092fe0c8fdae9bacdc28ed74e2701d06bc64b1a1
[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/ftrace_irq.h>
9 #include <linux/spinlock.h>
10 #include <linux/debugfs.h>
11 #include <linux/uaccess.h>
12 #include <linux/hardirq.h>
13 #include <linux/kmemcheck.h>
14 #include <linux/module.h>
15 #include <linux/percpu.h>
16 #include <linux/mutex.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 "trace.h"
24
25 /*
26  * The ring buffer header is special. We must manually up keep it.
27  */
28 int ring_buffer_print_entry_header(struct trace_seq *s)
29 {
30         int ret;
31
32         ret = trace_seq_printf(s, "# compressed entry header\n");
33         ret = trace_seq_printf(s, "\ttype_len    :    5 bits\n");
34         ret = trace_seq_printf(s, "\ttime_delta  :   27 bits\n");
35         ret = trace_seq_printf(s, "\tarray       :   32 bits\n");
36         ret = trace_seq_printf(s, "\n");
37         ret = trace_seq_printf(s, "\tpadding     : type == %d\n",
38                                RINGBUF_TYPE_PADDING);
39         ret = trace_seq_printf(s, "\ttime_extend : type == %d\n",
40                                RINGBUF_TYPE_TIME_EXTEND);
41         ret = trace_seq_printf(s, "\tdata max type_len  == %d\n",
42                                RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
43
44         return ret;
45 }
46
47 /*
48  * The ring buffer is made up of a list of pages. A separate list of pages is
49  * allocated for each CPU. A writer may only write to a buffer that is
50  * associated with the CPU it is currently executing on.  A reader may read
51  * from any per cpu buffer.
52  *
53  * The reader is special. For each per cpu buffer, the reader has its own
54  * reader page. When a reader has read the entire reader page, this reader
55  * page is swapped with another page in the ring buffer.
56  *
57  * Now, as long as the writer is off the reader page, the reader can do what
58  * ever it wants with that page. The writer will never write to that page
59  * again (as long as it is out of the ring buffer).
60  *
61  * Here's some silly ASCII art.
62  *
63  *   +------+
64  *   |reader|          RING BUFFER
65  *   |page  |
66  *   +------+        +---+   +---+   +---+
67  *                   |   |-->|   |-->|   |
68  *                   +---+   +---+   +---+
69  *                     ^               |
70  *                     |               |
71  *                     +---------------+
72  *
73  *
74  *   +------+
75  *   |reader|          RING BUFFER
76  *   |page  |------------------v
77  *   +------+        +---+   +---+   +---+
78  *                   |   |-->|   |-->|   |
79  *                   +---+   +---+   +---+
80  *                     ^               |
81  *                     |               |
82  *                     +---------------+
83  *
84  *
85  *   +------+
86  *   |reader|          RING BUFFER
87  *   |page  |------------------v
88  *   +------+        +---+   +---+   +---+
89  *      ^            |   |-->|   |-->|   |
90  *      |            +---+   +---+   +---+
91  *      |                              |
92  *      |                              |
93  *      +------------------------------+
94  *
95  *
96  *   +------+
97  *   |buffer|          RING BUFFER
98  *   |page  |------------------v
99  *   +------+        +---+   +---+   +---+
100  *      ^            |   |   |   |-->|   |
101  *      |   New      +---+   +---+   +---+
102  *      |  Reader------^               |
103  *      |   page                       |
104  *      +------------------------------+
105  *
106  *
107  * After we make this swap, the reader can hand this page off to the splice
108  * code and be done with it. It can even allocate a new page if it needs to
109  * and swap that into the ring buffer.
110  *
111  * We will be using cmpxchg soon to make all this lockless.
112  *
113  */
114
115 /*
116  * A fast way to enable or disable all ring buffers is to
117  * call tracing_on or tracing_off. Turning off the ring buffers
118  * prevents all ring buffers from being recorded to.
119  * Turning this switch on, makes it OK to write to the
120  * ring buffer, if the ring buffer is enabled itself.
121  *
122  * There's three layers that must be on in order to write
123  * to the ring buffer.
124  *
125  * 1) This global flag must be set.
126  * 2) The ring buffer must be enabled for recording.
127  * 3) The per cpu buffer must be enabled for recording.
128  *
129  * In case of an anomaly, this global flag has a bit set that
130  * will permantly disable all ring buffers.
131  */
132
133 /*
134  * Global flag to disable all recording to ring buffers
135  *  This has two bits: ON, DISABLED
136  *
137  *  ON   DISABLED
138  * ---- ----------
139  *   0      0        : ring buffers are off
140  *   1      0        : ring buffers are on
141  *   X      1        : ring buffers are permanently disabled
142  */
143
144 enum {
145         RB_BUFFERS_ON_BIT       = 0,
146         RB_BUFFERS_DISABLED_BIT = 1,
147 };
148
149 enum {
150         RB_BUFFERS_ON           = 1 << RB_BUFFERS_ON_BIT,
151         RB_BUFFERS_DISABLED     = 1 << RB_BUFFERS_DISABLED_BIT,
152 };
153
154 static unsigned long ring_buffer_flags __read_mostly = RB_BUFFERS_ON;
155
156 #define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
157
158 /**
159  * tracing_on - enable all tracing buffers
160  *
161  * This function enables all tracing buffers that may have been
162  * disabled with tracing_off.
163  */
164 void tracing_on(void)
165 {
166         set_bit(RB_BUFFERS_ON_BIT, &ring_buffer_flags);
167 }
168 EXPORT_SYMBOL_GPL(tracing_on);
169
170 /**
171  * tracing_off - turn off all tracing buffers
172  *
173  * This function stops all tracing buffers from recording data.
174  * It does not disable any overhead the tracers themselves may
175  * be causing. This function simply causes all recording to
176  * the ring buffers to fail.
177  */
178 void tracing_off(void)
179 {
180         clear_bit(RB_BUFFERS_ON_BIT, &ring_buffer_flags);
181 }
182 EXPORT_SYMBOL_GPL(tracing_off);
183
184 /**
185  * tracing_off_permanent - permanently disable ring buffers
186  *
187  * This function, once called, will disable all ring buffers
188  * permanently.
189  */
190 void tracing_off_permanent(void)
191 {
192         set_bit(RB_BUFFERS_DISABLED_BIT, &ring_buffer_flags);
193 }
194
195 /**
196  * tracing_is_on - show state of ring buffers enabled
197  */
198 int tracing_is_on(void)
199 {
200         return ring_buffer_flags == RB_BUFFERS_ON;
201 }
202 EXPORT_SYMBOL_GPL(tracing_is_on);
203
204 #include "trace.h"
205
206 #define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
207 #define RB_ALIGNMENT            4U
208 #define RB_MAX_SMALL_DATA       (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
209 #define RB_EVNT_MIN_SIZE        8U      /* two 32bit words */
210
211 /* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
212 #define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
213
214 enum {
215         RB_LEN_TIME_EXTEND = 8,
216         RB_LEN_TIME_STAMP = 16,
217 };
218
219 static inline int rb_null_event(struct ring_buffer_event *event)
220 {
221         return event->type_len == RINGBUF_TYPE_PADDING
222                         && event->time_delta == 0;
223 }
224
225 static inline int rb_discarded_event(struct ring_buffer_event *event)
226 {
227         return event->type_len == RINGBUF_TYPE_PADDING && event->time_delta;
228 }
229
230 static void rb_event_set_padding(struct ring_buffer_event *event)
231 {
232         event->type_len = RINGBUF_TYPE_PADDING;
233         event->time_delta = 0;
234 }
235
236 static unsigned
237 rb_event_data_length(struct ring_buffer_event *event)
238 {
239         unsigned length;
240
241         if (event->type_len)
242                 length = event->type_len * RB_ALIGNMENT;
243         else
244                 length = event->array[0];
245         return length + RB_EVNT_HDR_SIZE;
246 }
247
248 /* inline for ring buffer fast paths */
249 static unsigned
250 rb_event_length(struct ring_buffer_event *event)
251 {
252         switch (event->type_len) {
253         case RINGBUF_TYPE_PADDING:
254                 if (rb_null_event(event))
255                         /* undefined */
256                         return -1;
257                 return  event->array[0] + RB_EVNT_HDR_SIZE;
258
259         case RINGBUF_TYPE_TIME_EXTEND:
260                 return RB_LEN_TIME_EXTEND;
261
262         case RINGBUF_TYPE_TIME_STAMP:
263                 return RB_LEN_TIME_STAMP;
264
265         case RINGBUF_TYPE_DATA:
266                 return rb_event_data_length(event);
267         default:
268                 BUG();
269         }
270         /* not hit */
271         return 0;
272 }
273
274 /**
275  * ring_buffer_event_length - return the length of the event
276  * @event: the event to get the length of
277  */
278 unsigned ring_buffer_event_length(struct ring_buffer_event *event)
279 {
280         unsigned length = rb_event_length(event);
281         if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
282                 return length;
283         length -= RB_EVNT_HDR_SIZE;
284         if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
285                 length -= sizeof(event->array[0]);
286         return length;
287 }
288 EXPORT_SYMBOL_GPL(ring_buffer_event_length);
289
290 /* inline for ring buffer fast paths */
291 static void *
292 rb_event_data(struct ring_buffer_event *event)
293 {
294         BUG_ON(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
295         /* If length is in len field, then array[0] has the data */
296         if (event->type_len)
297                 return (void *)&event->array[0];
298         /* Otherwise length is in array[0] and array[1] has the data */
299         return (void *)&event->array[1];
300 }
301
302 /**
303  * ring_buffer_event_data - return the data of the event
304  * @event: the event to get the data from
305  */
306 void *ring_buffer_event_data(struct ring_buffer_event *event)
307 {
308         return rb_event_data(event);
309 }
310 EXPORT_SYMBOL_GPL(ring_buffer_event_data);
311
312 #define for_each_buffer_cpu(buffer, cpu)                \
313         for_each_cpu(cpu, buffer->cpumask)
314
315 #define TS_SHIFT        27
316 #define TS_MASK         ((1ULL << TS_SHIFT) - 1)
317 #define TS_DELTA_TEST   (~TS_MASK)
318
319 struct buffer_data_page {
320         u64              time_stamp;    /* page time stamp */
321         local_t          commit;        /* write committed index */
322         unsigned char    data[];        /* data of buffer page */
323 };
324
325 /*
326  * Note, the buffer_page list must be first. The buffer pages
327  * are allocated in cache lines, which means that each buffer
328  * page will be at the beginning of a cache line, and thus
329  * the least significant bits will be zero. We use this to
330  * add flags in the list struct pointers, to make the ring buffer
331  * lockless.
332  */
333 struct buffer_page {
334         struct list_head list;          /* list of buffer pages */
335         local_t          write;         /* index for next write */
336         unsigned         read;          /* index for next read */
337         local_t          entries;       /* entries on this page */
338         struct buffer_data_page *page;  /* Actual data page */
339 };
340
341 /*
342  * The buffer page counters, write and entries, must be reset
343  * atomically when crossing page boundaries. To synchronize this
344  * update, two counters are inserted into the number. One is
345  * the actual counter for the write position or count on the page.
346  *
347  * The other is a counter of updaters. Before an update happens
348  * the update partition of the counter is incremented. This will
349  * allow the updater to update the counter atomically.
350  *
351  * The counter is 20 bits, and the state data is 12.
352  */
353 #define RB_WRITE_MASK           0xfffff
354 #define RB_WRITE_INTCNT         (1 << 20)
355
356 static void rb_init_page(struct buffer_data_page *bpage)
357 {
358         local_set(&bpage->commit, 0);
359 }
360
361 /**
362  * ring_buffer_page_len - the size of data on the page.
363  * @page: The page to read
364  *
365  * Returns the amount of data on the page, including buffer page header.
366  */
367 size_t ring_buffer_page_len(void *page)
368 {
369         return local_read(&((struct buffer_data_page *)page)->commit)
370                 + BUF_PAGE_HDR_SIZE;
371 }
372
373 /*
374  * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
375  * this issue out.
376  */
377 static void free_buffer_page(struct buffer_page *bpage)
378 {
379         free_page((unsigned long)bpage->page);
380         kfree(bpage);
381 }
382
383 /*
384  * We need to fit the time_stamp delta into 27 bits.
385  */
386 static inline int test_time_stamp(u64 delta)
387 {
388         if (delta & TS_DELTA_TEST)
389                 return 1;
390         return 0;
391 }
392
393 #define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
394
395 /* Max payload is BUF_PAGE_SIZE - header (8bytes) */
396 #define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
397
398 /* Max number of timestamps that can fit on a page */
399 #define RB_TIMESTAMPS_PER_PAGE  (BUF_PAGE_SIZE / RB_LEN_TIME_STAMP)
400
401 int ring_buffer_print_page_header(struct trace_seq *s)
402 {
403         struct buffer_data_page field;
404         int ret;
405
406         ret = trace_seq_printf(s, "\tfield: u64 timestamp;\t"
407                                "offset:0;\tsize:%u;\n",
408                                (unsigned int)sizeof(field.time_stamp));
409
410         ret = trace_seq_printf(s, "\tfield: local_t commit;\t"
411                                "offset:%u;\tsize:%u;\n",
412                                (unsigned int)offsetof(typeof(field), commit),
413                                (unsigned int)sizeof(field.commit));
414
415         ret = trace_seq_printf(s, "\tfield: char data;\t"
416                                "offset:%u;\tsize:%u;\n",
417                                (unsigned int)offsetof(typeof(field), data),
418                                (unsigned int)BUF_PAGE_SIZE);
419
420         return ret;
421 }
422
423 /*
424  * head_page == tail_page && head == tail then buffer is empty.
425  */
426 struct ring_buffer_per_cpu {
427         int                             cpu;
428         struct ring_buffer              *buffer;
429         spinlock_t                      reader_lock;    /* serialize readers */
430         raw_spinlock_t                  lock;
431         struct lock_class_key           lock_key;
432         struct list_head                *pages;
433         struct buffer_page              *head_page;     /* read from head */
434         struct buffer_page              *tail_page;     /* write to tail */
435         struct buffer_page              *commit_page;   /* committed pages */
436         struct buffer_page              *reader_page;
437         local_t                         commit_overrun;
438         local_t                         overrun;
439         local_t                         entries;
440         local_t                         committing;
441         local_t                         commits;
442         unsigned long                   read;
443         u64                             write_stamp;
444         u64                             read_stamp;
445         atomic_t                        record_disabled;
446 };
447
448 struct ring_buffer {
449         unsigned                        pages;
450         unsigned                        flags;
451         int                             cpus;
452         atomic_t                        record_disabled;
453         cpumask_var_t                   cpumask;
454
455         struct lock_class_key           *reader_lock_key;
456
457         struct mutex                    mutex;
458
459         struct ring_buffer_per_cpu      **buffers;
460
461 #ifdef CONFIG_HOTPLUG_CPU
462         struct notifier_block           cpu_notify;
463 #endif
464         u64                             (*clock)(void);
465 };
466
467 struct ring_buffer_iter {
468         struct ring_buffer_per_cpu      *cpu_buffer;
469         unsigned long                   head;
470         struct buffer_page              *head_page;
471         u64                             read_stamp;
472 };
473
474 /* buffer may be either ring_buffer or ring_buffer_per_cpu */
475 #define RB_WARN_ON(buffer, cond)                                \
476         ({                                                      \
477                 int _____ret = unlikely(cond);                  \
478                 if (_____ret) {                                 \
479                         atomic_inc(&buffer->record_disabled);   \
480                         WARN_ON(1);                             \
481                 }                                               \
482                 _____ret;                                       \
483         })
484
485 /* Up this if you want to test the TIME_EXTENTS and normalization */
486 #define DEBUG_SHIFT 0
487
488 static inline u64 rb_time_stamp(struct ring_buffer *buffer, int cpu)
489 {
490         /* shift to debug/test normalization and TIME_EXTENTS */
491         return buffer->clock() << DEBUG_SHIFT;
492 }
493
494 u64 ring_buffer_time_stamp(struct ring_buffer *buffer, int cpu)
495 {
496         u64 time;
497
498         preempt_disable_notrace();
499         time = rb_time_stamp(buffer, cpu);
500         preempt_enable_no_resched_notrace();
501
502         return time;
503 }
504 EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
505
506 void ring_buffer_normalize_time_stamp(struct ring_buffer *buffer,
507                                       int cpu, u64 *ts)
508 {
509         /* Just stupid testing the normalize function and deltas */
510         *ts >>= DEBUG_SHIFT;
511 }
512 EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
513
514 /*
515  * Making the ring buffer lockless makes things tricky.
516  * Although writes only happen on the CPU that they are on,
517  * and they only need to worry about interrupts. Reads can
518  * happen on any CPU.
519  *
520  * The reader page is always off the ring buffer, but when the
521  * reader finishes with a page, it needs to swap its page with
522  * a new one from the buffer. The reader needs to take from
523  * the head (writes go to the tail). But if a writer is in overwrite
524  * mode and wraps, it must push the head page forward.
525  *
526  * Here lies the problem.
527  *
528  * The reader must be careful to replace only the head page, and
529  * not another one. As described at the top of the file in the
530  * ASCII art, the reader sets its old page to point to the next
531  * page after head. It then sets the page after head to point to
532  * the old reader page. But if the writer moves the head page
533  * during this operation, the reader could end up with the tail.
534  *
535  * We use cmpxchg to help prevent this race. We also do something
536  * special with the page before head. We set the LSB to 1.
537  *
538  * When the writer must push the page forward, it will clear the
539  * bit that points to the head page, move the head, and then set
540  * the bit that points to the new head page.
541  *
542  * We also don't want an interrupt coming in and moving the head
543  * page on another writer. Thus we use the second LSB to catch
544  * that too. Thus:
545  *
546  * head->list->prev->next        bit 1          bit 0
547  *                              -------        -------
548  * Normal page                     0              0
549  * Points to head page             0              1
550  * New head page                   1              0
551  *
552  * Note we can not trust the prev pointer of the head page, because:
553  *
554  * +----+       +-----+        +-----+
555  * |    |------>|  T  |---X--->|  N  |
556  * |    |<------|     |        |     |
557  * +----+       +-----+        +-----+
558  *   ^                           ^ |
559  *   |          +-----+          | |
560  *   +----------|  R  |----------+ |
561  *              |     |<-----------+
562  *              +-----+
563  *
564  * Key:  ---X-->  HEAD flag set in pointer
565  *         T      Tail page
566  *         R      Reader page
567  *         N      Next page
568  *
569  * (see __rb_reserve_next() to see where this happens)
570  *
571  *  What the above shows is that the reader just swapped out
572  *  the reader page with a page in the buffer, but before it
573  *  could make the new header point back to the new page added
574  *  it was preempted by a writer. The writer moved forward onto
575  *  the new page added by the reader and is about to move forward
576  *  again.
577  *
578  *  You can see, it is legitimate for the previous pointer of
579  *  the head (or any page) not to point back to itself. But only
580  *  temporarially.
581  */
582
583 #define RB_PAGE_NORMAL          0UL
584 #define RB_PAGE_HEAD            1UL
585 #define RB_PAGE_UPDATE          2UL
586
587
588 #define RB_FLAG_MASK            3UL
589
590 /* PAGE_MOVED is not part of the mask */
591 #define RB_PAGE_MOVED           4UL
592
593 /*
594  * rb_list_head - remove any bit
595  */
596 static struct list_head *rb_list_head(struct list_head *list)
597 {
598         unsigned long val = (unsigned long)list;
599
600         return (struct list_head *)(val & ~RB_FLAG_MASK);
601 }
602
603 /*
604  * rb_is_head_page - test if the give page is the head page
605  *
606  * Because the reader may move the head_page pointer, we can
607  * not trust what the head page is (it may be pointing to
608  * the reader page). But if the next page is a header page,
609  * its flags will be non zero.
610  */
611 static int inline
612 rb_is_head_page(struct ring_buffer_per_cpu *cpu_buffer,
613                 struct buffer_page *page, struct list_head *list)
614 {
615         unsigned long val;
616
617         val = (unsigned long)list->next;
618
619         if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
620                 return RB_PAGE_MOVED;
621
622         return val & RB_FLAG_MASK;
623 }
624
625 /*
626  * rb_is_reader_page
627  *
628  * The unique thing about the reader page, is that, if the
629  * writer is ever on it, the previous pointer never points
630  * back to the reader page.
631  */
632 static int rb_is_reader_page(struct buffer_page *page)
633 {
634         struct list_head *list = page->list.prev;
635
636         return rb_list_head(list->next) != &page->list;
637 }
638
639 /*
640  * rb_set_list_to_head - set a list_head to be pointing to head.
641  */
642 static void rb_set_list_to_head(struct ring_buffer_per_cpu *cpu_buffer,
643                                 struct list_head *list)
644 {
645         unsigned long *ptr;
646
647         ptr = (unsigned long *)&list->next;
648         *ptr |= RB_PAGE_HEAD;
649         *ptr &= ~RB_PAGE_UPDATE;
650 }
651
652 /*
653  * rb_head_page_activate - sets up head page
654  */
655 static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
656 {
657         struct buffer_page *head;
658
659         head = cpu_buffer->head_page;
660         if (!head)
661                 return;
662
663         /*
664          * Set the previous list pointer to have the HEAD flag.
665          */
666         rb_set_list_to_head(cpu_buffer, head->list.prev);
667 }
668
669 static void rb_list_head_clear(struct list_head *list)
670 {
671         unsigned long *ptr = (unsigned long *)&list->next;
672
673         *ptr &= ~RB_FLAG_MASK;
674 }
675
676 /*
677  * rb_head_page_dactivate - clears head page ptr (for free list)
678  */
679 static void
680 rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
681 {
682         struct list_head *hd;
683
684         /* Go through the whole list and clear any pointers found. */
685         rb_list_head_clear(cpu_buffer->pages);
686
687         list_for_each(hd, cpu_buffer->pages)
688                 rb_list_head_clear(hd);
689 }
690
691 static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
692                             struct buffer_page *head,
693                             struct buffer_page *prev,
694                             int old_flag, int new_flag)
695 {
696         struct list_head *list;
697         unsigned long val = (unsigned long)&head->list;
698         unsigned long ret;
699
700         list = &prev->list;
701
702         val &= ~RB_FLAG_MASK;
703
704         ret = (unsigned long)cmpxchg(&list->next,
705                                      val | old_flag, val | new_flag);
706
707         /* check if the reader took the page */
708         if ((ret & ~RB_FLAG_MASK) != val)
709                 return RB_PAGE_MOVED;
710
711         return ret & RB_FLAG_MASK;
712 }
713
714 static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
715                                    struct buffer_page *head,
716                                    struct buffer_page *prev,
717                                    int old_flag)
718 {
719         return rb_head_page_set(cpu_buffer, head, prev,
720                                 old_flag, RB_PAGE_UPDATE);
721 }
722
723 static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
724                                  struct buffer_page *head,
725                                  struct buffer_page *prev,
726                                  int old_flag)
727 {
728         return rb_head_page_set(cpu_buffer, head, prev,
729                                 old_flag, RB_PAGE_HEAD);
730 }
731
732 static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
733                                    struct buffer_page *head,
734                                    struct buffer_page *prev,
735                                    int old_flag)
736 {
737         return rb_head_page_set(cpu_buffer, head, prev,
738                                 old_flag, RB_PAGE_NORMAL);
739 }
740
741 static inline void rb_inc_page(struct ring_buffer_per_cpu *cpu_buffer,
742                                struct buffer_page **bpage)
743 {
744         struct list_head *p = rb_list_head((*bpage)->list.next);
745
746         *bpage = list_entry(p, struct buffer_page, list);
747 }
748
749 static struct buffer_page *
750 rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
751 {
752         struct buffer_page *head;
753         struct buffer_page *page;
754         struct list_head *list;
755         int i;
756
757         if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
758                 return NULL;
759
760         /* sanity check */
761         list = cpu_buffer->pages;
762         if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
763                 return NULL;
764
765         page = head = cpu_buffer->head_page;
766         /*
767          * It is possible that the writer moves the header behind
768          * where we started, and we miss in one loop.
769          * A second loop should grab the header, but we'll do
770          * three loops just because I'm paranoid.
771          */
772         for (i = 0; i < 3; i++) {
773                 do {
774                         if (rb_is_head_page(cpu_buffer, page, page->list.prev)) {
775                                 cpu_buffer->head_page = page;
776                                 return page;
777                         }
778                         rb_inc_page(cpu_buffer, &page);
779                 } while (page != head);
780         }
781
782         RB_WARN_ON(cpu_buffer, 1);
783
784         return NULL;
785 }
786
787 static int rb_head_page_replace(struct buffer_page *old,
788                                 struct buffer_page *new)
789 {
790         unsigned long *ptr = (unsigned long *)&old->list.prev->next;
791         unsigned long val;
792         unsigned long ret;
793
794         val = *ptr & ~RB_FLAG_MASK;
795         val |= RB_PAGE_HEAD;
796
797         ret = cmpxchg(ptr, val, &new->list);
798
799         return ret == val;
800 }
801
802 /*
803  * rb_tail_page_update - move the tail page forward
804  *
805  * Returns 1 if moved tail page, 0 if someone else did.
806  */
807 static int rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
808                                struct buffer_page *tail_page,
809                                struct buffer_page *next_page)
810 {
811         struct buffer_page *old_tail;
812         unsigned long old_entries;
813         unsigned long old_write;
814         int ret = 0;
815
816         /*
817          * The tail page now needs to be moved forward.
818          *
819          * We need to reset the tail page, but without messing
820          * with possible erasing of data brought in by interrupts
821          * that have moved the tail page and are currently on it.
822          *
823          * We add a counter to the write field to denote this.
824          */
825         old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
826         old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
827
828         /*
829          * Just make sure we have seen our old_write and synchronize
830          * with any interrupts that come in.
831          */
832         barrier();
833
834         /*
835          * If the tail page is still the same as what we think
836          * it is, then it is up to us to update the tail
837          * pointer.
838          */
839         if (tail_page == cpu_buffer->tail_page) {
840                 /* Zero the write counter */
841                 unsigned long val = old_write & ~RB_WRITE_MASK;
842                 unsigned long eval = old_entries & ~RB_WRITE_MASK;
843
844                 /*
845                  * This will only succeed if an interrupt did
846                  * not come in and change it. In which case, we
847                  * do not want to modify it.
848                  *
849                  * We add (void) to let the compiler know that we do not care
850                  * about the return value of these functions. We use the
851                  * cmpxchg to only update if an interrupt did not already
852                  * do it for us. If the cmpxchg fails, we don't care.
853                  */
854                 (void)local_cmpxchg(&next_page->write, old_write, val);
855                 (void)local_cmpxchg(&next_page->entries, old_entries, eval);
856
857                 /*
858                  * No need to worry about races with clearing out the commit.
859                  * it only can increment when a commit takes place. But that
860                  * only happens in the outer most nested commit.
861                  */
862                 local_set(&next_page->page->commit, 0);
863
864                 old_tail = cmpxchg(&cpu_buffer->tail_page,
865                                    tail_page, next_page);
866
867                 if (old_tail == tail_page)
868                         ret = 1;
869         }
870
871         return ret;
872 }
873
874 static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
875                           struct buffer_page *bpage)
876 {
877         unsigned long val = (unsigned long)bpage;
878
879         if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK))
880                 return 1;
881
882         return 0;
883 }
884
885 /**
886  * rb_check_list - make sure a pointer to a list has the last bits zero
887  */
888 static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer,
889                          struct list_head *list)
890 {
891         if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev))
892                 return 1;
893         if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next))
894                 return 1;
895         return 0;
896 }
897
898 /**
899  * check_pages - integrity check of buffer pages
900  * @cpu_buffer: CPU buffer with pages to test
901  *
902  * As a safety measure we check to make sure the data pages have not
903  * been corrupted.
904  */
905 static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
906 {
907         struct list_head *head = cpu_buffer->pages;
908         struct buffer_page *bpage, *tmp;
909
910         rb_head_page_deactivate(cpu_buffer);
911
912         if (RB_WARN_ON(cpu_buffer, head->next->prev != head))
913                 return -1;
914         if (RB_WARN_ON(cpu_buffer, head->prev->next != head))
915                 return -1;
916
917         if (rb_check_list(cpu_buffer, head))
918                 return -1;
919
920         list_for_each_entry_safe(bpage, tmp, head, list) {
921                 if (RB_WARN_ON(cpu_buffer,
922                                bpage->list.next->prev != &bpage->list))
923                         return -1;
924                 if (RB_WARN_ON(cpu_buffer,
925                                bpage->list.prev->next != &bpage->list))
926                         return -1;
927                 if (rb_check_list(cpu_buffer, &bpage->list))
928                         return -1;
929         }
930
931         rb_head_page_activate(cpu_buffer);
932
933         return 0;
934 }
935
936 static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
937                              unsigned nr_pages)
938 {
939         struct buffer_page *bpage, *tmp;
940         unsigned long addr;
941         LIST_HEAD(pages);
942         unsigned i;
943
944         WARN_ON(!nr_pages);
945
946         for (i = 0; i < nr_pages; i++) {
947                 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
948                                     GFP_KERNEL, cpu_to_node(cpu_buffer->cpu));
949                 if (!bpage)
950                         goto free_pages;
951
952                 rb_check_bpage(cpu_buffer, bpage);
953
954                 list_add(&bpage->list, &pages);
955
956                 addr = __get_free_page(GFP_KERNEL);
957                 if (!addr)
958                         goto free_pages;
959                 bpage->page = (void *)addr;
960                 rb_init_page(bpage->page);
961         }
962
963         /*
964          * The ring buffer page list is a circular list that does not
965          * start and end with a list head. All page list items point to
966          * other pages.
967          */
968         cpu_buffer->pages = pages.next;
969         list_del(&pages);
970
971         rb_check_pages(cpu_buffer);
972
973         return 0;
974
975  free_pages:
976         list_for_each_entry_safe(bpage, tmp, &pages, list) {
977                 list_del_init(&bpage->list);
978                 free_buffer_page(bpage);
979         }
980         return -ENOMEM;
981 }
982
983 static struct ring_buffer_per_cpu *
984 rb_allocate_cpu_buffer(struct ring_buffer *buffer, int cpu)
985 {
986         struct ring_buffer_per_cpu *cpu_buffer;
987         struct buffer_page *bpage;
988         unsigned long addr;
989         int ret;
990
991         cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
992                                   GFP_KERNEL, cpu_to_node(cpu));
993         if (!cpu_buffer)
994                 return NULL;
995
996         cpu_buffer->cpu = cpu;
997         cpu_buffer->buffer = buffer;
998         spin_lock_init(&cpu_buffer->reader_lock);
999         lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1000         cpu_buffer->lock = (raw_spinlock_t)__RAW_SPIN_LOCK_UNLOCKED;
1001
1002         bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1003                             GFP_KERNEL, cpu_to_node(cpu));
1004         if (!bpage)
1005                 goto fail_free_buffer;
1006
1007         rb_check_bpage(cpu_buffer, bpage);
1008
1009         cpu_buffer->reader_page = bpage;
1010         addr = __get_free_page(GFP_KERNEL);
1011         if (!addr)
1012                 goto fail_free_reader;
1013         bpage->page = (void *)addr;
1014         rb_init_page(bpage->page);
1015
1016         INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1017
1018         ret = rb_allocate_pages(cpu_buffer, buffer->pages);
1019         if (ret < 0)
1020                 goto fail_free_reader;
1021
1022         cpu_buffer->head_page
1023                 = list_entry(cpu_buffer->pages, struct buffer_page, list);
1024         cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1025
1026         rb_head_page_activate(cpu_buffer);
1027
1028         return cpu_buffer;
1029
1030  fail_free_reader:
1031         free_buffer_page(cpu_buffer->reader_page);
1032
1033  fail_free_buffer:
1034         kfree(cpu_buffer);
1035         return NULL;
1036 }
1037
1038 static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1039 {
1040         struct list_head *head = cpu_buffer->pages;
1041         struct buffer_page *bpage, *tmp;
1042
1043         free_buffer_page(cpu_buffer->reader_page);
1044
1045         rb_head_page_deactivate(cpu_buffer);
1046
1047         if (head) {
1048                 list_for_each_entry_safe(bpage, tmp, head, list) {
1049                         list_del_init(&bpage->list);
1050                         free_buffer_page(bpage);
1051                 }
1052                 bpage = list_entry(head, struct buffer_page, list);
1053                 free_buffer_page(bpage);
1054         }
1055
1056         kfree(cpu_buffer);
1057 }
1058
1059 #ifdef CONFIG_HOTPLUG_CPU
1060 static int rb_cpu_notify(struct notifier_block *self,
1061                          unsigned long action, void *hcpu);
1062 #endif
1063
1064 /**
1065  * ring_buffer_alloc - allocate a new ring_buffer
1066  * @size: the size in bytes per cpu that is needed.
1067  * @flags: attributes to set for the ring buffer.
1068  *
1069  * Currently the only flag that is available is the RB_FL_OVERWRITE
1070  * flag. This flag means that the buffer will overwrite old data
1071  * when the buffer wraps. If this flag is not set, the buffer will
1072  * drop data when the tail hits the head.
1073  */
1074 struct ring_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1075                                         struct lock_class_key *key)
1076 {
1077         struct ring_buffer *buffer;
1078         int bsize;
1079         int cpu;
1080
1081         /* keep it in its own cache line */
1082         buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1083                          GFP_KERNEL);
1084         if (!buffer)
1085                 return NULL;
1086
1087         if (!alloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1088                 goto fail_free_buffer;
1089
1090         buffer->pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1091         buffer->flags = flags;
1092         buffer->clock = trace_clock_local;
1093         buffer->reader_lock_key = key;
1094
1095         /* need at least two pages */
1096         if (buffer->pages < 2)
1097                 buffer->pages = 2;
1098
1099         /*
1100          * In case of non-hotplug cpu, if the ring-buffer is allocated
1101          * in early initcall, it will not be notified of secondary cpus.
1102          * In that off case, we need to allocate for all possible cpus.
1103          */
1104 #ifdef CONFIG_HOTPLUG_CPU
1105         get_online_cpus();
1106         cpumask_copy(buffer->cpumask, cpu_online_mask);
1107 #else
1108         cpumask_copy(buffer->cpumask, cpu_possible_mask);
1109 #endif
1110         buffer->cpus = nr_cpu_ids;
1111
1112         bsize = sizeof(void *) * nr_cpu_ids;
1113         buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1114                                   GFP_KERNEL);
1115         if (!buffer->buffers)
1116                 goto fail_free_cpumask;
1117
1118         for_each_buffer_cpu(buffer, cpu) {
1119                 buffer->buffers[cpu] =
1120                         rb_allocate_cpu_buffer(buffer, cpu);
1121                 if (!buffer->buffers[cpu])
1122                         goto fail_free_buffers;
1123         }
1124
1125 #ifdef CONFIG_HOTPLUG_CPU
1126         buffer->cpu_notify.notifier_call = rb_cpu_notify;
1127         buffer->cpu_notify.priority = 0;
1128         register_cpu_notifier(&buffer->cpu_notify);
1129 #endif
1130
1131         put_online_cpus();
1132         mutex_init(&buffer->mutex);
1133
1134         return buffer;
1135
1136  fail_free_buffers:
1137         for_each_buffer_cpu(buffer, cpu) {
1138                 if (buffer->buffers[cpu])
1139                         rb_free_cpu_buffer(buffer->buffers[cpu]);
1140         }
1141         kfree(buffer->buffers);
1142
1143  fail_free_cpumask:
1144         free_cpumask_var(buffer->cpumask);
1145         put_online_cpus();
1146
1147  fail_free_buffer:
1148         kfree(buffer);
1149         return NULL;
1150 }
1151 EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1152
1153 /**
1154  * ring_buffer_free - free a ring buffer.
1155  * @buffer: the buffer to free.
1156  */
1157 void
1158 ring_buffer_free(struct ring_buffer *buffer)
1159 {
1160         int cpu;
1161
1162         get_online_cpus();
1163
1164 #ifdef CONFIG_HOTPLUG_CPU
1165         unregister_cpu_notifier(&buffer->cpu_notify);
1166 #endif
1167
1168         for_each_buffer_cpu(buffer, cpu)
1169                 rb_free_cpu_buffer(buffer->buffers[cpu]);
1170
1171         put_online_cpus();
1172
1173         kfree(buffer->buffers);
1174         free_cpumask_var(buffer->cpumask);
1175
1176         kfree(buffer);
1177 }
1178 EXPORT_SYMBOL_GPL(ring_buffer_free);
1179
1180 void ring_buffer_set_clock(struct ring_buffer *buffer,
1181                            u64 (*clock)(void))
1182 {
1183         buffer->clock = clock;
1184 }
1185
1186 static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1187
1188 static void
1189 rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned nr_pages)
1190 {
1191         struct buffer_page *bpage;
1192         struct list_head *p;
1193         unsigned i;
1194
1195         atomic_inc(&cpu_buffer->record_disabled);
1196         synchronize_sched();
1197
1198         rb_head_page_deactivate(cpu_buffer);
1199
1200         for (i = 0; i < nr_pages; i++) {
1201                 if (RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages)))
1202                         return;
1203                 p = cpu_buffer->pages->next;
1204                 bpage = list_entry(p, struct buffer_page, list);
1205                 list_del_init(&bpage->list);
1206                 free_buffer_page(bpage);
1207         }
1208         if (RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages)))
1209                 return;
1210
1211         rb_reset_cpu(cpu_buffer);
1212
1213         rb_check_pages(cpu_buffer);
1214
1215         atomic_dec(&cpu_buffer->record_disabled);
1216
1217 }
1218
1219 static void
1220 rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer,
1221                 struct list_head *pages, unsigned nr_pages)
1222 {
1223         struct buffer_page *bpage;
1224         struct list_head *p;
1225         unsigned i;
1226
1227         atomic_inc(&cpu_buffer->record_disabled);
1228         synchronize_sched();
1229
1230         spin_lock_irq(&cpu_buffer->reader_lock);
1231         rb_head_page_deactivate(cpu_buffer);
1232
1233         for (i = 0; i < nr_pages; i++) {
1234                 if (RB_WARN_ON(cpu_buffer, list_empty(pages)))
1235                         return;
1236                 p = pages->next;
1237                 bpage = list_entry(p, struct buffer_page, list);
1238                 list_del_init(&bpage->list);
1239                 list_add_tail(&bpage->list, cpu_buffer->pages);
1240         }
1241         rb_reset_cpu(cpu_buffer);
1242         spin_unlock_irq(&cpu_buffer->reader_lock);
1243
1244         rb_check_pages(cpu_buffer);
1245
1246         atomic_dec(&cpu_buffer->record_disabled);
1247 }
1248
1249 /**
1250  * ring_buffer_resize - resize the ring buffer
1251  * @buffer: the buffer to resize.
1252  * @size: the new size.
1253  *
1254  * The tracer is responsible for making sure that the buffer is
1255  * not being used while changing the size.
1256  * Note: We may be able to change the above requirement by using
1257  *  RCU synchronizations.
1258  *
1259  * Minimum size is 2 * BUF_PAGE_SIZE.
1260  *
1261  * Returns -1 on failure.
1262  */
1263 int ring_buffer_resize(struct ring_buffer *buffer, unsigned long size)
1264 {
1265         struct ring_buffer_per_cpu *cpu_buffer;
1266         unsigned nr_pages, rm_pages, new_pages;
1267         struct buffer_page *bpage, *tmp;
1268         unsigned long buffer_size;
1269         unsigned long addr;
1270         LIST_HEAD(pages);
1271         int i, cpu;
1272
1273         /*
1274          * Always succeed at resizing a non-existent buffer:
1275          */
1276         if (!buffer)
1277                 return size;
1278
1279         size = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1280         size *= BUF_PAGE_SIZE;
1281         buffer_size = buffer->pages * BUF_PAGE_SIZE;
1282
1283         /* we need a minimum of two pages */
1284         if (size < BUF_PAGE_SIZE * 2)
1285                 size = BUF_PAGE_SIZE * 2;
1286
1287         if (size == buffer_size)
1288                 return size;
1289
1290         mutex_lock(&buffer->mutex);
1291         get_online_cpus();
1292
1293         nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1294
1295         if (size < buffer_size) {
1296
1297                 /* easy case, just free pages */
1298                 if (RB_WARN_ON(buffer, nr_pages >= buffer->pages))
1299                         goto out_fail;
1300
1301                 rm_pages = buffer->pages - nr_pages;
1302
1303                 for_each_buffer_cpu(buffer, cpu) {
1304                         cpu_buffer = buffer->buffers[cpu];
1305                         rb_remove_pages(cpu_buffer, rm_pages);
1306                 }
1307                 goto out;
1308         }
1309
1310         /*
1311          * This is a bit more difficult. We only want to add pages
1312          * when we can allocate enough for all CPUs. We do this
1313          * by allocating all the pages and storing them on a local
1314          * link list. If we succeed in our allocation, then we
1315          * add these pages to the cpu_buffers. Otherwise we just free
1316          * them all and return -ENOMEM;
1317          */
1318         if (RB_WARN_ON(buffer, nr_pages <= buffer->pages))
1319                 goto out_fail;
1320
1321         new_pages = nr_pages - buffer->pages;
1322
1323         for_each_buffer_cpu(buffer, cpu) {
1324                 for (i = 0; i < new_pages; i++) {
1325                         bpage = kzalloc_node(ALIGN(sizeof(*bpage),
1326                                                   cache_line_size()),
1327                                             GFP_KERNEL, cpu_to_node(cpu));
1328                         if (!bpage)
1329                                 goto free_pages;
1330                         list_add(&bpage->list, &pages);
1331                         addr = __get_free_page(GFP_KERNEL);
1332                         if (!addr)
1333                                 goto free_pages;
1334                         bpage->page = (void *)addr;
1335                         rb_init_page(bpage->page);
1336                 }
1337         }
1338
1339         for_each_buffer_cpu(buffer, cpu) {
1340                 cpu_buffer = buffer->buffers[cpu];
1341                 rb_insert_pages(cpu_buffer, &pages, new_pages);
1342         }
1343
1344         if (RB_WARN_ON(buffer, !list_empty(&pages)))
1345                 goto out_fail;
1346
1347  out:
1348         buffer->pages = nr_pages;
1349         put_online_cpus();
1350         mutex_unlock(&buffer->mutex);
1351
1352         return size;
1353
1354  free_pages:
1355         list_for_each_entry_safe(bpage, tmp, &pages, list) {
1356                 list_del_init(&bpage->list);
1357                 free_buffer_page(bpage);
1358         }
1359         put_online_cpus();
1360         mutex_unlock(&buffer->mutex);
1361         return -ENOMEM;
1362
1363         /*
1364          * Something went totally wrong, and we are too paranoid
1365          * to even clean up the mess.
1366          */
1367  out_fail:
1368         put_online_cpus();
1369         mutex_unlock(&buffer->mutex);
1370         return -1;
1371 }
1372 EXPORT_SYMBOL_GPL(ring_buffer_resize);
1373
1374 static inline void *
1375 __rb_data_page_index(struct buffer_data_page *bpage, unsigned index)
1376 {
1377         return bpage->data + index;
1378 }
1379
1380 static inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
1381 {
1382         return bpage->page->data + index;
1383 }
1384
1385 static inline struct ring_buffer_event *
1386 rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
1387 {
1388         return __rb_page_index(cpu_buffer->reader_page,
1389                                cpu_buffer->reader_page->read);
1390 }
1391
1392 static inline struct ring_buffer_event *
1393 rb_iter_head_event(struct ring_buffer_iter *iter)
1394 {
1395         return __rb_page_index(iter->head_page, iter->head);
1396 }
1397
1398 static inline unsigned long rb_page_write(struct buffer_page *bpage)
1399 {
1400         return local_read(&bpage->write) & RB_WRITE_MASK;
1401 }
1402
1403 static inline unsigned rb_page_commit(struct buffer_page *bpage)
1404 {
1405         return local_read(&bpage->page->commit);
1406 }
1407
1408 static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1409 {
1410         return local_read(&bpage->entries) & RB_WRITE_MASK;
1411 }
1412
1413 /* Size is determined by what has been commited */
1414 static inline unsigned rb_page_size(struct buffer_page *bpage)
1415 {
1416         return rb_page_commit(bpage);
1417 }
1418
1419 static inline unsigned
1420 rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
1421 {
1422         return rb_page_commit(cpu_buffer->commit_page);
1423 }
1424
1425 static inline unsigned
1426 rb_event_index(struct ring_buffer_event *event)
1427 {
1428         unsigned long addr = (unsigned long)event;
1429
1430         return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
1431 }
1432
1433 static inline int
1434 rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
1435                    struct ring_buffer_event *event)
1436 {
1437         unsigned long addr = (unsigned long)event;
1438         unsigned long index;
1439
1440         index = rb_event_index(event);
1441         addr &= PAGE_MASK;
1442
1443         return cpu_buffer->commit_page->page == (void *)addr &&
1444                 rb_commit_index(cpu_buffer) == index;
1445 }
1446
1447 static void
1448 rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
1449 {
1450         unsigned long max_count;
1451
1452         /*
1453          * We only race with interrupts and NMIs on this CPU.
1454          * If we own the commit event, then we can commit
1455          * all others that interrupted us, since the interruptions
1456          * are in stack format (they finish before they come
1457          * back to us). This allows us to do a simple loop to
1458          * assign the commit to the tail.
1459          */
1460  again:
1461         max_count = cpu_buffer->buffer->pages * 100;
1462
1463         while (cpu_buffer->commit_page != cpu_buffer->tail_page) {
1464                 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
1465                         return;
1466                 if (RB_WARN_ON(cpu_buffer,
1467                                rb_is_reader_page(cpu_buffer->tail_page)))
1468                         return;
1469                 local_set(&cpu_buffer->commit_page->page->commit,
1470                           rb_page_write(cpu_buffer->commit_page));
1471                 rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
1472                 cpu_buffer->write_stamp =
1473                         cpu_buffer->commit_page->page->time_stamp;
1474                 /* add barrier to keep gcc from optimizing too much */
1475                 barrier();
1476         }
1477         while (rb_commit_index(cpu_buffer) !=
1478                rb_page_write(cpu_buffer->commit_page)) {
1479
1480                 local_set(&cpu_buffer->commit_page->page->commit,
1481                           rb_page_write(cpu_buffer->commit_page));
1482                 RB_WARN_ON(cpu_buffer,
1483                            local_read(&cpu_buffer->commit_page->page->commit) &
1484                            ~RB_WRITE_MASK);
1485                 barrier();
1486         }
1487
1488         /* again, keep gcc from optimizing */
1489         barrier();
1490
1491         /*
1492          * If an interrupt came in just after the first while loop
1493          * and pushed the tail page forward, we will be left with
1494          * a dangling commit that will never go forward.
1495          */
1496         if (unlikely(cpu_buffer->commit_page != cpu_buffer->tail_page))
1497                 goto again;
1498 }
1499
1500 static void rb_reset_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
1501 {
1502         cpu_buffer->read_stamp = cpu_buffer->reader_page->page->time_stamp;
1503         cpu_buffer->reader_page->read = 0;
1504 }
1505
1506 static void rb_inc_iter(struct ring_buffer_iter *iter)
1507 {
1508         struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
1509
1510         /*
1511          * The iterator could be on the reader page (it starts there).
1512          * But the head could have moved, since the reader was
1513          * found. Check for this case and assign the iterator
1514          * to the head page instead of next.
1515          */
1516         if (iter->head_page == cpu_buffer->reader_page)
1517                 iter->head_page = rb_set_head_page(cpu_buffer);
1518         else
1519                 rb_inc_page(cpu_buffer, &iter->head_page);
1520
1521         iter->read_stamp = iter->head_page->page->time_stamp;
1522         iter->head = 0;
1523 }
1524
1525 /**
1526  * ring_buffer_update_event - update event type and data
1527  * @event: the even to update
1528  * @type: the type of event
1529  * @length: the size of the event field in the ring buffer
1530  *
1531  * Update the type and data fields of the event. The length
1532  * is the actual size that is written to the ring buffer,
1533  * and with this, we can determine what to place into the
1534  * data field.
1535  */
1536 static void
1537 rb_update_event(struct ring_buffer_event *event,
1538                          unsigned type, unsigned length)
1539 {
1540         event->type_len = type;
1541
1542         switch (type) {
1543
1544         case RINGBUF_TYPE_PADDING:
1545         case RINGBUF_TYPE_TIME_EXTEND:
1546         case RINGBUF_TYPE_TIME_STAMP:
1547                 break;
1548
1549         case 0:
1550                 length -= RB_EVNT_HDR_SIZE;
1551                 if (length > RB_MAX_SMALL_DATA)
1552                         event->array[0] = length;
1553                 else
1554                         event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
1555                 break;
1556         default:
1557                 BUG();
1558         }
1559 }
1560
1561 /*
1562  * rb_handle_head_page - writer hit the head page
1563  *
1564  * Returns: +1 to retry page
1565  *           0 to continue
1566  *          -1 on error
1567  */
1568 static int
1569 rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
1570                     struct buffer_page *tail_page,
1571                     struct buffer_page *next_page)
1572 {
1573         struct buffer_page *new_head;
1574         int entries;
1575         int type;
1576         int ret;
1577
1578         entries = rb_page_entries(next_page);
1579
1580         /*
1581          * The hard part is here. We need to move the head
1582          * forward, and protect against both readers on
1583          * other CPUs and writers coming in via interrupts.
1584          */
1585         type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
1586                                        RB_PAGE_HEAD);
1587
1588         /*
1589          * type can be one of four:
1590          *  NORMAL - an interrupt already moved it for us
1591          *  HEAD   - we are the first to get here.
1592          *  UPDATE - we are the interrupt interrupting
1593          *           a current move.
1594          *  MOVED  - a reader on another CPU moved the next
1595          *           pointer to its reader page. Give up
1596          *           and try again.
1597          */
1598
1599         switch (type) {
1600         case RB_PAGE_HEAD:
1601                 /*
1602                  * We changed the head to UPDATE, thus
1603                  * it is our responsibility to update
1604                  * the counters.
1605                  */
1606                 local_add(entries, &cpu_buffer->overrun);
1607
1608                 /*
1609                  * The entries will be zeroed out when we move the
1610                  * tail page.
1611                  */
1612
1613                 /* still more to do */
1614                 break;
1615
1616         case RB_PAGE_UPDATE:
1617                 /*
1618                  * This is an interrupt that interrupt the
1619                  * previous update. Still more to do.
1620                  */
1621                 break;
1622         case RB_PAGE_NORMAL:
1623                 /*
1624                  * An interrupt came in before the update
1625                  * and processed this for us.
1626                  * Nothing left to do.
1627                  */
1628                 return 1;
1629         case RB_PAGE_MOVED:
1630                 /*
1631                  * The reader is on another CPU and just did
1632                  * a swap with our next_page.
1633                  * Try again.
1634                  */
1635                 return 1;
1636         default:
1637                 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
1638                 return -1;
1639         }
1640
1641         /*
1642          * Now that we are here, the old head pointer is
1643          * set to UPDATE. This will keep the reader from
1644          * swapping the head page with the reader page.
1645          * The reader (on another CPU) will spin till
1646          * we are finished.
1647          *
1648          * We just need to protect against interrupts
1649          * doing the job. We will set the next pointer
1650          * to HEAD. After that, we set the old pointer
1651          * to NORMAL, but only if it was HEAD before.
1652          * otherwise we are an interrupt, and only
1653          * want the outer most commit to reset it.
1654          */
1655         new_head = next_page;
1656         rb_inc_page(cpu_buffer, &new_head);
1657
1658         ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
1659                                     RB_PAGE_NORMAL);
1660
1661         /*
1662          * Valid returns are:
1663          *  HEAD   - an interrupt came in and already set it.
1664          *  NORMAL - One of two things:
1665          *            1) We really set it.
1666          *            2) A bunch of interrupts came in and moved
1667          *               the page forward again.
1668          */
1669         switch (ret) {
1670         case RB_PAGE_HEAD:
1671         case RB_PAGE_NORMAL:
1672                 /* OK */
1673                 break;
1674         default:
1675                 RB_WARN_ON(cpu_buffer, 1);
1676                 return -1;
1677         }
1678
1679         /*
1680          * It is possible that an interrupt came in,
1681          * set the head up, then more interrupts came in
1682          * and moved it again. When we get back here,
1683          * the page would have been set to NORMAL but we
1684          * just set it back to HEAD.
1685          *
1686          * How do you detect this? Well, if that happened
1687          * the tail page would have moved.
1688          */
1689         if (ret == RB_PAGE_NORMAL) {
1690                 /*
1691                  * If the tail had moved passed next, then we need
1692                  * to reset the pointer.
1693                  */
1694                 if (cpu_buffer->tail_page != tail_page &&
1695                     cpu_buffer->tail_page != next_page)
1696                         rb_head_page_set_normal(cpu_buffer, new_head,
1697                                                 next_page,
1698                                                 RB_PAGE_HEAD);
1699         }
1700
1701         /*
1702          * If this was the outer most commit (the one that
1703          * changed the original pointer from HEAD to UPDATE),
1704          * then it is up to us to reset it to NORMAL.
1705          */
1706         if (type == RB_PAGE_HEAD) {
1707                 ret = rb_head_page_set_normal(cpu_buffer, next_page,
1708                                               tail_page,
1709                                               RB_PAGE_UPDATE);
1710                 if (RB_WARN_ON(cpu_buffer,
1711                                ret != RB_PAGE_UPDATE))
1712                         return -1;
1713         }
1714
1715         return 0;
1716 }
1717
1718 static unsigned rb_calculate_event_length(unsigned length)
1719 {
1720         struct ring_buffer_event event; /* Used only for sizeof array */
1721
1722         /* zero length can cause confusions */
1723         if (!length)
1724                 length = 1;
1725
1726         if (length > RB_MAX_SMALL_DATA)
1727                 length += sizeof(event.array[0]);
1728
1729         length += RB_EVNT_HDR_SIZE;
1730         length = ALIGN(length, RB_ALIGNMENT);
1731
1732         return length;
1733 }
1734
1735 static inline void
1736 rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
1737               struct buffer_page *tail_page,
1738               unsigned long tail, unsigned long length)
1739 {
1740         struct ring_buffer_event *event;
1741
1742         /*
1743          * Only the event that crossed the page boundary
1744          * must fill the old tail_page with padding.
1745          */
1746         if (tail >= BUF_PAGE_SIZE) {
1747                 local_sub(length, &tail_page->write);
1748                 return;
1749         }
1750
1751         event = __rb_page_index(tail_page, tail);
1752         kmemcheck_annotate_bitfield(event, bitfield);
1753
1754         /*
1755          * If this event is bigger than the minimum size, then
1756          * we need to be careful that we don't subtract the
1757          * write counter enough to allow another writer to slip
1758          * in on this page.
1759          * We put in a discarded commit instead, to make sure
1760          * that this space is not used again.
1761          *
1762          * If we are less than the minimum size, we don't need to
1763          * worry about it.
1764          */
1765         if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
1766                 /* No room for any events */
1767
1768                 /* Mark the rest of the page with padding */
1769                 rb_event_set_padding(event);
1770
1771                 /* Set the write back to the previous setting */
1772                 local_sub(length, &tail_page->write);
1773                 return;
1774         }
1775
1776         /* Put in a discarded event */
1777         event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
1778         event->type_len = RINGBUF_TYPE_PADDING;
1779         /* time delta must be non zero */
1780         event->time_delta = 1;
1781         /* Account for this as an entry */
1782         local_inc(&tail_page->entries);
1783         local_inc(&cpu_buffer->entries);
1784
1785         /* Set write to end of buffer */
1786         length = (tail + length) - BUF_PAGE_SIZE;
1787         local_sub(length, &tail_page->write);
1788 }
1789
1790 static struct ring_buffer_event *
1791 rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
1792              unsigned long length, unsigned long tail,
1793              struct buffer_page *commit_page,
1794              struct buffer_page *tail_page, u64 *ts)
1795 {
1796         struct ring_buffer *buffer = cpu_buffer->buffer;
1797         struct buffer_page *next_page;
1798         int ret;
1799
1800         next_page = tail_page;
1801
1802         rb_inc_page(cpu_buffer, &next_page);
1803
1804         /*
1805          * If for some reason, we had an interrupt storm that made
1806          * it all the way around the buffer, bail, and warn
1807          * about it.
1808          */
1809         if (unlikely(next_page == commit_page)) {
1810                 local_inc(&cpu_buffer->commit_overrun);
1811                 goto out_reset;
1812         }
1813
1814         /*
1815          * This is where the fun begins!
1816          *
1817          * We are fighting against races between a reader that
1818          * could be on another CPU trying to swap its reader
1819          * page with the buffer head.
1820          *
1821          * We are also fighting against interrupts coming in and
1822          * moving the head or tail on us as well.
1823          *
1824          * If the next page is the head page then we have filled
1825          * the buffer, unless the commit page is still on the
1826          * reader page.
1827          */
1828         if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) {
1829
1830                 /*
1831                  * If the commit is not on the reader page, then
1832                  * move the header page.
1833                  */
1834                 if (!rb_is_reader_page(cpu_buffer->commit_page)) {
1835                         /*
1836                          * If we are not in overwrite mode,
1837                          * this is easy, just stop here.
1838                          */
1839                         if (!(buffer->flags & RB_FL_OVERWRITE))
1840                                 goto out_reset;
1841
1842                         ret = rb_handle_head_page(cpu_buffer,
1843                                                   tail_page,
1844                                                   next_page);
1845                         if (ret < 0)
1846                                 goto out_reset;
1847                         if (ret)
1848                                 goto out_again;
1849                 } else {
1850                         /*
1851                          * We need to be careful here too. The
1852                          * commit page could still be on the reader
1853                          * page. We could have a small buffer, and
1854                          * have filled up the buffer with events
1855                          * from interrupts and such, and wrapped.
1856                          *
1857                          * Note, if the tail page is also the on the
1858                          * reader_page, we let it move out.
1859                          */
1860                         if (unlikely((cpu_buffer->commit_page !=
1861                                       cpu_buffer->tail_page) &&
1862                                      (cpu_buffer->commit_page ==
1863                                       cpu_buffer->reader_page))) {
1864                                 local_inc(&cpu_buffer->commit_overrun);
1865                                 goto out_reset;
1866                         }
1867                 }
1868         }
1869
1870         ret = rb_tail_page_update(cpu_buffer, tail_page, next_page);
1871         if (ret) {
1872                 /*
1873                  * Nested commits always have zero deltas, so
1874                  * just reread the time stamp
1875                  */
1876                 *ts = rb_time_stamp(buffer, cpu_buffer->cpu);
1877                 next_page->page->time_stamp = *ts;
1878         }
1879
1880  out_again:
1881
1882         rb_reset_tail(cpu_buffer, tail_page, tail, length);
1883
1884         /* fail and let the caller try again */
1885         return ERR_PTR(-EAGAIN);
1886
1887  out_reset:
1888         /* reset write */
1889         rb_reset_tail(cpu_buffer, tail_page, tail, length);
1890
1891         return NULL;
1892 }
1893
1894 static struct ring_buffer_event *
1895 __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
1896                   unsigned type, unsigned long length, u64 *ts)
1897 {
1898         struct buffer_page *tail_page, *commit_page;
1899         struct ring_buffer_event *event;
1900         unsigned long tail, write;
1901
1902         commit_page = cpu_buffer->commit_page;
1903         /* we just need to protect against interrupts */
1904         barrier();
1905         tail_page = cpu_buffer->tail_page;
1906         write = local_add_return(length, &tail_page->write);
1907
1908         /* set write to only the index of the write */
1909         write &= RB_WRITE_MASK;
1910         tail = write - length;
1911
1912         /* See if we shot pass the end of this buffer page */
1913         if (write > BUF_PAGE_SIZE)
1914                 return rb_move_tail(cpu_buffer, length, tail,
1915                                     commit_page, tail_page, ts);
1916
1917         /* We reserved something on the buffer */
1918
1919         event = __rb_page_index(tail_page, tail);
1920         kmemcheck_annotate_bitfield(event, bitfield);
1921         rb_update_event(event, type, length);
1922
1923         /* The passed in type is zero for DATA */
1924         if (likely(!type))
1925                 local_inc(&tail_page->entries);
1926
1927         /*
1928          * If this is the first commit on the page, then update
1929          * its timestamp.
1930          */
1931         if (!tail)
1932                 tail_page->page->time_stamp = *ts;
1933
1934         return event;
1935 }
1936
1937 static inline int
1938 rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
1939                   struct ring_buffer_event *event)
1940 {
1941         unsigned long new_index, old_index;
1942         struct buffer_page *bpage;
1943         unsigned long index;
1944         unsigned long addr;
1945
1946         new_index = rb_event_index(event);
1947         old_index = new_index + rb_event_length(event);
1948         addr = (unsigned long)event;
1949         addr &= PAGE_MASK;
1950
1951         bpage = cpu_buffer->tail_page;
1952
1953         if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
1954                 unsigned long write_mask =
1955                         local_read(&bpage->write) & ~RB_WRITE_MASK;
1956                 /*
1957                  * This is on the tail page. It is possible that
1958                  * a write could come in and move the tail page
1959                  * and write to the next page. That is fine
1960                  * because we just shorten what is on this page.
1961                  */
1962                 old_index += write_mask;
1963                 new_index += write_mask;
1964                 index = local_cmpxchg(&bpage->write, old_index, new_index);
1965                 if (index == old_index)
1966                         return 1;
1967         }
1968
1969         /* could not discard */
1970         return 0;
1971 }
1972
1973 static int
1974 rb_add_time_stamp(struct ring_buffer_per_cpu *cpu_buffer,
1975                   u64 *ts, u64 *delta)
1976 {
1977         struct ring_buffer_event *event;
1978         static int once;
1979         int ret;
1980
1981         if (unlikely(*delta > (1ULL << 59) && !once++)) {
1982                 printk(KERN_WARNING "Delta way too big! %llu"
1983                        " ts=%llu write stamp = %llu\n",
1984                        (unsigned long long)*delta,
1985                        (unsigned long long)*ts,
1986                        (unsigned long long)cpu_buffer->write_stamp);
1987                 WARN_ON(1);
1988         }
1989
1990         /*
1991          * The delta is too big, we to add a
1992          * new timestamp.
1993          */
1994         event = __rb_reserve_next(cpu_buffer,
1995                                   RINGBUF_TYPE_TIME_EXTEND,
1996                                   RB_LEN_TIME_EXTEND,
1997                                   ts);
1998         if (!event)
1999                 return -EBUSY;
2000
2001         if (PTR_ERR(event) == -EAGAIN)
2002                 return -EAGAIN;
2003
2004         /* Only a commited time event can update the write stamp */
2005         if (rb_event_is_commit(cpu_buffer, event)) {
2006                 /*
2007                  * If this is the first on the page, then it was
2008                  * updated with the page itself. Try to discard it
2009                  * and if we can't just make it zero.
2010                  */
2011                 if (rb_event_index(event)) {
2012                         event->time_delta = *delta & TS_MASK;
2013                         event->array[0] = *delta >> TS_SHIFT;
2014                 } else {
2015                         /* try to discard, since we do not need this */
2016                         if (!rb_try_to_discard(cpu_buffer, event)) {
2017                                 /* nope, just zero it */
2018                                 event->time_delta = 0;
2019                                 event->array[0] = 0;
2020                         }
2021                 }
2022                 cpu_buffer->write_stamp = *ts;
2023                 /* let the caller know this was the commit */
2024                 ret = 1;
2025         } else {
2026                 /* Try to discard the event */
2027                 if (!rb_try_to_discard(cpu_buffer, event)) {
2028                         /* Darn, this is just wasted space */
2029                         event->time_delta = 0;
2030                         event->array[0] = 0;
2031                 }
2032                 ret = 0;
2033         }
2034
2035         *delta = 0;
2036
2037         return ret;
2038 }
2039
2040 static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2041 {
2042         local_inc(&cpu_buffer->committing);
2043         local_inc(&cpu_buffer->commits);
2044 }
2045
2046 static void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
2047 {
2048         unsigned long commits;
2049
2050         if (RB_WARN_ON(cpu_buffer,
2051                        !local_read(&cpu_buffer->committing)))
2052                 return;
2053
2054  again:
2055         commits = local_read(&cpu_buffer->commits);
2056         /* synchronize with interrupts */
2057         barrier();
2058         if (local_read(&cpu_buffer->committing) == 1)
2059                 rb_set_commit_to_write(cpu_buffer);
2060
2061         local_dec(&cpu_buffer->committing);
2062
2063         /* synchronize with interrupts */
2064         barrier();
2065
2066         /*
2067          * Need to account for interrupts coming in between the
2068          * updating of the commit page and the clearing of the
2069          * committing counter.
2070          */
2071         if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
2072             !local_read(&cpu_buffer->committing)) {
2073                 local_inc(&cpu_buffer->committing);
2074                 goto again;
2075         }
2076 }
2077
2078 static struct ring_buffer_event *
2079 rb_reserve_next_event(struct ring_buffer_per_cpu *cpu_buffer,
2080                       unsigned long length)
2081 {
2082         struct ring_buffer_event *event;
2083         u64 ts, delta = 0;
2084         int commit = 0;
2085         int nr_loops = 0;
2086
2087         rb_start_commit(cpu_buffer);
2088
2089         length = rb_calculate_event_length(length);
2090  again:
2091         /*
2092          * We allow for interrupts to reenter here and do a trace.
2093          * If one does, it will cause this original code to loop
2094          * back here. Even with heavy interrupts happening, this
2095          * should only happen a few times in a row. If this happens
2096          * 1000 times in a row, there must be either an interrupt
2097          * storm or we have something buggy.
2098          * Bail!
2099          */
2100         if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
2101                 goto out_fail;
2102
2103         ts = rb_time_stamp(cpu_buffer->buffer, cpu_buffer->cpu);
2104
2105         /*
2106          * Only the first commit can update the timestamp.
2107          * Yes there is a race here. If an interrupt comes in
2108          * just after the conditional and it traces too, then it
2109          * will also check the deltas. More than one timestamp may
2110          * also be made. But only the entry that did the actual
2111          * commit will be something other than zero.
2112          */
2113         if (likely(cpu_buffer->tail_page == cpu_buffer->commit_page &&
2114                    rb_page_write(cpu_buffer->tail_page) ==
2115                    rb_commit_index(cpu_buffer))) {
2116                 u64 diff;
2117
2118                 diff = ts - cpu_buffer->write_stamp;
2119
2120                 /* make sure this diff is calculated here */
2121                 barrier();
2122
2123                 /* Did the write stamp get updated already? */
2124                 if (unlikely(ts < cpu_buffer->write_stamp))
2125                         goto get_event;
2126
2127                 delta = diff;
2128                 if (unlikely(test_time_stamp(delta))) {
2129
2130                         commit = rb_add_time_stamp(cpu_buffer, &ts, &delta);
2131                         if (commit == -EBUSY)
2132                                 goto out_fail;
2133
2134                         if (commit == -EAGAIN)
2135                                 goto again;
2136
2137                         RB_WARN_ON(cpu_buffer, commit < 0);
2138                 }
2139         }
2140
2141  get_event:
2142         event = __rb_reserve_next(cpu_buffer, 0, length, &ts);
2143         if (unlikely(PTR_ERR(event) == -EAGAIN))
2144                 goto again;
2145
2146         if (!event)
2147                 goto out_fail;
2148
2149         if (!rb_event_is_commit(cpu_buffer, event))
2150                 delta = 0;
2151
2152         event->time_delta = delta;
2153
2154         return event;
2155
2156  out_fail:
2157         rb_end_commit(cpu_buffer);
2158         return NULL;
2159 }
2160
2161 #ifdef CONFIG_TRACING
2162
2163 #define TRACE_RECURSIVE_DEPTH 16
2164
2165 static int trace_recursive_lock(void)
2166 {
2167         current->trace_recursion++;
2168
2169         if (likely(current->trace_recursion < TRACE_RECURSIVE_DEPTH))
2170                 return 0;
2171
2172         /* Disable all tracing before we do anything else */
2173         tracing_off_permanent();
2174
2175         printk_once(KERN_WARNING "Tracing recursion: depth[%ld]:"
2176                     "HC[%lu]:SC[%lu]:NMI[%lu]\n",
2177                     current->trace_recursion,
2178                     hardirq_count() >> HARDIRQ_SHIFT,
2179                     softirq_count() >> SOFTIRQ_SHIFT,
2180                     in_nmi());
2181
2182         WARN_ON_ONCE(1);
2183         return -1;
2184 }
2185
2186 static void trace_recursive_unlock(void)
2187 {
2188         WARN_ON_ONCE(!current->trace_recursion);
2189
2190         current->trace_recursion--;
2191 }
2192
2193 #else
2194
2195 #define trace_recursive_lock()          (0)
2196 #define trace_recursive_unlock()        do { } while (0)
2197
2198 #endif
2199
2200 static DEFINE_PER_CPU(int, rb_need_resched);
2201
2202 /**
2203  * ring_buffer_lock_reserve - reserve a part of the buffer
2204  * @buffer: the ring buffer to reserve from
2205  * @length: the length of the data to reserve (excluding event header)
2206  *
2207  * Returns a reseverd event on the ring buffer to copy directly to.
2208  * The user of this interface will need to get the body to write into
2209  * and can use the ring_buffer_event_data() interface.
2210  *
2211  * The length is the length of the data needed, not the event length
2212  * which also includes the event header.
2213  *
2214  * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2215  * If NULL is returned, then nothing has been allocated or locked.
2216  */
2217 struct ring_buffer_event *
2218 ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length)
2219 {
2220         struct ring_buffer_per_cpu *cpu_buffer;
2221         struct ring_buffer_event *event;
2222         int cpu, resched;
2223
2224         if (ring_buffer_flags != RB_BUFFERS_ON)
2225                 return NULL;
2226
2227         if (atomic_read(&buffer->record_disabled))
2228                 return NULL;
2229
2230         /* If we are tracing schedule, we don't want to recurse */
2231         resched = ftrace_preempt_disable();
2232
2233         if (trace_recursive_lock())
2234                 goto out_nocheck;
2235
2236         cpu = raw_smp_processor_id();
2237
2238         if (!cpumask_test_cpu(cpu, buffer->cpumask))
2239                 goto out;
2240
2241         cpu_buffer = buffer->buffers[cpu];
2242
2243         if (atomic_read(&cpu_buffer->record_disabled))
2244                 goto out;
2245
2246         if (length > BUF_MAX_DATA_SIZE)
2247                 goto out;
2248
2249         event = rb_reserve_next_event(cpu_buffer, length);
2250         if (!event)
2251                 goto out;
2252
2253         /*
2254          * Need to store resched state on this cpu.
2255          * Only the first needs to.
2256          */
2257
2258         if (preempt_count() == 1)
2259                 per_cpu(rb_need_resched, cpu) = resched;
2260
2261         return event;
2262
2263  out:
2264         trace_recursive_unlock();
2265
2266  out_nocheck:
2267         ftrace_preempt_enable(resched);
2268         return NULL;
2269 }
2270 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
2271
2272 static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
2273                       struct ring_buffer_event *event)
2274 {
2275         local_inc(&cpu_buffer->entries);
2276
2277         /*
2278          * The event first in the commit queue updates the
2279          * time stamp.
2280          */
2281         if (rb_event_is_commit(cpu_buffer, event))
2282                 cpu_buffer->write_stamp += event->time_delta;
2283
2284         rb_end_commit(cpu_buffer);
2285 }
2286
2287 /**
2288  * ring_buffer_unlock_commit - commit a reserved
2289  * @buffer: The buffer to commit to
2290  * @event: The event pointer to commit.
2291  *
2292  * This commits the data to the ring buffer, and releases any locks held.
2293  *
2294  * Must be paired with ring_buffer_lock_reserve.
2295  */
2296 int ring_buffer_unlock_commit(struct ring_buffer *buffer,
2297                               struct ring_buffer_event *event)
2298 {
2299         struct ring_buffer_per_cpu *cpu_buffer;
2300         int cpu = raw_smp_processor_id();
2301
2302         cpu_buffer = buffer->buffers[cpu];
2303
2304         rb_commit(cpu_buffer, event);
2305
2306         trace_recursive_unlock();
2307
2308         /*
2309          * Only the last preempt count needs to restore preemption.
2310          */
2311         if (preempt_count() == 1)
2312                 ftrace_preempt_enable(per_cpu(rb_need_resched, cpu));
2313         else
2314                 preempt_enable_no_resched_notrace();
2315
2316         return 0;
2317 }
2318 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
2319
2320 static inline void rb_event_discard(struct ring_buffer_event *event)
2321 {
2322         /* array[0] holds the actual length for the discarded event */
2323         event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
2324         event->type_len = RINGBUF_TYPE_PADDING;
2325         /* time delta must be non zero */
2326         if (!event->time_delta)
2327                 event->time_delta = 1;
2328 }
2329
2330 /**
2331  * ring_buffer_commit_discard - discard an event that has not been committed
2332  * @buffer: the ring buffer
2333  * @event: non committed event to discard
2334  *
2335  * Sometimes an event that is in the ring buffer needs to be ignored.
2336  * This function lets the user discard an event in the ring buffer
2337  * and then that event will not be read later.
2338  *
2339  * This function only works if it is called before the the item has been
2340  * committed. It will try to free the event from the ring buffer
2341  * if another event has not been added behind it.
2342  *
2343  * If another event has been added behind it, it will set the event
2344  * up as discarded, and perform the commit.
2345  *
2346  * If this function is called, do not call ring_buffer_unlock_commit on
2347  * the event.
2348  */
2349 void ring_buffer_discard_commit(struct ring_buffer *buffer,
2350                                 struct ring_buffer_event *event)
2351 {
2352         struct ring_buffer_per_cpu *cpu_buffer;
2353         int cpu;
2354
2355         /* The event is discarded regardless */
2356         rb_event_discard(event);
2357
2358         cpu = smp_processor_id();
2359         cpu_buffer = buffer->buffers[cpu];
2360
2361         /*
2362          * This must only be called if the event has not been
2363          * committed yet. Thus we can assume that preemption
2364          * is still disabled.
2365          */
2366         RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
2367
2368         if (rb_try_to_discard(cpu_buffer, event))
2369                 goto out;
2370
2371         /*
2372          * The commit is still visible by the reader, so we
2373          * must increment entries.
2374          */
2375         local_inc(&cpu_buffer->entries);
2376  out:
2377         rb_end_commit(cpu_buffer);
2378
2379         trace_recursive_unlock();
2380
2381         /*
2382          * Only the last preempt count needs to restore preemption.
2383          */
2384         if (preempt_count() == 1)
2385                 ftrace_preempt_enable(per_cpu(rb_need_resched, cpu));
2386         else
2387                 preempt_enable_no_resched_notrace();
2388
2389 }
2390 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
2391
2392 /**
2393  * ring_buffer_write - write data to the buffer without reserving
2394  * @buffer: The ring buffer to write to.
2395  * @length: The length of the data being written (excluding the event header)
2396  * @data: The data to write to the buffer.
2397  *
2398  * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
2399  * one function. If you already have the data to write to the buffer, it
2400  * may be easier to simply call this function.
2401  *
2402  * Note, like ring_buffer_lock_reserve, the length is the length of the data
2403  * and not the length of the event which would hold the header.
2404  */
2405 int ring_buffer_write(struct ring_buffer *buffer,
2406                         unsigned long length,
2407                         void *data)
2408 {
2409         struct ring_buffer_per_cpu *cpu_buffer;
2410         struct ring_buffer_event *event;
2411         void *body;
2412         int ret = -EBUSY;
2413         int cpu, resched;
2414
2415         if (ring_buffer_flags != RB_BUFFERS_ON)
2416                 return -EBUSY;
2417
2418         if (atomic_read(&buffer->record_disabled))
2419                 return -EBUSY;
2420
2421         resched = ftrace_preempt_disable();
2422
2423         cpu = raw_smp_processor_id();
2424
2425         if (!cpumask_test_cpu(cpu, buffer->cpumask))
2426                 goto out;
2427
2428         cpu_buffer = buffer->buffers[cpu];
2429
2430         if (atomic_read(&cpu_buffer->record_disabled))
2431                 goto out;
2432
2433         if (length > BUF_MAX_DATA_SIZE)
2434                 goto out;
2435
2436         event = rb_reserve_next_event(cpu_buffer, length);
2437         if (!event)
2438                 goto out;
2439
2440         body = rb_event_data(event);
2441
2442         memcpy(body, data, length);
2443
2444         rb_commit(cpu_buffer, event);
2445
2446         ret = 0;
2447  out:
2448         ftrace_preempt_enable(resched);
2449
2450         return ret;
2451 }
2452 EXPORT_SYMBOL_GPL(ring_buffer_write);
2453
2454 static int rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
2455 {
2456         struct buffer_page *reader = cpu_buffer->reader_page;
2457         struct buffer_page *head = rb_set_head_page(cpu_buffer);
2458         struct buffer_page *commit = cpu_buffer->commit_page;
2459
2460         /* In case of error, head will be NULL */
2461         if (unlikely(!head))
2462                 return 1;
2463
2464         return reader->read == rb_page_commit(reader) &&
2465                 (commit == reader ||
2466                  (commit == head &&
2467                   head->read == rb_page_commit(commit)));
2468 }
2469
2470 /**
2471  * ring_buffer_record_disable - stop all writes into the buffer
2472  * @buffer: The ring buffer to stop writes to.
2473  *
2474  * This prevents all writes to the buffer. Any attempt to write
2475  * to the buffer after this will fail and return NULL.
2476  *
2477  * The caller should call synchronize_sched() after this.
2478  */
2479 void ring_buffer_record_disable(struct ring_buffer *buffer)
2480 {
2481         atomic_inc(&buffer->record_disabled);
2482 }
2483 EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
2484
2485 /**
2486  * ring_buffer_record_enable - enable writes to the buffer
2487  * @buffer: The ring buffer to enable writes
2488  *
2489  * Note, multiple disables will need the same number of enables
2490  * to truely enable the writing (much like preempt_disable).
2491  */
2492 void ring_buffer_record_enable(struct ring_buffer *buffer)
2493 {
2494         atomic_dec(&buffer->record_disabled);
2495 }
2496 EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
2497
2498 /**
2499  * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
2500  * @buffer: The ring buffer to stop writes to.
2501  * @cpu: The CPU buffer to stop
2502  *
2503  * This prevents all writes to the buffer. Any attempt to write
2504  * to the buffer after this will fail and return NULL.
2505  *
2506  * The caller should call synchronize_sched() after this.
2507  */
2508 void ring_buffer_record_disable_cpu(struct ring_buffer *buffer, int cpu)
2509 {
2510         struct ring_buffer_per_cpu *cpu_buffer;
2511
2512         if (!cpumask_test_cpu(cpu, buffer->cpumask))
2513                 return;
2514
2515         cpu_buffer = buffer->buffers[cpu];
2516         atomic_inc(&cpu_buffer->record_disabled);
2517 }
2518 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
2519
2520 /**
2521  * ring_buffer_record_enable_cpu - enable writes to the buffer
2522  * @buffer: The ring buffer to enable writes
2523  * @cpu: The CPU to enable.
2524  *
2525  * Note, multiple disables will need the same number of enables
2526  * to truely enable the writing (much like preempt_disable).
2527  */
2528 void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu)
2529 {
2530         struct ring_buffer_per_cpu *cpu_buffer;
2531
2532         if (!cpumask_test_cpu(cpu, buffer->cpumask))
2533                 return;
2534
2535         cpu_buffer = buffer->buffers[cpu];
2536         atomic_dec(&cpu_buffer->record_disabled);
2537 }
2538 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
2539
2540 /**
2541  * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
2542  * @buffer: The ring buffer
2543  * @cpu: The per CPU buffer to get the entries from.
2544  */
2545 unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu)
2546 {
2547         struct ring_buffer_per_cpu *cpu_buffer;
2548         unsigned long ret;
2549
2550         if (!cpumask_test_cpu(cpu, buffer->cpumask))
2551                 return 0;
2552
2553         cpu_buffer = buffer->buffers[cpu];
2554         ret = (local_read(&cpu_buffer->entries) - local_read(&cpu_buffer->overrun))
2555                 - cpu_buffer->read;
2556
2557         return ret;
2558 }
2559 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
2560
2561 /**
2562  * ring_buffer_overrun_cpu - get the number of overruns in a cpu_buffer
2563  * @buffer: The ring buffer
2564  * @cpu: The per CPU buffer to get the number of overruns from
2565  */
2566 unsigned long ring_buffer_overrun_cpu(struct ring_buffer *buffer, int cpu)
2567 {
2568         struct ring_buffer_per_cpu *cpu_buffer;
2569         unsigned long ret;
2570
2571         if (!cpumask_test_cpu(cpu, buffer->cpumask))
2572                 return 0;
2573
2574         cpu_buffer = buffer->buffers[cpu];
2575         ret = local_read(&cpu_buffer->overrun);
2576
2577         return ret;
2578 }
2579 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
2580
2581 /**
2582  * ring_buffer_commit_overrun_cpu - get the number of overruns caused by commits
2583  * @buffer: The ring buffer
2584  * @cpu: The per CPU buffer to get the number of overruns from
2585  */
2586 unsigned long
2587 ring_buffer_commit_overrun_cpu(struct ring_buffer *buffer, int cpu)
2588 {
2589         struct ring_buffer_per_cpu *cpu_buffer;
2590         unsigned long ret;
2591
2592         if (!cpumask_test_cpu(cpu, buffer->cpumask))
2593                 return 0;
2594
2595         cpu_buffer = buffer->buffers[cpu];
2596         ret = local_read(&cpu_buffer->commit_overrun);
2597
2598         return ret;
2599 }
2600 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
2601
2602 /**
2603  * ring_buffer_entries - get the number of entries in a buffer
2604  * @buffer: The ring buffer
2605  *
2606  * Returns the total number of entries in the ring buffer
2607  * (all CPU entries)
2608  */
2609 unsigned long ring_buffer_entries(struct ring_buffer *buffer)
2610 {
2611         struct ring_buffer_per_cpu *cpu_buffer;
2612         unsigned long entries = 0;
2613         int cpu;
2614
2615         /* if you care about this being correct, lock the buffer */
2616         for_each_buffer_cpu(buffer, cpu) {
2617                 cpu_buffer = buffer->buffers[cpu];
2618                 entries += (local_read(&cpu_buffer->entries) -
2619                             local_read(&cpu_buffer->overrun)) - cpu_buffer->read;
2620         }
2621
2622         return entries;
2623 }
2624 EXPORT_SYMBOL_GPL(ring_buffer_entries);
2625
2626 /**
2627  * ring_buffer_overrun_cpu - get the number of overruns in buffer
2628  * @buffer: The ring buffer
2629  *
2630  * Returns the total number of overruns in the ring buffer
2631  * (all CPU entries)
2632  */
2633 unsigned long ring_buffer_overruns(struct ring_buffer *buffer)
2634 {
2635         struct ring_buffer_per_cpu *cpu_buffer;
2636         unsigned long overruns = 0;
2637         int cpu;
2638
2639         /* if you care about this being correct, lock the buffer */
2640         for_each_buffer_cpu(buffer, cpu) {
2641                 cpu_buffer = buffer->buffers[cpu];
2642                 overruns += local_read(&cpu_buffer->overrun);
2643         }
2644
2645         return overruns;
2646 }
2647 EXPORT_SYMBOL_GPL(ring_buffer_overruns);
2648
2649 static void rb_iter_reset(struct ring_buffer_iter *iter)
2650 {
2651         struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
2652
2653         /* Iterator usage is expected to have record disabled */
2654         if (list_empty(&cpu_buffer->reader_page->list)) {
2655                 iter->head_page = rb_set_head_page(cpu_buffer);
2656                 if (unlikely(!iter->head_page))
2657                         return;
2658                 iter->head = iter->head_page->read;
2659         } else {
2660                 iter->head_page = cpu_buffer->reader_page;
2661                 iter->head = cpu_buffer->reader_page->read;
2662         }
2663         if (iter->head)
2664                 iter->read_stamp = cpu_buffer->read_stamp;
2665         else
2666                 iter->read_stamp = iter->head_page->page->time_stamp;
2667 }
2668
2669 /**
2670  * ring_buffer_iter_reset - reset an iterator
2671  * @iter: The iterator to reset
2672  *
2673  * Resets the iterator, so that it will start from the beginning
2674  * again.
2675  */
2676 void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
2677 {
2678         struct ring_buffer_per_cpu *cpu_buffer;
2679         unsigned long flags;
2680
2681         if (!iter)
2682                 return;
2683
2684         cpu_buffer = iter->cpu_buffer;
2685
2686         spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
2687         rb_iter_reset(iter);
2688         spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
2689 }
2690 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
2691
2692 /**
2693  * ring_buffer_iter_empty - check if an iterator has no more to read
2694  * @iter: The iterator to check
2695  */
2696 int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
2697 {
2698         struct ring_buffer_per_cpu *cpu_buffer;
2699
2700         cpu_buffer = iter->cpu_buffer;
2701
2702         return iter->head_page == cpu_buffer->commit_page &&
2703                 iter->head == rb_commit_index(cpu_buffer);
2704 }
2705 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
2706
2707 static void
2708 rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2709                      struct ring_buffer_event *event)
2710 {
2711         u64 delta;
2712
2713         switch (event->type_len) {
2714         case RINGBUF_TYPE_PADDING:
2715                 return;
2716
2717         case RINGBUF_TYPE_TIME_EXTEND:
2718                 delta = event->array[0];
2719                 delta <<= TS_SHIFT;
2720                 delta += event->time_delta;
2721                 cpu_buffer->read_stamp += delta;
2722                 return;
2723
2724         case RINGBUF_TYPE_TIME_STAMP:
2725                 /* FIXME: not implemented */
2726                 return;
2727
2728         case RINGBUF_TYPE_DATA:
2729                 cpu_buffer->read_stamp += event->time_delta;
2730                 return;
2731
2732         default:
2733                 BUG();
2734         }
2735         return;
2736 }
2737
2738 static void
2739 rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
2740                           struct ring_buffer_event *event)
2741 {
2742         u64 delta;
2743
2744         switch (event->type_len) {
2745         case RINGBUF_TYPE_PADDING:
2746                 return;
2747
2748         case RINGBUF_TYPE_TIME_EXTEND:
2749                 delta = event->array[0];
2750                 delta <<= TS_SHIFT;
2751                 delta += event->time_delta;
2752                 iter->read_stamp += delta;
2753                 return;
2754
2755         case RINGBUF_TYPE_TIME_STAMP:
2756                 /* FIXME: not implemented */
2757                 return;
2758
2759         case RINGBUF_TYPE_DATA:
2760                 iter->read_stamp += event->time_delta;
2761                 return;
2762
2763         default:
2764                 BUG();
2765         }
2766         return;
2767 }
2768
2769 static struct buffer_page *
2770 rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
2771 {
2772         struct buffer_page *reader = NULL;
2773         unsigned long flags;
2774         int nr_loops = 0;
2775         int ret;
2776
2777         local_irq_save(flags);
2778         __raw_spin_lock(&cpu_buffer->lock);
2779
2780  again:
2781         /*
2782          * This should normally only loop twice. But because the
2783          * start of the reader inserts an empty page, it causes
2784          * a case where we will loop three times. There should be no
2785          * reason to loop four times (that I know of).
2786          */
2787         if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
2788                 reader = NULL;
2789                 goto out;
2790         }
2791
2792         reader = cpu_buffer->reader_page;
2793
2794         /* If there's more to read, return this page */
2795         if (cpu_buffer->reader_page->read < rb_page_size(reader))
2796                 goto out;
2797
2798         /* Never should we have an index greater than the size */
2799         if (RB_WARN_ON(cpu_buffer,
2800                        cpu_buffer->reader_page->read > rb_page_size(reader)))
2801                 goto out;
2802
2803         /* check if we caught up to the tail */
2804         reader = NULL;
2805         if (cpu_buffer->commit_page == cpu_buffer->reader_page)
2806                 goto out;
2807
2808         /*
2809          * Reset the reader page to size zero.
2810          */
2811         local_set(&cpu_buffer->reader_page->write, 0);
2812         local_set(&cpu_buffer->reader_page->entries, 0);
2813         local_set(&cpu_buffer->reader_page->page->commit, 0);
2814
2815  spin:
2816         /*
2817          * Splice the empty reader page into the list around the head.
2818          */
2819         reader = rb_set_head_page(cpu_buffer);
2820         cpu_buffer->reader_page->list.next = reader->list.next;
2821         cpu_buffer->reader_page->list.prev = reader->list.prev;
2822
2823         /*
2824          * cpu_buffer->pages just needs to point to the buffer, it
2825          *  has no specific buffer page to point to. Lets move it out
2826          *  of our way so we don't accidently swap it.
2827          */
2828         cpu_buffer->pages = reader->list.prev;
2829
2830         /* The reader page will be pointing to the new head */
2831         rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list);
2832
2833         /*
2834          * Here's the tricky part.
2835          *
2836          * We need to move the pointer past the header page.
2837          * But we can only do that if a writer is not currently
2838          * moving it. The page before the header page has the
2839          * flag bit '1' set if it is pointing to the page we want.
2840          * but if the writer is in the process of moving it
2841          * than it will be '2' or already moved '0'.
2842          */
2843
2844         ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
2845
2846         /*
2847          * If we did not convert it, then we must try again.
2848          */
2849         if (!ret)
2850                 goto spin;
2851
2852         /*
2853          * Yeah! We succeeded in replacing the page.
2854          *
2855          * Now make the new head point back to the reader page.
2856          */
2857         reader->list.next->prev = &cpu_buffer->reader_page->list;
2858         rb_inc_page(cpu_buffer, &cpu_buffer->head_page);
2859
2860         /* Finally update the reader page to the new head */
2861         cpu_buffer->reader_page = reader;
2862         rb_reset_reader_page(cpu_buffer);
2863
2864         goto again;
2865
2866  out:
2867         __raw_spin_unlock(&cpu_buffer->lock);
2868         local_irq_restore(flags);
2869
2870         return reader;
2871 }
2872
2873 static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
2874 {
2875         struct ring_buffer_event *event;
2876         struct buffer_page *reader;
2877         unsigned length;
2878
2879         reader = rb_get_reader_page(cpu_buffer);
2880
2881         /* This function should not be called when buffer is empty */
2882         if (RB_WARN_ON(cpu_buffer, !reader))
2883                 return;
2884
2885         event = rb_reader_event(cpu_buffer);
2886
2887         if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX
2888                         || rb_discarded_event(event))
2889                 cpu_buffer->read++;
2890
2891         rb_update_read_stamp(cpu_buffer, event);
2892
2893         length = rb_event_length(event);
2894         cpu_buffer->reader_page->read += length;
2895 }
2896
2897 static void rb_advance_iter(struct ring_buffer_iter *iter)
2898 {
2899         struct ring_buffer *buffer;
2900         struct ring_buffer_per_cpu *cpu_buffer;
2901         struct ring_buffer_event *event;
2902         unsigned length;
2903
2904         cpu_buffer = iter->cpu_buffer;
2905         buffer = cpu_buffer->buffer;
2906
2907         /*
2908          * Check if we are at the end of the buffer.
2909          */
2910         if (iter->head >= rb_page_size(iter->head_page)) {
2911                 /* discarded commits can make the page empty */
2912                 if (iter->head_page == cpu_buffer->commit_page)
2913                         return;
2914                 rb_inc_iter(iter);
2915                 return;
2916         }
2917
2918         event = rb_iter_head_event(iter);
2919
2920         length = rb_event_length(event);
2921
2922         /*
2923          * This should not be called to advance the header if we are
2924          * at the tail of the buffer.
2925          */
2926         if (RB_WARN_ON(cpu_buffer,
2927                        (iter->head_page == cpu_buffer->commit_page) &&
2928                        (iter->head + length > rb_commit_index(cpu_buffer))))
2929                 return;
2930
2931         rb_update_iter_read_stamp(iter, event);
2932
2933         iter->head += length;
2934
2935         /* check for end of page padding */
2936         if ((iter->head >= rb_page_size(iter->head_page)) &&
2937             (iter->head_page != cpu_buffer->commit_page))
2938                 rb_advance_iter(iter);
2939 }
2940
2941 static struct ring_buffer_event *
2942 rb_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts)
2943 {
2944         struct ring_buffer_per_cpu *cpu_buffer;
2945         struct ring_buffer_event *event;
2946         struct buffer_page *reader;
2947         int nr_loops = 0;
2948
2949         cpu_buffer = buffer->buffers[cpu];
2950
2951  again:
2952         /*
2953          * We repeat when a timestamp is encountered. It is possible
2954          * to get multiple timestamps from an interrupt entering just
2955          * as one timestamp is about to be written, or from discarded
2956          * commits. The most that we can have is the number on a single page.
2957          */
2958         if (RB_WARN_ON(cpu_buffer, ++nr_loops > RB_TIMESTAMPS_PER_PAGE))
2959                 return NULL;
2960
2961         reader = rb_get_reader_page(cpu_buffer);
2962         if (!reader)
2963                 return NULL;
2964
2965         event = rb_reader_event(cpu_buffer);
2966
2967         switch (event->type_len) {
2968         case RINGBUF_TYPE_PADDING:
2969                 if (rb_null_event(event))
2970                         RB_WARN_ON(cpu_buffer, 1);
2971                 /*
2972                  * Because the writer could be discarding every
2973                  * event it creates (which would probably be bad)
2974                  * if we were to go back to "again" then we may never
2975                  * catch up, and will trigger the warn on, or lock
2976                  * the box. Return the padding, and we will release
2977                  * the current locks, and try again.
2978                  */
2979                 return event;
2980
2981         case RINGBUF_TYPE_TIME_EXTEND:
2982                 /* Internal data, OK to advance */
2983                 rb_advance_reader(cpu_buffer);
2984                 goto again;
2985
2986         case RINGBUF_TYPE_TIME_STAMP:
2987                 /* FIXME: not implemented */
2988                 rb_advance_reader(cpu_buffer);
2989                 goto again;
2990
2991         case RINGBUF_TYPE_DATA:
2992                 if (ts) {
2993                         *ts = cpu_buffer->read_stamp + event->time_delta;
2994                         ring_buffer_normalize_time_stamp(buffer,
2995                                                          cpu_buffer->cpu, ts);
2996                 }
2997                 return event;
2998
2999         default:
3000                 BUG();
3001         }
3002
3003         return NULL;
3004 }
3005 EXPORT_SYMBOL_GPL(ring_buffer_peek);
3006
3007 static struct ring_buffer_event *
3008 rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3009 {
3010         struct ring_buffer *buffer;
3011         struct ring_buffer_per_cpu *cpu_buffer;
3012         struct ring_buffer_event *event;
3013         int nr_loops = 0;
3014
3015         if (ring_buffer_iter_empty(iter))
3016                 return NULL;
3017
3018         cpu_buffer = iter->cpu_buffer;
3019         buffer = cpu_buffer->buffer;
3020
3021  again:
3022         /*
3023          * We repeat when a timestamp is encountered.
3024          * We can get multiple timestamps by nested interrupts or also
3025          * if filtering is on (discarding commits). Since discarding
3026          * commits can be frequent we can get a lot of timestamps.
3027          * But we limit them by not adding timestamps if they begin
3028          * at the start of a page.
3029          */
3030         if (RB_WARN_ON(cpu_buffer, ++nr_loops > RB_TIMESTAMPS_PER_PAGE))
3031                 return NULL;
3032
3033         if (rb_per_cpu_empty(cpu_buffer))
3034                 return NULL;
3035
3036         event = rb_iter_head_event(iter);
3037
3038         switch (event->type_len) {
3039         case RINGBUF_TYPE_PADDING:
3040                 if (rb_null_event(event)) {
3041                         rb_inc_iter(iter);
3042                         goto again;
3043                 }
3044                 rb_advance_iter(iter);
3045                 return event;
3046
3047         case RINGBUF_TYPE_TIME_EXTEND:
3048                 /* Internal data, OK to advance */
3049                 rb_advance_iter(iter);
3050                 goto again;
3051
3052         case RINGBUF_TYPE_TIME_STAMP:
3053                 /* FIXME: not implemented */
3054                 rb_advance_iter(iter);
3055                 goto again;
3056
3057         case RINGBUF_TYPE_DATA:
3058                 if (ts) {
3059                         *ts = iter->read_stamp + event->time_delta;
3060                         ring_buffer_normalize_time_stamp(buffer,
3061                                                          cpu_buffer->cpu, ts);
3062                 }
3063                 return event;
3064
3065         default:
3066                 BUG();
3067         }
3068
3069         return NULL;
3070 }
3071 EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
3072
3073 static inline int rb_ok_to_lock(void)
3074 {
3075         /*
3076          * If an NMI die dumps out the content of the ring buffer
3077          * do not grab locks. We also permanently disable the ring
3078          * buffer too. A one time deal is all you get from reading
3079          * the ring buffer from an NMI.
3080          */
3081         if (likely(!in_nmi()))
3082                 return 1;
3083
3084         tracing_off_permanent();
3085         return 0;
3086 }
3087
3088 /**
3089  * ring_buffer_peek - peek at the next event to be read
3090  * @buffer: The ring buffer to read
3091  * @cpu: The cpu to peak at
3092  * @ts: The timestamp counter of this event.
3093  *
3094  * This will return the event that will be read next, but does
3095  * not consume the data.
3096  */
3097 struct ring_buffer_event *
3098 ring_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts)
3099 {
3100         struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3101         struct ring_buffer_event *event;
3102         unsigned long flags;
3103         int dolock;
3104
3105         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3106                 return NULL;
3107
3108         dolock = rb_ok_to_lock();
3109  again:
3110         local_irq_save(flags);
3111         if (dolock)
3112                 spin_lock(&cpu_buffer->reader_lock);
3113         event = rb_buffer_peek(buffer, cpu, ts);
3114         if (event && event->type_len == RINGBUF_TYPE_PADDING)
3115                 rb_advance_reader(cpu_buffer);
3116         if (dolock)
3117                 spin_unlock(&cpu_buffer->reader_lock);
3118         local_irq_restore(flags);
3119
3120         if (event && event->type_len == RINGBUF_TYPE_PADDING)
3121                 goto again;
3122
3123         return event;
3124 }
3125
3126 /**
3127  * ring_buffer_iter_peek - peek at the next event to be read
3128  * @iter: The ring buffer iterator
3129  * @ts: The timestamp counter of this event.
3130  *
3131  * This will return the event that will be read next, but does
3132  * not increment the iterator.
3133  */
3134 struct ring_buffer_event *
3135 ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3136 {
3137         struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3138         struct ring_buffer_event *event;
3139         unsigned long flags;
3140
3141  again:
3142         spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3143         event = rb_iter_peek(iter, ts);
3144         spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3145
3146         if (event && event->type_len == RINGBUF_TYPE_PADDING)
3147                 goto again;
3148
3149         return event;
3150 }
3151
3152 /**
3153  * ring_buffer_consume - return an event and consume it
3154  * @buffer: The ring buffer to get the next event from
3155  *
3156  * Returns the next event in the ring buffer, and that event is consumed.
3157  * Meaning, that sequential reads will keep returning a different event,
3158  * and eventually empty the ring buffer if the producer is slower.
3159  */
3160 struct ring_buffer_event *
3161 ring_buffer_consume(struct ring_buffer *buffer, int cpu, u64 *ts)
3162 {
3163         struct ring_buffer_per_cpu *cpu_buffer;
3164         struct ring_buffer_event *event = NULL;
3165         unsigned long flags;
3166         int dolock;
3167
3168         dolock = rb_ok_to_lock();
3169
3170  again:
3171         /* might be called in atomic */
3172         preempt_disable();
3173
3174         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3175                 goto out;
3176
3177         cpu_buffer = buffer->buffers[cpu];
3178         local_irq_save(flags);
3179         if (dolock)
3180                 spin_lock(&cpu_buffer->reader_lock);
3181
3182         event = rb_buffer_peek(buffer, cpu, ts);
3183         if (event)
3184                 rb_advance_reader(cpu_buffer);
3185
3186         if (dolock)
3187                 spin_unlock(&cpu_buffer->reader_lock);
3188         local_irq_restore(flags);
3189
3190  out:
3191         preempt_enable();
3192
3193         if (event && event->type_len == RINGBUF_TYPE_PADDING)
3194                 goto again;
3195
3196         return event;
3197 }
3198 EXPORT_SYMBOL_GPL(ring_buffer_consume);
3199
3200 /**
3201  * ring_buffer_read_start - start a non consuming read of the buffer
3202  * @buffer: The ring buffer to read from
3203  * @cpu: The cpu buffer to iterate over
3204  *
3205  * This starts up an iteration through the buffer. It also disables
3206  * the recording to the buffer until the reading is finished.
3207  * This prevents the reading from being corrupted. This is not
3208  * a consuming read, so a producer is not expected.
3209  *
3210  * Must be paired with ring_buffer_finish.
3211  */
3212 struct ring_buffer_iter *
3213 ring_buffer_read_start(struct ring_buffer *buffer, int cpu)
3214 {
3215         struct ring_buffer_per_cpu *cpu_buffer;
3216         struct ring_buffer_iter *iter;
3217         unsigned long flags;
3218
3219         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3220                 return NULL;
3221
3222         iter = kmalloc(sizeof(*iter), GFP_KERNEL);
3223         if (!iter)
3224                 return NULL;
3225
3226         cpu_buffer = buffer->buffers[cpu];
3227
3228         iter->cpu_buffer = cpu_buffer;
3229
3230         atomic_inc(&cpu_buffer->record_disabled);
3231         synchronize_sched();
3232
3233         spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3234         __raw_spin_lock(&cpu_buffer->lock);
3235         rb_iter_reset(iter);
3236         __raw_spin_unlock(&cpu_buffer->lock);
3237         spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3238
3239         return iter;
3240 }
3241 EXPORT_SYMBOL_GPL(ring_buffer_read_start);
3242
3243 /**
3244  * ring_buffer_finish - finish reading the iterator of the buffer
3245  * @iter: The iterator retrieved by ring_buffer_start
3246  *
3247  * This re-enables the recording to the buffer, and frees the
3248  * iterator.
3249  */
3250 void
3251 ring_buffer_read_finish(struct ring_buffer_iter *iter)
3252 {
3253         struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3254
3255         atomic_dec(&cpu_buffer->record_disabled);
3256         kfree(iter);
3257 }
3258 EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
3259
3260 /**
3261  * ring_buffer_read - read the next item in the ring buffer by the iterator
3262  * @iter: The ring buffer iterator
3263  * @ts: The time stamp of the event read.
3264  *
3265  * This reads the next event in the ring buffer and increments the iterator.
3266  */
3267 struct ring_buffer_event *
3268 ring_buffer_read(struct ring_buffer_iter *iter, u64 *ts)
3269 {
3270         struct ring_buffer_event *event;
3271         struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3272         unsigned long flags;
3273
3274         spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3275  again:
3276         event = rb_iter_peek(iter, ts);
3277         if (!event)
3278                 goto out;
3279
3280         if (event->type_len == RINGBUF_TYPE_PADDING)
3281                 goto again;
3282
3283         rb_advance_iter(iter);
3284  out:
3285         spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3286
3287         return event;
3288 }
3289 EXPORT_SYMBOL_GPL(ring_buffer_read);
3290
3291 /**
3292  * ring_buffer_size - return the size of the ring buffer (in bytes)
3293  * @buffer: The ring buffer.
3294  */
3295 unsigned long ring_buffer_size(struct ring_buffer *buffer)
3296 {
3297         return BUF_PAGE_SIZE * buffer->pages;
3298 }
3299 EXPORT_SYMBOL_GPL(ring_buffer_size);
3300
3301 static void
3302 rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
3303 {
3304         rb_head_page_deactivate(cpu_buffer);
3305
3306         cpu_buffer->head_page
3307                 = list_entry(cpu_buffer->pages, struct buffer_page, list);
3308         local_set(&cpu_buffer->head_page->write, 0);
3309         local_set(&cpu_buffer->head_page->entries, 0);
3310         local_set(&cpu_buffer->head_page->page->commit, 0);
3311
3312         cpu_buffer->head_page->read = 0;
3313
3314         cpu_buffer->tail_page = cpu_buffer->head_page;
3315         cpu_buffer->commit_page = cpu_buffer->head_page;
3316
3317         INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
3318         local_set(&cpu_buffer->reader_page->write, 0);
3319         local_set(&cpu_buffer->reader_page->entries, 0);
3320         local_set(&cpu_buffer->reader_page->page->commit, 0);
3321         cpu_buffer->reader_page->read = 0;
3322
3323         local_set(&cpu_buffer->commit_overrun, 0);
3324         local_set(&cpu_buffer->overrun, 0);
3325         local_set(&cpu_buffer->entries, 0);
3326         local_set(&cpu_buffer->committing, 0);
3327         local_set(&cpu_buffer->commits, 0);
3328         cpu_buffer->read = 0;
3329
3330         cpu_buffer->write_stamp = 0;
3331         cpu_buffer->read_stamp = 0;
3332
3333         rb_head_page_activate(cpu_buffer);
3334 }
3335
3336 /**
3337  * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
3338  * @buffer: The ring buffer to reset a per cpu buffer of
3339  * @cpu: The CPU buffer to be reset
3340  */
3341 void ring_buffer_reset_cpu(struct ring_buffer *buffer, int cpu)
3342 {
3343         struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3344         unsigned long flags;
3345
3346         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3347                 return;
3348
3349         atomic_inc(&cpu_buffer->record_disabled);
3350
3351         spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3352
3353         if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
3354                 goto out;
3355
3356         __raw_spin_lock(&cpu_buffer->lock);
3357
3358         rb_reset_cpu(cpu_buffer);
3359
3360         __raw_spin_unlock(&cpu_buffer->lock);
3361
3362  out:
3363         spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3364
3365         atomic_dec(&cpu_buffer->record_disabled);
3366 }
3367 EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
3368
3369 /**
3370  * ring_buffer_reset - reset a ring buffer
3371  * @buffer: The ring buffer to reset all cpu buffers
3372  */
3373 void ring_buffer_reset(struct ring_buffer *buffer)
3374 {
3375         int cpu;
3376
3377         for_each_buffer_cpu(buffer, cpu)
3378                 ring_buffer_reset_cpu(buffer, cpu);
3379 }
3380 EXPORT_SYMBOL_GPL(ring_buffer_reset);
3381
3382 /**
3383  * rind_buffer_empty - is the ring buffer empty?
3384  * @buffer: The ring buffer to test
3385  */
3386 int ring_buffer_empty(struct ring_buffer *buffer)
3387 {
3388         struct ring_buffer_per_cpu *cpu_buffer;
3389         unsigned long flags;
3390         int dolock;
3391         int cpu;
3392         int ret;
3393
3394         dolock = rb_ok_to_lock();
3395
3396         /* yes this is racy, but if you don't like the race, lock the buffer */
3397         for_each_buffer_cpu(buffer, cpu) {
3398                 cpu_buffer = buffer->buffers[cpu];
3399                 local_irq_save(flags);
3400                 if (dolock)
3401                         spin_lock(&cpu_buffer->reader_lock);
3402                 ret = rb_per_cpu_empty(cpu_buffer);
3403                 if (dolock)
3404                         spin_unlock(&cpu_buffer->reader_lock);
3405                 local_irq_restore(flags);
3406
3407                 if (!ret)
3408                         return 0;
3409         }
3410
3411         return 1;
3412 }
3413 EXPORT_SYMBOL_GPL(ring_buffer_empty);
3414
3415 /**
3416  * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
3417  * @buffer: The ring buffer
3418  * @cpu: The CPU buffer to test
3419  */
3420 int ring_buffer_empty_cpu(struct ring_buffer *buffer, int cpu)
3421 {
3422         struct ring_buffer_per_cpu *cpu_buffer;
3423         unsigned long flags;
3424         int dolock;
3425         int ret;
3426
3427         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3428                 return 1;
3429
3430         dolock = rb_ok_to_lock();
3431
3432         cpu_buffer = buffer->buffers[cpu];
3433         local_irq_save(flags);
3434         if (dolock)
3435                 spin_lock(&cpu_buffer->reader_lock);
3436         ret = rb_per_cpu_empty(cpu_buffer);
3437         if (dolock)
3438                 spin_unlock(&cpu_buffer->reader_lock);
3439         local_irq_restore(flags);
3440
3441         return ret;
3442 }
3443 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
3444
3445 /**
3446  * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
3447  * @buffer_a: One buffer to swap with
3448  * @buffer_b: The other buffer to swap with
3449  *
3450  * This function is useful for tracers that want to take a "snapshot"
3451  * of a CPU buffer and has another back up buffer lying around.
3452  * it is expected that the tracer handles the cpu buffer not being
3453  * used at the moment.
3454  */
3455 int ring_buffer_swap_cpu(struct ring_buffer *buffer_a,
3456                          struct ring_buffer *buffer_b, int cpu)
3457 {
3458         struct ring_buffer_per_cpu *cpu_buffer_a;
3459         struct ring_buffer_per_cpu *cpu_buffer_b;
3460         int ret = -EINVAL;
3461
3462         if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
3463             !cpumask_test_cpu(cpu, buffer_b->cpumask))
3464                 goto out;
3465
3466         /* At least make sure the two buffers are somewhat the same */
3467         if (buffer_a->pages != buffer_b->pages)
3468                 goto out;
3469
3470         ret = -EAGAIN;
3471
3472         if (ring_buffer_flags != RB_BUFFERS_ON)
3473                 goto out;
3474
3475         if (atomic_read(&buffer_a->record_disabled))
3476                 goto out;
3477
3478         if (atomic_read(&buffer_b->record_disabled))
3479                 goto out;
3480
3481         cpu_buffer_a = buffer_a->buffers[cpu];
3482         cpu_buffer_b = buffer_b->buffers[cpu];
3483
3484         if (atomic_read(&cpu_buffer_a->record_disabled))
3485                 goto out;
3486
3487         if (atomic_read(&cpu_buffer_b->record_disabled))
3488                 goto out;
3489
3490         /*
3491          * We can't do a synchronize_sched here because this
3492          * function can be called in atomic context.
3493          * Normally this will be called from the same CPU as cpu.
3494          * If not it's up to the caller to protect this.
3495          */
3496         atomic_inc(&cpu_buffer_a->record_disabled);
3497         atomic_inc(&cpu_buffer_b->record_disabled);
3498
3499         ret = -EBUSY;
3500         if (local_read(&cpu_buffer_a->committing))
3501                 goto out_dec;
3502         if (local_read(&cpu_buffer_b->committing))
3503                 goto out_dec;
3504
3505         buffer_a->buffers[cpu] = cpu_buffer_b;
3506         buffer_b->buffers[cpu] = cpu_buffer_a;
3507
3508         cpu_buffer_b->buffer = buffer_a;
3509         cpu_buffer_a->buffer = buffer_b;
3510
3511         ret = 0;
3512
3513 out_dec:
3514         atomic_dec(&cpu_buffer_a->record_disabled);
3515         atomic_dec(&cpu_buffer_b->record_disabled);
3516 out:
3517         return ret;
3518 }
3519 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
3520
3521 /**
3522  * ring_buffer_alloc_read_page - allocate a page to read from buffer
3523  * @buffer: the buffer to allocate for.
3524  *
3525  * This function is used in conjunction with ring_buffer_read_page.
3526  * When reading a full page from the ring buffer, these functions
3527  * can be used to speed up the process. The calling function should
3528  * allocate a few pages first with this function. Then when it
3529  * needs to get pages from the ring buffer, it passes the result
3530  * of this function into ring_buffer_read_page, which will swap
3531  * the page that was allocated, with the read page of the buffer.
3532  *
3533  * Returns:
3534  *  The page allocated, or NULL on error.
3535  */
3536 void *ring_buffer_alloc_read_page(struct ring_buffer *buffer)
3537 {
3538         struct buffer_data_page *bpage;
3539         unsigned long addr;
3540
3541         addr = __get_free_page(GFP_KERNEL);
3542         if (!addr)
3543                 return NULL;
3544
3545         bpage = (void *)addr;
3546
3547         rb_init_page(bpage);
3548
3549         return bpage;
3550 }
3551 EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
3552
3553 /**
3554  * ring_buffer_free_read_page - free an allocated read page
3555  * @buffer: the buffer the page was allocate for
3556  * @data: the page to free
3557  *
3558  * Free a page allocated from ring_buffer_alloc_read_page.
3559  */
3560 void ring_buffer_free_read_page(struct ring_buffer *buffer, void *data)
3561 {
3562         free_page((unsigned long)data);
3563 }
3564 EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
3565
3566 /**
3567  * ring_buffer_read_page - extract a page from the ring buffer
3568  * @buffer: buffer to extract from
3569  * @data_page: the page to use allocated from ring_buffer_alloc_read_page
3570  * @len: amount to extract
3571  * @cpu: the cpu of the buffer to extract
3572  * @full: should the extraction only happen when the page is full.
3573  *
3574  * This function will pull out a page from the ring buffer and consume it.
3575  * @data_page must be the address of the variable that was returned
3576  * from ring_buffer_alloc_read_page. This is because the page might be used
3577  * to swap with a page in the ring buffer.
3578  *
3579  * for example:
3580  *      rpage = ring_buffer_alloc_read_page(buffer);
3581  *      if (!rpage)
3582  *              return error;
3583  *      ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
3584  *      if (ret >= 0)
3585  *              process_page(rpage, ret);
3586  *
3587  * When @full is set, the function will not return true unless
3588  * the writer is off the reader page.
3589  *
3590  * Note: it is up to the calling functions to handle sleeps and wakeups.
3591  *  The ring buffer can be used anywhere in the kernel and can not
3592  *  blindly call wake_up. The layer that uses the ring buffer must be
3593  *  responsible for that.
3594  *
3595  * Returns:
3596  *  >=0 if data has been transferred, returns the offset of consumed data.
3597  *  <0 if no data has been transferred.
3598  */
3599 int ring_buffer_read_page(struct ring_buffer *buffer,
3600                           void **data_page, size_t len, int cpu, int full)
3601 {
3602         struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3603         struct ring_buffer_event *event;
3604         struct buffer_data_page *bpage;
3605         struct buffer_page *reader;
3606         unsigned long flags;
3607         unsigned int commit;
3608         unsigned int read;
3609         u64 save_timestamp;
3610         int ret = -1;
3611
3612         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3613                 goto out;
3614
3615         /*
3616          * If len is not big enough to hold the page header, then
3617          * we can not copy anything.
3618          */
3619         if (len <= BUF_PAGE_HDR_SIZE)
3620                 goto out;
3621
3622         len -= BUF_PAGE_HDR_SIZE;
3623
3624         if (!data_page)
3625                 goto out;
3626
3627         bpage = *data_page;
3628         if (!bpage)
3629                 goto out;
3630
3631         spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3632
3633         reader = rb_get_reader_page(cpu_buffer);
3634         if (!reader)
3635                 goto out_unlock;
3636
3637         event = rb_reader_event(cpu_buffer);
3638
3639         read = reader->read;
3640         commit = rb_page_commit(reader);
3641
3642         /*
3643          * If this page has been partially read or
3644          * if len is not big enough to read the rest of the page or
3645          * a writer is still on the page, then
3646          * we must copy the data from the page to the buffer.
3647          * Otherwise, we can simply swap the page with the one passed in.
3648          */
3649         if (read || (len < (commit - read)) ||
3650             cpu_buffer->reader_page == cpu_buffer->commit_page) {
3651                 struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
3652                 unsigned int rpos = read;
3653                 unsigned int pos = 0;
3654                 unsigned int size;
3655
3656                 if (full)
3657                         goto out_unlock;
3658
3659                 if (len > (commit - read))
3660                         len = (commit - read);
3661
3662                 size = rb_event_length(event);
3663
3664                 if (len < size)
3665                         goto out_unlock;
3666
3667                 /* save the current timestamp, since the user will need it */
3668                 save_timestamp = cpu_buffer->read_stamp;
3669
3670                 /* Need to copy one event at a time */
3671                 do {
3672                         memcpy(bpage->data + pos, rpage->data + rpos, size);
3673
3674                         len -= size;
3675
3676                         rb_advance_reader(cpu_buffer);
3677                         rpos = reader->read;
3678                         pos += size;
3679
3680                         event = rb_reader_event(cpu_buffer);
3681                         size = rb_event_length(event);
3682                 } while (len > size);
3683
3684                 /* update bpage */
3685                 local_set(&bpage->commit, pos);
3686                 bpage->time_stamp = save_timestamp;
3687
3688                 /* we copied everything to the beginning */
3689                 read = 0;
3690         } else {
3691                 /* update the entry counter */
3692                 cpu_buffer->read += rb_page_entries(reader);
3693
3694                 /* swap the pages */
3695                 rb_init_page(bpage);
3696                 bpage = reader->page;
3697                 reader->page = *data_page;
3698                 local_set(&reader->write, 0);
3699                 local_set(&reader->entries, 0);
3700                 reader->read = 0;
3701                 *data_page = bpage;
3702         }
3703         ret = read;
3704
3705  out_unlock:
3706         spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3707
3708  out:
3709         return ret;
3710 }
3711 EXPORT_SYMBOL_GPL(ring_buffer_read_page);
3712
3713 #ifdef CONFIG_TRACING
3714 static ssize_t
3715 rb_simple_read(struct file *filp, char __user *ubuf,
3716                size_t cnt, loff_t *ppos)
3717 {
3718         unsigned long *p = filp->private_data;
3719         char buf[64];
3720         int r;
3721
3722         if (test_bit(RB_BUFFERS_DISABLED_BIT, p))
3723                 r = sprintf(buf, "permanently disabled\n");
3724         else
3725                 r = sprintf(buf, "%d\n", test_bit(RB_BUFFERS_ON_BIT, p));
3726
3727         return simple_read_from_buffer(ubuf, cnt, ppos, buf, r);
3728 }
3729
3730 static ssize_t
3731 rb_simple_write(struct file *filp, const char __user *ubuf,
3732                 size_t cnt, loff_t *ppos)
3733 {
3734         unsigned long *p = filp->private_data;
3735         char buf[64];
3736         unsigned long val;
3737         int ret;
3738
3739         if (cnt >= sizeof(buf))
3740                 return -EINVAL;
3741
3742         if (copy_from_user(&buf, ubuf, cnt))
3743                 return -EFAULT;
3744
3745         buf[cnt] = 0;
3746
3747         ret = strict_strtoul(buf, 10, &val);
3748         if (ret < 0)
3749                 return ret;
3750
3751         if (val)
3752                 set_bit(RB_BUFFERS_ON_BIT, p);
3753         else
3754                 clear_bit(RB_BUFFERS_ON_BIT, p);
3755
3756         (*ppos)++;
3757
3758         return cnt;
3759 }
3760
3761 static const struct file_operations rb_simple_fops = {
3762         .open           = tracing_open_generic,
3763         .read           = rb_simple_read,
3764         .write          = rb_simple_write,
3765 };
3766
3767
3768 static __init int rb_init_debugfs(void)
3769 {
3770         struct dentry *d_tracer;
3771
3772         d_tracer = tracing_init_dentry();
3773
3774         trace_create_file("tracing_on", 0644, d_tracer,
3775                             &ring_buffer_flags, &rb_simple_fops);
3776
3777         return 0;
3778 }
3779
3780 fs_initcall(rb_init_debugfs);
3781 #endif
3782
3783 #ifdef CONFIG_HOTPLUG_CPU
3784 static int rb_cpu_notify(struct notifier_block *self,
3785                          unsigned long action, void *hcpu)
3786 {
3787         struct ring_buffer *buffer =
3788                 container_of(self, struct ring_buffer, cpu_notify);
3789         long cpu = (long)hcpu;
3790
3791         switch (action) {
3792         case CPU_UP_PREPARE:
3793         case CPU_UP_PREPARE_FROZEN:
3794                 if (cpumask_test_cpu(cpu, buffer->cpumask))
3795                         return NOTIFY_OK;
3796
3797                 buffer->buffers[cpu] =
3798                         rb_allocate_cpu_buffer(buffer, cpu);
3799                 if (!buffer->buffers[cpu]) {
3800                         WARN(1, "failed to allocate ring buffer on CPU %ld\n",
3801                              cpu);
3802                         return NOTIFY_OK;
3803                 }
3804                 smp_wmb();
3805                 cpumask_set_cpu(cpu, buffer->cpumask);
3806                 break;
3807         case CPU_DOWN_PREPARE:
3808         case CPU_DOWN_PREPARE_FROZEN:
3809                 /*
3810                  * Do nothing.
3811                  *  If we were to free the buffer, then the user would
3812                  *  lose any trace that was in the buffer.
3813                  */
3814                 break;
3815         default:
3816                 break;
3817         }
3818         return NOTIFY_OK;
3819 }
3820 #endif