132 * writing to the per cpu ring buffer various sizes of data.
8133 * Some of the data will be large items, some small.
8134 *
8135 * Another thread is created that goes into a spin, sending out
8136 * IPIs to the other CPUs to also write into the ring buffer.
8137 * this is to test the nesting ability of the buffer.
8138 *
8139 * Basic stats are recorded and reported. If something in the
8140 * ring buffer should happen that's not expected, a big warning
8141 * is displayed and all ring buffers are disabled.
8142 */
8143static struct task_struct *rb_threads[NR_CPUS] __initdata;
8144
8145struct rb_test_data {
8146 struct trace_buffer *buffer;
8147 unsigned long events;
8148 unsigned long bytes_written;
8149 unsigned long bytes_alloc;
8150 unsigned long bytes_dropped;
8151 unsigned long events_nested;
8152 unsigned long bytes_written_nested;
8153 unsigned long bytes_alloc_nested;
8154 unsigned long bytes_dropped_nested;
8155 int min_size_nested;
8156 int max_size_nested;
8157 int max_size;
8158 int min_size;
8159 int cpu;
8160 int cnt;
8161};
8162
8163static struct rb_test_data rb_data[NR_CPUS] __initdata;
8164
8165/* 1 meg per cpu */
8166#define RB_TEST_BUFFER_SIZE 1048576
8167
8168static char rb_string[] __initdata =
8169 "abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
8170 "?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
8171 "!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
8172
8173static bool rb_test_started __initdata;
8174
8175struct rb_item {
8176 int size;
8177 char str[];
8178};
8179
8180static __init int rb_write_something(struct rb_test_data *data, bool nested)
8181{
8182 struct ring_buffer_event *event;
8183 struct rb_item *item;
8184 bool started;
8185 int event_len;
8186 int size;
8187 int len;
8188 int cnt;
8189
8190 /* Have nested writes different that what is written */
8191 cnt = data->cnt + (nested ? 27 : 0);
8192
8193 /* Multiply cnt by ~e, to make some unique increment */
8194 size = (cnt * 68 / 25) % (sizeof(rb_string) - 1);
8195
8196 len = size + sizeof(struct rb_item);
8197
8198 started = rb_test_started;
8199 /* read rb_test_started before checking buffer enabled */
8200 smp_rmb();
8201
8202 event = ring_buffer_lock_reserve(data->buffer, len);
8203 if (!event) {
8204 /* Ignore dropped events before test starts. */
8205 if (started) {
8206 if (nested)
8207 data->bytes_dropped_nested += len;
8208 else
8209 data->bytes_dropped += len;
8210 }
8211 return len;
8212 }
8213
8214 event_len = ring_buffer_event_length(event);
8215
8216 if (RB_WARN_ON(data->buffer, event_len < len))
8217 goto out;
8218
8219 item = ring_buffer_event_data(event);
8220 item->size = size;
8221 memcpy(item->str, rb_string, size);
8222
8223 if (nested) {
8224 data->bytes_alloc_nested += event_len;
8225 data->bytes_written_nested += len;
8226 data->events_nested++;
8227 if (!data->min_size_nested || len < data->min_size_nested)
8228 data->min_size_nested = len;
8229 if (len > data->max_size_nested)
8230 data->max_size_nested = len;
8231 } else {
8232 data->bytes_alloc += event_len;
8233 data->bytes_written += len;
8234 data->events++;
8235 if (!data->min_size || len < data->min_size)
8236 data->max_size = len;
8237 if (len > data->max_size)
8238 data->max_size = len;
8239 }
8240
8241 out:
8242 ring_buffer_unlock_commit(data->buffer);
8243
8244 return 0;
8245}
8246
8247static __init int rb_test(void *arg)
8248{
8249 struct rb_test_data *data = arg;
8250
8251 while (!kthread_should_stop()) {
8252 rb_write_something(data, false);
8253 data->cnt++;
8254
8255 set_current_state(TASK_INTERRUPTIBLE);
8256 /* Now sleep between a min of 100-300us and a max of 1ms */
8257 usleep_range(((data->cnt % 3) + 1) * 100, 1000);
8258 }
8259
8260 return 0;
8261}
8262
8263static __init void rb_ipi(void *ignore)
8264{
8265 struct rb_test_data *data;
8266 int cpu = smp_processor_id();
8267
8268 data = &rb_data[cpu];
8269 rb_write_something(data, true);
8270}
8271
8272static __init int rb_hammer_test(void *arg)
8273{
8274 while (!kthread_should_stop()) {
8275
8276 /* Send an IPI to all cpus to write data! */
8277 smp_call_function(rb_ipi, NULL, 1);
8278 /* No sleep, but for non preempt, let others run */
8279 schedule();
8280 }
8281
8282 return 0;
8283}
8284
8285static __init int test_ringbuffer(void)
8286{
8287 struct task_struct *rb_hammer;
8288 struct trace_buffer *buffer;
8289 int cpu;
8290 int ret = 0;
8291
8292 if (security_locked_down(LOCKDOWN_TRACEFS)) {
8293 pr_warn("Lockdown is enabled, skipping ring buffer tests\n");
8294 return 0;
8295 }
8296
8297 pr_info("Running ring buffer tests...\n");
8298
8299 buffer = ring_buffer_alloc(RB_TEST_BUFFER_SIZE, RB_FL_OVERWRITE);
8300 if (WARN_ON(!buffer))
8301 return 0;
8302
8303 /* Disable buffer so that threads can't write to it yet */
8304 ring_buffer_record_off(buffer);
8305
8306 for_each_online_cpu(cpu) {
8307 rb_data[cpu].buffer = buffer;
8308 rb_data[cpu].cpu = cpu;
8309 rb_data[cpu].cnt = cpu;
8310 rb_threads[cpu] = kthread_run_on_cpu(rb_test, &rb_data[cpu],
8311 cpu, "rbtester/%u");
8312 if (WARN_ON(IS_ERR(rb_threads[cpu]))) {
8313 pr_cont("FAILED\n");
8314 ret = PTR_ERR(rb_threads[cpu]);
8315 goto out_free;
8316 }
8317 }
8318
8319 /* Now create the rb hammer! */
8320 rb_hammer = kthread_run(rb_hammer_test, NULL, "rbhammer");
8321 if (WARN_ON(IS_ERR(rb_hammer))) {
8322 pr_cont("FAILED\n");
8323 ret = PTR_ERR(rb_hammer);
8324 goto out_free;
8325 }
8326
8327 ring_buffer_record_on(buffer);
8328 /*
8329 * Show buffer is enabled before setting rb_test_started.
8330 * Yes there's a small race window where events could be
8331 * dropped and the thread won't catch it. But when a ring
8332 * buffer gets enabled, there will always be some kind of
8333 * delay before other CPUs see it. Thus, we don't care about
8334 * those dropped events. We care about events dropped after
8335 * the threads see that the buffer is active.
8336 */
8337 smp_wmb();
8338 rb_test_started = true;
8339
8340 set_current_state(TASK_INTERRUPTIBLE);
8341 /* Just run for 10 seconds */
8342 schedule_timeout(10 * HZ);
8343
8344 kthread_stop(rb_hammer);
8345
8346 out_free:
8347 for_each_online_cpu(cpu) {
8348 if (IS_ERR_OR_NULL(rb_threads[cpu]))
8349 break;
8350 kthread_stop(rb_threads[cpu]);
8351 }
8352 if (ret) {
8353 ring_buffer_free(buffer);
8354 return ret;
8355 }
8356
8357 /* Report! */
8358 pr_info("finished\n");
8359 for_each_online_cpu(cpu) {
8360 struct ring_buffer_event *event;
8361 struct rb_test_data *data = &rb_data[cpu];
8362 struct rb_item *item;
8363 unsigned long total_events;
8364 unsigned long total_dropped;
8365 unsigned long total_written;
8366 unsigned long total_alloc;
8367 unsigned long total_read = 0;
8368 unsigned long total_size = 0;
8369 unsigned long total_len = 0;
8370 unsigned long total_lost = 0;
8371 unsigned long lost;
8372 int big_event_size;
8373 int small_event_size;
8374
8375 ret = -1;
8376
8377 total_events = data->events + data->events_nested;
8378 total_written = data->bytes_written + data->bytes_written_nested;
8379 total_alloc = data->bytes_alloc + data->bytes_alloc_nested;
8380 total_dropped = data->bytes_dropped + data->bytes_dropped_nested;
8381
8382 big_event_size = data->max_size + data->max_size_nested;
8383 small_event_size = data->min_size + data->min_size_nested;
8384
8385 pr_info("CPU %d:\n", cpu);
8386 pr_info(" events: %ld\n", total_events);
8387 pr_info(" dropped bytes: %ld\n", total_dropped);
8388 pr_info(" alloced bytes: %ld\n", total_alloc);
8389 pr_info(" written bytes: %ld\n", total_written);
8390 pr_info(" biggest event: %d\n", big_event_size);
8391 pr_info(" smallest event: %d\n", small_event_size);
8392
8393 if (RB_WARN_ON(buffer, total_dropped))
8394 break;
8395
8396 ret = 0;
8397
8398 while ((event = ring_buffer_consume(buffer, cpu, NULL, &lost))) {
8399 total_lost += lost;
8400 item = ring_buffer_event_data(event);
8401 total_len += ring_buffer_event_length(event);
8402 total_size += item->size + sizeof(struct rb_item);
8403 if (memcmp(&item->str[0], rb_string, item->size) != 0) {
8404 pr_info("FAILED!\n");
8405 pr_info("buffer had: %.*s\n", item->size, item->str);
8406 pr_info("expected: %.*s\n", item->size, rb_string);
8407 RB_WARN_ON(buffer, 1);
8408 ret = -1;
8409 break;
8410 }
8411 total_read++;
8412 }
8413 if (ret)
8414 break;
8415
8416 ret = -1;
8417
8418 pr_info(" read events: %ld\n", total_read);
8419 pr_info(" lost events: %ld\n", total_lost);
8420 pr_info(" total events: %ld\n", total_lost + total_read);
8421 pr_info(" recorded len bytes: %ld\n", total_len);
8422 pr_info(" recorded size bytes: %ld\n", total_size);
8423 if (total_lost) {
8424 pr_info(" With dropped events, record len and size may not match\n"
8425 " alloced and written from above\n");
8426 } else {
8427 if (RB_WARN_ON(buffer, total_len != total_alloc ||
8428 total_size != total_written))
8429 break;
8430 }
8431 if (RB_WARN_ON(buffer, total_lost + total_read != total_events))
8432 break;
8433
8434 ret = 0;
8435 }
8436 if (!ret)
8437 pr_info("Ring buffer PASSED!\n");
8438
8439 ring_buffer_free(buffer);
8440 return 0;
8441}
8442
8443late_initcall(test_ringbuffer);
8444#endif /* CONFIG_RING_BUFFER_STARTUP_TEST */