Merge branch 'tip/perf/core' of git://git.kernel.org/pub/scm/linux/kernel/git/rostedt...
[linux-2.6.git] / kernel / trace / ring_buffer.c
1 /*
2  * Generic ring buffer
3  *
4  * Copyright (C) 2008 Steven Rostedt <srostedt@redhat.com>
5  */
6 #include <linux/ring_buffer.h>
7 #include <linux/trace_clock.h>
8 #include <linux/spinlock.h>
9 #include <linux/debugfs.h>
10 #include <linux/uaccess.h>
11 #include <linux/hardirq.h>
12 #include <linux/kmemcheck.h>
13 #include <linux/module.h>
14 #include <linux/percpu.h>
15 #include <linux/mutex.h>
16 #include <linux/slab.h>
17 #include <linux/init.h>
18 #include <linux/hash.h>
19 #include <linux/list.h>
20 #include <linux/cpu.h>
21 #include <linux/fs.h>
22
23 #include <asm/local.h>
24 #include "trace.h"
25
26 /*
27  * The ring buffer header is special. We must manually up keep it.
28  */
29 int ring_buffer_print_entry_header(struct trace_seq *s)
30 {
31         int ret;
32
33         ret = trace_seq_printf(s, "# compressed entry header\n");
34         ret = trace_seq_printf(s, "\ttype_len    :    5 bits\n");
35         ret = trace_seq_printf(s, "\ttime_delta  :   27 bits\n");
36         ret = trace_seq_printf(s, "\tarray       :   32 bits\n");
37         ret = trace_seq_printf(s, "\n");
38         ret = trace_seq_printf(s, "\tpadding     : type == %d\n",
39                                RINGBUF_TYPE_PADDING);
40         ret = trace_seq_printf(s, "\ttime_extend : type == %d\n",
41                                RINGBUF_TYPE_TIME_EXTEND);
42         ret = trace_seq_printf(s, "\tdata max type_len  == %d\n",
43                                RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
44
45         return ret;
46 }
47
48 /*
49  * The ring buffer is made up of a list of pages. A separate list of pages is
50  * allocated for each CPU. A writer may only write to a buffer that is
51  * associated with the CPU it is currently executing on.  A reader may read
52  * from any per cpu buffer.
53  *
54  * The reader is special. For each per cpu buffer, the reader has its own
55  * reader page. When a reader has read the entire reader page, this reader
56  * page is swapped with another page in the ring buffer.
57  *
58  * Now, as long as the writer is off the reader page, the reader can do what
59  * ever it wants with that page. The writer will never write to that page
60  * again (as long as it is out of the ring buffer).
61  *
62  * Here's some silly ASCII art.
63  *
64  *   +------+
65  *   |reader|          RING BUFFER
66  *   |page  |
67  *   +------+        +---+   +---+   +---+
68  *                   |   |-->|   |-->|   |
69  *                   +---+   +---+   +---+
70  *                     ^               |
71  *                     |               |
72  *                     +---------------+
73  *
74  *
75  *   +------+
76  *   |reader|          RING BUFFER
77  *   |page  |------------------v
78  *   +------+        +---+   +---+   +---+
79  *                   |   |-->|   |-->|   |
80  *                   +---+   +---+   +---+
81  *                     ^               |
82  *                     |               |
83  *                     +---------------+
84  *
85  *
86  *   +------+
87  *   |reader|          RING BUFFER
88  *   |page  |------------------v
89  *   +------+        +---+   +---+   +---+
90  *      ^            |   |-->|   |-->|   |
91  *      |            +---+   +---+   +---+
92  *      |                              |
93  *      |                              |
94  *      +------------------------------+
95  *
96  *
97  *   +------+
98  *   |buffer|          RING BUFFER
99  *   |page  |------------------v
100  *   +------+        +---+   +---+   +---+
101  *      ^            |   |   |   |-->|   |
102  *      |   New      +---+   +---+   +---+
103  *      |  Reader------^               |
104  *      |   page                       |
105  *      +------------------------------+
106  *
107  *
108  * After we make this swap, the reader can hand this page off to the splice
109  * code and be done with it. It can even allocate a new page if it needs to
110  * and swap that into the ring buffer.
111  *
112  * We will be using cmpxchg soon to make all this lockless.
113  *
114  */
115
116 /*
117  * A fast way to enable or disable all ring buffers is to
118  * call tracing_on or tracing_off. Turning off the ring buffers
119  * prevents all ring buffers from being recorded to.
120  * Turning this switch on, makes it OK to write to the
121  * ring buffer, if the ring buffer is enabled itself.
122  *
123  * There's three layers that must be on in order to write
124  * to the ring buffer.
125  *
126  * 1) This global flag must be set.
127  * 2) The ring buffer must be enabled for recording.
128  * 3) The per cpu buffer must be enabled for recording.
129  *
130  * In case of an anomaly, this global flag has a bit set that
131  * will permantly disable all ring buffers.
132  */
133
134 /*
135  * Global flag to disable all recording to ring buffers
136  *  This has two bits: ON, DISABLED
137  *
138  *  ON   DISABLED
139  * ---- ----------
140  *   0      0        : ring buffers are off
141  *   1      0        : ring buffers are on
142  *   X      1        : ring buffers are permanently disabled
143  */
144
145 enum {
146         RB_BUFFERS_ON_BIT       = 0,
147         RB_BUFFERS_DISABLED_BIT = 1,
148 };
149
150 enum {
151         RB_BUFFERS_ON           = 1 << RB_BUFFERS_ON_BIT,
152         RB_BUFFERS_DISABLED     = 1 << RB_BUFFERS_DISABLED_BIT,
153 };
154
155 static unsigned long ring_buffer_flags __read_mostly = RB_BUFFERS_ON;
156
157 #define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
158
159 /**
160  * tracing_on - enable all tracing buffers
161  *
162  * This function enables all tracing buffers that may have been
163  * disabled with tracing_off.
164  */
165 void tracing_on(void)
166 {
167         set_bit(RB_BUFFERS_ON_BIT, &ring_buffer_flags);
168 }
169 EXPORT_SYMBOL_GPL(tracing_on);
170
171 /**
172  * tracing_off - turn off all tracing buffers
173  *
174  * This function stops all tracing buffers from recording data.
175  * It does not disable any overhead the tracers themselves may
176  * be causing. This function simply causes all recording to
177  * the ring buffers to fail.
178  */
179 void tracing_off(void)
180 {
181         clear_bit(RB_BUFFERS_ON_BIT, &ring_buffer_flags);
182 }
183 EXPORT_SYMBOL_GPL(tracing_off);
184
185 /**
186  * tracing_off_permanent - permanently disable ring buffers
187  *
188  * This function, once called, will disable all ring buffers
189  * permanently.
190  */
191 void tracing_off_permanent(void)
192 {
193         set_bit(RB_BUFFERS_DISABLED_BIT, &ring_buffer_flags);
194 }
195
196 /**
197  * tracing_is_on - show state of ring buffers enabled
198  */
199 int tracing_is_on(void)
200 {
201         return ring_buffer_flags == RB_BUFFERS_ON;
202 }
203 EXPORT_SYMBOL_GPL(tracing_is_on);
204
205 #define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
206 #define RB_ALIGNMENT            4U
207 #define RB_MAX_SMALL_DATA       (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
208 #define RB_EVNT_MIN_SIZE        8U      /* two 32bit words */
209
210 #if !defined(CONFIG_64BIT) || defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS)
211 # define RB_FORCE_8BYTE_ALIGNMENT       0
212 # define RB_ARCH_ALIGNMENT              RB_ALIGNMENT
213 #else
214 # define RB_FORCE_8BYTE_ALIGNMENT       1
215 # define RB_ARCH_ALIGNMENT              8U
216 #endif
217
218 /* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
219 #define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
220
221 enum {
222         RB_LEN_TIME_EXTEND = 8,
223         RB_LEN_TIME_STAMP = 16,
224 };
225
226 #define skip_time_extend(event) \
227         ((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
228
229 static inline int rb_null_event(struct ring_buffer_event *event)
230 {
231         return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
232 }
233
234 static void rb_event_set_padding(struct ring_buffer_event *event)
235 {
236         /* padding has a NULL time_delta */
237         event->type_len = RINGBUF_TYPE_PADDING;
238         event->time_delta = 0;
239 }
240
241 static unsigned
242 rb_event_data_length(struct ring_buffer_event *event)
243 {
244         unsigned length;
245
246         if (event->type_len)
247                 length = event->type_len * RB_ALIGNMENT;
248         else
249                 length = event->array[0];
250         return length + RB_EVNT_HDR_SIZE;
251 }
252
253 /*
254  * Return the length of the given event. Will return
255  * the length of the time extend if the event is a
256  * time extend.
257  */
258 static inline unsigned
259 rb_event_length(struct ring_buffer_event *event)
260 {
261         switch (event->type_len) {
262         case RINGBUF_TYPE_PADDING:
263                 if (rb_null_event(event))
264                         /* undefined */
265                         return -1;
266                 return  event->array[0] + RB_EVNT_HDR_SIZE;
267
268         case RINGBUF_TYPE_TIME_EXTEND:
269                 return RB_LEN_TIME_EXTEND;
270
271         case RINGBUF_TYPE_TIME_STAMP:
272                 return RB_LEN_TIME_STAMP;
273
274         case RINGBUF_TYPE_DATA:
275                 return rb_event_data_length(event);
276         default:
277                 BUG();
278         }
279         /* not hit */
280         return 0;
281 }
282
283 /*
284  * Return total length of time extend and data,
285  *   or just the event length for all other events.
286  */
287 static inline unsigned
288 rb_event_ts_length(struct ring_buffer_event *event)
289 {
290         unsigned len = 0;
291
292         if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
293                 /* time extends include the data event after it */
294                 len = RB_LEN_TIME_EXTEND;
295                 event = skip_time_extend(event);
296         }
297         return len + rb_event_length(event);
298 }
299
300 /**
301  * ring_buffer_event_length - return the length of the event
302  * @event: the event to get the length of
303  *
304  * Returns the size of the data load of a data event.
305  * If the event is something other than a data event, it
306  * returns the size of the event itself. With the exception
307  * of a TIME EXTEND, where it still returns the size of the
308  * data load of the data event after it.
309  */
310 unsigned ring_buffer_event_length(struct ring_buffer_event *event)
311 {
312         unsigned length;
313
314         if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
315                 event = skip_time_extend(event);
316
317         length = rb_event_length(event);
318         if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
319                 return length;
320         length -= RB_EVNT_HDR_SIZE;
321         if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
322                 length -= sizeof(event->array[0]);
323         return length;
324 }
325 EXPORT_SYMBOL_GPL(ring_buffer_event_length);
326
327 /* inline for ring buffer fast paths */
328 static void *
329 rb_event_data(struct ring_buffer_event *event)
330 {
331         if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
332                 event = skip_time_extend(event);
333         BUG_ON(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
334         /* If length is in len field, then array[0] has the data */
335         if (event->type_len)
336                 return (void *)&event->array[0];
337         /* Otherwise length is in array[0] and array[1] has the data */
338         return (void *)&event->array[1];
339 }
340
341 /**
342  * ring_buffer_event_data - return the data of the event
343  * @event: the event to get the data from
344  */
345 void *ring_buffer_event_data(struct ring_buffer_event *event)
346 {
347         return rb_event_data(event);
348 }
349 EXPORT_SYMBOL_GPL(ring_buffer_event_data);
350
351 #define for_each_buffer_cpu(buffer, cpu)                \
352         for_each_cpu(cpu, buffer->cpumask)
353
354 #define TS_SHIFT        27
355 #define TS_MASK         ((1ULL << TS_SHIFT) - 1)
356 #define TS_DELTA_TEST   (~TS_MASK)
357
358 /* Flag when events were overwritten */
359 #define RB_MISSED_EVENTS        (1 << 31)
360 /* Missed count stored at end */
361 #define RB_MISSED_STORED        (1 << 30)
362
363 struct buffer_data_page {
364         u64              time_stamp;    /* page time stamp */
365         local_t          commit;        /* write committed index */
366         unsigned char    data[];        /* data of buffer page */
367 };
368
369 /*
370  * Note, the buffer_page list must be first. The buffer pages
371  * are allocated in cache lines, which means that each buffer
372  * page will be at the beginning of a cache line, and thus
373  * the least significant bits will be zero. We use this to
374  * add flags in the list struct pointers, to make the ring buffer
375  * lockless.
376  */
377 struct buffer_page {
378         struct list_head list;          /* list of buffer pages */
379         local_t          write;         /* index for next write */
380         unsigned         read;          /* index for next read */
381         local_t          entries;       /* entries on this page */
382         unsigned long    real_end;      /* real end of data */
383         struct buffer_data_page *page;  /* Actual data page */
384 };
385
386 /*
387  * The buffer page counters, write and entries, must be reset
388  * atomically when crossing page boundaries. To synchronize this
389  * update, two counters are inserted into the number. One is
390  * the actual counter for the write position or count on the page.
391  *
392  * The other is a counter of updaters. Before an update happens
393  * the update partition of the counter is incremented. This will
394  * allow the updater to update the counter atomically.
395  *
396  * The counter is 20 bits, and the state data is 12.
397  */
398 #define RB_WRITE_MASK           0xfffff
399 #define RB_WRITE_INTCNT         (1 << 20)
400
401 static void rb_init_page(struct buffer_data_page *bpage)
402 {
403         local_set(&bpage->commit, 0);
404 }
405
406 /**
407  * ring_buffer_page_len - the size of data on the page.
408  * @page: The page to read
409  *
410  * Returns the amount of data on the page, including buffer page header.
411  */
412 size_t ring_buffer_page_len(void *page)
413 {
414         return local_read(&((struct buffer_data_page *)page)->commit)
415                 + BUF_PAGE_HDR_SIZE;
416 }
417
418 /*
419  * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
420  * this issue out.
421  */
422 static void free_buffer_page(struct buffer_page *bpage)
423 {
424         free_page((unsigned long)bpage->page);
425         kfree(bpage);
426 }
427
428 /*
429  * We need to fit the time_stamp delta into 27 bits.
430  */
431 static inline int test_time_stamp(u64 delta)
432 {
433         if (delta & TS_DELTA_TEST)
434                 return 1;
435         return 0;
436 }
437
438 #define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
439
440 /* Max payload is BUF_PAGE_SIZE - header (8bytes) */
441 #define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
442
443 int ring_buffer_print_page_header(struct trace_seq *s)
444 {
445         struct buffer_data_page field;
446         int ret;
447
448         ret = trace_seq_printf(s, "\tfield: u64 timestamp;\t"
449                                "offset:0;\tsize:%u;\tsigned:%u;\n",
450                                (unsigned int)sizeof(field.time_stamp),
451                                (unsigned int)is_signed_type(u64));
452
453         ret = trace_seq_printf(s, "\tfield: local_t commit;\t"
454                                "offset:%u;\tsize:%u;\tsigned:%u;\n",
455                                (unsigned int)offsetof(typeof(field), commit),
456                                (unsigned int)sizeof(field.commit),
457                                (unsigned int)is_signed_type(long));
458
459         ret = trace_seq_printf(s, "\tfield: int overwrite;\t"
460                                "offset:%u;\tsize:%u;\tsigned:%u;\n",
461                                (unsigned int)offsetof(typeof(field), commit),
462                                1,
463                                (unsigned int)is_signed_type(long));
464
465         ret = trace_seq_printf(s, "\tfield: char data;\t"
466                                "offset:%u;\tsize:%u;\tsigned:%u;\n",
467                                (unsigned int)offsetof(typeof(field), data),
468                                (unsigned int)BUF_PAGE_SIZE,
469                                (unsigned int)is_signed_type(char));
470
471         return ret;
472 }
473
474 /*
475  * head_page == tail_page && head == tail then buffer is empty.
476  */
477 struct ring_buffer_per_cpu {
478         int                             cpu;
479         atomic_t                        record_disabled;
480         struct ring_buffer              *buffer;
481         spinlock_t                      reader_lock;    /* serialize readers */
482         arch_spinlock_t                 lock;
483         struct lock_class_key           lock_key;
484         struct list_head                *pages;
485         struct buffer_page              *head_page;     /* read from head */
486         struct buffer_page              *tail_page;     /* write to tail */
487         struct buffer_page              *commit_page;   /* committed pages */
488         struct buffer_page              *reader_page;
489         unsigned long                   lost_events;
490         unsigned long                   last_overrun;
491         local_t                         commit_overrun;
492         local_t                         overrun;
493         local_t                         entries;
494         local_t                         committing;
495         local_t                         commits;
496         unsigned long                   read;
497         u64                             write_stamp;
498         u64                             read_stamp;
499 };
500
501 struct ring_buffer {
502         unsigned                        pages;
503         unsigned                        flags;
504         int                             cpus;
505         atomic_t                        record_disabled;
506         cpumask_var_t                   cpumask;
507
508         struct lock_class_key           *reader_lock_key;
509
510         struct mutex                    mutex;
511
512         struct ring_buffer_per_cpu      **buffers;
513
514 #ifdef CONFIG_HOTPLUG_CPU
515         struct notifier_block           cpu_notify;
516 #endif
517         u64                             (*clock)(void);
518 };
519
520 struct ring_buffer_iter {
521         struct ring_buffer_per_cpu      *cpu_buffer;
522         unsigned long                   head;
523         struct buffer_page              *head_page;
524         struct buffer_page              *cache_reader_page;
525         unsigned long                   cache_read;
526         u64                             read_stamp;
527 };
528
529 /* buffer may be either ring_buffer or ring_buffer_per_cpu */
530 #define RB_WARN_ON(b, cond)                                             \
531         ({                                                              \
532                 int _____ret = unlikely(cond);                          \
533                 if (_____ret) {                                         \
534                         if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
535                                 struct ring_buffer_per_cpu *__b =       \
536                                         (void *)b;                      \
537                                 atomic_inc(&__b->buffer->record_disabled); \
538                         } else                                          \
539                                 atomic_inc(&b->record_disabled);        \
540                         WARN_ON(1);                                     \
541                 }                                                       \
542                 _____ret;                                               \
543         })
544
545 /* Up this if you want to test the TIME_EXTENTS and normalization */
546 #define DEBUG_SHIFT 0
547
548 static inline u64 rb_time_stamp(struct ring_buffer *buffer)
549 {
550         /* shift to debug/test normalization and TIME_EXTENTS */
551         return buffer->clock() << DEBUG_SHIFT;
552 }
553
554 u64 ring_buffer_time_stamp(struct ring_buffer *buffer, int cpu)
555 {
556         u64 time;
557
558         preempt_disable_notrace();
559         time = rb_time_stamp(buffer);
560         preempt_enable_no_resched_notrace();
561
562         return time;
563 }
564 EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
565
566 void ring_buffer_normalize_time_stamp(struct ring_buffer *buffer,
567                                       int cpu, u64 *ts)
568 {
569         /* Just stupid testing the normalize function and deltas */
570         *ts >>= DEBUG_SHIFT;
571 }
572 EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
573
574 /*
575  * Making the ring buffer lockless makes things tricky.
576  * Although writes only happen on the CPU that they are on,
577  * and they only need to worry about interrupts. Reads can
578  * happen on any CPU.
579  *
580  * The reader page is always off the ring buffer, but when the
581  * reader finishes with a page, it needs to swap its page with
582  * a new one from the buffer. The reader needs to take from
583  * the head (writes go to the tail). But if a writer is in overwrite
584  * mode and wraps, it must push the head page forward.
585  *
586  * Here lies the problem.
587  *
588  * The reader must be careful to replace only the head page, and
589  * not another one. As described at the top of the file in the
590  * ASCII art, the reader sets its old page to point to the next
591  * page after head. It then sets the page after head to point to
592  * the old reader page. But if the writer moves the head page
593  * during this operation, the reader could end up with the tail.
594  *
595  * We use cmpxchg to help prevent this race. We also do something
596  * special with the page before head. We set the LSB to 1.
597  *
598  * When the writer must push the page forward, it will clear the
599  * bit that points to the head page, move the head, and then set
600  * the bit that points to the new head page.
601  *
602  * We also don't want an interrupt coming in and moving the head
603  * page on another writer. Thus we use the second LSB to catch
604  * that too. Thus:
605  *
606  * head->list->prev->next        bit 1          bit 0
607  *                              -------        -------
608  * Normal page                     0              0
609  * Points to head page             0              1
610  * New head page                   1              0
611  *
612  * Note we can not trust the prev pointer of the head page, because:
613  *
614  * +----+       +-----+        +-----+
615  * |    |------>|  T  |---X--->|  N  |
616  * |    |<------|     |        |     |
617  * +----+       +-----+        +-----+
618  *   ^                           ^ |
619  *   |          +-----+          | |
620  *   +----------|  R  |----------+ |
621  *              |     |<-----------+
622  *              +-----+
623  *
624  * Key:  ---X-->  HEAD flag set in pointer
625  *         T      Tail page
626  *         R      Reader page
627  *         N      Next page
628  *
629  * (see __rb_reserve_next() to see where this happens)
630  *
631  *  What the above shows is that the reader just swapped out
632  *  the reader page with a page in the buffer, but before it
633  *  could make the new header point back to the new page added
634  *  it was preempted by a writer. The writer moved forward onto
635  *  the new page added by the reader and is about to move forward
636  *  again.
637  *
638  *  You can see, it is legitimate for the previous pointer of
639  *  the head (or any page) not to point back to itself. But only
640  *  temporarially.
641  */
642
643 #define RB_PAGE_NORMAL          0UL
644 #define RB_PAGE_HEAD            1UL
645 #define RB_PAGE_UPDATE          2UL
646
647
648 #define RB_FLAG_MASK            3UL
649
650 /* PAGE_MOVED is not part of the mask */
651 #define RB_PAGE_MOVED           4UL
652
653 /*
654  * rb_list_head - remove any bit
655  */
656 static struct list_head *rb_list_head(struct list_head *list)
657 {
658         unsigned long val = (unsigned long)list;
659
660         return (struct list_head *)(val & ~RB_FLAG_MASK);
661 }
662
663 /*
664  * rb_is_head_page - test if the given page is the head page
665  *
666  * Because the reader may move the head_page pointer, we can
667  * not trust what the head page is (it may be pointing to
668  * the reader page). But if the next page is a header page,
669  * its flags will be non zero.
670  */
671 static inline int
672 rb_is_head_page(struct ring_buffer_per_cpu *cpu_buffer,
673                 struct buffer_page *page, struct list_head *list)
674 {
675         unsigned long val;
676
677         val = (unsigned long)list->next;
678
679         if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
680                 return RB_PAGE_MOVED;
681
682         return val & RB_FLAG_MASK;
683 }
684
685 /*
686  * rb_is_reader_page
687  *
688  * The unique thing about the reader page, is that, if the
689  * writer is ever on it, the previous pointer never points
690  * back to the reader page.
691  */
692 static int rb_is_reader_page(struct buffer_page *page)
693 {
694         struct list_head *list = page->list.prev;
695
696         return rb_list_head(list->next) != &page->list;
697 }
698
699 /*
700  * rb_set_list_to_head - set a list_head to be pointing to head.
701  */
702 static void rb_set_list_to_head(struct ring_buffer_per_cpu *cpu_buffer,
703                                 struct list_head *list)
704 {
705         unsigned long *ptr;
706
707         ptr = (unsigned long *)&list->next;
708         *ptr |= RB_PAGE_HEAD;
709         *ptr &= ~RB_PAGE_UPDATE;
710 }
711
712 /*
713  * rb_head_page_activate - sets up head page
714  */
715 static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
716 {
717         struct buffer_page *head;
718
719         head = cpu_buffer->head_page;
720         if (!head)
721                 return;
722
723         /*
724          * Set the previous list pointer to have the HEAD flag.
725          */
726         rb_set_list_to_head(cpu_buffer, head->list.prev);
727 }
728
729 static void rb_list_head_clear(struct list_head *list)
730 {
731         unsigned long *ptr = (unsigned long *)&list->next;
732
733         *ptr &= ~RB_FLAG_MASK;
734 }
735
736 /*
737  * rb_head_page_dactivate - clears head page ptr (for free list)
738  */
739 static void
740 rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
741 {
742         struct list_head *hd;
743
744         /* Go through the whole list and clear any pointers found. */
745         rb_list_head_clear(cpu_buffer->pages);
746
747         list_for_each(hd, cpu_buffer->pages)
748                 rb_list_head_clear(hd);
749 }
750
751 static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
752                             struct buffer_page *head,
753                             struct buffer_page *prev,
754                             int old_flag, int new_flag)
755 {
756         struct list_head *list;
757         unsigned long val = (unsigned long)&head->list;
758         unsigned long ret;
759
760         list = &prev->list;
761
762         val &= ~RB_FLAG_MASK;
763
764         ret = cmpxchg((unsigned long *)&list->next,
765                       val | old_flag, val | new_flag);
766
767         /* check if the reader took the page */
768         if ((ret & ~RB_FLAG_MASK) != val)
769                 return RB_PAGE_MOVED;
770
771         return ret & RB_FLAG_MASK;
772 }
773
774 static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
775                                    struct buffer_page *head,
776                                    struct buffer_page *prev,
777                                    int old_flag)
778 {
779         return rb_head_page_set(cpu_buffer, head, prev,
780                                 old_flag, RB_PAGE_UPDATE);
781 }
782
783 static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
784                                  struct buffer_page *head,
785                                  struct buffer_page *prev,
786                                  int old_flag)
787 {
788         return rb_head_page_set(cpu_buffer, head, prev,
789                                 old_flag, RB_PAGE_HEAD);
790 }
791
792 static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
793                                    struct buffer_page *head,
794                                    struct buffer_page *prev,
795                                    int old_flag)
796 {
797         return rb_head_page_set(cpu_buffer, head, prev,
798                                 old_flag, RB_PAGE_NORMAL);
799 }
800
801 static inline void rb_inc_page(struct ring_buffer_per_cpu *cpu_buffer,
802                                struct buffer_page **bpage)
803 {
804         struct list_head *p = rb_list_head((*bpage)->list.next);
805
806         *bpage = list_entry(p, struct buffer_page, list);
807 }
808
809 static struct buffer_page *
810 rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
811 {
812         struct buffer_page *head;
813         struct buffer_page *page;
814         struct list_head *list;
815         int i;
816
817         if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
818                 return NULL;
819
820         /* sanity check */
821         list = cpu_buffer->pages;
822         if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
823                 return NULL;
824
825         page = head = cpu_buffer->head_page;
826         /*
827          * It is possible that the writer moves the header behind
828          * where we started, and we miss in one loop.
829          * A second loop should grab the header, but we'll do
830          * three loops just because I'm paranoid.
831          */
832         for (i = 0; i < 3; i++) {
833                 do {
834                         if (rb_is_head_page(cpu_buffer, page, page->list.prev)) {
835                                 cpu_buffer->head_page = page;
836                                 return page;
837                         }
838                         rb_inc_page(cpu_buffer, &page);
839                 } while (page != head);
840         }
841
842         RB_WARN_ON(cpu_buffer, 1);
843
844         return NULL;
845 }
846
847 static int rb_head_page_replace(struct buffer_page *old,
848                                 struct buffer_page *new)
849 {
850         unsigned long *ptr = (unsigned long *)&old->list.prev->next;
851         unsigned long val;
852         unsigned long ret;
853
854         val = *ptr & ~RB_FLAG_MASK;
855         val |= RB_PAGE_HEAD;
856
857         ret = cmpxchg(ptr, val, (unsigned long)&new->list);
858
859         return ret == val;
860 }
861
862 /*
863  * rb_tail_page_update - move the tail page forward
864  *
865  * Returns 1 if moved tail page, 0 if someone else did.
866  */
867 static int rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
868                                struct buffer_page *tail_page,
869                                struct buffer_page *next_page)
870 {
871         struct buffer_page *old_tail;
872         unsigned long old_entries;
873         unsigned long old_write;
874         int ret = 0;
875
876         /*
877          * The tail page now needs to be moved forward.
878          *
879          * We need to reset the tail page, but without messing
880          * with possible erasing of data brought in by interrupts
881          * that have moved the tail page and are currently on it.
882          *
883          * We add a counter to the write field to denote this.
884          */
885         old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
886         old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
887
888         /*
889          * Just make sure we have seen our old_write and synchronize
890          * with any interrupts that come in.
891          */
892         barrier();
893
894         /*
895          * If the tail page is still the same as what we think
896          * it is, then it is up to us to update the tail
897          * pointer.
898          */
899         if (tail_page == cpu_buffer->tail_page) {
900                 /* Zero the write counter */
901                 unsigned long val = old_write & ~RB_WRITE_MASK;
902                 unsigned long eval = old_entries & ~RB_WRITE_MASK;
903
904                 /*
905                  * This will only succeed if an interrupt did
906                  * not come in and change it. In which case, we
907                  * do not want to modify it.
908                  *
909                  * We add (void) to let the compiler know that we do not care
910                  * about the return value of these functions. We use the
911                  * cmpxchg to only update if an interrupt did not already
912                  * do it for us. If the cmpxchg fails, we don't care.
913                  */
914                 (void)local_cmpxchg(&next_page->write, old_write, val);
915                 (void)local_cmpxchg(&next_page->entries, old_entries, eval);
916
917                 /*
918                  * No need to worry about races with clearing out the commit.
919                  * it only can increment when a commit takes place. But that
920                  * only happens in the outer most nested commit.
921                  */
922                 local_set(&next_page->page->commit, 0);
923
924                 old_tail = cmpxchg(&cpu_buffer->tail_page,
925                                    tail_page, next_page);
926
927                 if (old_tail == tail_page)
928                         ret = 1;
929         }
930
931         return ret;
932 }
933
934 static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
935                           struct buffer_page *bpage)
936 {
937         unsigned long val = (unsigned long)bpage;
938
939         if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK))
940                 return 1;
941
942         return 0;
943 }
944
945 /**
946  * rb_check_list - make sure a pointer to a list has the last bits zero
947  */
948 static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer,
949                          struct list_head *list)
950 {
951         if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev))
952                 return 1;
953         if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next))
954                 return 1;
955         return 0;
956 }
957
958 /**
959  * check_pages - integrity check of buffer pages
960  * @cpu_buffer: CPU buffer with pages to test
961  *
962  * As a safety measure we check to make sure the data pages have not
963  * been corrupted.
964  */
965 static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
966 {
967         struct list_head *head = cpu_buffer->pages;
968         struct buffer_page *bpage, *tmp;
969
970         rb_head_page_deactivate(cpu_buffer);
971
972         if (RB_WARN_ON(cpu_buffer, head->next->prev != head))
973                 return -1;
974         if (RB_WARN_ON(cpu_buffer, head->prev->next != head))
975                 return -1;
976
977         if (rb_check_list(cpu_buffer, head))
978                 return -1;
979
980         list_for_each_entry_safe(bpage, tmp, head, list) {
981                 if (RB_WARN_ON(cpu_buffer,
982                                bpage->list.next->prev != &bpage->list))
983                         return -1;
984                 if (RB_WARN_ON(cpu_buffer,
985                                bpage->list.prev->next != &bpage->list))
986                         return -1;
987                 if (rb_check_list(cpu_buffer, &bpage->list))
988                         return -1;
989         }
990
991         rb_head_page_activate(cpu_buffer);
992
993         return 0;
994 }
995
996 static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
997                              unsigned nr_pages)
998 {
999         struct buffer_page *bpage, *tmp;
1000         unsigned long addr;
1001         LIST_HEAD(pages);
1002         unsigned i;
1003
1004         WARN_ON(!nr_pages);
1005
1006         for (i = 0; i < nr_pages; i++) {
1007                 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1008                                     GFP_KERNEL, cpu_to_node(cpu_buffer->cpu));
1009                 if (!bpage)
1010                         goto free_pages;
1011
1012                 rb_check_bpage(cpu_buffer, bpage);
1013
1014                 list_add(&bpage->list, &pages);
1015
1016                 addr = __get_free_page(GFP_KERNEL);
1017                 if (!addr)
1018                         goto free_pages;
1019                 bpage->page = (void *)addr;
1020                 rb_init_page(bpage->page);
1021         }
1022
1023         /*
1024          * The ring buffer page list is a circular list that does not
1025          * start and end with a list head. All page list items point to
1026          * other pages.
1027          */
1028         cpu_buffer->pages = pages.next;
1029         list_del(&pages);
1030
1031         rb_check_pages(cpu_buffer);
1032
1033         return 0;
1034
1035  free_pages:
1036         list_for_each_entry_safe(bpage, tmp, &pages, list) {
1037                 list_del_init(&bpage->list);
1038                 free_buffer_page(bpage);
1039         }
1040         return -ENOMEM;
1041 }
1042
1043 static struct ring_buffer_per_cpu *
1044 rb_allocate_cpu_buffer(struct ring_buffer *buffer, int cpu)
1045 {
1046         struct ring_buffer_per_cpu *cpu_buffer;
1047         struct buffer_page *bpage;
1048         unsigned long addr;
1049         int ret;
1050
1051         cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1052                                   GFP_KERNEL, cpu_to_node(cpu));
1053         if (!cpu_buffer)
1054                 return NULL;
1055
1056         cpu_buffer->cpu = cpu;
1057         cpu_buffer->buffer = buffer;
1058         spin_lock_init(&cpu_buffer->reader_lock);
1059         lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1060         cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1061
1062         bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1063                             GFP_KERNEL, cpu_to_node(cpu));
1064         if (!bpage)
1065                 goto fail_free_buffer;
1066
1067         rb_check_bpage(cpu_buffer, bpage);
1068
1069         cpu_buffer->reader_page = bpage;
1070         addr = __get_free_page(GFP_KERNEL);
1071         if (!addr)
1072                 goto fail_free_reader;
1073         bpage->page = (void *)addr;
1074         rb_init_page(bpage->page);
1075
1076         INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1077
1078         ret = rb_allocate_pages(cpu_buffer, buffer->pages);
1079         if (ret < 0)
1080                 goto fail_free_reader;
1081
1082         cpu_buffer->head_page
1083                 = list_entry(cpu_buffer->pages, struct buffer_page, list);
1084         cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1085
1086         rb_head_page_activate(cpu_buffer);
1087
1088         return cpu_buffer;
1089
1090  fail_free_reader:
1091         free_buffer_page(cpu_buffer->reader_page);
1092
1093  fail_free_buffer:
1094         kfree(cpu_buffer);
1095         return NULL;
1096 }
1097
1098 static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1099 {
1100         struct list_head *head = cpu_buffer->pages;
1101         struct buffer_page *bpage, *tmp;
1102
1103         free_buffer_page(cpu_buffer->reader_page);
1104
1105         rb_head_page_deactivate(cpu_buffer);
1106
1107         if (head) {
1108                 list_for_each_entry_safe(bpage, tmp, head, list) {
1109                         list_del_init(&bpage->list);
1110                         free_buffer_page(bpage);
1111                 }
1112                 bpage = list_entry(head, struct buffer_page, list);
1113                 free_buffer_page(bpage);
1114         }
1115
1116         kfree(cpu_buffer);
1117 }
1118
1119 #ifdef CONFIG_HOTPLUG_CPU
1120 static int rb_cpu_notify(struct notifier_block *self,
1121                          unsigned long action, void *hcpu);
1122 #endif
1123
1124 /**
1125  * ring_buffer_alloc - allocate a new ring_buffer
1126  * @size: the size in bytes per cpu that is needed.
1127  * @flags: attributes to set for the ring buffer.
1128  *
1129  * Currently the only flag that is available is the RB_FL_OVERWRITE
1130  * flag. This flag means that the buffer will overwrite old data
1131  * when the buffer wraps. If this flag is not set, the buffer will
1132  * drop data when the tail hits the head.
1133  */
1134 struct ring_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1135                                         struct lock_class_key *key)
1136 {
1137         struct ring_buffer *buffer;
1138         int bsize;
1139         int cpu;
1140
1141         /* keep it in its own cache line */
1142         buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1143                          GFP_KERNEL);
1144         if (!buffer)
1145                 return NULL;
1146
1147         if (!alloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1148                 goto fail_free_buffer;
1149
1150         buffer->pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1151         buffer->flags = flags;
1152         buffer->clock = trace_clock_local;
1153         buffer->reader_lock_key = key;
1154
1155         /* need at least two pages */
1156         if (buffer->pages < 2)
1157                 buffer->pages = 2;
1158
1159         /*
1160          * In case of non-hotplug cpu, if the ring-buffer is allocated
1161          * in early initcall, it will not be notified of secondary cpus.
1162          * In that off case, we need to allocate for all possible cpus.
1163          */
1164 #ifdef CONFIG_HOTPLUG_CPU
1165         get_online_cpus();
1166         cpumask_copy(buffer->cpumask, cpu_online_mask);
1167 #else
1168         cpumask_copy(buffer->cpumask, cpu_possible_mask);
1169 #endif
1170         buffer->cpus = nr_cpu_ids;
1171
1172         bsize = sizeof(void *) * nr_cpu_ids;
1173         buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1174                                   GFP_KERNEL);
1175         if (!buffer->buffers)
1176                 goto fail_free_cpumask;
1177
1178         for_each_buffer_cpu(buffer, cpu) {
1179                 buffer->buffers[cpu] =
1180                         rb_allocate_cpu_buffer(buffer, cpu);
1181                 if (!buffer->buffers[cpu])
1182                         goto fail_free_buffers;
1183         }
1184
1185 #ifdef CONFIG_HOTPLUG_CPU
1186         buffer->cpu_notify.notifier_call = rb_cpu_notify;
1187         buffer->cpu_notify.priority = 0;
1188         register_cpu_notifier(&buffer->cpu_notify);
1189 #endif
1190
1191         put_online_cpus();
1192         mutex_init(&buffer->mutex);
1193
1194         return buffer;
1195
1196  fail_free_buffers:
1197         for_each_buffer_cpu(buffer, cpu) {
1198                 if (buffer->buffers[cpu])
1199                         rb_free_cpu_buffer(buffer->buffers[cpu]);
1200         }
1201         kfree(buffer->buffers);
1202
1203  fail_free_cpumask:
1204         free_cpumask_var(buffer->cpumask);
1205         put_online_cpus();
1206
1207  fail_free_buffer:
1208         kfree(buffer);
1209         return NULL;
1210 }
1211 EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1212
1213 /**
1214  * ring_buffer_free - free a ring buffer.
1215  * @buffer: the buffer to free.
1216  */
1217 void
1218 ring_buffer_free(struct ring_buffer *buffer)
1219 {
1220         int cpu;
1221
1222         get_online_cpus();
1223
1224 #ifdef CONFIG_HOTPLUG_CPU
1225         unregister_cpu_notifier(&buffer->cpu_notify);
1226 #endif
1227
1228         for_each_buffer_cpu(buffer, cpu)
1229                 rb_free_cpu_buffer(buffer->buffers[cpu]);
1230
1231         put_online_cpus();
1232
1233         kfree(buffer->buffers);
1234         free_cpumask_var(buffer->cpumask);
1235
1236         kfree(buffer);
1237 }
1238 EXPORT_SYMBOL_GPL(ring_buffer_free);
1239
1240 void ring_buffer_set_clock(struct ring_buffer *buffer,
1241                            u64 (*clock)(void))
1242 {
1243         buffer->clock = clock;
1244 }
1245
1246 static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1247
1248 static void
1249 rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned nr_pages)
1250 {
1251         struct buffer_page *bpage;
1252         struct list_head *p;
1253         unsigned i;
1254
1255         spin_lock_irq(&cpu_buffer->reader_lock);
1256         rb_head_page_deactivate(cpu_buffer);
1257
1258         for (i = 0; i < nr_pages; i++) {
1259                 if (RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages)))
1260                         goto out;
1261                 p = cpu_buffer->pages->next;
1262                 bpage = list_entry(p, struct buffer_page, list);
1263                 list_del_init(&bpage->list);
1264                 free_buffer_page(bpage);
1265         }
1266         if (RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages)))
1267                 goto out;
1268
1269         rb_reset_cpu(cpu_buffer);
1270         rb_check_pages(cpu_buffer);
1271
1272 out:
1273         spin_unlock_irq(&cpu_buffer->reader_lock);
1274 }
1275
1276 static void
1277 rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer,
1278                 struct list_head *pages, unsigned nr_pages)
1279 {
1280         struct buffer_page *bpage;
1281         struct list_head *p;
1282         unsigned i;
1283
1284         spin_lock_irq(&cpu_buffer->reader_lock);
1285         rb_head_page_deactivate(cpu_buffer);
1286
1287         for (i = 0; i < nr_pages; i++) {
1288                 if (RB_WARN_ON(cpu_buffer, list_empty(pages)))
1289                         goto out;
1290                 p = pages->next;
1291                 bpage = list_entry(p, struct buffer_page, list);
1292                 list_del_init(&bpage->list);
1293                 list_add_tail(&bpage->list, cpu_buffer->pages);
1294         }
1295         rb_reset_cpu(cpu_buffer);
1296         rb_check_pages(cpu_buffer);
1297
1298 out:
1299         spin_unlock_irq(&cpu_buffer->reader_lock);
1300 }
1301
1302 /**
1303  * ring_buffer_resize - resize the ring buffer
1304  * @buffer: the buffer to resize.
1305  * @size: the new size.
1306  *
1307  * Minimum size is 2 * BUF_PAGE_SIZE.
1308  *
1309  * Returns -1 on failure.
1310  */
1311 int ring_buffer_resize(struct ring_buffer *buffer, unsigned long size)
1312 {
1313         struct ring_buffer_per_cpu *cpu_buffer;
1314         unsigned nr_pages, rm_pages, new_pages;
1315         struct buffer_page *bpage, *tmp;
1316         unsigned long buffer_size;
1317         unsigned long addr;
1318         LIST_HEAD(pages);
1319         int i, cpu;
1320
1321         /*
1322          * Always succeed at resizing a non-existent buffer:
1323          */
1324         if (!buffer)
1325                 return size;
1326
1327         size = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1328         size *= BUF_PAGE_SIZE;
1329         buffer_size = buffer->pages * BUF_PAGE_SIZE;
1330
1331         /* we need a minimum of two pages */
1332         if (size < BUF_PAGE_SIZE * 2)
1333                 size = BUF_PAGE_SIZE * 2;
1334
1335         if (size == buffer_size)
1336                 return size;
1337
1338         atomic_inc(&buffer->record_disabled);
1339
1340         /* Make sure all writers are done with this buffer. */
1341         synchronize_sched();
1342
1343         mutex_lock(&buffer->mutex);
1344         get_online_cpus();
1345
1346         nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1347
1348         if (size < buffer_size) {
1349
1350                 /* easy case, just free pages */
1351                 if (RB_WARN_ON(buffer, nr_pages >= buffer->pages))
1352                         goto out_fail;
1353
1354                 rm_pages = buffer->pages - nr_pages;
1355
1356                 for_each_buffer_cpu(buffer, cpu) {
1357                         cpu_buffer = buffer->buffers[cpu];
1358                         rb_remove_pages(cpu_buffer, rm_pages);
1359                 }
1360                 goto out;
1361         }
1362
1363         /*
1364          * This is a bit more difficult. We only want to add pages
1365          * when we can allocate enough for all CPUs. We do this
1366          * by allocating all the pages and storing them on a local
1367          * link list. If we succeed in our allocation, then we
1368          * add these pages to the cpu_buffers. Otherwise we just free
1369          * them all and return -ENOMEM;
1370          */
1371         if (RB_WARN_ON(buffer, nr_pages <= buffer->pages))
1372                 goto out_fail;
1373
1374         new_pages = nr_pages - buffer->pages;
1375
1376         for_each_buffer_cpu(buffer, cpu) {
1377                 for (i = 0; i < new_pages; i++) {
1378                         bpage = kzalloc_node(ALIGN(sizeof(*bpage),
1379                                                   cache_line_size()),
1380                                             GFP_KERNEL, cpu_to_node(cpu));
1381                         if (!bpage)
1382                                 goto free_pages;
1383                         list_add(&bpage->list, &pages);
1384                         addr = __get_free_page(GFP_KERNEL);
1385                         if (!addr)
1386                                 goto free_pages;
1387                         bpage->page = (void *)addr;
1388                         rb_init_page(bpage->page);
1389                 }
1390         }
1391
1392         for_each_buffer_cpu(buffer, cpu) {
1393                 cpu_buffer = buffer->buffers[cpu];
1394                 rb_insert_pages(cpu_buffer, &pages, new_pages);
1395         }
1396
1397         if (RB_WARN_ON(buffer, !list_empty(&pages)))
1398                 goto out_fail;
1399
1400  out:
1401         buffer->pages = nr_pages;
1402         put_online_cpus();
1403         mutex_unlock(&buffer->mutex);
1404
1405         atomic_dec(&buffer->record_disabled);
1406
1407         return size;
1408
1409  free_pages:
1410         list_for_each_entry_safe(bpage, tmp, &pages, list) {
1411                 list_del_init(&bpage->list);
1412                 free_buffer_page(bpage);
1413         }
1414         put_online_cpus();
1415         mutex_unlock(&buffer->mutex);
1416         atomic_dec(&buffer->record_disabled);
1417         return -ENOMEM;
1418
1419         /*
1420          * Something went totally wrong, and we are too paranoid
1421          * to even clean up the mess.
1422          */
1423  out_fail:
1424         put_online_cpus();
1425         mutex_unlock(&buffer->mutex);
1426         atomic_dec(&buffer->record_disabled);
1427         return -1;
1428 }
1429 EXPORT_SYMBOL_GPL(ring_buffer_resize);
1430
1431 void ring_buffer_change_overwrite(struct ring_buffer *buffer, int val)
1432 {
1433         mutex_lock(&buffer->mutex);
1434         if (val)
1435                 buffer->flags |= RB_FL_OVERWRITE;
1436         else
1437                 buffer->flags &= ~RB_FL_OVERWRITE;
1438         mutex_unlock(&buffer->mutex);
1439 }
1440 EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
1441
1442 static inline void *
1443 __rb_data_page_index(struct buffer_data_page *bpage, unsigned index)
1444 {
1445         return bpage->data + index;
1446 }
1447
1448 static inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
1449 {
1450         return bpage->page->data + index;
1451 }
1452
1453 static inline struct ring_buffer_event *
1454 rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
1455 {
1456         return __rb_page_index(cpu_buffer->reader_page,
1457                                cpu_buffer->reader_page->read);
1458 }
1459
1460 static inline struct ring_buffer_event *
1461 rb_iter_head_event(struct ring_buffer_iter *iter)
1462 {
1463         return __rb_page_index(iter->head_page, iter->head);
1464 }
1465
1466 static inline unsigned long rb_page_write(struct buffer_page *bpage)
1467 {
1468         return local_read(&bpage->write) & RB_WRITE_MASK;
1469 }
1470
1471 static inline unsigned rb_page_commit(struct buffer_page *bpage)
1472 {
1473         return local_read(&bpage->page->commit);
1474 }
1475
1476 static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1477 {
1478         return local_read(&bpage->entries) & RB_WRITE_MASK;
1479 }
1480
1481 /* Size is determined by what has been committed */
1482 static inline unsigned rb_page_size(struct buffer_page *bpage)
1483 {
1484         return rb_page_commit(bpage);
1485 }
1486
1487 static inline unsigned
1488 rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
1489 {
1490         return rb_page_commit(cpu_buffer->commit_page);
1491 }
1492
1493 static inline unsigned
1494 rb_event_index(struct ring_buffer_event *event)
1495 {
1496         unsigned long addr = (unsigned long)event;
1497
1498         return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
1499 }
1500
1501 static inline int
1502 rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
1503                    struct ring_buffer_event *event)
1504 {
1505         unsigned long addr = (unsigned long)event;
1506         unsigned long index;
1507
1508         index = rb_event_index(event);
1509         addr &= PAGE_MASK;
1510
1511         return cpu_buffer->commit_page->page == (void *)addr &&
1512                 rb_commit_index(cpu_buffer) == index;
1513 }
1514
1515 static void
1516 rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
1517 {
1518         unsigned long max_count;
1519
1520         /*
1521          * We only race with interrupts and NMIs on this CPU.
1522          * If we own the commit event, then we can commit
1523          * all others that interrupted us, since the interruptions
1524          * are in stack format (they finish before they come
1525          * back to us). This allows us to do a simple loop to
1526          * assign the commit to the tail.
1527          */
1528  again:
1529         max_count = cpu_buffer->buffer->pages * 100;
1530
1531         while (cpu_buffer->commit_page != cpu_buffer->tail_page) {
1532                 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
1533                         return;
1534                 if (RB_WARN_ON(cpu_buffer,
1535                                rb_is_reader_page(cpu_buffer->tail_page)))
1536                         return;
1537                 local_set(&cpu_buffer->commit_page->page->commit,
1538                           rb_page_write(cpu_buffer->commit_page));
1539                 rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
1540                 cpu_buffer->write_stamp =
1541                         cpu_buffer->commit_page->page->time_stamp;
1542                 /* add barrier to keep gcc from optimizing too much */
1543                 barrier();
1544         }
1545         while (rb_commit_index(cpu_buffer) !=
1546                rb_page_write(cpu_buffer->commit_page)) {
1547
1548                 local_set(&cpu_buffer->commit_page->page->commit,
1549                           rb_page_write(cpu_buffer->commit_page));
1550                 RB_WARN_ON(cpu_buffer,
1551                            local_read(&cpu_buffer->commit_page->page->commit) &
1552                            ~RB_WRITE_MASK);
1553                 barrier();
1554         }
1555
1556         /* again, keep gcc from optimizing */
1557         barrier();
1558
1559         /*
1560          * If an interrupt came in just after the first while loop
1561          * and pushed the tail page forward, we will be left with
1562          * a dangling commit that will never go forward.
1563          */
1564         if (unlikely(cpu_buffer->commit_page != cpu_buffer->tail_page))
1565                 goto again;
1566 }
1567
1568 static void rb_reset_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
1569 {
1570         cpu_buffer->read_stamp = cpu_buffer->reader_page->page->time_stamp;
1571         cpu_buffer->reader_page->read = 0;
1572 }
1573
1574 static void rb_inc_iter(struct ring_buffer_iter *iter)
1575 {
1576         struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
1577
1578         /*
1579          * The iterator could be on the reader page (it starts there).
1580          * But the head could have moved, since the reader was
1581          * found. Check for this case and assign the iterator
1582          * to the head page instead of next.
1583          */
1584         if (iter->head_page == cpu_buffer->reader_page)
1585                 iter->head_page = rb_set_head_page(cpu_buffer);
1586         else
1587                 rb_inc_page(cpu_buffer, &iter->head_page);
1588
1589         iter->read_stamp = iter->head_page->page->time_stamp;
1590         iter->head = 0;
1591 }
1592
1593 /* Slow path, do not inline */
1594 static noinline struct ring_buffer_event *
1595 rb_add_time_stamp(struct ring_buffer_event *event, u64 delta)
1596 {
1597         event->type_len = RINGBUF_TYPE_TIME_EXTEND;
1598
1599         /* Not the first event on the page? */
1600         if (rb_event_index(event)) {
1601                 event->time_delta = delta & TS_MASK;
1602                 event->array[0] = delta >> TS_SHIFT;
1603         } else {
1604                 /* nope, just zero it */
1605                 event->time_delta = 0;
1606                 event->array[0] = 0;
1607         }
1608
1609         return skip_time_extend(event);
1610 }
1611
1612 /**
1613  * ring_buffer_update_event - update event type and data
1614  * @event: the even to update
1615  * @type: the type of event
1616  * @length: the size of the event field in the ring buffer
1617  *
1618  * Update the type and data fields of the event. The length
1619  * is the actual size that is written to the ring buffer,
1620  * and with this, we can determine what to place into the
1621  * data field.
1622  */
1623 static void
1624 rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
1625                 struct ring_buffer_event *event, unsigned length,
1626                 int add_timestamp, u64 delta)
1627 {
1628         /* Only a commit updates the timestamp */
1629         if (unlikely(!rb_event_is_commit(cpu_buffer, event)))
1630                 delta = 0;
1631
1632         /*
1633          * If we need to add a timestamp, then we
1634          * add it to the start of the resevered space.
1635          */
1636         if (unlikely(add_timestamp)) {
1637                 event = rb_add_time_stamp(event, delta);
1638                 length -= RB_LEN_TIME_EXTEND;
1639                 delta = 0;
1640         }
1641
1642         event->time_delta = delta;
1643         length -= RB_EVNT_HDR_SIZE;
1644         if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
1645                 event->type_len = 0;
1646                 event->array[0] = length;
1647         } else
1648                 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
1649 }
1650
1651 /*
1652  * rb_handle_head_page - writer hit the head page
1653  *
1654  * Returns: +1 to retry page
1655  *           0 to continue
1656  *          -1 on error
1657  */
1658 static int
1659 rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
1660                     struct buffer_page *tail_page,
1661                     struct buffer_page *next_page)
1662 {
1663         struct buffer_page *new_head;
1664         int entries;
1665         int type;
1666         int ret;
1667
1668         entries = rb_page_entries(next_page);
1669
1670         /*
1671          * The hard part is here. We need to move the head
1672          * forward, and protect against both readers on
1673          * other CPUs and writers coming in via interrupts.
1674          */
1675         type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
1676                                        RB_PAGE_HEAD);
1677
1678         /*
1679          * type can be one of four:
1680          *  NORMAL - an interrupt already moved it for us
1681          *  HEAD   - we are the first to get here.
1682          *  UPDATE - we are the interrupt interrupting
1683          *           a current move.
1684          *  MOVED  - a reader on another CPU moved the next
1685          *           pointer to its reader page. Give up
1686          *           and try again.
1687          */
1688
1689         switch (type) {
1690         case RB_PAGE_HEAD:
1691                 /*
1692                  * We changed the head to UPDATE, thus
1693                  * it is our responsibility to update
1694                  * the counters.
1695                  */
1696                 local_add(entries, &cpu_buffer->overrun);
1697
1698                 /*
1699                  * The entries will be zeroed out when we move the
1700                  * tail page.
1701                  */
1702
1703                 /* still more to do */
1704                 break;
1705
1706         case RB_PAGE_UPDATE:
1707                 /*
1708                  * This is an interrupt that interrupt the
1709                  * previous update. Still more to do.
1710                  */
1711                 break;
1712         case RB_PAGE_NORMAL:
1713                 /*
1714                  * An interrupt came in before the update
1715                  * and processed this for us.
1716                  * Nothing left to do.
1717                  */
1718                 return 1;
1719         case RB_PAGE_MOVED:
1720                 /*
1721                  * The reader is on another CPU and just did
1722                  * a swap with our next_page.
1723                  * Try again.
1724                  */
1725                 return 1;
1726         default:
1727                 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
1728                 return -1;
1729         }
1730
1731         /*
1732          * Now that we are here, the old head pointer is
1733          * set to UPDATE. This will keep the reader from
1734          * swapping the head page with the reader page.
1735          * The reader (on another CPU) will spin till
1736          * we are finished.
1737          *
1738          * We just need to protect against interrupts
1739          * doing the job. We will set the next pointer
1740          * to HEAD. After that, we set the old pointer
1741          * to NORMAL, but only if it was HEAD before.
1742          * otherwise we are an interrupt, and only
1743          * want the outer most commit to reset it.
1744          */
1745         new_head = next_page;
1746         rb_inc_page(cpu_buffer, &new_head);
1747
1748         ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
1749                                     RB_PAGE_NORMAL);
1750
1751         /*
1752          * Valid returns are:
1753          *  HEAD   - an interrupt came in and already set it.
1754          *  NORMAL - One of two things:
1755          *            1) We really set it.
1756          *            2) A bunch of interrupts came in and moved
1757          *               the page forward again.
1758          */
1759         switch (ret) {
1760         case RB_PAGE_HEAD:
1761         case RB_PAGE_NORMAL:
1762                 /* OK */
1763                 break;
1764         default:
1765                 RB_WARN_ON(cpu_buffer, 1);
1766                 return -1;
1767         }
1768
1769         /*
1770          * It is possible that an interrupt came in,
1771          * set the head up, then more interrupts came in
1772          * and moved it again. When we get back here,
1773          * the page would have been set to NORMAL but we
1774          * just set it back to HEAD.
1775          *
1776          * How do you detect this? Well, if that happened
1777          * the tail page would have moved.
1778          */
1779         if (ret == RB_PAGE_NORMAL) {
1780                 /*
1781                  * If the tail had moved passed next, then we need
1782                  * to reset the pointer.
1783                  */
1784                 if (cpu_buffer->tail_page != tail_page &&
1785                     cpu_buffer->tail_page != next_page)
1786                         rb_head_page_set_normal(cpu_buffer, new_head,
1787                                                 next_page,
1788                                                 RB_PAGE_HEAD);
1789         }
1790
1791         /*
1792          * If this was the outer most commit (the one that
1793          * changed the original pointer from HEAD to UPDATE),
1794          * then it is up to us to reset it to NORMAL.
1795          */
1796         if (type == RB_PAGE_HEAD) {
1797                 ret = rb_head_page_set_normal(cpu_buffer, next_page,
1798                                               tail_page,
1799                                               RB_PAGE_UPDATE);
1800                 if (RB_WARN_ON(cpu_buffer,
1801                                ret != RB_PAGE_UPDATE))
1802                         return -1;
1803         }
1804
1805         return 0;
1806 }
1807
1808 static unsigned rb_calculate_event_length(unsigned length)
1809 {
1810         struct ring_buffer_event event; /* Used only for sizeof array */
1811
1812         /* zero length can cause confusions */
1813         if (!length)
1814                 length = 1;
1815
1816         if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
1817                 length += sizeof(event.array[0]);
1818
1819         length += RB_EVNT_HDR_SIZE;
1820         length = ALIGN(length, RB_ARCH_ALIGNMENT);
1821
1822         return length;
1823 }
1824
1825 static inline void
1826 rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
1827               struct buffer_page *tail_page,
1828               unsigned long tail, unsigned long length)
1829 {
1830         struct ring_buffer_event *event;
1831
1832         /*
1833          * Only the event that crossed the page boundary
1834          * must fill the old tail_page with padding.
1835          */
1836         if (tail >= BUF_PAGE_SIZE) {
1837                 /*
1838                  * If the page was filled, then we still need
1839                  * to update the real_end. Reset it to zero
1840                  * and the reader will ignore it.
1841                  */
1842                 if (tail == BUF_PAGE_SIZE)
1843                         tail_page->real_end = 0;
1844
1845                 local_sub(length, &tail_page->write);
1846                 return;
1847         }
1848
1849         event = __rb_page_index(tail_page, tail);
1850         kmemcheck_annotate_bitfield(event, bitfield);
1851
1852         /*
1853          * Save the original length to the meta data.
1854          * This will be used by the reader to add lost event
1855          * counter.
1856          */
1857         tail_page->real_end = tail;
1858
1859         /*
1860          * If this event is bigger than the minimum size, then
1861          * we need to be careful that we don't subtract the
1862          * write counter enough to allow another writer to slip
1863          * in on this page.
1864          * We put in a discarded commit instead, to make sure
1865          * that this space is not used again.
1866          *
1867          * If we are less than the minimum size, we don't need to
1868          * worry about it.
1869          */
1870         if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
1871                 /* No room for any events */
1872
1873                 /* Mark the rest of the page with padding */
1874                 rb_event_set_padding(event);
1875
1876                 /* Set the write back to the previous setting */
1877                 local_sub(length, &tail_page->write);
1878                 return;
1879         }
1880
1881         /* Put in a discarded event */
1882         event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
1883         event->type_len = RINGBUF_TYPE_PADDING;
1884         /* time delta must be non zero */
1885         event->time_delta = 1;
1886
1887         /* Set write to end of buffer */
1888         length = (tail + length) - BUF_PAGE_SIZE;
1889         local_sub(length, &tail_page->write);
1890 }
1891
1892 /*
1893  * This is the slow path, force gcc not to inline it.
1894  */
1895 static noinline struct ring_buffer_event *
1896 rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
1897              unsigned long length, unsigned long tail,
1898              struct buffer_page *tail_page, u64 ts)
1899 {
1900         struct buffer_page *commit_page = cpu_buffer->commit_page;
1901         struct ring_buffer *buffer = cpu_buffer->buffer;
1902         struct buffer_page *next_page;
1903         int ret;
1904
1905         next_page = tail_page;
1906
1907         rb_inc_page(cpu_buffer, &next_page);
1908
1909         /*
1910          * If for some reason, we had an interrupt storm that made
1911          * it all the way around the buffer, bail, and warn
1912          * about it.
1913          */
1914         if (unlikely(next_page == commit_page)) {
1915                 local_inc(&cpu_buffer->commit_overrun);
1916                 goto out_reset;
1917         }
1918
1919         /*
1920          * This is where the fun begins!
1921          *
1922          * We are fighting against races between a reader that
1923          * could be on another CPU trying to swap its reader
1924          * page with the buffer head.
1925          *
1926          * We are also fighting against interrupts coming in and
1927          * moving the head or tail on us as well.
1928          *
1929          * If the next page is the head page then we have filled
1930          * the buffer, unless the commit page is still on the
1931          * reader page.
1932          */
1933         if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) {
1934
1935                 /*
1936                  * If the commit is not on the reader page, then
1937                  * move the header page.
1938                  */
1939                 if (!rb_is_reader_page(cpu_buffer->commit_page)) {
1940                         /*
1941                          * If we are not in overwrite mode,
1942                          * this is easy, just stop here.
1943                          */
1944                         if (!(buffer->flags & RB_FL_OVERWRITE))
1945                                 goto out_reset;
1946
1947                         ret = rb_handle_head_page(cpu_buffer,
1948                                                   tail_page,
1949                                                   next_page);
1950                         if (ret < 0)
1951                                 goto out_reset;
1952                         if (ret)
1953                                 goto out_again;
1954                 } else {
1955                         /*
1956                          * We need to be careful here too. The
1957                          * commit page could still be on the reader
1958                          * page. We could have a small buffer, and
1959                          * have filled up the buffer with events
1960                          * from interrupts and such, and wrapped.
1961                          *
1962                          * Note, if the tail page is also the on the
1963                          * reader_page, we let it move out.
1964                          */
1965                         if (unlikely((cpu_buffer->commit_page !=
1966                                       cpu_buffer->tail_page) &&
1967                                      (cpu_buffer->commit_page ==
1968                                       cpu_buffer->reader_page))) {
1969                                 local_inc(&cpu_buffer->commit_overrun);
1970                                 goto out_reset;
1971                         }
1972                 }
1973         }
1974
1975         ret = rb_tail_page_update(cpu_buffer, tail_page, next_page);
1976         if (ret) {
1977                 /*
1978                  * Nested commits always have zero deltas, so
1979                  * just reread the time stamp
1980                  */
1981                 ts = rb_time_stamp(buffer);
1982                 next_page->page->time_stamp = ts;
1983         }
1984
1985  out_again:
1986
1987         rb_reset_tail(cpu_buffer, tail_page, tail, length);
1988
1989         /* fail and let the caller try again */
1990         return ERR_PTR(-EAGAIN);
1991
1992  out_reset:
1993         /* reset write */
1994         rb_reset_tail(cpu_buffer, tail_page, tail, length);
1995
1996         return NULL;
1997 }
1998
1999 static struct ring_buffer_event *
2000 __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
2001                   unsigned long length, u64 ts,
2002                   u64 delta, int add_timestamp)
2003 {
2004         struct buffer_page *tail_page;
2005         struct ring_buffer_event *event;
2006         unsigned long tail, write;
2007
2008         /*
2009          * If the time delta since the last event is too big to
2010          * hold in the time field of the event, then we append a
2011          * TIME EXTEND event ahead of the data event.
2012          */
2013         if (unlikely(add_timestamp))
2014                 length += RB_LEN_TIME_EXTEND;
2015
2016         tail_page = cpu_buffer->tail_page;
2017         write = local_add_return(length, &tail_page->write);
2018
2019         /* set write to only the index of the write */
2020         write &= RB_WRITE_MASK;
2021         tail = write - length;
2022
2023         /* See if we shot pass the end of this buffer page */
2024         if (unlikely(write > BUF_PAGE_SIZE))
2025                 return rb_move_tail(cpu_buffer, length, tail,
2026                                     tail_page, ts);
2027
2028         /* We reserved something on the buffer */
2029
2030         event = __rb_page_index(tail_page, tail);
2031         kmemcheck_annotate_bitfield(event, bitfield);
2032         rb_update_event(cpu_buffer, event, length, add_timestamp, delta);
2033
2034         local_inc(&tail_page->entries);
2035
2036         /*
2037          * If this is the first commit on the page, then update
2038          * its timestamp.
2039          */
2040         if (!tail)
2041                 tail_page->page->time_stamp = ts;
2042
2043         return event;
2044 }
2045
2046 static inline int
2047 rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2048                   struct ring_buffer_event *event)
2049 {
2050         unsigned long new_index, old_index;
2051         struct buffer_page *bpage;
2052         unsigned long index;
2053         unsigned long addr;
2054
2055         new_index = rb_event_index(event);
2056         old_index = new_index + rb_event_ts_length(event);
2057         addr = (unsigned long)event;
2058         addr &= PAGE_MASK;
2059
2060         bpage = cpu_buffer->tail_page;
2061
2062         if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2063                 unsigned long write_mask =
2064                         local_read(&bpage->write) & ~RB_WRITE_MASK;
2065                 /*
2066                  * This is on the tail page. It is possible that
2067                  * a write could come in and move the tail page
2068                  * and write to the next page. That is fine
2069                  * because we just shorten what is on this page.
2070                  */
2071                 old_index += write_mask;
2072                 new_index += write_mask;
2073                 index = local_cmpxchg(&bpage->write, old_index, new_index);
2074                 if (index == old_index)
2075                         return 1;
2076         }
2077
2078         /* could not discard */
2079         return 0;
2080 }
2081
2082 static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2083 {
2084         local_inc(&cpu_buffer->committing);
2085         local_inc(&cpu_buffer->commits);
2086 }
2087
2088 static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
2089 {
2090         unsigned long commits;
2091
2092         if (RB_WARN_ON(cpu_buffer,
2093                        !local_read(&cpu_buffer->committing)))
2094                 return;
2095
2096  again:
2097         commits = local_read(&cpu_buffer->commits);
2098         /* synchronize with interrupts */
2099         barrier();
2100         if (local_read(&cpu_buffer->committing) == 1)
2101                 rb_set_commit_to_write(cpu_buffer);
2102
2103         local_dec(&cpu_buffer->committing);
2104
2105         /* synchronize with interrupts */
2106         barrier();
2107
2108         /*
2109          * Need to account for interrupts coming in between the
2110          * updating of the commit page and the clearing of the
2111          * committing counter.
2112          */
2113         if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
2114             !local_read(&cpu_buffer->committing)) {
2115                 local_inc(&cpu_buffer->committing);
2116                 goto again;
2117         }
2118 }
2119
2120 static struct ring_buffer_event *
2121 rb_reserve_next_event(struct ring_buffer *buffer,
2122                       struct ring_buffer_per_cpu *cpu_buffer,
2123                       unsigned long length)
2124 {
2125         struct ring_buffer_event *event;
2126         u64 ts, delta;
2127         int nr_loops = 0;
2128         int add_timestamp;
2129         u64 diff;
2130
2131         rb_start_commit(cpu_buffer);
2132
2133 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2134         /*
2135          * Due to the ability to swap a cpu buffer from a buffer
2136          * it is possible it was swapped before we committed.
2137          * (committing stops a swap). We check for it here and
2138          * if it happened, we have to fail the write.
2139          */
2140         barrier();
2141         if (unlikely(ACCESS_ONCE(cpu_buffer->buffer) != buffer)) {
2142                 local_dec(&cpu_buffer->committing);
2143                 local_dec(&cpu_buffer->commits);
2144                 return NULL;
2145         }
2146 #endif
2147
2148         length = rb_calculate_event_length(length);
2149  again:
2150         add_timestamp = 0;
2151         delta = 0;
2152
2153         /*
2154          * We allow for interrupts to reenter here and do a trace.
2155          * If one does, it will cause this original code to loop
2156          * back here. Even with heavy interrupts happening, this
2157          * should only happen a few times in a row. If this happens
2158          * 1000 times in a row, there must be either an interrupt
2159          * storm or we have something buggy.
2160          * Bail!
2161          */
2162         if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
2163                 goto out_fail;
2164
2165         ts = rb_time_stamp(cpu_buffer->buffer);
2166         diff = ts - cpu_buffer->write_stamp;
2167
2168         /* make sure this diff is calculated here */
2169         barrier();
2170
2171         /* Did the write stamp get updated already? */
2172         if (likely(ts >= cpu_buffer->write_stamp)) {
2173                 delta = diff;
2174                 if (unlikely(test_time_stamp(delta))) {
2175                         int local_clock_stable = 1;
2176 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2177                         local_clock_stable = sched_clock_stable;
2178 #endif
2179                         WARN_ONCE(delta > (1ULL << 59),
2180                                   KERN_WARNING "Delta way too big! %llu ts=%llu write stamp = %llu\n%s",
2181                                   (unsigned long long)delta,
2182                                   (unsigned long long)ts,
2183                                   (unsigned long long)cpu_buffer->write_stamp,
2184                                   local_clock_stable ? "" :
2185                                   "If you just came from a suspend/resume,\n"
2186                                   "please switch to the trace global clock:\n"
2187                                   "  echo global > /sys/kernel/debug/tracing/trace_clock\n");
2188                         add_timestamp = 1;
2189                 }
2190         }
2191
2192         event = __rb_reserve_next(cpu_buffer, length, ts,
2193                                   delta, add_timestamp);
2194         if (unlikely(PTR_ERR(event) == -EAGAIN))
2195                 goto again;
2196
2197         if (!event)
2198                 goto out_fail;
2199
2200         return event;
2201
2202  out_fail:
2203         rb_end_commit(cpu_buffer);
2204         return NULL;
2205 }
2206
2207 #ifdef CONFIG_TRACING
2208
2209 #define TRACE_RECURSIVE_DEPTH 16
2210
2211 /* Keep this code out of the fast path cache */
2212 static noinline void trace_recursive_fail(void)
2213 {
2214         /* Disable all tracing before we do anything else */
2215         tracing_off_permanent();
2216
2217         printk_once(KERN_WARNING "Tracing recursion: depth[%ld]:"
2218                     "HC[%lu]:SC[%lu]:NMI[%lu]\n",
2219                     current->trace_recursion,
2220                     hardirq_count() >> HARDIRQ_SHIFT,
2221                     softirq_count() >> SOFTIRQ_SHIFT,
2222                     in_nmi());
2223
2224         WARN_ON_ONCE(1);
2225 }
2226
2227 static inline int trace_recursive_lock(void)
2228 {
2229         current->trace_recursion++;
2230
2231         if (likely(current->trace_recursion < TRACE_RECURSIVE_DEPTH))
2232                 return 0;
2233
2234         trace_recursive_fail();
2235
2236         return -1;
2237 }
2238
2239 static inline void trace_recursive_unlock(void)
2240 {
2241         WARN_ON_ONCE(!current->trace_recursion);
2242
2243         current->trace_recursion--;
2244 }
2245
2246 #else
2247
2248 #define trace_recursive_lock()          (0)
2249 #define trace_recursive_unlock()        do { } while (0)
2250
2251 #endif
2252
2253 /**
2254  * ring_buffer_lock_reserve - reserve a part of the buffer
2255  * @buffer: the ring buffer to reserve from
2256  * @length: the length of the data to reserve (excluding event header)
2257  *
2258  * Returns a reseverd event on the ring buffer to copy directly to.
2259  * The user of this interface will need to get the body to write into
2260  * and can use the ring_buffer_event_data() interface.
2261  *
2262  * The length is the length of the data needed, not the event length
2263  * which also includes the event header.
2264  *
2265  * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2266  * If NULL is returned, then nothing has been allocated or locked.
2267  */
2268 struct ring_buffer_event *
2269 ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length)
2270 {
2271         struct ring_buffer_per_cpu *cpu_buffer;
2272         struct ring_buffer_event *event;
2273         int cpu;
2274
2275         if (ring_buffer_flags != RB_BUFFERS_ON)
2276                 return NULL;
2277
2278         /* If we are tracing schedule, we don't want to recurse */
2279         preempt_disable_notrace();
2280
2281         if (atomic_read(&buffer->record_disabled))
2282                 goto out_nocheck;
2283
2284         if (trace_recursive_lock())
2285                 goto out_nocheck;
2286
2287         cpu = raw_smp_processor_id();
2288
2289         if (!cpumask_test_cpu(cpu, buffer->cpumask))
2290                 goto out;
2291
2292         cpu_buffer = buffer->buffers[cpu];
2293
2294         if (atomic_read(&cpu_buffer->record_disabled))
2295                 goto out;
2296
2297         if (length > BUF_MAX_DATA_SIZE)
2298                 goto out;
2299
2300         event = rb_reserve_next_event(buffer, cpu_buffer, length);
2301         if (!event)
2302                 goto out;
2303
2304         return event;
2305
2306  out:
2307         trace_recursive_unlock();
2308
2309  out_nocheck:
2310         preempt_enable_notrace();
2311         return NULL;
2312 }
2313 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
2314
2315 static void
2316 rb_update_write_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2317                       struct ring_buffer_event *event)
2318 {
2319         u64 delta;
2320
2321         /*
2322          * The event first in the commit queue updates the
2323          * time stamp.
2324          */
2325         if (rb_event_is_commit(cpu_buffer, event)) {
2326                 /*
2327                  * A commit event that is first on a page
2328                  * updates the write timestamp with the page stamp
2329                  */
2330                 if (!rb_event_index(event))
2331                         cpu_buffer->write_stamp =
2332                                 cpu_buffer->commit_page->page->time_stamp;
2333                 else if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
2334                         delta = event->array[0];
2335                         delta <<= TS_SHIFT;
2336                         delta += event->time_delta;
2337                         cpu_buffer->write_stamp += delta;
2338                 } else
2339                         cpu_buffer->write_stamp += event->time_delta;
2340         }
2341 }
2342
2343 static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
2344                       struct ring_buffer_event *event)
2345 {
2346         local_inc(&cpu_buffer->entries);
2347         rb_update_write_stamp(cpu_buffer, event);
2348         rb_end_commit(cpu_buffer);
2349 }
2350
2351 /**
2352  * ring_buffer_unlock_commit - commit a reserved
2353  * @buffer: The buffer to commit to
2354  * @event: The event pointer to commit.
2355  *
2356  * This commits the data to the ring buffer, and releases any locks held.
2357  *
2358  * Must be paired with ring_buffer_lock_reserve.
2359  */
2360 int ring_buffer_unlock_commit(struct ring_buffer *buffer,
2361                               struct ring_buffer_event *event)
2362 {
2363         struct ring_buffer_per_cpu *cpu_buffer;
2364         int cpu = raw_smp_processor_id();
2365
2366         cpu_buffer = buffer->buffers[cpu];
2367
2368         rb_commit(cpu_buffer, event);
2369
2370         trace_recursive_unlock();
2371
2372         preempt_enable_notrace();
2373
2374         return 0;
2375 }
2376 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
2377
2378 static inline void rb_event_discard(struct ring_buffer_event *event)
2379 {
2380         if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
2381                 event = skip_time_extend(event);
2382
2383         /* array[0] holds the actual length for the discarded event */
2384         event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
2385         event->type_len = RINGBUF_TYPE_PADDING;
2386         /* time delta must be non zero */
2387         if (!event->time_delta)
2388                 event->time_delta = 1;
2389 }
2390
2391 /*
2392  * Decrement the entries to the page that an event is on.
2393  * The event does not even need to exist, only the pointer
2394  * to the page it is on. This may only be called before the commit
2395  * takes place.
2396  */
2397 static inline void
2398 rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
2399                    struct ring_buffer_event *event)
2400 {
2401         unsigned long addr = (unsigned long)event;
2402         struct buffer_page *bpage = cpu_buffer->commit_page;
2403         struct buffer_page *start;
2404
2405         addr &= PAGE_MASK;
2406
2407         /* Do the likely case first */
2408         if (likely(bpage->page == (void *)addr)) {
2409                 local_dec(&bpage->entries);
2410                 return;
2411         }
2412
2413         /*
2414          * Because the commit page may be on the reader page we
2415          * start with the next page and check the end loop there.
2416          */
2417         rb_inc_page(cpu_buffer, &bpage);
2418         start = bpage;
2419         do {
2420                 if (bpage->page == (void *)addr) {
2421                         local_dec(&bpage->entries);
2422                         return;
2423                 }
2424                 rb_inc_page(cpu_buffer, &bpage);
2425         } while (bpage != start);
2426
2427         /* commit not part of this buffer?? */
2428         RB_WARN_ON(cpu_buffer, 1);
2429 }
2430
2431 /**
2432  * ring_buffer_commit_discard - discard an event that has not been committed
2433  * @buffer: the ring buffer
2434  * @event: non committed event to discard
2435  *
2436  * Sometimes an event that is in the ring buffer needs to be ignored.
2437  * This function lets the user discard an event in the ring buffer
2438  * and then that event will not be read later.
2439  *
2440  * This function only works if it is called before the the item has been
2441  * committed. It will try to free the event from the ring buffer
2442  * if another event has not been added behind it.
2443  *
2444  * If another event has been added behind it, it will set the event
2445  * up as discarded, and perform the commit.
2446  *
2447  * If this function is called, do not call ring_buffer_unlock_commit on
2448  * the event.
2449  */
2450 void ring_buffer_discard_commit(struct ring_buffer *buffer,
2451                                 struct ring_buffer_event *event)
2452 {
2453         struct ring_buffer_per_cpu *cpu_buffer;
2454         int cpu;
2455
2456         /* The event is discarded regardless */
2457         rb_event_discard(event);
2458
2459         cpu = smp_processor_id();
2460         cpu_buffer = buffer->buffers[cpu];
2461
2462         /*
2463          * This must only be called if the event has not been
2464          * committed yet. Thus we can assume that preemption
2465          * is still disabled.
2466          */
2467         RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
2468
2469         rb_decrement_entry(cpu_buffer, event);
2470         if (rb_try_to_discard(cpu_buffer, event))
2471                 goto out;
2472
2473         /*
2474          * The commit is still visible by the reader, so we
2475          * must still update the timestamp.
2476          */
2477         rb_update_write_stamp(cpu_buffer, event);
2478  out:
2479         rb_end_commit(cpu_buffer);
2480
2481         trace_recursive_unlock();
2482
2483         preempt_enable_notrace();
2484
2485 }
2486 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
2487
2488 /**
2489  * ring_buffer_write - write data to the buffer without reserving
2490  * @buffer: The ring buffer to write to.
2491  * @length: The length of the data being written (excluding the event header)
2492  * @data: The data to write to the buffer.
2493  *
2494  * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
2495  * one function. If you already have the data to write to the buffer, it
2496  * may be easier to simply call this function.
2497  *
2498  * Note, like ring_buffer_lock_reserve, the length is the length of the data
2499  * and not the length of the event which would hold the header.
2500  */
2501 int ring_buffer_write(struct ring_buffer *buffer,
2502                         unsigned long length,
2503                         void *data)
2504 {
2505         struct ring_buffer_per_cpu *cpu_buffer;
2506         struct ring_buffer_event *event;
2507         void *body;
2508         int ret = -EBUSY;
2509         int cpu;
2510
2511         if (ring_buffer_flags != RB_BUFFERS_ON)
2512                 return -EBUSY;
2513
2514         preempt_disable_notrace();
2515
2516         if (atomic_read(&buffer->record_disabled))
2517                 goto out;
2518
2519         cpu = raw_smp_processor_id();
2520
2521         if (!cpumask_test_cpu(cpu, buffer->cpumask))
2522                 goto out;
2523
2524         cpu_buffer = buffer->buffers[cpu];
2525
2526         if (atomic_read(&cpu_buffer->record_disabled))
2527                 goto out;
2528
2529         if (length > BUF_MAX_DATA_SIZE)
2530                 goto out;
2531
2532         event = rb_reserve_next_event(buffer, cpu_buffer, length);
2533         if (!event)
2534                 goto out;
2535
2536         body = rb_event_data(event);
2537
2538         memcpy(body, data, length);
2539
2540         rb_commit(cpu_buffer, event);
2541
2542         ret = 0;
2543  out:
2544         preempt_enable_notrace();
2545
2546         return ret;
2547 }
2548 EXPORT_SYMBOL_GPL(ring_buffer_write);
2549
2550 static int rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
2551 {
2552         struct buffer_page *reader = cpu_buffer->reader_page;
2553         struct buffer_page *head = rb_set_head_page(cpu_buffer);
2554         struct buffer_page *commit = cpu_buffer->commit_page;
2555
2556         /* In case of error, head will be NULL */
2557         if (unlikely(!head))
2558                 return 1;
2559
2560         return reader->read == rb_page_commit(reader) &&
2561                 (commit == reader ||
2562                  (commit == head &&
2563                   head->read == rb_page_commit(commit)));
2564 }
2565
2566 /**
2567  * ring_buffer_record_disable - stop all writes into the buffer
2568  * @buffer: The ring buffer to stop writes to.
2569  *
2570  * This prevents all writes to the buffer. Any attempt to write
2571  * to the buffer after this will fail and return NULL.
2572  *
2573  * The caller should call synchronize_sched() after this.
2574  */
2575 void ring_buffer_record_disable(struct ring_buffer *buffer)
2576 {
2577         atomic_inc(&buffer->record_disabled);
2578 }
2579 EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
2580
2581 /**
2582  * ring_buffer_record_enable - enable writes to the buffer
2583  * @buffer: The ring buffer to enable writes
2584  *
2585  * Note, multiple disables will need the same number of enables
2586  * to truly enable the writing (much like preempt_disable).
2587  */
2588 void ring_buffer_record_enable(struct ring_buffer *buffer)
2589 {
2590         atomic_dec(&buffer->record_disabled);
2591 }
2592 EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
2593
2594 /**
2595  * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
2596  * @buffer: The ring buffer to stop writes to.
2597  * @cpu: The CPU buffer to stop
2598  *
2599  * This prevents all writes to the buffer. Any attempt to write
2600  * to the buffer after this will fail and return NULL.
2601  *
2602  * The caller should call synchronize_sched() after this.
2603  */
2604 void ring_buffer_record_disable_cpu(struct ring_buffer *buffer, int cpu)
2605 {
2606         struct ring_buffer_per_cpu *cpu_buffer;
2607
2608         if (!cpumask_test_cpu(cpu, buffer->cpumask))
2609                 return;
2610
2611         cpu_buffer = buffer->buffers[cpu];
2612         atomic_inc(&cpu_buffer->record_disabled);
2613 }
2614 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
2615
2616 /**
2617  * ring_buffer_record_enable_cpu - enable writes to the buffer
2618  * @buffer: The ring buffer to enable writes
2619  * @cpu: The CPU to enable.
2620  *
2621  * Note, multiple disables will need the same number of enables
2622  * to truly enable the writing (much like preempt_disable).
2623  */
2624 void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu)
2625 {
2626         struct ring_buffer_per_cpu *cpu_buffer;
2627
2628         if (!cpumask_test_cpu(cpu, buffer->cpumask))
2629                 return;
2630
2631         cpu_buffer = buffer->buffers[cpu];
2632         atomic_dec(&cpu_buffer->record_disabled);
2633 }
2634 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
2635
2636 /*
2637  * The total entries in the ring buffer is the running counter
2638  * of entries entered into the ring buffer, minus the sum of
2639  * the entries read from the ring buffer and the number of
2640  * entries that were overwritten.
2641  */
2642 static inline unsigned long
2643 rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
2644 {
2645         return local_read(&cpu_buffer->entries) -
2646                 (local_read(&cpu_buffer->overrun) + cpu_buffer->read);
2647 }
2648
2649 /**
2650  * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
2651  * @buffer: The ring buffer
2652  * @cpu: The per CPU buffer to get the entries from.
2653  */
2654 unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu)
2655 {
2656         struct ring_buffer_per_cpu *cpu_buffer;
2657
2658         if (!cpumask_test_cpu(cpu, buffer->cpumask))
2659                 return 0;
2660
2661         cpu_buffer = buffer->buffers[cpu];
2662
2663         return rb_num_of_entries(cpu_buffer);
2664 }
2665 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
2666
2667 /**
2668  * ring_buffer_overrun_cpu - get the number of overruns in a cpu_buffer
2669  * @buffer: The ring buffer
2670  * @cpu: The per CPU buffer to get the number of overruns from
2671  */
2672 unsigned long ring_buffer_overrun_cpu(struct ring_buffer *buffer, int cpu)
2673 {
2674         struct ring_buffer_per_cpu *cpu_buffer;
2675         unsigned long ret;
2676
2677         if (!cpumask_test_cpu(cpu, buffer->cpumask))
2678                 return 0;
2679
2680         cpu_buffer = buffer->buffers[cpu];
2681         ret = local_read(&cpu_buffer->overrun);
2682
2683         return ret;
2684 }
2685 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
2686
2687 /**
2688  * ring_buffer_commit_overrun_cpu - get the number of overruns caused by commits
2689  * @buffer: The ring buffer
2690  * @cpu: The per CPU buffer to get the number of overruns from
2691  */
2692 unsigned long
2693 ring_buffer_commit_overrun_cpu(struct ring_buffer *buffer, int cpu)
2694 {
2695         struct ring_buffer_per_cpu *cpu_buffer;
2696         unsigned long ret;
2697
2698         if (!cpumask_test_cpu(cpu, buffer->cpumask))
2699                 return 0;
2700
2701         cpu_buffer = buffer->buffers[cpu];
2702         ret = local_read(&cpu_buffer->commit_overrun);
2703
2704         return ret;
2705 }
2706 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
2707
2708 /**
2709  * ring_buffer_entries - get the number of entries in a buffer
2710  * @buffer: The ring buffer
2711  *
2712  * Returns the total number of entries in the ring buffer
2713  * (all CPU entries)
2714  */
2715 unsigned long ring_buffer_entries(struct ring_buffer *buffer)
2716 {
2717         struct ring_buffer_per_cpu *cpu_buffer;
2718         unsigned long entries = 0;
2719         int cpu;
2720
2721         /* if you care about this being correct, lock the buffer */
2722         for_each_buffer_cpu(buffer, cpu) {
2723                 cpu_buffer = buffer->buffers[cpu];
2724                 entries += rb_num_of_entries(cpu_buffer);
2725         }
2726
2727         return entries;
2728 }
2729 EXPORT_SYMBOL_GPL(ring_buffer_entries);
2730
2731 /**
2732  * ring_buffer_overruns - get the number of overruns in buffer
2733  * @buffer: The ring buffer
2734  *
2735  * Returns the total number of overruns in the ring buffer
2736  * (all CPU entries)
2737  */
2738 unsigned long ring_buffer_overruns(struct ring_buffer *buffer)
2739 {
2740         struct ring_buffer_per_cpu *cpu_buffer;
2741         unsigned long overruns = 0;
2742         int cpu;
2743
2744         /* if you care about this being correct, lock the buffer */
2745         for_each_buffer_cpu(buffer, cpu) {
2746                 cpu_buffer = buffer->buffers[cpu];
2747                 overruns += local_read(&cpu_buffer->overrun);
2748         }
2749
2750         return overruns;
2751 }
2752 EXPORT_SYMBOL_GPL(ring_buffer_overruns);
2753
2754 static void rb_iter_reset(struct ring_buffer_iter *iter)
2755 {
2756         struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
2757
2758         /* Iterator usage is expected to have record disabled */
2759         if (list_empty(&cpu_buffer->reader_page->list)) {
2760                 iter->head_page = rb_set_head_page(cpu_buffer);
2761                 if (unlikely(!iter->head_page))
2762                         return;
2763                 iter->head = iter->head_page->read;
2764         } else {
2765                 iter->head_page = cpu_buffer->reader_page;
2766                 iter->head = cpu_buffer->reader_page->read;
2767         }
2768         if (iter->head)
2769                 iter->read_stamp = cpu_buffer->read_stamp;
2770         else
2771                 iter->read_stamp = iter->head_page->page->time_stamp;
2772         iter->cache_reader_page = cpu_buffer->reader_page;
2773         iter->cache_read = cpu_buffer->read;
2774 }
2775
2776 /**
2777  * ring_buffer_iter_reset - reset an iterator
2778  * @iter: The iterator to reset
2779  *
2780  * Resets the iterator, so that it will start from the beginning
2781  * again.
2782  */
2783 void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
2784 {
2785         struct ring_buffer_per_cpu *cpu_buffer;
2786         unsigned long flags;
2787
2788         if (!iter)
2789                 return;
2790
2791         cpu_buffer = iter->cpu_buffer;
2792
2793         spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
2794         rb_iter_reset(iter);
2795         spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
2796 }
2797 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
2798
2799 /**
2800  * ring_buffer_iter_empty - check if an iterator has no more to read
2801  * @iter: The iterator to check
2802  */
2803 int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
2804 {
2805         struct ring_buffer_per_cpu *cpu_buffer;
2806
2807         cpu_buffer = iter->cpu_buffer;
2808
2809         return iter->head_page == cpu_buffer->commit_page &&
2810                 iter->head == rb_commit_index(cpu_buffer);
2811 }
2812 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
2813
2814 static void
2815 rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2816                      struct ring_buffer_event *event)
2817 {
2818         u64 delta;
2819
2820         switch (event->type_len) {
2821         case RINGBUF_TYPE_PADDING:
2822                 return;
2823
2824         case RINGBUF_TYPE_TIME_EXTEND:
2825                 delta = event->array[0];
2826                 delta <<= TS_SHIFT;
2827                 delta += event->time_delta;
2828                 cpu_buffer->read_stamp += delta;
2829                 return;
2830
2831         case RINGBUF_TYPE_TIME_STAMP:
2832                 /* FIXME: not implemented */
2833                 return;
2834
2835         case RINGBUF_TYPE_DATA:
2836                 cpu_buffer->read_stamp += event->time_delta;
2837                 return;
2838
2839         default:
2840                 BUG();
2841         }
2842         return;
2843 }
2844
2845 static void
2846 rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
2847                           struct ring_buffer_event *event)
2848 {
2849         u64 delta;
2850
2851         switch (event->type_len) {
2852         case RINGBUF_TYPE_PADDING:
2853                 return;
2854
2855         case RINGBUF_TYPE_TIME_EXTEND:
2856                 delta = event->array[0];
2857                 delta <<= TS_SHIFT;
2858                 delta += event->time_delta;
2859                 iter->read_stamp += delta;
2860                 return;
2861
2862         case RINGBUF_TYPE_TIME_STAMP:
2863                 /* FIXME: not implemented */
2864                 return;
2865
2866         case RINGBUF_TYPE_DATA:
2867                 iter->read_stamp += event->time_delta;
2868                 return;
2869
2870         default:
2871                 BUG();
2872         }
2873         return;
2874 }
2875
2876 static struct buffer_page *
2877 rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
2878 {
2879         struct buffer_page *reader = NULL;
2880         unsigned long overwrite;
2881         unsigned long flags;
2882         int nr_loops = 0;
2883         int ret;
2884
2885         local_irq_save(flags);
2886         arch_spin_lock(&cpu_buffer->lock);
2887
2888  again:
2889         /*
2890          * This should normally only loop twice. But because the
2891          * start of the reader inserts an empty page, it causes
2892          * a case where we will loop three times. There should be no
2893          * reason to loop four times (that I know of).
2894          */
2895         if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
2896                 reader = NULL;
2897                 goto out;
2898         }
2899
2900         reader = cpu_buffer->reader_page;
2901
2902         /* If there's more to read, return this page */
2903         if (cpu_buffer->reader_page->read < rb_page_size(reader))
2904                 goto out;
2905
2906         /* Never should we have an index greater than the size */
2907         if (RB_WARN_ON(cpu_buffer,
2908                        cpu_buffer->reader_page->read > rb_page_size(reader)))
2909                 goto out;
2910
2911         /* check if we caught up to the tail */
2912         reader = NULL;
2913         if (cpu_buffer->commit_page == cpu_buffer->reader_page)
2914                 goto out;
2915
2916         /*
2917          * Reset the reader page to size zero.
2918          */
2919         local_set(&cpu_buffer->reader_page->write, 0);
2920         local_set(&cpu_buffer->reader_page->entries, 0);
2921         local_set(&cpu_buffer->reader_page->page->commit, 0);
2922         cpu_buffer->reader_page->real_end = 0;
2923
2924  spin:
2925         /*
2926          * Splice the empty reader page into the list around the head.
2927          */
2928         reader = rb_set_head_page(cpu_buffer);
2929         cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
2930         cpu_buffer->reader_page->list.prev = reader->list.prev;
2931
2932         /*
2933          * cpu_buffer->pages just needs to point to the buffer, it
2934          *  has no specific buffer page to point to. Lets move it out
2935          *  of our way so we don't accidentally swap it.
2936          */
2937         cpu_buffer->pages = reader->list.prev;
2938
2939         /* The reader page will be pointing to the new head */
2940         rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list);
2941
2942         /*
2943          * We want to make sure we read the overruns after we set up our
2944          * pointers to the next object. The writer side does a
2945          * cmpxchg to cross pages which acts as the mb on the writer
2946          * side. Note, the reader will constantly fail the swap
2947          * while the writer is updating the pointers, so this
2948          * guarantees that the overwrite recorded here is the one we
2949          * want to compare with the last_overrun.
2950          */
2951         smp_mb();
2952         overwrite = local_read(&(cpu_buffer->overrun));
2953
2954         /*
2955          * Here's the tricky part.
2956          *
2957          * We need to move the pointer past the header page.
2958          * But we can only do that if a writer is not currently
2959          * moving it. The page before the header page has the
2960          * flag bit '1' set if it is pointing to the page we want.
2961          * but if the writer is in the process of moving it
2962          * than it will be '2' or already moved '0'.
2963          */
2964
2965         ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
2966
2967         /*
2968          * If we did not convert it, then we must try again.
2969          */
2970         if (!ret)
2971                 goto spin;
2972
2973         /*
2974          * Yeah! We succeeded in replacing the page.
2975          *
2976          * Now make the new head point back to the reader page.
2977          */
2978         rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
2979         rb_inc_page(cpu_buffer, &cpu_buffer->head_page);
2980
2981         /* Finally update the reader page to the new head */
2982         cpu_buffer->reader_page = reader;
2983         rb_reset_reader_page(cpu_buffer);
2984
2985         if (overwrite != cpu_buffer->last_overrun) {
2986                 cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
2987                 cpu_buffer->last_overrun = overwrite;
2988         }
2989
2990         goto again;
2991
2992  out:
2993         arch_spin_unlock(&cpu_buffer->lock);
2994         local_irq_restore(flags);
2995
2996         return reader;
2997 }
2998
2999 static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
3000 {
3001         struct ring_buffer_event *event;
3002         struct buffer_page *reader;
3003         unsigned length;
3004
3005         reader = rb_get_reader_page(cpu_buffer);
3006
3007         /* This function should not be called when buffer is empty */
3008         if (RB_WARN_ON(cpu_buffer, !reader))
3009                 return;
3010
3011         event = rb_reader_event(cpu_buffer);
3012
3013         if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
3014                 cpu_buffer->read++;
3015
3016         rb_update_read_stamp(cpu_buffer, event);
3017
3018         length = rb_event_length(event);
3019         cpu_buffer->reader_page->read += length;
3020 }
3021
3022 static void rb_advance_iter(struct ring_buffer_iter *iter)
3023 {
3024         struct ring_buffer_per_cpu *cpu_buffer;
3025         struct ring_buffer_event *event;
3026         unsigned length;
3027
3028         cpu_buffer = iter->cpu_buffer;
3029
3030         /*
3031          * Check if we are at the end of the buffer.
3032          */
3033         if (iter->head >= rb_page_size(iter->head_page)) {
3034                 /* discarded commits can make the page empty */
3035                 if (iter->head_page == cpu_buffer->commit_page)
3036                         return;
3037                 rb_inc_iter(iter);
3038                 return;
3039         }
3040
3041         event = rb_iter_head_event(iter);
3042
3043         length = rb_event_length(event);
3044
3045         /*
3046          * This should not be called to advance the header if we are
3047          * at the tail of the buffer.
3048          */
3049         if (RB_WARN_ON(cpu_buffer,
3050                        (iter->head_page == cpu_buffer->commit_page) &&
3051                        (iter->head + length > rb_commit_index(cpu_buffer))))
3052                 return;
3053
3054         rb_update_iter_read_stamp(iter, event);
3055
3056         iter->head += length;
3057
3058         /* check for end of page padding */
3059         if ((iter->head >= rb_page_size(iter->head_page)) &&
3060             (iter->head_page != cpu_buffer->commit_page))
3061                 rb_advance_iter(iter);
3062 }
3063
3064 static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
3065 {
3066         return cpu_buffer->lost_events;
3067 }
3068
3069 static struct ring_buffer_event *
3070 rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
3071                unsigned long *lost_events)
3072 {
3073         struct ring_buffer_event *event;
3074         struct buffer_page *reader;
3075         int nr_loops = 0;
3076
3077  again:
3078         /*
3079          * We repeat when a time extend is encountered.
3080          * Since the time extend is always attached to a data event,
3081          * we should never loop more than once.
3082          * (We never hit the following condition more than twice).
3083          */
3084         if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3085                 return NULL;
3086
3087         reader = rb_get_reader_page(cpu_buffer);
3088         if (!reader)
3089                 return NULL;
3090
3091         event = rb_reader_event(cpu_buffer);
3092
3093         switch (event->type_len) {
3094         case RINGBUF_TYPE_PADDING:
3095                 if (rb_null_event(event))
3096                         RB_WARN_ON(cpu_buffer, 1);
3097                 /*
3098                  * Because the writer could be discarding every
3099                  * event it creates (which would probably be bad)
3100                  * if we were to go back to "again" then we may never
3101                  * catch up, and will trigger the warn on, or lock
3102                  * the box. Return the padding, and we will release
3103                  * the current locks, and try again.
3104                  */
3105                 return event;
3106
3107         case RINGBUF_TYPE_TIME_EXTEND:
3108                 /* Internal data, OK to advance */
3109                 rb_advance_reader(cpu_buffer);
3110                 goto again;
3111
3112         case RINGBUF_TYPE_TIME_STAMP:
3113                 /* FIXME: not implemented */
3114                 rb_advance_reader(cpu_buffer);
3115                 goto again;
3116
3117         case RINGBUF_TYPE_DATA:
3118                 if (ts) {
3119                         *ts = cpu_buffer->read_stamp + event->time_delta;
3120                         ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3121                                                          cpu_buffer->cpu, ts);
3122                 }
3123                 if (lost_events)
3124                         *lost_events = rb_lost_events(cpu_buffer);
3125                 return event;
3126
3127         default:
3128                 BUG();
3129         }
3130
3131         return NULL;
3132 }
3133 EXPORT_SYMBOL_GPL(ring_buffer_peek);
3134
3135 static struct ring_buffer_event *
3136 rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3137 {
3138         struct ring_buffer *buffer;
3139         struct ring_buffer_per_cpu *cpu_buffer;
3140         struct ring_buffer_event *event;
3141         int nr_loops = 0;
3142
3143         cpu_buffer = iter->cpu_buffer;
3144         buffer = cpu_buffer->buffer;
3145
3146         /*
3147          * Check if someone performed a consuming read to
3148          * the buffer. A consuming read invalidates the iterator
3149          * and we need to reset the iterator in this case.
3150          */
3151         if (unlikely(iter->cache_read != cpu_buffer->read ||
3152                      iter->cache_reader_page != cpu_buffer->reader_page))
3153                 rb_iter_reset(iter);
3154
3155  again:
3156         if (ring_buffer_iter_empty(iter))
3157                 return NULL;
3158
3159         /*
3160          * We repeat when a time extend is encountered.
3161          * Since the time extend is always attached to a data event,
3162          * we should never loop more than once.
3163          * (We never hit the following condition more than twice).
3164          */
3165         if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3166                 return NULL;
3167
3168         if (rb_per_cpu_empty(cpu_buffer))
3169                 return NULL;
3170
3171         if (iter->head >= local_read(&iter->head_page->page->commit)) {
3172                 rb_inc_iter(iter);
3173                 goto again;
3174         }
3175
3176         event = rb_iter_head_event(iter);
3177
3178         switch (event->type_len) {
3179         case RINGBUF_TYPE_PADDING:
3180                 if (rb_null_event(event)) {
3181                         rb_inc_iter(iter);
3182                         goto again;
3183                 }
3184                 rb_advance_iter(iter);
3185                 return event;
3186
3187         case RINGBUF_TYPE_TIME_EXTEND:
3188                 /* Internal data, OK to advance */
3189                 rb_advance_iter(iter);
3190                 goto again;
3191
3192         case RINGBUF_TYPE_TIME_STAMP:
3193                 /* FIXME: not implemented */
3194                 rb_advance_iter(iter);
3195                 goto again;
3196
3197         case RINGBUF_TYPE_DATA:
3198                 if (ts) {
3199                         *ts = iter->read_stamp + event->time_delta;
3200                         ring_buffer_normalize_time_stamp(buffer,
3201                                                          cpu_buffer->cpu, ts);
3202                 }
3203                 return event;
3204
3205         default:
3206                 BUG();
3207         }
3208
3209         return NULL;
3210 }
3211 EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
3212
3213 static inline int rb_ok_to_lock(void)
3214 {
3215         /*
3216          * If an NMI die dumps out the content of the ring buffer
3217          * do not grab locks. We also permanently disable the ring
3218          * buffer too. A one time deal is all you get from reading
3219          * the ring buffer from an NMI.
3220          */
3221         if (likely(!in_nmi()))
3222                 return 1;
3223
3224         tracing_off_permanent();
3225         return 0;
3226 }
3227
3228 /**
3229  * ring_buffer_peek - peek at the next event to be read
3230  * @buffer: The ring buffer to read
3231  * @cpu: The cpu to peak at
3232  * @ts: The timestamp counter of this event.
3233  * @lost_events: a variable to store if events were lost (may be NULL)
3234  *
3235  * This will return the event that will be read next, but does
3236  * not consume the data.
3237  */
3238 struct ring_buffer_event *
3239 ring_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts,
3240                  unsigned long *lost_events)
3241 {
3242         struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3243         struct ring_buffer_event *event;
3244         unsigned long flags;
3245         int dolock;
3246
3247         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3248                 return NULL;
3249
3250         dolock = rb_ok_to_lock();
3251  again:
3252         local_irq_save(flags);
3253         if (dolock)
3254                 spin_lock(&cpu_buffer->reader_lock);
3255         event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3256         if (event && event->type_len == RINGBUF_TYPE_PADDING)
3257                 rb_advance_reader(cpu_buffer);
3258         if (dolock)
3259                 spin_unlock(&cpu_buffer->reader_lock);
3260         local_irq_restore(flags);
3261
3262         if (event && event->type_len == RINGBUF_TYPE_PADDING)
3263                 goto again;
3264
3265         return event;
3266 }
3267
3268 /**
3269  * ring_buffer_iter_peek - peek at the next event to be read
3270  * @iter: The ring buffer iterator
3271  * @ts: The timestamp counter of this event.
3272  *
3273  * This will return the event that will be read next, but does
3274  * not increment the iterator.
3275  */
3276 struct ring_buffer_event *
3277 ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3278 {
3279         struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3280         struct ring_buffer_event *event;
3281         unsigned long flags;
3282
3283  again:
3284         spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3285         event = rb_iter_peek(iter, ts);
3286         spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3287
3288         if (event && event->type_len == RINGBUF_TYPE_PADDING)
3289                 goto again;
3290
3291         return event;
3292 }
3293
3294 /**
3295  * ring_buffer_consume - return an event and consume it
3296  * @buffer: The ring buffer to get the next event from
3297  * @cpu: the cpu to read the buffer from
3298  * @ts: a variable to store the timestamp (may be NULL)
3299  * @lost_events: a variable to store if events were lost (may be NULL)
3300  *
3301  * Returns the next event in the ring buffer, and that event is consumed.
3302  * Meaning, that sequential reads will keep returning a different event,
3303  * and eventually empty the ring buffer if the producer is slower.
3304  */
3305 struct ring_buffer_event *
3306 ring_buffer_consume(struct ring_buffer *buffer, int cpu, u64 *ts,
3307                     unsigned long *lost_events)
3308 {
3309         struct ring_buffer_per_cpu *cpu_buffer;
3310         struct ring_buffer_event *event = NULL;
3311         unsigned long flags;
3312         int dolock;
3313
3314         dolock = rb_ok_to_lock();
3315
3316  again:
3317         /* might be called in atomic */
3318         preempt_disable();
3319
3320         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3321                 goto out;
3322
3323         cpu_buffer = buffer->buffers[cpu];
3324         local_irq_save(flags);
3325         if (dolock)
3326                 spin_lock(&cpu_buffer->reader_lock);
3327
3328         event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3329         if (event) {
3330                 cpu_buffer->lost_events = 0;
3331                 rb_advance_reader(cpu_buffer);
3332         }
3333
3334         if (dolock)
3335                 spin_unlock(&cpu_buffer->reader_lock);
3336         local_irq_restore(flags);
3337
3338  out:
3339         preempt_enable();
3340
3341         if (event && event->type_len == RINGBUF_TYPE_PADDING)
3342                 goto again;
3343
3344         return event;
3345 }
3346 EXPORT_SYMBOL_GPL(ring_buffer_consume);
3347
3348 /**
3349  * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
3350  * @buffer: The ring buffer to read from
3351  * @cpu: The cpu buffer to iterate over
3352  *
3353  * This performs the initial preparations necessary to iterate
3354  * through the buffer.  Memory is allocated, buffer recording
3355  * is disabled, and the iterator pointer is returned to the caller.
3356  *
3357  * Disabling buffer recordng prevents the reading from being
3358  * corrupted. This is not a consuming read, so a producer is not
3359  * expected.
3360  *
3361  * After a sequence of ring_buffer_read_prepare calls, the user is
3362  * expected to make at least one call to ring_buffer_prepare_sync.
3363  * Afterwards, ring_buffer_read_start is invoked to get things going
3364  * for real.
3365  *
3366  * This overall must be paired with ring_buffer_finish.
3367  */
3368 struct ring_buffer_iter *
3369 ring_buffer_read_prepare(struct ring_buffer *buffer, int cpu)
3370 {
3371         struct ring_buffer_per_cpu *cpu_buffer;
3372         struct ring_buffer_iter *iter;
3373
3374         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3375                 return NULL;
3376
3377         iter = kmalloc(sizeof(*iter), GFP_KERNEL);
3378         if (!iter)
3379                 return NULL;
3380
3381         cpu_buffer = buffer->buffers[cpu];
3382
3383         iter->cpu_buffer = cpu_buffer;
3384
3385         atomic_inc(&cpu_buffer->record_disabled);
3386
3387         return iter;
3388 }
3389 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
3390
3391 /**
3392  * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
3393  *
3394  * All previously invoked ring_buffer_read_prepare calls to prepare
3395  * iterators will be synchronized.  Afterwards, read_buffer_read_start
3396  * calls on those iterators are allowed.
3397  */
3398 void
3399 ring_buffer_read_prepare_sync(void)
3400 {
3401         synchronize_sched();
3402 }
3403 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
3404
3405 /**
3406  * ring_buffer_read_start - start a non consuming read of the buffer
3407  * @iter: The iterator returned by ring_buffer_read_prepare
3408  *
3409  * This finalizes the startup of an iteration through the buffer.
3410  * The iterator comes from a call to ring_buffer_read_prepare and
3411  * an intervening ring_buffer_read_prepare_sync must have been
3412  * performed.
3413  *
3414  * Must be paired with ring_buffer_finish.
3415  */
3416 void
3417 ring_buffer_read_start(struct ring_buffer_iter *iter)
3418 {
3419         struct ring_buffer_per_cpu *cpu_buffer;
3420         unsigned long flags;
3421
3422         if (!iter)
3423                 return;
3424
3425         cpu_buffer = iter->cpu_buffer;
3426
3427         spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3428         arch_spin_lock(&cpu_buffer->lock);
3429         rb_iter_reset(iter);
3430         arch_spin_unlock(&cpu_buffer->lock);
3431         spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3432 }
3433 EXPORT_SYMBOL_GPL(ring_buffer_read_start);
3434
3435 /**
3436  * ring_buffer_finish - finish reading the iterator of the buffer
3437  * @iter: The iterator retrieved by ring_buffer_start
3438  *
3439  * This re-enables the recording to the buffer, and frees the
3440  * iterator.
3441  */
3442 void
3443 ring_buffer_read_finish(struct ring_buffer_iter *iter)
3444 {
3445         struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3446
3447         atomic_dec(&cpu_buffer->record_disabled);
3448         kfree(iter);
3449 }
3450 EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
3451
3452 /**
3453  * ring_buffer_read - read the next item in the ring buffer by the iterator
3454  * @iter: The ring buffer iterator
3455  * @ts: The time stamp of the event read.
3456  *
3457  * This reads the next event in the ring buffer and increments the iterator.
3458  */
3459 struct ring_buffer_event *
3460 ring_buffer_read(struct ring_buffer_iter *iter, u64 *ts)
3461 {
3462         struct ring_buffer_event *event;
3463         struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3464         unsigned long flags;
3465
3466         spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3467  again:
3468         event = rb_iter_peek(iter, ts);
3469         if (!event)
3470                 goto out;
3471
3472         if (event->type_len == RINGBUF_TYPE_PADDING)
3473                 goto again;
3474
3475         rb_advance_iter(iter);
3476  out:
3477         spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3478
3479         return event;
3480 }
3481 EXPORT_SYMBOL_GPL(ring_buffer_read);
3482
3483 /**
3484  * ring_buffer_size - return the size of the ring buffer (in bytes)
3485  * @buffer: The ring buffer.
3486  */
3487 unsigned long ring_buffer_size(struct ring_buffer *buffer)
3488 {
3489         return BUF_PAGE_SIZE * buffer->pages;
3490 }
3491 EXPORT_SYMBOL_GPL(ring_buffer_size);
3492
3493 static void
3494 rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
3495 {
3496         rb_head_page_deactivate(cpu_buffer);
3497
3498         cpu_buffer->head_page
3499                 = list_entry(cpu_buffer->pages, struct buffer_page, list);
3500         local_set(&cpu_buffer->head_page->write, 0);
3501         local_set(&cpu_buffer->head_page->entries, 0);
3502         local_set(&cpu_buffer->head_page->page->commit, 0);
3503
3504         cpu_buffer->head_page->read = 0;
3505
3506         cpu_buffer->tail_page = cpu_buffer->head_page;
3507         cpu_buffer->commit_page = cpu_buffer->head_page;
3508
3509         INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
3510         local_set(&cpu_buffer->reader_page->write, 0);
3511         local_set(&cpu_buffer->reader_page->entries, 0);
3512         local_set(&cpu_buffer->reader_page->page->commit, 0);
3513         cpu_buffer->reader_page->read = 0;
3514
3515         local_set(&cpu_buffer->commit_overrun, 0);
3516         local_set(&cpu_buffer->overrun, 0);
3517         local_set(&cpu_buffer->entries, 0);
3518         local_set(&cpu_buffer->committing, 0);
3519         local_set(&cpu_buffer->commits, 0);
3520         cpu_buffer->read = 0;
3521
3522         cpu_buffer->write_stamp = 0;
3523         cpu_buffer->read_stamp = 0;
3524
3525         cpu_buffer->lost_events = 0;
3526         cpu_buffer->last_overrun = 0;
3527
3528         rb_head_page_activate(cpu_buffer);
3529 }
3530
3531 /**
3532  * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
3533  * @buffer: The ring buffer to reset a per cpu buffer of
3534  * @cpu: The CPU buffer to be reset
3535  */
3536 void ring_buffer_reset_cpu(struct ring_buffer *buffer, int cpu)
3537 {
3538         struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3539         unsigned long flags;
3540
3541         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3542                 return;
3543
3544         atomic_inc(&cpu_buffer->record_disabled);
3545
3546         spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3547
3548         if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
3549                 goto out;
3550
3551         arch_spin_lock(&cpu_buffer->lock);
3552
3553         rb_reset_cpu(cpu_buffer);
3554
3555         arch_spin_unlock(&cpu_buffer->lock);
3556
3557  out:
3558         spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3559
3560         atomic_dec(&cpu_buffer->record_disabled);
3561 }
3562 EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
3563
3564 /**
3565  * ring_buffer_reset - reset a ring buffer
3566  * @buffer: The ring buffer to reset all cpu buffers
3567  */
3568 void ring_buffer_reset(struct ring_buffer *buffer)
3569 {
3570         int cpu;
3571
3572         for_each_buffer_cpu(buffer, cpu)
3573                 ring_buffer_reset_cpu(buffer, cpu);
3574 }
3575 EXPORT_SYMBOL_GPL(ring_buffer_reset);
3576
3577 /**
3578  * rind_buffer_empty - is the ring buffer empty?
3579  * @buffer: The ring buffer to test
3580  */
3581 int ring_buffer_empty(struct ring_buffer *buffer)
3582 {
3583         struct ring_buffer_per_cpu *cpu_buffer;
3584         unsigned long flags;
3585         int dolock;
3586         int cpu;
3587         int ret;
3588
3589         dolock = rb_ok_to_lock();
3590
3591         /* yes this is racy, but if you don't like the race, lock the buffer */
3592         for_each_buffer_cpu(buffer, cpu) {
3593                 cpu_buffer = buffer->buffers[cpu];
3594                 local_irq_save(flags);
3595                 if (dolock)
3596                         spin_lock(&cpu_buffer->reader_lock);
3597                 ret = rb_per_cpu_empty(cpu_buffer);
3598                 if (dolock)
3599                         spin_unlock(&cpu_buffer->reader_lock);
3600                 local_irq_restore(flags);
3601
3602                 if (!ret)
3603                         return 0;
3604         }
3605
3606         return 1;
3607 }
3608 EXPORT_SYMBOL_GPL(ring_buffer_empty);
3609
3610 /**
3611  * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
3612  * @buffer: The ring buffer
3613  * @cpu: The CPU buffer to test
3614  */
3615 int ring_buffer_empty_cpu(struct ring_buffer *buffer, int cpu)
3616 {
3617         struct ring_buffer_per_cpu *cpu_buffer;
3618         unsigned long flags;
3619         int dolock;
3620         int ret;
3621
3622         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3623                 return 1;
3624
3625         dolock = rb_ok_to_lock();
3626
3627         cpu_buffer = buffer->buffers[cpu];
3628         local_irq_save(flags);
3629         if (dolock)
3630                 spin_lock(&cpu_buffer->reader_lock);
3631         ret = rb_per_cpu_empty(cpu_buffer);
3632         if (dolock)
3633                 spin_unlock(&cpu_buffer->reader_lock);
3634         local_irq_restore(flags);
3635
3636         return ret;
3637 }
3638 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
3639
3640 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
3641 /**
3642  * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
3643  * @buffer_a: One buffer to swap with
3644  * @buffer_b: The other buffer to swap with
3645  *
3646  * This function is useful for tracers that want to take a "snapshot"
3647  * of a CPU buffer and has another back up buffer lying around.
3648  * it is expected that the tracer handles the cpu buffer not being
3649  * used at the moment.
3650  */
3651 int ring_buffer_swap_cpu(struct ring_buffer *buffer_a,
3652                          struct ring_buffer *buffer_b, int cpu)
3653 {
3654         struct ring_buffer_per_cpu *cpu_buffer_a;
3655         struct ring_buffer_per_cpu *cpu_buffer_b;
3656         int ret = -EINVAL;
3657
3658         if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
3659             !cpumask_test_cpu(cpu, buffer_b->cpumask))
3660                 goto out;
3661
3662         /* At least make sure the two buffers are somewhat the same */
3663         if (buffer_a->pages != buffer_b->pages)
3664                 goto out;
3665
3666         ret = -EAGAIN;
3667
3668         if (ring_buffer_flags != RB_BUFFERS_ON)
3669                 goto out;
3670
3671         if (atomic_read(&buffer_a->record_disabled))
3672                 goto out;
3673
3674         if (atomic_read(&buffer_b->record_disabled))
3675                 goto out;
3676
3677         cpu_buffer_a = buffer_a->buffers[cpu];
3678         cpu_buffer_b = buffer_b->buffers[cpu];
3679
3680         if (atomic_read(&cpu_buffer_a->record_disabled))
3681                 goto out;
3682
3683         if (atomic_read(&cpu_buffer_b->record_disabled))
3684                 goto out;
3685
3686         /*
3687          * We can't do a synchronize_sched here because this
3688          * function can be called in atomic context.
3689          * Normally this will be called from the same CPU as cpu.
3690          * If not it's up to the caller to protect this.
3691          */
3692         atomic_inc(&cpu_buffer_a->record_disabled);
3693         atomic_inc(&cpu_buffer_b->record_disabled);
3694
3695         ret = -EBUSY;
3696         if (local_read(&cpu_buffer_a->committing))
3697                 goto out_dec;
3698         if (local_read(&cpu_buffer_b->committing))
3699                 goto out_dec;
3700
3701         buffer_a->buffers[cpu] = cpu_buffer_b;
3702         buffer_b->buffers[cpu] = cpu_buffer_a;
3703
3704         cpu_buffer_b->buffer = buffer_a;
3705         cpu_buffer_a->buffer = buffer_b;
3706
3707         ret = 0;
3708
3709 out_dec:
3710         atomic_dec(&cpu_buffer_a->record_disabled);
3711         atomic_dec(&cpu_buffer_b->record_disabled);
3712 out:
3713         return ret;
3714 }
3715 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
3716 #endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
3717
3718 /**
3719  * ring_buffer_alloc_read_page - allocate a page to read from buffer
3720  * @buffer: the buffer to allocate for.
3721  *
3722  * This function is used in conjunction with ring_buffer_read_page.
3723  * When reading a full page from the ring buffer, these functions
3724  * can be used to speed up the process. The calling function should
3725  * allocate a few pages first with this function. Then when it
3726  * needs to get pages from the ring buffer, it passes the result
3727  * of this function into ring_buffer_read_page, which will swap
3728  * the page that was allocated, with the read page of the buffer.
3729  *
3730  * Returns:
3731  *  The page allocated, or NULL on error.
3732  */
3733 void *ring_buffer_alloc_read_page(struct ring_buffer *buffer)
3734 {
3735         struct buffer_data_page *bpage;
3736         unsigned long addr;
3737
3738         addr = __get_free_page(GFP_KERNEL);
3739         if (!addr)
3740                 return NULL;
3741
3742         bpage = (void *)addr;
3743
3744         rb_init_page(bpage);
3745
3746         return bpage;
3747 }
3748 EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
3749
3750 /**
3751  * ring_buffer_free_read_page - free an allocated read page
3752  * @buffer: the buffer the page was allocate for
3753  * @data: the page to free
3754  *
3755  * Free a page allocated from ring_buffer_alloc_read_page.
3756  */
3757 void ring_buffer_free_read_page(struct ring_buffer *buffer, void *data)
3758 {
3759         free_page((unsigned long)data);
3760 }
3761 EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
3762
3763 /**
3764  * ring_buffer_read_page - extract a page from the ring buffer
3765  * @buffer: buffer to extract from
3766  * @data_page: the page to use allocated from ring_buffer_alloc_read_page
3767  * @len: amount to extract
3768  * @cpu: the cpu of the buffer to extract
3769  * @full: should the extraction only happen when the page is full.
3770  *
3771  * This function will pull out a page from the ring buffer and consume it.
3772  * @data_page must be the address of the variable that was returned
3773  * from ring_buffer_alloc_read_page. This is because the page might be used
3774  * to swap with a page in the ring buffer.
3775  *
3776  * for example:
3777  *      rpage = ring_buffer_alloc_read_page(buffer);
3778  *      if (!rpage)
3779  *              return error;
3780  *      ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
3781  *      if (ret >= 0)
3782  *              process_page(rpage, ret);
3783  *
3784  * When @full is set, the function will not return true unless
3785  * the writer is off the reader page.
3786  *
3787  * Note: it is up to the calling functions to handle sleeps and wakeups.
3788  *  The ring buffer can be used anywhere in the kernel and can not
3789  *  blindly call wake_up. The layer that uses the ring buffer must be
3790  *  responsible for that.
3791  *
3792  * Returns:
3793  *  >=0 if data has been transferred, returns the offset of consumed data.
3794  *  <0 if no data has been transferred.
3795  */
3796 int ring_buffer_read_page(struct ring_buffer *buffer,
3797                           void **data_page, size_t len, int cpu, int full)
3798 {
3799         struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3800         struct ring_buffer_event *event;
3801         struct buffer_data_page *bpage;
3802         struct buffer_page *reader;
3803         unsigned long missed_events;
3804         unsigned long flags;
3805         unsigned int commit;
3806         unsigned int read;
3807         u64 save_timestamp;
3808         int ret = -1;
3809
3810         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3811                 goto out;
3812
3813         /*
3814          * If len is not big enough to hold the page header, then
3815          * we can not copy anything.
3816          */
3817         if (len <= BUF_PAGE_HDR_SIZE)
3818                 goto out;
3819
3820         len -= BUF_PAGE_HDR_SIZE;
3821
3822         if (!data_page)
3823                 goto out;
3824
3825         bpage = *data_page;
3826         if (!bpage)
3827                 goto out;
3828
3829         spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3830
3831         reader = rb_get_reader_page(cpu_buffer);
3832         if (!reader)
3833                 goto out_unlock;
3834
3835         event = rb_reader_event(cpu_buffer);
3836
3837         read = reader->read;
3838         commit = rb_page_commit(reader);
3839
3840         /* Check if any events were dropped */
3841         missed_events = cpu_buffer->lost_events;
3842
3843         /*
3844          * If this page has been partially read or
3845          * if len is not big enough to read the rest of the page or
3846          * a writer is still on the page, then
3847          * we must copy the data from the page to the buffer.
3848          * Otherwise, we can simply swap the page with the one passed in.
3849          */
3850         if (read || (len < (commit - read)) ||
3851             cpu_buffer->reader_page == cpu_buffer->commit_page) {
3852                 struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
3853                 unsigned int rpos = read;
3854                 unsigned int pos = 0;
3855                 unsigned int size;
3856
3857                 if (full)
3858                         goto out_unlock;
3859
3860                 if (len > (commit - read))
3861                         len = (commit - read);
3862
3863                 /* Always keep the time extend and data together */
3864                 size = rb_event_ts_length(event);
3865
3866                 if (len < size)
3867                         goto out_unlock;
3868
3869                 /* save the current timestamp, since the user will need it */
3870                 save_timestamp = cpu_buffer->read_stamp;
3871
3872                 /* Need to copy one event at a time */
3873                 do {
3874                         /* We need the size of one event, because
3875                          * rb_advance_reader only advances by one event,
3876                          * whereas rb_event_ts_length may include the size of
3877                          * one or two events.
3878                          * We have already ensured there's enough space if this
3879                          * is a time extend. */
3880                         size = rb_event_length(event);
3881                         memcpy(bpage->data + pos, rpage->data + rpos, size);
3882
3883                         len -= size;
3884
3885                         rb_advance_reader(cpu_buffer);
3886                         rpos = reader->read;
3887                         pos += size;
3888
3889                         if (rpos >= commit)
3890                                 break;
3891
3892                         event = rb_reader_event(cpu_buffer);
3893                         /* Always keep the time extend and data together */
3894                         size = rb_event_ts_length(event);
3895                 } while (len >= size);
3896
3897                 /* update bpage */
3898                 local_set(&bpage->commit, pos);
3899                 bpage->time_stamp = save_timestamp;
3900
3901                 /* we copied everything to the beginning */
3902                 read = 0;
3903         } else {
3904                 /* update the entry counter */
3905                 cpu_buffer->read += rb_page_entries(reader);
3906
3907                 /* swap the pages */
3908                 rb_init_page(bpage);
3909                 bpage = reader->page;
3910                 reader->page = *data_page;
3911                 local_set(&reader->write, 0);
3912                 local_set(&reader->entries, 0);
3913                 reader->read = 0;
3914                 *data_page = bpage;
3915
3916                 /*
3917                  * Use the real_end for the data size,
3918                  * This gives us a chance to store the lost events
3919                  * on the page.
3920                  */
3921                 if (reader->real_end)
3922                         local_set(&bpage->commit, reader->real_end);
3923         }
3924         ret = read;
3925
3926         cpu_buffer->lost_events = 0;
3927
3928         commit = local_read(&bpage->commit);
3929         /*
3930          * Set a flag in the commit field if we lost events
3931          */
3932         if (missed_events) {
3933                 /* If there is room at the end of the page to save the
3934                  * missed events, then record it there.
3935                  */
3936                 if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
3937                         memcpy(&bpage->data[commit], &missed_events,
3938                                sizeof(missed_events));
3939                         local_add(RB_MISSED_STORED, &bpage->commit);
3940                         commit += sizeof(missed_events);
3941                 }
3942                 local_add(RB_MISSED_EVENTS, &bpage->commit);
3943         }
3944
3945         /*
3946          * This page may be off to user land. Zero it out here.
3947          */
3948         if (commit < BUF_PAGE_SIZE)
3949                 memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
3950
3951  out_unlock:
3952         spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3953
3954  out:
3955         return ret;
3956 }
3957 EXPORT_SYMBOL_GPL(ring_buffer_read_page);
3958
3959 #ifdef CONFIG_TRACING
3960 static ssize_t
3961 rb_simple_read(struct file *filp, char __user *ubuf,
3962                size_t cnt, loff_t *ppos)
3963 {
3964         unsigned long *p = filp->private_data;
3965         char buf[64];
3966         int r;
3967
3968         if (test_bit(RB_BUFFERS_DISABLED_BIT, p))
3969                 r = sprintf(buf, "permanently disabled\n");
3970         else
3971                 r = sprintf(buf, "%d\n", test_bit(RB_BUFFERS_ON_BIT, p));
3972
3973         return simple_read_from_buffer(ubuf, cnt, ppos, buf, r);
3974 }
3975
3976 static ssize_t
3977 rb_simple_write(struct file *filp, const char __user *ubuf,
3978                 size_t cnt, loff_t *ppos)
3979 {
3980         unsigned long *p = filp->private_data;
3981         char buf[64];
3982         unsigned long val;
3983         int ret;
3984
3985         if (cnt >= sizeof(buf))
3986                 return -EINVAL;
3987
3988         if (copy_from_user(&buf, ubuf, cnt))
3989                 return -EFAULT;
3990
3991         buf[cnt] = 0;
3992
3993         ret = strict_strtoul(buf, 10, &val);
3994         if (ret < 0)
3995                 return ret;
3996
3997         if (val)
3998                 set_bit(RB_BUFFERS_ON_BIT, p);
3999         else
4000                 clear_bit(RB_BUFFERS_ON_BIT, p);
4001
4002         (*ppos)++;
4003
4004         return cnt;
4005 }
4006
4007 static const struct file_operations rb_simple_fops = {
4008         .open           = tracing_open_generic,
4009         .read           = rb_simple_read,
4010         .write          = rb_simple_write,
4011         .llseek         = default_llseek,
4012 };
4013
4014
4015 static __init int rb_init_debugfs(void)
4016 {
4017         struct dentry *d_tracer;
4018
4019         d_tracer = tracing_init_dentry();
4020
4021         trace_create_file("tracing_on", 0644, d_tracer,
4022                             &ring_buffer_flags, &rb_simple_fops);
4023
4024         return 0;
4025 }
4026
4027 fs_initcall(rb_init_debugfs);
4028 #endif
4029
4030 #ifdef CONFIG_HOTPLUG_CPU
4031 static int rb_cpu_notify(struct notifier_block *self,
4032                          unsigned long action, void *hcpu)
4033 {
4034         struct ring_buffer *buffer =
4035                 container_of(self, struct ring_buffer, cpu_notify);
4036         long cpu = (long)hcpu;
4037
4038         switch (action) {
4039         case CPU_UP_PREPARE:
4040         case CPU_UP_PREPARE_FROZEN:
4041                 if (cpumask_test_cpu(cpu, buffer->cpumask))
4042                         return NOTIFY_OK;
4043
4044                 buffer->buffers[cpu] =
4045                         rb_allocate_cpu_buffer(buffer, cpu);
4046                 if (!buffer->buffers[cpu]) {
4047                         WARN(1, "failed to allocate ring buffer on CPU %ld\n",
4048                              cpu);
4049                         return NOTIFY_OK;
4050                 }
4051                 smp_wmb();
4052                 cpumask_set_cpu(cpu, buffer->cpumask);
4053                 break;
4054         case CPU_DOWN_PREPARE:
4055         case CPU_DOWN_PREPARE_FROZEN:
4056                 /*
4057                  * Do nothing.
4058                  *  If we were to free the buffer, then the user would
4059                  *  lose any trace that was in the buffer.
4060                  */
4061                 break;
4062         default:
4063                 break;
4064         }
4065         return NOTIFY_OK;
4066 }
4067 #endif