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