ring-buffer: Do no reuse reader page if still in use
[muen/linux.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/trace_events.h>
7 #include <linux/ring_buffer.h>
8 #include <linux/trace_clock.h>
9 #include <linux/sched/clock.h>
10 #include <linux/trace_seq.h>
11 #include <linux/spinlock.h>
12 #include <linux/irq_work.h>
13 #include <linux/uaccess.h>
14 #include <linux/hardirq.h>
15 #include <linux/kthread.h>      /* for self test */
16 #include <linux/module.h>
17 #include <linux/percpu.h>
18 #include <linux/mutex.h>
19 #include <linux/delay.h>
20 #include <linux/slab.h>
21 #include <linux/init.h>
22 #include <linux/hash.h>
23 #include <linux/list.h>
24 #include <linux/cpu.h>
25
26 #include <asm/local.h>
27
28 static void update_pages_handler(struct work_struct *work);
29
30 /*
31  * The ring buffer header is special. We must manually up keep it.
32  */
33 int ring_buffer_print_entry_header(struct trace_seq *s)
34 {
35         trace_seq_puts(s, "# compressed entry header\n");
36         trace_seq_puts(s, "\ttype_len    :    5 bits\n");
37         trace_seq_puts(s, "\ttime_delta  :   27 bits\n");
38         trace_seq_puts(s, "\tarray       :   32 bits\n");
39         trace_seq_putc(s, '\n');
40         trace_seq_printf(s, "\tpadding     : type == %d\n",
41                          RINGBUF_TYPE_PADDING);
42         trace_seq_printf(s, "\ttime_extend : type == %d\n",
43                          RINGBUF_TYPE_TIME_EXTEND);
44         trace_seq_printf(s, "\tdata max type_len  == %d\n",
45                          RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
46
47         return !trace_seq_has_overflowed(s);
48 }
49
50 /*
51  * The ring buffer is made up of a list of pages. A separate list of pages is
52  * allocated for each CPU. A writer may only write to a buffer that is
53  * associated with the CPU it is currently executing on.  A reader may read
54  * from any per cpu buffer.
55  *
56  * The reader is special. For each per cpu buffer, the reader has its own
57  * reader page. When a reader has read the entire reader page, this reader
58  * page is swapped with another page in the ring buffer.
59  *
60  * Now, as long as the writer is off the reader page, the reader can do what
61  * ever it wants with that page. The writer will never write to that page
62  * again (as long as it is out of the ring buffer).
63  *
64  * Here's some silly ASCII art.
65  *
66  *   +------+
67  *   |reader|          RING BUFFER
68  *   |page  |
69  *   +------+        +---+   +---+   +---+
70  *                   |   |-->|   |-->|   |
71  *                   +---+   +---+   +---+
72  *                     ^               |
73  *                     |               |
74  *                     +---------------+
75  *
76  *
77  *   +------+
78  *   |reader|          RING BUFFER
79  *   |page  |------------------v
80  *   +------+        +---+   +---+   +---+
81  *                   |   |-->|   |-->|   |
82  *                   +---+   +---+   +---+
83  *                     ^               |
84  *                     |               |
85  *                     +---------------+
86  *
87  *
88  *   +------+
89  *   |reader|          RING BUFFER
90  *   |page  |------------------v
91  *   +------+        +---+   +---+   +---+
92  *      ^            |   |-->|   |-->|   |
93  *      |            +---+   +---+   +---+
94  *      |                              |
95  *      |                              |
96  *      +------------------------------+
97  *
98  *
99  *   +------+
100  *   |buffer|          RING BUFFER
101  *   |page  |------------------v
102  *   +------+        +---+   +---+   +---+
103  *      ^            |   |   |   |-->|   |
104  *      |   New      +---+   +---+   +---+
105  *      |  Reader------^               |
106  *      |   page                       |
107  *      +------------------------------+
108  *
109  *
110  * After we make this swap, the reader can hand this page off to the splice
111  * code and be done with it. It can even allocate a new page if it needs to
112  * and swap that into the ring buffer.
113  *
114  * We will be using cmpxchg soon to make all this lockless.
115  *
116  */
117
118 /* Used for individual buffers (after the counter) */
119 #define RB_BUFFER_OFF           (1 << 20)
120
121 #define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
122
123 #define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
124 #define RB_ALIGNMENT            4U
125 #define RB_MAX_SMALL_DATA       (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
126 #define RB_EVNT_MIN_SIZE        8U      /* two 32bit words */
127
128 #ifndef CONFIG_HAVE_64BIT_ALIGNED_ACCESS
129 # define RB_FORCE_8BYTE_ALIGNMENT       0
130 # define RB_ARCH_ALIGNMENT              RB_ALIGNMENT
131 #else
132 # define RB_FORCE_8BYTE_ALIGNMENT       1
133 # define RB_ARCH_ALIGNMENT              8U
134 #endif
135
136 #define RB_ALIGN_DATA           __aligned(RB_ARCH_ALIGNMENT)
137
138 /* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
139 #define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
140
141 enum {
142         RB_LEN_TIME_EXTEND = 8,
143         RB_LEN_TIME_STAMP = 16,
144 };
145
146 #define skip_time_extend(event) \
147         ((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
148
149 static inline int rb_null_event(struct ring_buffer_event *event)
150 {
151         return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
152 }
153
154 static void rb_event_set_padding(struct ring_buffer_event *event)
155 {
156         /* padding has a NULL time_delta */
157         event->type_len = RINGBUF_TYPE_PADDING;
158         event->time_delta = 0;
159 }
160
161 static unsigned
162 rb_event_data_length(struct ring_buffer_event *event)
163 {
164         unsigned length;
165
166         if (event->type_len)
167                 length = event->type_len * RB_ALIGNMENT;
168         else
169                 length = event->array[0];
170         return length + RB_EVNT_HDR_SIZE;
171 }
172
173 /*
174  * Return the length of the given event. Will return
175  * the length of the time extend if the event is a
176  * time extend.
177  */
178 static inline unsigned
179 rb_event_length(struct ring_buffer_event *event)
180 {
181         switch (event->type_len) {
182         case RINGBUF_TYPE_PADDING:
183                 if (rb_null_event(event))
184                         /* undefined */
185                         return -1;
186                 return  event->array[0] + RB_EVNT_HDR_SIZE;
187
188         case RINGBUF_TYPE_TIME_EXTEND:
189                 return RB_LEN_TIME_EXTEND;
190
191         case RINGBUF_TYPE_TIME_STAMP:
192                 return RB_LEN_TIME_STAMP;
193
194         case RINGBUF_TYPE_DATA:
195                 return rb_event_data_length(event);
196         default:
197                 BUG();
198         }
199         /* not hit */
200         return 0;
201 }
202
203 /*
204  * Return total length of time extend and data,
205  *   or just the event length for all other events.
206  */
207 static inline unsigned
208 rb_event_ts_length(struct ring_buffer_event *event)
209 {
210         unsigned len = 0;
211
212         if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
213                 /* time extends include the data event after it */
214                 len = RB_LEN_TIME_EXTEND;
215                 event = skip_time_extend(event);
216         }
217         return len + rb_event_length(event);
218 }
219
220 /**
221  * ring_buffer_event_length - return the length of the event
222  * @event: the event to get the length of
223  *
224  * Returns the size of the data load of a data event.
225  * If the event is something other than a data event, it
226  * returns the size of the event itself. With the exception
227  * of a TIME EXTEND, where it still returns the size of the
228  * data load of the data event after it.
229  */
230 unsigned ring_buffer_event_length(struct ring_buffer_event *event)
231 {
232         unsigned length;
233
234         if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
235                 event = skip_time_extend(event);
236
237         length = rb_event_length(event);
238         if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
239                 return length;
240         length -= RB_EVNT_HDR_SIZE;
241         if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
242                 length -= sizeof(event->array[0]);
243         return length;
244 }
245 EXPORT_SYMBOL_GPL(ring_buffer_event_length);
246
247 /* inline for ring buffer fast paths */
248 static __always_inline void *
249 rb_event_data(struct ring_buffer_event *event)
250 {
251         if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
252                 event = skip_time_extend(event);
253         BUG_ON(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
254         /* If length is in len field, then array[0] has the data */
255         if (event->type_len)
256                 return (void *)&event->array[0];
257         /* Otherwise length is in array[0] and array[1] has the data */
258         return (void *)&event->array[1];
259 }
260
261 /**
262  * ring_buffer_event_data - return the data of the event
263  * @event: the event to get the data from
264  */
265 void *ring_buffer_event_data(struct ring_buffer_event *event)
266 {
267         return rb_event_data(event);
268 }
269 EXPORT_SYMBOL_GPL(ring_buffer_event_data);
270
271 #define for_each_buffer_cpu(buffer, cpu)                \
272         for_each_cpu(cpu, buffer->cpumask)
273
274 #define TS_SHIFT        27
275 #define TS_MASK         ((1ULL << TS_SHIFT) - 1)
276 #define TS_DELTA_TEST   (~TS_MASK)
277
278 /* Flag when events were overwritten */
279 #define RB_MISSED_EVENTS        (1 << 31)
280 /* Missed count stored at end */
281 #define RB_MISSED_STORED        (1 << 30)
282
283 #define RB_MISSED_FLAGS         (RB_MISSED_EVENTS|RB_MISSED_STORED)
284
285 struct buffer_data_page {
286         u64              time_stamp;    /* page time stamp */
287         local_t          commit;        /* write committed index */
288         unsigned char    data[] RB_ALIGN_DATA;  /* data of buffer page */
289 };
290
291 /*
292  * Note, the buffer_page list must be first. The buffer pages
293  * are allocated in cache lines, which means that each buffer
294  * page will be at the beginning of a cache line, and thus
295  * the least significant bits will be zero. We use this to
296  * add flags in the list struct pointers, to make the ring buffer
297  * lockless.
298  */
299 struct buffer_page {
300         struct list_head list;          /* list of buffer pages */
301         local_t          write;         /* index for next write */
302         unsigned         read;          /* index for next read */
303         local_t          entries;       /* entries on this page */
304         unsigned long    real_end;      /* real end of data */
305         struct buffer_data_page *page;  /* Actual data page */
306 };
307
308 /*
309  * The buffer page counters, write and entries, must be reset
310  * atomically when crossing page boundaries. To synchronize this
311  * update, two counters are inserted into the number. One is
312  * the actual counter for the write position or count on the page.
313  *
314  * The other is a counter of updaters. Before an update happens
315  * the update partition of the counter is incremented. This will
316  * allow the updater to update the counter atomically.
317  *
318  * The counter is 20 bits, and the state data is 12.
319  */
320 #define RB_WRITE_MASK           0xfffff
321 #define RB_WRITE_INTCNT         (1 << 20)
322
323 static void rb_init_page(struct buffer_data_page *bpage)
324 {
325         local_set(&bpage->commit, 0);
326 }
327
328 /**
329  * ring_buffer_page_len - the size of data on the page.
330  * @page: The page to read
331  *
332  * Returns the amount of data on the page, including buffer page header.
333  */
334 size_t ring_buffer_page_len(void *page)
335 {
336         struct buffer_data_page *bpage = page;
337
338         return (local_read(&bpage->commit) & ~RB_MISSED_FLAGS)
339                 + BUF_PAGE_HDR_SIZE;
340 }
341
342 /*
343  * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
344  * this issue out.
345  */
346 static void free_buffer_page(struct buffer_page *bpage)
347 {
348         free_page((unsigned long)bpage->page);
349         kfree(bpage);
350 }
351
352 /*
353  * We need to fit the time_stamp delta into 27 bits.
354  */
355 static inline int test_time_stamp(u64 delta)
356 {
357         if (delta & TS_DELTA_TEST)
358                 return 1;
359         return 0;
360 }
361
362 #define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
363
364 /* Max payload is BUF_PAGE_SIZE - header (8bytes) */
365 #define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
366
367 int ring_buffer_print_page_header(struct trace_seq *s)
368 {
369         struct buffer_data_page field;
370
371         trace_seq_printf(s, "\tfield: u64 timestamp;\t"
372                          "offset:0;\tsize:%u;\tsigned:%u;\n",
373                          (unsigned int)sizeof(field.time_stamp),
374                          (unsigned int)is_signed_type(u64));
375
376         trace_seq_printf(s, "\tfield: local_t commit;\t"
377                          "offset:%u;\tsize:%u;\tsigned:%u;\n",
378                          (unsigned int)offsetof(typeof(field), commit),
379                          (unsigned int)sizeof(field.commit),
380                          (unsigned int)is_signed_type(long));
381
382         trace_seq_printf(s, "\tfield: int overwrite;\t"
383                          "offset:%u;\tsize:%u;\tsigned:%u;\n",
384                          (unsigned int)offsetof(typeof(field), commit),
385                          1,
386                          (unsigned int)is_signed_type(long));
387
388         trace_seq_printf(s, "\tfield: char data;\t"
389                          "offset:%u;\tsize:%u;\tsigned:%u;\n",
390                          (unsigned int)offsetof(typeof(field), data),
391                          (unsigned int)BUF_PAGE_SIZE,
392                          (unsigned int)is_signed_type(char));
393
394         return !trace_seq_has_overflowed(s);
395 }
396
397 struct rb_irq_work {
398         struct irq_work                 work;
399         wait_queue_head_t               waiters;
400         wait_queue_head_t               full_waiters;
401         bool                            waiters_pending;
402         bool                            full_waiters_pending;
403         bool                            wakeup_full;
404 };
405
406 /*
407  * Structure to hold event state and handle nested events.
408  */
409 struct rb_event_info {
410         u64                     ts;
411         u64                     delta;
412         unsigned long           length;
413         struct buffer_page      *tail_page;
414         int                     add_timestamp;
415 };
416
417 /*
418  * Used for which event context the event is in.
419  *  NMI     = 0
420  *  IRQ     = 1
421  *  SOFTIRQ = 2
422  *  NORMAL  = 3
423  *
424  * See trace_recursive_lock() comment below for more details.
425  */
426 enum {
427         RB_CTX_NMI,
428         RB_CTX_IRQ,
429         RB_CTX_SOFTIRQ,
430         RB_CTX_NORMAL,
431         RB_CTX_MAX
432 };
433
434 /*
435  * head_page == tail_page && head == tail then buffer is empty.
436  */
437 struct ring_buffer_per_cpu {
438         int                             cpu;
439         atomic_t                        record_disabled;
440         struct ring_buffer              *buffer;
441         raw_spinlock_t                  reader_lock;    /* serialize readers */
442         arch_spinlock_t                 lock;
443         struct lock_class_key           lock_key;
444         struct buffer_data_page         *free_page;
445         unsigned long                   nr_pages;
446         unsigned int                    current_context;
447         struct list_head                *pages;
448         struct buffer_page              *head_page;     /* read from head */
449         struct buffer_page              *tail_page;     /* write to tail */
450         struct buffer_page              *commit_page;   /* committed pages */
451         struct buffer_page              *reader_page;
452         unsigned long                   lost_events;
453         unsigned long                   last_overrun;
454         local_t                         entries_bytes;
455         local_t                         entries;
456         local_t                         overrun;
457         local_t                         commit_overrun;
458         local_t                         dropped_events;
459         local_t                         committing;
460         local_t                         commits;
461         unsigned long                   read;
462         unsigned long                   read_bytes;
463         u64                             write_stamp;
464         u64                             read_stamp;
465         /* ring buffer pages to update, > 0 to add, < 0 to remove */
466         long                            nr_pages_to_update;
467         struct list_head                new_pages; /* new pages to add */
468         struct work_struct              update_pages_work;
469         struct completion               update_done;
470
471         struct rb_irq_work              irq_work;
472 };
473
474 struct ring_buffer {
475         unsigned                        flags;
476         int                             cpus;
477         atomic_t                        record_disabled;
478         atomic_t                        resize_disabled;
479         cpumask_var_t                   cpumask;
480
481         struct lock_class_key           *reader_lock_key;
482
483         struct mutex                    mutex;
484
485         struct ring_buffer_per_cpu      **buffers;
486
487         struct hlist_node               node;
488         u64                             (*clock)(void);
489
490         struct rb_irq_work              irq_work;
491 };
492
493 struct ring_buffer_iter {
494         struct ring_buffer_per_cpu      *cpu_buffer;
495         unsigned long                   head;
496         struct buffer_page              *head_page;
497         struct buffer_page              *cache_reader_page;
498         unsigned long                   cache_read;
499         u64                             read_stamp;
500 };
501
502 /*
503  * rb_wake_up_waiters - wake up tasks waiting for ring buffer input
504  *
505  * Schedules a delayed work to wake up any task that is blocked on the
506  * ring buffer waiters queue.
507  */
508 static void rb_wake_up_waiters(struct irq_work *work)
509 {
510         struct rb_irq_work *rbwork = container_of(work, struct rb_irq_work, work);
511
512         wake_up_all(&rbwork->waiters);
513         if (rbwork->wakeup_full) {
514                 rbwork->wakeup_full = false;
515                 wake_up_all(&rbwork->full_waiters);
516         }
517 }
518
519 /**
520  * ring_buffer_wait - wait for input to the ring buffer
521  * @buffer: buffer to wait on
522  * @cpu: the cpu buffer to wait on
523  * @full: wait until a full page is available, if @cpu != RING_BUFFER_ALL_CPUS
524  *
525  * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
526  * as data is added to any of the @buffer's cpu buffers. Otherwise
527  * it will wait for data to be added to a specific cpu buffer.
528  */
529 int ring_buffer_wait(struct ring_buffer *buffer, int cpu, bool full)
530 {
531         struct ring_buffer_per_cpu *uninitialized_var(cpu_buffer);
532         DEFINE_WAIT(wait);
533         struct rb_irq_work *work;
534         int ret = 0;
535
536         /*
537          * Depending on what the caller is waiting for, either any
538          * data in any cpu buffer, or a specific buffer, put the
539          * caller on the appropriate wait queue.
540          */
541         if (cpu == RING_BUFFER_ALL_CPUS) {
542                 work = &buffer->irq_work;
543                 /* Full only makes sense on per cpu reads */
544                 full = false;
545         } else {
546                 if (!cpumask_test_cpu(cpu, buffer->cpumask))
547                         return -ENODEV;
548                 cpu_buffer = buffer->buffers[cpu];
549                 work = &cpu_buffer->irq_work;
550         }
551
552
553         while (true) {
554                 if (full)
555                         prepare_to_wait(&work->full_waiters, &wait, TASK_INTERRUPTIBLE);
556                 else
557                         prepare_to_wait(&work->waiters, &wait, TASK_INTERRUPTIBLE);
558
559                 /*
560                  * The events can happen in critical sections where
561                  * checking a work queue can cause deadlocks.
562                  * After adding a task to the queue, this flag is set
563                  * only to notify events to try to wake up the queue
564                  * using irq_work.
565                  *
566                  * We don't clear it even if the buffer is no longer
567                  * empty. The flag only causes the next event to run
568                  * irq_work to do the work queue wake up. The worse
569                  * that can happen if we race with !trace_empty() is that
570                  * an event will cause an irq_work to try to wake up
571                  * an empty queue.
572                  *
573                  * There's no reason to protect this flag either, as
574                  * the work queue and irq_work logic will do the necessary
575                  * synchronization for the wake ups. The only thing
576                  * that is necessary is that the wake up happens after
577                  * a task has been queued. It's OK for spurious wake ups.
578                  */
579                 if (full)
580                         work->full_waiters_pending = true;
581                 else
582                         work->waiters_pending = true;
583
584                 if (signal_pending(current)) {
585                         ret = -EINTR;
586                         break;
587                 }
588
589                 if (cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer))
590                         break;
591
592                 if (cpu != RING_BUFFER_ALL_CPUS &&
593                     !ring_buffer_empty_cpu(buffer, cpu)) {
594                         unsigned long flags;
595                         bool pagebusy;
596
597                         if (!full)
598                                 break;
599
600                         raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
601                         pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
602                         raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
603
604                         if (!pagebusy)
605                                 break;
606                 }
607
608                 schedule();
609         }
610
611         if (full)
612                 finish_wait(&work->full_waiters, &wait);
613         else
614                 finish_wait(&work->waiters, &wait);
615
616         return ret;
617 }
618
619 /**
620  * ring_buffer_poll_wait - poll on buffer input
621  * @buffer: buffer to wait on
622  * @cpu: the cpu buffer to wait on
623  * @filp: the file descriptor
624  * @poll_table: The poll descriptor
625  *
626  * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
627  * as data is added to any of the @buffer's cpu buffers. Otherwise
628  * it will wait for data to be added to a specific cpu buffer.
629  *
630  * Returns POLLIN | POLLRDNORM if data exists in the buffers,
631  * zero otherwise.
632  */
633 int ring_buffer_poll_wait(struct ring_buffer *buffer, int cpu,
634                           struct file *filp, poll_table *poll_table)
635 {
636         struct ring_buffer_per_cpu *cpu_buffer;
637         struct rb_irq_work *work;
638
639         if (cpu == RING_BUFFER_ALL_CPUS)
640                 work = &buffer->irq_work;
641         else {
642                 if (!cpumask_test_cpu(cpu, buffer->cpumask))
643                         return -EINVAL;
644
645                 cpu_buffer = buffer->buffers[cpu];
646                 work = &cpu_buffer->irq_work;
647         }
648
649         poll_wait(filp, &work->waiters, poll_table);
650         work->waiters_pending = true;
651         /*
652          * There's a tight race between setting the waiters_pending and
653          * checking if the ring buffer is empty.  Once the waiters_pending bit
654          * is set, the next event will wake the task up, but we can get stuck
655          * if there's only a single event in.
656          *
657          * FIXME: Ideally, we need a memory barrier on the writer side as well,
658          * but adding a memory barrier to all events will cause too much of a
659          * performance hit in the fast path.  We only need a memory barrier when
660          * the buffer goes from empty to having content.  But as this race is
661          * extremely small, and it's not a problem if another event comes in, we
662          * will fix it later.
663          */
664         smp_mb();
665
666         if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
667             (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
668                 return POLLIN | POLLRDNORM;
669         return 0;
670 }
671
672 /* buffer may be either ring_buffer or ring_buffer_per_cpu */
673 #define RB_WARN_ON(b, cond)                                             \
674         ({                                                              \
675                 int _____ret = unlikely(cond);                          \
676                 if (_____ret) {                                         \
677                         if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
678                                 struct ring_buffer_per_cpu *__b =       \
679                                         (void *)b;                      \
680                                 atomic_inc(&__b->buffer->record_disabled); \
681                         } else                                          \
682                                 atomic_inc(&b->record_disabled);        \
683                         WARN_ON(1);                                     \
684                 }                                                       \
685                 _____ret;                                               \
686         })
687
688 /* Up this if you want to test the TIME_EXTENTS and normalization */
689 #define DEBUG_SHIFT 0
690
691 static inline u64 rb_time_stamp(struct ring_buffer *buffer)
692 {
693         /* shift to debug/test normalization and TIME_EXTENTS */
694         return buffer->clock() << DEBUG_SHIFT;
695 }
696
697 u64 ring_buffer_time_stamp(struct ring_buffer *buffer, int cpu)
698 {
699         u64 time;
700
701         preempt_disable_notrace();
702         time = rb_time_stamp(buffer);
703         preempt_enable_no_resched_notrace();
704
705         return time;
706 }
707 EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
708
709 void ring_buffer_normalize_time_stamp(struct ring_buffer *buffer,
710                                       int cpu, u64 *ts)
711 {
712         /* Just stupid testing the normalize function and deltas */
713         *ts >>= DEBUG_SHIFT;
714 }
715 EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
716
717 /*
718  * Making the ring buffer lockless makes things tricky.
719  * Although writes only happen on the CPU that they are on,
720  * and they only need to worry about interrupts. Reads can
721  * happen on any CPU.
722  *
723  * The reader page is always off the ring buffer, but when the
724  * reader finishes with a page, it needs to swap its page with
725  * a new one from the buffer. The reader needs to take from
726  * the head (writes go to the tail). But if a writer is in overwrite
727  * mode and wraps, it must push the head page forward.
728  *
729  * Here lies the problem.
730  *
731  * The reader must be careful to replace only the head page, and
732  * not another one. As described at the top of the file in the
733  * ASCII art, the reader sets its old page to point to the next
734  * page after head. It then sets the page after head to point to
735  * the old reader page. But if the writer moves the head page
736  * during this operation, the reader could end up with the tail.
737  *
738  * We use cmpxchg to help prevent this race. We also do something
739  * special with the page before head. We set the LSB to 1.
740  *
741  * When the writer must push the page forward, it will clear the
742  * bit that points to the head page, move the head, and then set
743  * the bit that points to the new head page.
744  *
745  * We also don't want an interrupt coming in and moving the head
746  * page on another writer. Thus we use the second LSB to catch
747  * that too. Thus:
748  *
749  * head->list->prev->next        bit 1          bit 0
750  *                              -------        -------
751  * Normal page                     0              0
752  * Points to head page             0              1
753  * New head page                   1              0
754  *
755  * Note we can not trust the prev pointer of the head page, because:
756  *
757  * +----+       +-----+        +-----+
758  * |    |------>|  T  |---X--->|  N  |
759  * |    |<------|     |        |     |
760  * +----+       +-----+        +-----+
761  *   ^                           ^ |
762  *   |          +-----+          | |
763  *   +----------|  R  |----------+ |
764  *              |     |<-----------+
765  *              +-----+
766  *
767  * Key:  ---X-->  HEAD flag set in pointer
768  *         T      Tail page
769  *         R      Reader page
770  *         N      Next page
771  *
772  * (see __rb_reserve_next() to see where this happens)
773  *
774  *  What the above shows is that the reader just swapped out
775  *  the reader page with a page in the buffer, but before it
776  *  could make the new header point back to the new page added
777  *  it was preempted by a writer. The writer moved forward onto
778  *  the new page added by the reader and is about to move forward
779  *  again.
780  *
781  *  You can see, it is legitimate for the previous pointer of
782  *  the head (or any page) not to point back to itself. But only
783  *  temporarially.
784  */
785
786 #define RB_PAGE_NORMAL          0UL
787 #define RB_PAGE_HEAD            1UL
788 #define RB_PAGE_UPDATE          2UL
789
790
791 #define RB_FLAG_MASK            3UL
792
793 /* PAGE_MOVED is not part of the mask */
794 #define RB_PAGE_MOVED           4UL
795
796 /*
797  * rb_list_head - remove any bit
798  */
799 static struct list_head *rb_list_head(struct list_head *list)
800 {
801         unsigned long val = (unsigned long)list;
802
803         return (struct list_head *)(val & ~RB_FLAG_MASK);
804 }
805
806 /*
807  * rb_is_head_page - test if the given page is the head page
808  *
809  * Because the reader may move the head_page pointer, we can
810  * not trust what the head page is (it may be pointing to
811  * the reader page). But if the next page is a header page,
812  * its flags will be non zero.
813  */
814 static inline int
815 rb_is_head_page(struct ring_buffer_per_cpu *cpu_buffer,
816                 struct buffer_page *page, struct list_head *list)
817 {
818         unsigned long val;
819
820         val = (unsigned long)list->next;
821
822         if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
823                 return RB_PAGE_MOVED;
824
825         return val & RB_FLAG_MASK;
826 }
827
828 /*
829  * rb_is_reader_page
830  *
831  * The unique thing about the reader page, is that, if the
832  * writer is ever on it, the previous pointer never points
833  * back to the reader page.
834  */
835 static bool rb_is_reader_page(struct buffer_page *page)
836 {
837         struct list_head *list = page->list.prev;
838
839         return rb_list_head(list->next) != &page->list;
840 }
841
842 /*
843  * rb_set_list_to_head - set a list_head to be pointing to head.
844  */
845 static void rb_set_list_to_head(struct ring_buffer_per_cpu *cpu_buffer,
846                                 struct list_head *list)
847 {
848         unsigned long *ptr;
849
850         ptr = (unsigned long *)&list->next;
851         *ptr |= RB_PAGE_HEAD;
852         *ptr &= ~RB_PAGE_UPDATE;
853 }
854
855 /*
856  * rb_head_page_activate - sets up head page
857  */
858 static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
859 {
860         struct buffer_page *head;
861
862         head = cpu_buffer->head_page;
863         if (!head)
864                 return;
865
866         /*
867          * Set the previous list pointer to have the HEAD flag.
868          */
869         rb_set_list_to_head(cpu_buffer, head->list.prev);
870 }
871
872 static void rb_list_head_clear(struct list_head *list)
873 {
874         unsigned long *ptr = (unsigned long *)&list->next;
875
876         *ptr &= ~RB_FLAG_MASK;
877 }
878
879 /*
880  * rb_head_page_dactivate - clears head page ptr (for free list)
881  */
882 static void
883 rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
884 {
885         struct list_head *hd;
886
887         /* Go through the whole list and clear any pointers found. */
888         rb_list_head_clear(cpu_buffer->pages);
889
890         list_for_each(hd, cpu_buffer->pages)
891                 rb_list_head_clear(hd);
892 }
893
894 static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
895                             struct buffer_page *head,
896                             struct buffer_page *prev,
897                             int old_flag, int new_flag)
898 {
899         struct list_head *list;
900         unsigned long val = (unsigned long)&head->list;
901         unsigned long ret;
902
903         list = &prev->list;
904
905         val &= ~RB_FLAG_MASK;
906
907         ret = cmpxchg((unsigned long *)&list->next,
908                       val | old_flag, val | new_flag);
909
910         /* check if the reader took the page */
911         if ((ret & ~RB_FLAG_MASK) != val)
912                 return RB_PAGE_MOVED;
913
914         return ret & RB_FLAG_MASK;
915 }
916
917 static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
918                                    struct buffer_page *head,
919                                    struct buffer_page *prev,
920                                    int old_flag)
921 {
922         return rb_head_page_set(cpu_buffer, head, prev,
923                                 old_flag, RB_PAGE_UPDATE);
924 }
925
926 static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
927                                  struct buffer_page *head,
928                                  struct buffer_page *prev,
929                                  int old_flag)
930 {
931         return rb_head_page_set(cpu_buffer, head, prev,
932                                 old_flag, RB_PAGE_HEAD);
933 }
934
935 static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
936                                    struct buffer_page *head,
937                                    struct buffer_page *prev,
938                                    int old_flag)
939 {
940         return rb_head_page_set(cpu_buffer, head, prev,
941                                 old_flag, RB_PAGE_NORMAL);
942 }
943
944 static inline void rb_inc_page(struct ring_buffer_per_cpu *cpu_buffer,
945                                struct buffer_page **bpage)
946 {
947         struct list_head *p = rb_list_head((*bpage)->list.next);
948
949         *bpage = list_entry(p, struct buffer_page, list);
950 }
951
952 static struct buffer_page *
953 rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
954 {
955         struct buffer_page *head;
956         struct buffer_page *page;
957         struct list_head *list;
958         int i;
959
960         if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
961                 return NULL;
962
963         /* sanity check */
964         list = cpu_buffer->pages;
965         if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
966                 return NULL;
967
968         page = head = cpu_buffer->head_page;
969         /*
970          * It is possible that the writer moves the header behind
971          * where we started, and we miss in one loop.
972          * A second loop should grab the header, but we'll do
973          * three loops just because I'm paranoid.
974          */
975         for (i = 0; i < 3; i++) {
976                 do {
977                         if (rb_is_head_page(cpu_buffer, page, page->list.prev)) {
978                                 cpu_buffer->head_page = page;
979                                 return page;
980                         }
981                         rb_inc_page(cpu_buffer, &page);
982                 } while (page != head);
983         }
984
985         RB_WARN_ON(cpu_buffer, 1);
986
987         return NULL;
988 }
989
990 static int rb_head_page_replace(struct buffer_page *old,
991                                 struct buffer_page *new)
992 {
993         unsigned long *ptr = (unsigned long *)&old->list.prev->next;
994         unsigned long val;
995         unsigned long ret;
996
997         val = *ptr & ~RB_FLAG_MASK;
998         val |= RB_PAGE_HEAD;
999
1000         ret = cmpxchg(ptr, val, (unsigned long)&new->list);
1001
1002         return ret == val;
1003 }
1004
1005 /*
1006  * rb_tail_page_update - move the tail page forward
1007  */
1008 static void rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
1009                                struct buffer_page *tail_page,
1010                                struct buffer_page *next_page)
1011 {
1012         unsigned long old_entries;
1013         unsigned long old_write;
1014
1015         /*
1016          * The tail page now needs to be moved forward.
1017          *
1018          * We need to reset the tail page, but without messing
1019          * with possible erasing of data brought in by interrupts
1020          * that have moved the tail page and are currently on it.
1021          *
1022          * We add a counter to the write field to denote this.
1023          */
1024         old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
1025         old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
1026
1027         /*
1028          * Just make sure we have seen our old_write and synchronize
1029          * with any interrupts that come in.
1030          */
1031         barrier();
1032
1033         /*
1034          * If the tail page is still the same as what we think
1035          * it is, then it is up to us to update the tail
1036          * pointer.
1037          */
1038         if (tail_page == READ_ONCE(cpu_buffer->tail_page)) {
1039                 /* Zero the write counter */
1040                 unsigned long val = old_write & ~RB_WRITE_MASK;
1041                 unsigned long eval = old_entries & ~RB_WRITE_MASK;
1042
1043                 /*
1044                  * This will only succeed if an interrupt did
1045                  * not come in and change it. In which case, we
1046                  * do not want to modify it.
1047                  *
1048                  * We add (void) to let the compiler know that we do not care
1049                  * about the return value of these functions. We use the
1050                  * cmpxchg to only update if an interrupt did not already
1051                  * do it for us. If the cmpxchg fails, we don't care.
1052                  */
1053                 (void)local_cmpxchg(&next_page->write, old_write, val);
1054                 (void)local_cmpxchg(&next_page->entries, old_entries, eval);
1055
1056                 /*
1057                  * No need to worry about races with clearing out the commit.
1058                  * it only can increment when a commit takes place. But that
1059                  * only happens in the outer most nested commit.
1060                  */
1061                 local_set(&next_page->page->commit, 0);
1062
1063                 /* Again, either we update tail_page or an interrupt does */
1064                 (void)cmpxchg(&cpu_buffer->tail_page, tail_page, next_page);
1065         }
1066 }
1067
1068 static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
1069                           struct buffer_page *bpage)
1070 {
1071         unsigned long val = (unsigned long)bpage;
1072
1073         if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK))
1074                 return 1;
1075
1076         return 0;
1077 }
1078
1079 /**
1080  * rb_check_list - make sure a pointer to a list has the last bits zero
1081  */
1082 static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer,
1083                          struct list_head *list)
1084 {
1085         if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev))
1086                 return 1;
1087         if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next))
1088                 return 1;
1089         return 0;
1090 }
1091
1092 /**
1093  * rb_check_pages - integrity check of buffer pages
1094  * @cpu_buffer: CPU buffer with pages to test
1095  *
1096  * As a safety measure we check to make sure the data pages have not
1097  * been corrupted.
1098  */
1099 static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
1100 {
1101         struct list_head *head = cpu_buffer->pages;
1102         struct buffer_page *bpage, *tmp;
1103
1104         /* Reset the head page if it exists */
1105         if (cpu_buffer->head_page)
1106                 rb_set_head_page(cpu_buffer);
1107
1108         rb_head_page_deactivate(cpu_buffer);
1109
1110         if (RB_WARN_ON(cpu_buffer, head->next->prev != head))
1111                 return -1;
1112         if (RB_WARN_ON(cpu_buffer, head->prev->next != head))
1113                 return -1;
1114
1115         if (rb_check_list(cpu_buffer, head))
1116                 return -1;
1117
1118         list_for_each_entry_safe(bpage, tmp, head, list) {
1119                 if (RB_WARN_ON(cpu_buffer,
1120                                bpage->list.next->prev != &bpage->list))
1121                         return -1;
1122                 if (RB_WARN_ON(cpu_buffer,
1123                                bpage->list.prev->next != &bpage->list))
1124                         return -1;
1125                 if (rb_check_list(cpu_buffer, &bpage->list))
1126                         return -1;
1127         }
1128
1129         rb_head_page_activate(cpu_buffer);
1130
1131         return 0;
1132 }
1133
1134 static int __rb_allocate_pages(long nr_pages, struct list_head *pages, int cpu)
1135 {
1136         struct buffer_page *bpage, *tmp;
1137         long i;
1138
1139         for (i = 0; i < nr_pages; i++) {
1140                 struct page *page;
1141                 /*
1142                  * __GFP_RETRY_MAYFAIL flag makes sure that the allocation fails
1143                  * gracefully without invoking oom-killer and the system is not
1144                  * destabilized.
1145                  */
1146                 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1147                                     GFP_KERNEL | __GFP_RETRY_MAYFAIL,
1148                                     cpu_to_node(cpu));
1149                 if (!bpage)
1150                         goto free_pages;
1151
1152                 list_add(&bpage->list, pages);
1153
1154                 page = alloc_pages_node(cpu_to_node(cpu),
1155                                         GFP_KERNEL | __GFP_RETRY_MAYFAIL, 0);
1156                 if (!page)
1157                         goto free_pages;
1158                 bpage->page = page_address(page);
1159                 rb_init_page(bpage->page);
1160         }
1161
1162         return 0;
1163
1164 free_pages:
1165         list_for_each_entry_safe(bpage, tmp, pages, list) {
1166                 list_del_init(&bpage->list);
1167                 free_buffer_page(bpage);
1168         }
1169
1170         return -ENOMEM;
1171 }
1172
1173 static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1174                              unsigned long nr_pages)
1175 {
1176         LIST_HEAD(pages);
1177
1178         WARN_ON(!nr_pages);
1179
1180         if (__rb_allocate_pages(nr_pages, &pages, cpu_buffer->cpu))
1181                 return -ENOMEM;
1182
1183         /*
1184          * The ring buffer page list is a circular list that does not
1185          * start and end with a list head. All page list items point to
1186          * other pages.
1187          */
1188         cpu_buffer->pages = pages.next;
1189         list_del(&pages);
1190
1191         cpu_buffer->nr_pages = nr_pages;
1192
1193         rb_check_pages(cpu_buffer);
1194
1195         return 0;
1196 }
1197
1198 static struct ring_buffer_per_cpu *
1199 rb_allocate_cpu_buffer(struct ring_buffer *buffer, long nr_pages, int cpu)
1200 {
1201         struct ring_buffer_per_cpu *cpu_buffer;
1202         struct buffer_page *bpage;
1203         struct page *page;
1204         int ret;
1205
1206         cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1207                                   GFP_KERNEL, cpu_to_node(cpu));
1208         if (!cpu_buffer)
1209                 return NULL;
1210
1211         cpu_buffer->cpu = cpu;
1212         cpu_buffer->buffer = buffer;
1213         raw_spin_lock_init(&cpu_buffer->reader_lock);
1214         lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1215         cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1216         INIT_WORK(&cpu_buffer->update_pages_work, update_pages_handler);
1217         init_completion(&cpu_buffer->update_done);
1218         init_irq_work(&cpu_buffer->irq_work.work, rb_wake_up_waiters);
1219         init_waitqueue_head(&cpu_buffer->irq_work.waiters);
1220         init_waitqueue_head(&cpu_buffer->irq_work.full_waiters);
1221
1222         bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1223                             GFP_KERNEL, cpu_to_node(cpu));
1224         if (!bpage)
1225                 goto fail_free_buffer;
1226
1227         rb_check_bpage(cpu_buffer, bpage);
1228
1229         cpu_buffer->reader_page = bpage;
1230         page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0);
1231         if (!page)
1232                 goto fail_free_reader;
1233         bpage->page = page_address(page);
1234         rb_init_page(bpage->page);
1235
1236         INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1237         INIT_LIST_HEAD(&cpu_buffer->new_pages);
1238
1239         ret = rb_allocate_pages(cpu_buffer, nr_pages);
1240         if (ret < 0)
1241                 goto fail_free_reader;
1242
1243         cpu_buffer->head_page
1244                 = list_entry(cpu_buffer->pages, struct buffer_page, list);
1245         cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1246
1247         rb_head_page_activate(cpu_buffer);
1248
1249         return cpu_buffer;
1250
1251  fail_free_reader:
1252         free_buffer_page(cpu_buffer->reader_page);
1253
1254  fail_free_buffer:
1255         kfree(cpu_buffer);
1256         return NULL;
1257 }
1258
1259 static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1260 {
1261         struct list_head *head = cpu_buffer->pages;
1262         struct buffer_page *bpage, *tmp;
1263
1264         free_buffer_page(cpu_buffer->reader_page);
1265
1266         rb_head_page_deactivate(cpu_buffer);
1267
1268         if (head) {
1269                 list_for_each_entry_safe(bpage, tmp, head, list) {
1270                         list_del_init(&bpage->list);
1271                         free_buffer_page(bpage);
1272                 }
1273                 bpage = list_entry(head, struct buffer_page, list);
1274                 free_buffer_page(bpage);
1275         }
1276
1277         kfree(cpu_buffer);
1278 }
1279
1280 /**
1281  * __ring_buffer_alloc - allocate a new ring_buffer
1282  * @size: the size in bytes per cpu that is needed.
1283  * @flags: attributes to set for the ring buffer.
1284  *
1285  * Currently the only flag that is available is the RB_FL_OVERWRITE
1286  * flag. This flag means that the buffer will overwrite old data
1287  * when the buffer wraps. If this flag is not set, the buffer will
1288  * drop data when the tail hits the head.
1289  */
1290 struct ring_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1291                                         struct lock_class_key *key)
1292 {
1293         struct ring_buffer *buffer;
1294         long nr_pages;
1295         int bsize;
1296         int cpu;
1297         int ret;
1298
1299         /* keep it in its own cache line */
1300         buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1301                          GFP_KERNEL);
1302         if (!buffer)
1303                 return NULL;
1304
1305         if (!zalloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1306                 goto fail_free_buffer;
1307
1308         nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1309         buffer->flags = flags;
1310         buffer->clock = trace_clock_local;
1311         buffer->reader_lock_key = key;
1312
1313         init_irq_work(&buffer->irq_work.work, rb_wake_up_waiters);
1314         init_waitqueue_head(&buffer->irq_work.waiters);
1315
1316         /* need at least two pages */
1317         if (nr_pages < 2)
1318                 nr_pages = 2;
1319
1320         buffer->cpus = nr_cpu_ids;
1321
1322         bsize = sizeof(void *) * nr_cpu_ids;
1323         buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1324                                   GFP_KERNEL);
1325         if (!buffer->buffers)
1326                 goto fail_free_cpumask;
1327
1328         cpu = raw_smp_processor_id();
1329         cpumask_set_cpu(cpu, buffer->cpumask);
1330         buffer->buffers[cpu] = rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
1331         if (!buffer->buffers[cpu])
1332                 goto fail_free_buffers;
1333
1334         ret = cpuhp_state_add_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1335         if (ret < 0)
1336                 goto fail_free_buffers;
1337
1338         mutex_init(&buffer->mutex);
1339
1340         return buffer;
1341
1342  fail_free_buffers:
1343         for_each_buffer_cpu(buffer, cpu) {
1344                 if (buffer->buffers[cpu])
1345                         rb_free_cpu_buffer(buffer->buffers[cpu]);
1346         }
1347         kfree(buffer->buffers);
1348
1349  fail_free_cpumask:
1350         free_cpumask_var(buffer->cpumask);
1351
1352  fail_free_buffer:
1353         kfree(buffer);
1354         return NULL;
1355 }
1356 EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1357
1358 /**
1359  * ring_buffer_free - free a ring buffer.
1360  * @buffer: the buffer to free.
1361  */
1362 void
1363 ring_buffer_free(struct ring_buffer *buffer)
1364 {
1365         int cpu;
1366
1367         cpuhp_state_remove_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1368
1369         for_each_buffer_cpu(buffer, cpu)
1370                 rb_free_cpu_buffer(buffer->buffers[cpu]);
1371
1372         kfree(buffer->buffers);
1373         free_cpumask_var(buffer->cpumask);
1374
1375         kfree(buffer);
1376 }
1377 EXPORT_SYMBOL_GPL(ring_buffer_free);
1378
1379 void ring_buffer_set_clock(struct ring_buffer *buffer,
1380                            u64 (*clock)(void))
1381 {
1382         buffer->clock = clock;
1383 }
1384
1385 static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1386
1387 static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1388 {
1389         return local_read(&bpage->entries) & RB_WRITE_MASK;
1390 }
1391
1392 static inline unsigned long rb_page_write(struct buffer_page *bpage)
1393 {
1394         return local_read(&bpage->write) & RB_WRITE_MASK;
1395 }
1396
1397 static int
1398 rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned long nr_pages)
1399 {
1400         struct list_head *tail_page, *to_remove, *next_page;
1401         struct buffer_page *to_remove_page, *tmp_iter_page;
1402         struct buffer_page *last_page, *first_page;
1403         unsigned long nr_removed;
1404         unsigned long head_bit;
1405         int page_entries;
1406
1407         head_bit = 0;
1408
1409         raw_spin_lock_irq(&cpu_buffer->reader_lock);
1410         atomic_inc(&cpu_buffer->record_disabled);
1411         /*
1412          * We don't race with the readers since we have acquired the reader
1413          * lock. We also don't race with writers after disabling recording.
1414          * This makes it easy to figure out the first and the last page to be
1415          * removed from the list. We unlink all the pages in between including
1416          * the first and last pages. This is done in a busy loop so that we
1417          * lose the least number of traces.
1418          * The pages are freed after we restart recording and unlock readers.
1419          */
1420         tail_page = &cpu_buffer->tail_page->list;
1421
1422         /*
1423          * tail page might be on reader page, we remove the next page
1424          * from the ring buffer
1425          */
1426         if (cpu_buffer->tail_page == cpu_buffer->reader_page)
1427                 tail_page = rb_list_head(tail_page->next);
1428         to_remove = tail_page;
1429
1430         /* start of pages to remove */
1431         first_page = list_entry(rb_list_head(to_remove->next),
1432                                 struct buffer_page, list);
1433
1434         for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
1435                 to_remove = rb_list_head(to_remove)->next;
1436                 head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
1437         }
1438
1439         next_page = rb_list_head(to_remove)->next;
1440
1441         /*
1442          * Now we remove all pages between tail_page and next_page.
1443          * Make sure that we have head_bit value preserved for the
1444          * next page
1445          */
1446         tail_page->next = (struct list_head *)((unsigned long)next_page |
1447                                                 head_bit);
1448         next_page = rb_list_head(next_page);
1449         next_page->prev = tail_page;
1450
1451         /* make sure pages points to a valid page in the ring buffer */
1452         cpu_buffer->pages = next_page;
1453
1454         /* update head page */
1455         if (head_bit)
1456                 cpu_buffer->head_page = list_entry(next_page,
1457                                                 struct buffer_page, list);
1458
1459         /*
1460          * change read pointer to make sure any read iterators reset
1461          * themselves
1462          */
1463         cpu_buffer->read = 0;
1464
1465         /* pages are removed, resume tracing and then free the pages */
1466         atomic_dec(&cpu_buffer->record_disabled);
1467         raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1468
1469         RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
1470
1471         /* last buffer page to remove */
1472         last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
1473                                 list);
1474         tmp_iter_page = first_page;
1475
1476         do {
1477                 to_remove_page = tmp_iter_page;
1478                 rb_inc_page(cpu_buffer, &tmp_iter_page);
1479
1480                 /* update the counters */
1481                 page_entries = rb_page_entries(to_remove_page);
1482                 if (page_entries) {
1483                         /*
1484                          * If something was added to this page, it was full
1485                          * since it is not the tail page. So we deduct the
1486                          * bytes consumed in ring buffer from here.
1487                          * Increment overrun to account for the lost events.
1488                          */
1489                         local_add(page_entries, &cpu_buffer->overrun);
1490                         local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1491                 }
1492
1493                 /*
1494                  * We have already removed references to this list item, just
1495                  * free up the buffer_page and its page
1496                  */
1497                 free_buffer_page(to_remove_page);
1498                 nr_removed--;
1499
1500         } while (to_remove_page != last_page);
1501
1502         RB_WARN_ON(cpu_buffer, nr_removed);
1503
1504         return nr_removed == 0;
1505 }
1506
1507 static int
1508 rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
1509 {
1510         struct list_head *pages = &cpu_buffer->new_pages;
1511         int retries, success;
1512
1513         raw_spin_lock_irq(&cpu_buffer->reader_lock);
1514         /*
1515          * We are holding the reader lock, so the reader page won't be swapped
1516          * in the ring buffer. Now we are racing with the writer trying to
1517          * move head page and the tail page.
1518          * We are going to adapt the reader page update process where:
1519          * 1. We first splice the start and end of list of new pages between
1520          *    the head page and its previous page.
1521          * 2. We cmpxchg the prev_page->next to point from head page to the
1522          *    start of new pages list.
1523          * 3. Finally, we update the head->prev to the end of new list.
1524          *
1525          * We will try this process 10 times, to make sure that we don't keep
1526          * spinning.
1527          */
1528         retries = 10;
1529         success = 0;
1530         while (retries--) {
1531                 struct list_head *head_page, *prev_page, *r;
1532                 struct list_head *last_page, *first_page;
1533                 struct list_head *head_page_with_bit;
1534
1535                 head_page = &rb_set_head_page(cpu_buffer)->list;
1536                 if (!head_page)
1537                         break;
1538                 prev_page = head_page->prev;
1539
1540                 first_page = pages->next;
1541                 last_page  = pages->prev;
1542
1543                 head_page_with_bit = (struct list_head *)
1544                                      ((unsigned long)head_page | RB_PAGE_HEAD);
1545
1546                 last_page->next = head_page_with_bit;
1547                 first_page->prev = prev_page;
1548
1549                 r = cmpxchg(&prev_page->next, head_page_with_bit, first_page);
1550
1551                 if (r == head_page_with_bit) {
1552                         /*
1553                          * yay, we replaced the page pointer to our new list,
1554                          * now, we just have to update to head page's prev
1555                          * pointer to point to end of list
1556                          */
1557                         head_page->prev = last_page;
1558                         success = 1;
1559                         break;
1560                 }
1561         }
1562
1563         if (success)
1564                 INIT_LIST_HEAD(pages);
1565         /*
1566          * If we weren't successful in adding in new pages, warn and stop
1567          * tracing
1568          */
1569         RB_WARN_ON(cpu_buffer, !success);
1570         raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1571
1572         /* free pages if they weren't inserted */
1573         if (!success) {
1574                 struct buffer_page *bpage, *tmp;
1575                 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1576                                          list) {
1577                         list_del_init(&bpage->list);
1578                         free_buffer_page(bpage);
1579                 }
1580         }
1581         return success;
1582 }
1583
1584 static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
1585 {
1586         int success;
1587
1588         if (cpu_buffer->nr_pages_to_update > 0)
1589                 success = rb_insert_pages(cpu_buffer);
1590         else
1591                 success = rb_remove_pages(cpu_buffer,
1592                                         -cpu_buffer->nr_pages_to_update);
1593
1594         if (success)
1595                 cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
1596 }
1597
1598 static void update_pages_handler(struct work_struct *work)
1599 {
1600         struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
1601                         struct ring_buffer_per_cpu, update_pages_work);
1602         rb_update_pages(cpu_buffer);
1603         complete(&cpu_buffer->update_done);
1604 }
1605
1606 /**
1607  * ring_buffer_resize - resize the ring buffer
1608  * @buffer: the buffer to resize.
1609  * @size: the new size.
1610  * @cpu_id: the cpu buffer to resize
1611  *
1612  * Minimum size is 2 * BUF_PAGE_SIZE.
1613  *
1614  * Returns 0 on success and < 0 on failure.
1615  */
1616 int ring_buffer_resize(struct ring_buffer *buffer, unsigned long size,
1617                         int cpu_id)
1618 {
1619         struct ring_buffer_per_cpu *cpu_buffer;
1620         unsigned long nr_pages;
1621         int cpu, err = 0;
1622
1623         /*
1624          * Always succeed at resizing a non-existent buffer:
1625          */
1626         if (!buffer)
1627                 return size;
1628
1629         /* Make sure the requested buffer exists */
1630         if (cpu_id != RING_BUFFER_ALL_CPUS &&
1631             !cpumask_test_cpu(cpu_id, buffer->cpumask))
1632                 return size;
1633
1634         nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1635
1636         /* we need a minimum of two pages */
1637         if (nr_pages < 2)
1638                 nr_pages = 2;
1639
1640         size = nr_pages * BUF_PAGE_SIZE;
1641
1642         /*
1643          * Don't succeed if resizing is disabled, as a reader might be
1644          * manipulating the ring buffer and is expecting a sane state while
1645          * this is true.
1646          */
1647         if (atomic_read(&buffer->resize_disabled))
1648                 return -EBUSY;
1649
1650         /* prevent another thread from changing buffer sizes */
1651         mutex_lock(&buffer->mutex);
1652
1653         if (cpu_id == RING_BUFFER_ALL_CPUS) {
1654                 /* calculate the pages to update */
1655                 for_each_buffer_cpu(buffer, cpu) {
1656                         cpu_buffer = buffer->buffers[cpu];
1657
1658                         cpu_buffer->nr_pages_to_update = nr_pages -
1659                                                         cpu_buffer->nr_pages;
1660                         /*
1661                          * nothing more to do for removing pages or no update
1662                          */
1663                         if (cpu_buffer->nr_pages_to_update <= 0)
1664                                 continue;
1665                         /*
1666                          * to add pages, make sure all new pages can be
1667                          * allocated without receiving ENOMEM
1668                          */
1669                         INIT_LIST_HEAD(&cpu_buffer->new_pages);
1670                         if (__rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1671                                                 &cpu_buffer->new_pages, cpu)) {
1672                                 /* not enough memory for new pages */
1673                                 err = -ENOMEM;
1674                                 goto out_err;
1675                         }
1676                 }
1677
1678                 get_online_cpus();
1679                 /*
1680                  * Fire off all the required work handlers
1681                  * We can't schedule on offline CPUs, but it's not necessary
1682                  * since we can change their buffer sizes without any race.
1683                  */
1684                 for_each_buffer_cpu(buffer, cpu) {
1685                         cpu_buffer = buffer->buffers[cpu];
1686                         if (!cpu_buffer->nr_pages_to_update)
1687                                 continue;
1688
1689                         /* Can't run something on an offline CPU. */
1690                         if (!cpu_online(cpu)) {
1691                                 rb_update_pages(cpu_buffer);
1692                                 cpu_buffer->nr_pages_to_update = 0;
1693                         } else {
1694                                 schedule_work_on(cpu,
1695                                                 &cpu_buffer->update_pages_work);
1696                         }
1697                 }
1698
1699                 /* wait for all the updates to complete */
1700                 for_each_buffer_cpu(buffer, cpu) {
1701                         cpu_buffer = buffer->buffers[cpu];
1702                         if (!cpu_buffer->nr_pages_to_update)
1703                                 continue;
1704
1705                         if (cpu_online(cpu))
1706                                 wait_for_completion(&cpu_buffer->update_done);
1707                         cpu_buffer->nr_pages_to_update = 0;
1708                 }
1709
1710                 put_online_cpus();
1711         } else {
1712                 /* Make sure this CPU has been intitialized */
1713                 if (!cpumask_test_cpu(cpu_id, buffer->cpumask))
1714                         goto out;
1715
1716                 cpu_buffer = buffer->buffers[cpu_id];
1717
1718                 if (nr_pages == cpu_buffer->nr_pages)
1719                         goto out;
1720
1721                 cpu_buffer->nr_pages_to_update = nr_pages -
1722                                                 cpu_buffer->nr_pages;
1723
1724                 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1725                 if (cpu_buffer->nr_pages_to_update > 0 &&
1726                         __rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1727                                             &cpu_buffer->new_pages, cpu_id)) {
1728                         err = -ENOMEM;
1729                         goto out_err;
1730                 }
1731
1732                 get_online_cpus();
1733
1734                 /* Can't run something on an offline CPU. */
1735                 if (!cpu_online(cpu_id))
1736                         rb_update_pages(cpu_buffer);
1737                 else {
1738                         schedule_work_on(cpu_id,
1739                                          &cpu_buffer->update_pages_work);
1740                         wait_for_completion(&cpu_buffer->update_done);
1741                 }
1742
1743                 cpu_buffer->nr_pages_to_update = 0;
1744                 put_online_cpus();
1745         }
1746
1747  out:
1748         /*
1749          * The ring buffer resize can happen with the ring buffer
1750          * enabled, so that the update disturbs the tracing as little
1751          * as possible. But if the buffer is disabled, we do not need
1752          * to worry about that, and we can take the time to verify
1753          * that the buffer is not corrupt.
1754          */
1755         if (atomic_read(&buffer->record_disabled)) {
1756                 atomic_inc(&buffer->record_disabled);
1757                 /*
1758                  * Even though the buffer was disabled, we must make sure
1759                  * that it is truly disabled before calling rb_check_pages.
1760                  * There could have been a race between checking
1761                  * record_disable and incrementing it.
1762                  */
1763                 synchronize_sched();
1764                 for_each_buffer_cpu(buffer, cpu) {
1765                         cpu_buffer = buffer->buffers[cpu];
1766                         rb_check_pages(cpu_buffer);
1767                 }
1768                 atomic_dec(&buffer->record_disabled);
1769         }
1770
1771         mutex_unlock(&buffer->mutex);
1772         return size;
1773
1774  out_err:
1775         for_each_buffer_cpu(buffer, cpu) {
1776                 struct buffer_page *bpage, *tmp;
1777
1778                 cpu_buffer = buffer->buffers[cpu];
1779                 cpu_buffer->nr_pages_to_update = 0;
1780
1781                 if (list_empty(&cpu_buffer->new_pages))
1782                         continue;
1783
1784                 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1785                                         list) {
1786                         list_del_init(&bpage->list);
1787                         free_buffer_page(bpage);
1788                 }
1789         }
1790         mutex_unlock(&buffer->mutex);
1791         return err;
1792 }
1793 EXPORT_SYMBOL_GPL(ring_buffer_resize);
1794
1795 void ring_buffer_change_overwrite(struct ring_buffer *buffer, int val)
1796 {
1797         mutex_lock(&buffer->mutex);
1798         if (val)
1799                 buffer->flags |= RB_FL_OVERWRITE;
1800         else
1801                 buffer->flags &= ~RB_FL_OVERWRITE;
1802         mutex_unlock(&buffer->mutex);
1803 }
1804 EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
1805
1806 static __always_inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
1807 {
1808         return bpage->page->data + index;
1809 }
1810
1811 static __always_inline struct ring_buffer_event *
1812 rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
1813 {
1814         return __rb_page_index(cpu_buffer->reader_page,
1815                                cpu_buffer->reader_page->read);
1816 }
1817
1818 static __always_inline struct ring_buffer_event *
1819 rb_iter_head_event(struct ring_buffer_iter *iter)
1820 {
1821         return __rb_page_index(iter->head_page, iter->head);
1822 }
1823
1824 static __always_inline unsigned rb_page_commit(struct buffer_page *bpage)
1825 {
1826         return local_read(&bpage->page->commit);
1827 }
1828
1829 /* Size is determined by what has been committed */
1830 static __always_inline unsigned rb_page_size(struct buffer_page *bpage)
1831 {
1832         return rb_page_commit(bpage);
1833 }
1834
1835 static __always_inline unsigned
1836 rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
1837 {
1838         return rb_page_commit(cpu_buffer->commit_page);
1839 }
1840
1841 static __always_inline unsigned
1842 rb_event_index(struct ring_buffer_event *event)
1843 {
1844         unsigned long addr = (unsigned long)event;
1845
1846         return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
1847 }
1848
1849 static void rb_inc_iter(struct ring_buffer_iter *iter)
1850 {
1851         struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
1852
1853         /*
1854          * The iterator could be on the reader page (it starts there).
1855          * But the head could have moved, since the reader was
1856          * found. Check for this case and assign the iterator
1857          * to the head page instead of next.
1858          */
1859         if (iter->head_page == cpu_buffer->reader_page)
1860                 iter->head_page = rb_set_head_page(cpu_buffer);
1861         else
1862                 rb_inc_page(cpu_buffer, &iter->head_page);
1863
1864         iter->read_stamp = iter->head_page->page->time_stamp;
1865         iter->head = 0;
1866 }
1867
1868 /*
1869  * rb_handle_head_page - writer hit the head page
1870  *
1871  * Returns: +1 to retry page
1872  *           0 to continue
1873  *          -1 on error
1874  */
1875 static int
1876 rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
1877                     struct buffer_page *tail_page,
1878                     struct buffer_page *next_page)
1879 {
1880         struct buffer_page *new_head;
1881         int entries;
1882         int type;
1883         int ret;
1884
1885         entries = rb_page_entries(next_page);
1886
1887         /*
1888          * The hard part is here. We need to move the head
1889          * forward, and protect against both readers on
1890          * other CPUs and writers coming in via interrupts.
1891          */
1892         type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
1893                                        RB_PAGE_HEAD);
1894
1895         /*
1896          * type can be one of four:
1897          *  NORMAL - an interrupt already moved it for us
1898          *  HEAD   - we are the first to get here.
1899          *  UPDATE - we are the interrupt interrupting
1900          *           a current move.
1901          *  MOVED  - a reader on another CPU moved the next
1902          *           pointer to its reader page. Give up
1903          *           and try again.
1904          */
1905
1906         switch (type) {
1907         case RB_PAGE_HEAD:
1908                 /*
1909                  * We changed the head to UPDATE, thus
1910                  * it is our responsibility to update
1911                  * the counters.
1912                  */
1913                 local_add(entries, &cpu_buffer->overrun);
1914                 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1915
1916                 /*
1917                  * The entries will be zeroed out when we move the
1918                  * tail page.
1919                  */
1920
1921                 /* still more to do */
1922                 break;
1923
1924         case RB_PAGE_UPDATE:
1925                 /*
1926                  * This is an interrupt that interrupt the
1927                  * previous update. Still more to do.
1928                  */
1929                 break;
1930         case RB_PAGE_NORMAL:
1931                 /*
1932                  * An interrupt came in before the update
1933                  * and processed this for us.
1934                  * Nothing left to do.
1935                  */
1936                 return 1;
1937         case RB_PAGE_MOVED:
1938                 /*
1939                  * The reader is on another CPU and just did
1940                  * a swap with our next_page.
1941                  * Try again.
1942                  */
1943                 return 1;
1944         default:
1945                 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
1946                 return -1;
1947         }
1948
1949         /*
1950          * Now that we are here, the old head pointer is
1951          * set to UPDATE. This will keep the reader from
1952          * swapping the head page with the reader page.
1953          * The reader (on another CPU) will spin till
1954          * we are finished.
1955          *
1956          * We just need to protect against interrupts
1957          * doing the job. We will set the next pointer
1958          * to HEAD. After that, we set the old pointer
1959          * to NORMAL, but only if it was HEAD before.
1960          * otherwise we are an interrupt, and only
1961          * want the outer most commit to reset it.
1962          */
1963         new_head = next_page;
1964         rb_inc_page(cpu_buffer, &new_head);
1965
1966         ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
1967                                     RB_PAGE_NORMAL);
1968
1969         /*
1970          * Valid returns are:
1971          *  HEAD   - an interrupt came in and already set it.
1972          *  NORMAL - One of two things:
1973          *            1) We really set it.
1974          *            2) A bunch of interrupts came in and moved
1975          *               the page forward again.
1976          */
1977         switch (ret) {
1978         case RB_PAGE_HEAD:
1979         case RB_PAGE_NORMAL:
1980                 /* OK */
1981                 break;
1982         default:
1983                 RB_WARN_ON(cpu_buffer, 1);
1984                 return -1;
1985         }
1986
1987         /*
1988          * It is possible that an interrupt came in,
1989          * set the head up, then more interrupts came in
1990          * and moved it again. When we get back here,
1991          * the page would have been set to NORMAL but we
1992          * just set it back to HEAD.
1993          *
1994          * How do you detect this? Well, if that happened
1995          * the tail page would have moved.
1996          */
1997         if (ret == RB_PAGE_NORMAL) {
1998                 struct buffer_page *buffer_tail_page;
1999
2000                 buffer_tail_page = READ_ONCE(cpu_buffer->tail_page);
2001                 /*
2002                  * If the tail had moved passed next, then we need
2003                  * to reset the pointer.
2004                  */
2005                 if (buffer_tail_page != tail_page &&
2006                     buffer_tail_page != next_page)
2007                         rb_head_page_set_normal(cpu_buffer, new_head,
2008                                                 next_page,
2009                                                 RB_PAGE_HEAD);
2010         }
2011
2012         /*
2013          * If this was the outer most commit (the one that
2014          * changed the original pointer from HEAD to UPDATE),
2015          * then it is up to us to reset it to NORMAL.
2016          */
2017         if (type == RB_PAGE_HEAD) {
2018                 ret = rb_head_page_set_normal(cpu_buffer, next_page,
2019                                               tail_page,
2020                                               RB_PAGE_UPDATE);
2021                 if (RB_WARN_ON(cpu_buffer,
2022                                ret != RB_PAGE_UPDATE))
2023                         return -1;
2024         }
2025
2026         return 0;
2027 }
2028
2029 static inline void
2030 rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2031               unsigned long tail, struct rb_event_info *info)
2032 {
2033         struct buffer_page *tail_page = info->tail_page;
2034         struct ring_buffer_event *event;
2035         unsigned long length = info->length;
2036
2037         /*
2038          * Only the event that crossed the page boundary
2039          * must fill the old tail_page with padding.
2040          */
2041         if (tail >= BUF_PAGE_SIZE) {
2042                 /*
2043                  * If the page was filled, then we still need
2044                  * to update the real_end. Reset it to zero
2045                  * and the reader will ignore it.
2046                  */
2047                 if (tail == BUF_PAGE_SIZE)
2048                         tail_page->real_end = 0;
2049
2050                 local_sub(length, &tail_page->write);
2051                 return;
2052         }
2053
2054         event = __rb_page_index(tail_page, tail);
2055
2056         /* account for padding bytes */
2057         local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes);
2058
2059         /*
2060          * Save the original length to the meta data.
2061          * This will be used by the reader to add lost event
2062          * counter.
2063          */
2064         tail_page->real_end = tail;
2065
2066         /*
2067          * If this event is bigger than the minimum size, then
2068          * we need to be careful that we don't subtract the
2069          * write counter enough to allow another writer to slip
2070          * in on this page.
2071          * We put in a discarded commit instead, to make sure
2072          * that this space is not used again.
2073          *
2074          * If we are less than the minimum size, we don't need to
2075          * worry about it.
2076          */
2077         if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
2078                 /* No room for any events */
2079
2080                 /* Mark the rest of the page with padding */
2081                 rb_event_set_padding(event);
2082
2083                 /* Set the write back to the previous setting */
2084                 local_sub(length, &tail_page->write);
2085                 return;
2086         }
2087
2088         /* Put in a discarded event */
2089         event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
2090         event->type_len = RINGBUF_TYPE_PADDING;
2091         /* time delta must be non zero */
2092         event->time_delta = 1;
2093
2094         /* Set write to end of buffer */
2095         length = (tail + length) - BUF_PAGE_SIZE;
2096         local_sub(length, &tail_page->write);
2097 }
2098
2099 static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer);
2100
2101 /*
2102  * This is the slow path, force gcc not to inline it.
2103  */
2104 static noinline struct ring_buffer_event *
2105 rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2106              unsigned long tail, struct rb_event_info *info)
2107 {
2108         struct buffer_page *tail_page = info->tail_page;
2109         struct buffer_page *commit_page = cpu_buffer->commit_page;
2110         struct ring_buffer *buffer = cpu_buffer->buffer;
2111         struct buffer_page *next_page;
2112         int ret;
2113
2114         next_page = tail_page;
2115
2116         rb_inc_page(cpu_buffer, &next_page);
2117
2118         /*
2119          * If for some reason, we had an interrupt storm that made
2120          * it all the way around the buffer, bail, and warn
2121          * about it.
2122          */
2123         if (unlikely(next_page == commit_page)) {
2124                 local_inc(&cpu_buffer->commit_overrun);
2125                 goto out_reset;
2126         }
2127
2128         /*
2129          * This is where the fun begins!
2130          *
2131          * We are fighting against races between a reader that
2132          * could be on another CPU trying to swap its reader
2133          * page with the buffer head.
2134          *
2135          * We are also fighting against interrupts coming in and
2136          * moving the head or tail on us as well.
2137          *
2138          * If the next page is the head page then we have filled
2139          * the buffer, unless the commit page is still on the
2140          * reader page.
2141          */
2142         if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) {
2143
2144                 /*
2145                  * If the commit is not on the reader page, then
2146                  * move the header page.
2147                  */
2148                 if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2149                         /*
2150                          * If we are not in overwrite mode,
2151                          * this is easy, just stop here.
2152                          */
2153                         if (!(buffer->flags & RB_FL_OVERWRITE)) {
2154                                 local_inc(&cpu_buffer->dropped_events);
2155                                 goto out_reset;
2156                         }
2157
2158                         ret = rb_handle_head_page(cpu_buffer,
2159                                                   tail_page,
2160                                                   next_page);
2161                         if (ret < 0)
2162                                 goto out_reset;
2163                         if (ret)
2164                                 goto out_again;
2165                 } else {
2166                         /*
2167                          * We need to be careful here too. The
2168                          * commit page could still be on the reader
2169                          * page. We could have a small buffer, and
2170                          * have filled up the buffer with events
2171                          * from interrupts and such, and wrapped.
2172                          *
2173                          * Note, if the tail page is also the on the
2174                          * reader_page, we let it move out.
2175                          */
2176                         if (unlikely((cpu_buffer->commit_page !=
2177                                       cpu_buffer->tail_page) &&
2178                                      (cpu_buffer->commit_page ==
2179                                       cpu_buffer->reader_page))) {
2180                                 local_inc(&cpu_buffer->commit_overrun);
2181                                 goto out_reset;
2182                         }
2183                 }
2184         }
2185
2186         rb_tail_page_update(cpu_buffer, tail_page, next_page);
2187
2188  out_again:
2189
2190         rb_reset_tail(cpu_buffer, tail, info);
2191
2192         /* Commit what we have for now. */
2193         rb_end_commit(cpu_buffer);
2194         /* rb_end_commit() decs committing */
2195         local_inc(&cpu_buffer->committing);
2196
2197         /* fail and let the caller try again */
2198         return ERR_PTR(-EAGAIN);
2199
2200  out_reset:
2201         /* reset write */
2202         rb_reset_tail(cpu_buffer, tail, info);
2203
2204         return NULL;
2205 }
2206
2207 /* Slow path, do not inline */
2208 static noinline struct ring_buffer_event *
2209 rb_add_time_stamp(struct ring_buffer_event *event, u64 delta)
2210 {
2211         event->type_len = RINGBUF_TYPE_TIME_EXTEND;
2212
2213         /* Not the first event on the page? */
2214         if (rb_event_index(event)) {
2215                 event->time_delta = delta & TS_MASK;
2216                 event->array[0] = delta >> TS_SHIFT;
2217         } else {
2218                 /* nope, just zero it */
2219                 event->time_delta = 0;
2220                 event->array[0] = 0;
2221         }
2222
2223         return skip_time_extend(event);
2224 }
2225
2226 static inline bool rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2227                                      struct ring_buffer_event *event);
2228
2229 /**
2230  * rb_update_event - update event type and data
2231  * @event: the event to update
2232  * @type: the type of event
2233  * @length: the size of the event field in the ring buffer
2234  *
2235  * Update the type and data fields of the event. The length
2236  * is the actual size that is written to the ring buffer,
2237  * and with this, we can determine what to place into the
2238  * data field.
2239  */
2240 static void
2241 rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
2242                 struct ring_buffer_event *event,
2243                 struct rb_event_info *info)
2244 {
2245         unsigned length = info->length;
2246         u64 delta = info->delta;
2247
2248         /* Only a commit updates the timestamp */
2249         if (unlikely(!rb_event_is_commit(cpu_buffer, event)))
2250                 delta = 0;
2251
2252         /*
2253          * If we need to add a timestamp, then we
2254          * add it to the start of the resevered space.
2255          */
2256         if (unlikely(info->add_timestamp)) {
2257                 event = rb_add_time_stamp(event, delta);
2258                 length -= RB_LEN_TIME_EXTEND;
2259                 delta = 0;
2260         }
2261
2262         event->time_delta = delta;
2263         length -= RB_EVNT_HDR_SIZE;
2264         if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
2265                 event->type_len = 0;
2266                 event->array[0] = length;
2267         } else
2268                 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
2269 }
2270
2271 static unsigned rb_calculate_event_length(unsigned length)
2272 {
2273         struct ring_buffer_event event; /* Used only for sizeof array */
2274
2275         /* zero length can cause confusions */
2276         if (!length)
2277                 length++;
2278
2279         if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
2280                 length += sizeof(event.array[0]);
2281
2282         length += RB_EVNT_HDR_SIZE;
2283         length = ALIGN(length, RB_ARCH_ALIGNMENT);
2284
2285         /*
2286          * In case the time delta is larger than the 27 bits for it
2287          * in the header, we need to add a timestamp. If another
2288          * event comes in when trying to discard this one to increase
2289          * the length, then the timestamp will be added in the allocated
2290          * space of this event. If length is bigger than the size needed
2291          * for the TIME_EXTEND, then padding has to be used. The events
2292          * length must be either RB_LEN_TIME_EXTEND, or greater than or equal
2293          * to RB_LEN_TIME_EXTEND + 8, as 8 is the minimum size for padding.
2294          * As length is a multiple of 4, we only need to worry if it
2295          * is 12 (RB_LEN_TIME_EXTEND + 4).
2296          */
2297         if (length == RB_LEN_TIME_EXTEND + RB_ALIGNMENT)
2298                 length += RB_ALIGNMENT;
2299
2300         return length;
2301 }
2302
2303 #ifndef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2304 static inline bool sched_clock_stable(void)
2305 {
2306         return true;
2307 }
2308 #endif
2309
2310 static inline int
2311 rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2312                   struct ring_buffer_event *event)
2313 {
2314         unsigned long new_index, old_index;
2315         struct buffer_page *bpage;
2316         unsigned long index;
2317         unsigned long addr;
2318
2319         new_index = rb_event_index(event);
2320         old_index = new_index + rb_event_ts_length(event);
2321         addr = (unsigned long)event;
2322         addr &= PAGE_MASK;
2323
2324         bpage = READ_ONCE(cpu_buffer->tail_page);
2325
2326         if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2327                 unsigned long write_mask =
2328                         local_read(&bpage->write) & ~RB_WRITE_MASK;
2329                 unsigned long event_length = rb_event_length(event);
2330                 /*
2331                  * This is on the tail page. It is possible that
2332                  * a write could come in and move the tail page
2333                  * and write to the next page. That is fine
2334                  * because we just shorten what is on this page.
2335                  */
2336                 old_index += write_mask;
2337                 new_index += write_mask;
2338                 index = local_cmpxchg(&bpage->write, old_index, new_index);
2339                 if (index == old_index) {
2340                         /* update counters */
2341                         local_sub(event_length, &cpu_buffer->entries_bytes);
2342                         return 1;
2343                 }
2344         }
2345
2346         /* could not discard */
2347         return 0;
2348 }
2349
2350 static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2351 {
2352         local_inc(&cpu_buffer->committing);
2353         local_inc(&cpu_buffer->commits);
2354 }
2355
2356 static __always_inline void
2357 rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
2358 {
2359         unsigned long max_count;
2360
2361         /*
2362          * We only race with interrupts and NMIs on this CPU.
2363          * If we own the commit event, then we can commit
2364          * all others that interrupted us, since the interruptions
2365          * are in stack format (they finish before they come
2366          * back to us). This allows us to do a simple loop to
2367          * assign the commit to the tail.
2368          */
2369  again:
2370         max_count = cpu_buffer->nr_pages * 100;
2371
2372         while (cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)) {
2373                 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
2374                         return;
2375                 if (RB_WARN_ON(cpu_buffer,
2376                                rb_is_reader_page(cpu_buffer->tail_page)))
2377                         return;
2378                 local_set(&cpu_buffer->commit_page->page->commit,
2379                           rb_page_write(cpu_buffer->commit_page));
2380                 rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
2381                 /* Only update the write stamp if the page has an event */
2382                 if (rb_page_write(cpu_buffer->commit_page))
2383                         cpu_buffer->write_stamp =
2384                                 cpu_buffer->commit_page->page->time_stamp;
2385                 /* add barrier to keep gcc from optimizing too much */
2386                 barrier();
2387         }
2388         while (rb_commit_index(cpu_buffer) !=
2389                rb_page_write(cpu_buffer->commit_page)) {
2390
2391                 local_set(&cpu_buffer->commit_page->page->commit,
2392                           rb_page_write(cpu_buffer->commit_page));
2393                 RB_WARN_ON(cpu_buffer,
2394                            local_read(&cpu_buffer->commit_page->page->commit) &
2395                            ~RB_WRITE_MASK);
2396                 barrier();
2397         }
2398
2399         /* again, keep gcc from optimizing */
2400         barrier();
2401
2402         /*
2403          * If an interrupt came in just after the first while loop
2404          * and pushed the tail page forward, we will be left with
2405          * a dangling commit that will never go forward.
2406          */
2407         if (unlikely(cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)))
2408                 goto again;
2409 }
2410
2411 static __always_inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
2412 {
2413         unsigned long commits;
2414
2415         if (RB_WARN_ON(cpu_buffer,
2416                        !local_read(&cpu_buffer->committing)))
2417                 return;
2418
2419  again:
2420         commits = local_read(&cpu_buffer->commits);
2421         /* synchronize with interrupts */
2422         barrier();
2423         if (local_read(&cpu_buffer->committing) == 1)
2424                 rb_set_commit_to_write(cpu_buffer);
2425
2426         local_dec(&cpu_buffer->committing);
2427
2428         /* synchronize with interrupts */
2429         barrier();
2430
2431         /*
2432          * Need to account for interrupts coming in between the
2433          * updating of the commit page and the clearing of the
2434          * committing counter.
2435          */
2436         if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
2437             !local_read(&cpu_buffer->committing)) {
2438                 local_inc(&cpu_buffer->committing);
2439                 goto again;
2440         }
2441 }
2442
2443 static inline void rb_event_discard(struct ring_buffer_event *event)
2444 {
2445         if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
2446                 event = skip_time_extend(event);
2447
2448         /* array[0] holds the actual length for the discarded event */
2449         event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
2450         event->type_len = RINGBUF_TYPE_PADDING;
2451         /* time delta must be non zero */
2452         if (!event->time_delta)
2453                 event->time_delta = 1;
2454 }
2455
2456 static __always_inline bool
2457 rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2458                    struct ring_buffer_event *event)
2459 {
2460         unsigned long addr = (unsigned long)event;
2461         unsigned long index;
2462
2463         index = rb_event_index(event);
2464         addr &= PAGE_MASK;
2465
2466         return cpu_buffer->commit_page->page == (void *)addr &&
2467                 rb_commit_index(cpu_buffer) == index;
2468 }
2469
2470 static __always_inline void
2471 rb_update_write_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2472                       struct ring_buffer_event *event)
2473 {
2474         u64 delta;
2475
2476         /*
2477          * The event first in the commit queue updates the
2478          * time stamp.
2479          */
2480         if (rb_event_is_commit(cpu_buffer, event)) {
2481                 /*
2482                  * A commit event that is first on a page
2483                  * updates the write timestamp with the page stamp
2484                  */
2485                 if (!rb_event_index(event))
2486                         cpu_buffer->write_stamp =
2487                                 cpu_buffer->commit_page->page->time_stamp;
2488                 else if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
2489                         delta = event->array[0];
2490                         delta <<= TS_SHIFT;
2491                         delta += event->time_delta;
2492                         cpu_buffer->write_stamp += delta;
2493                 } else
2494                         cpu_buffer->write_stamp += event->time_delta;
2495         }
2496 }
2497
2498 static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
2499                       struct ring_buffer_event *event)
2500 {
2501         local_inc(&cpu_buffer->entries);
2502         rb_update_write_stamp(cpu_buffer, event);
2503         rb_end_commit(cpu_buffer);
2504 }
2505
2506 static __always_inline void
2507 rb_wakeups(struct ring_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer)
2508 {
2509         bool pagebusy;
2510
2511         if (buffer->irq_work.waiters_pending) {
2512                 buffer->irq_work.waiters_pending = false;
2513                 /* irq_work_queue() supplies it's own memory barriers */
2514                 irq_work_queue(&buffer->irq_work.work);
2515         }
2516
2517         if (cpu_buffer->irq_work.waiters_pending) {
2518                 cpu_buffer->irq_work.waiters_pending = false;
2519                 /* irq_work_queue() supplies it's own memory barriers */
2520                 irq_work_queue(&cpu_buffer->irq_work.work);
2521         }
2522
2523         pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
2524
2525         if (!pagebusy && cpu_buffer->irq_work.full_waiters_pending) {
2526                 cpu_buffer->irq_work.wakeup_full = true;
2527                 cpu_buffer->irq_work.full_waiters_pending = false;
2528                 /* irq_work_queue() supplies it's own memory barriers */
2529                 irq_work_queue(&cpu_buffer->irq_work.work);
2530         }
2531 }
2532
2533 /*
2534  * The lock and unlock are done within a preempt disable section.
2535  * The current_context per_cpu variable can only be modified
2536  * by the current task between lock and unlock. But it can
2537  * be modified more than once via an interrupt. There are four
2538  * different contexts that we need to consider.
2539  *
2540  *  Normal context.
2541  *  SoftIRQ context
2542  *  IRQ context
2543  *  NMI context
2544  *
2545  * If for some reason the ring buffer starts to recurse, we
2546  * only allow that to happen at most 4 times (one for each
2547  * context). If it happens 5 times, then we consider this a
2548  * recusive loop and do not let it go further.
2549  */
2550
2551 static __always_inline int
2552 trace_recursive_lock(struct ring_buffer_per_cpu *cpu_buffer)
2553 {
2554         if (cpu_buffer->current_context >= 4)
2555                 return 1;
2556
2557         cpu_buffer->current_context++;
2558         /* Interrupts must see this update */
2559         barrier();
2560
2561         return 0;
2562 }
2563
2564 static __always_inline void
2565 trace_recursive_unlock(struct ring_buffer_per_cpu *cpu_buffer)
2566 {
2567         /* Don't let the dec leak out */
2568         barrier();
2569         cpu_buffer->current_context--;
2570 }
2571
2572 /**
2573  * ring_buffer_unlock_commit - commit a reserved
2574  * @buffer: The buffer to commit to
2575  * @event: The event pointer to commit.
2576  *
2577  * This commits the data to the ring buffer, and releases any locks held.
2578  *
2579  * Must be paired with ring_buffer_lock_reserve.
2580  */
2581 int ring_buffer_unlock_commit(struct ring_buffer *buffer,
2582                               struct ring_buffer_event *event)
2583 {
2584         struct ring_buffer_per_cpu *cpu_buffer;
2585         int cpu = raw_smp_processor_id();
2586
2587         cpu_buffer = buffer->buffers[cpu];
2588
2589         rb_commit(cpu_buffer, event);
2590
2591         rb_wakeups(buffer, cpu_buffer);
2592
2593         trace_recursive_unlock(cpu_buffer);
2594
2595         preempt_enable_notrace();
2596
2597         return 0;
2598 }
2599 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
2600
2601 static noinline void
2602 rb_handle_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2603                     struct rb_event_info *info)
2604 {
2605         WARN_ONCE(info->delta > (1ULL << 59),
2606                   KERN_WARNING "Delta way too big! %llu ts=%llu write stamp = %llu\n%s",
2607                   (unsigned long long)info->delta,
2608                   (unsigned long long)info->ts,
2609                   (unsigned long long)cpu_buffer->write_stamp,
2610                   sched_clock_stable() ? "" :
2611                   "If you just came from a suspend/resume,\n"
2612                   "please switch to the trace global clock:\n"
2613                   "  echo global > /sys/kernel/debug/tracing/trace_clock\n");
2614         info->add_timestamp = 1;
2615 }
2616
2617 static struct ring_buffer_event *
2618 __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
2619                   struct rb_event_info *info)
2620 {
2621         struct ring_buffer_event *event;
2622         struct buffer_page *tail_page;
2623         unsigned long tail, write;
2624
2625         /*
2626          * If the time delta since the last event is too big to
2627          * hold in the time field of the event, then we append a
2628          * TIME EXTEND event ahead of the data event.
2629          */
2630         if (unlikely(info->add_timestamp))
2631                 info->length += RB_LEN_TIME_EXTEND;
2632
2633         /* Don't let the compiler play games with cpu_buffer->tail_page */
2634         tail_page = info->tail_page = READ_ONCE(cpu_buffer->tail_page);
2635         write = local_add_return(info->length, &tail_page->write);
2636
2637         /* set write to only the index of the write */
2638         write &= RB_WRITE_MASK;
2639         tail = write - info->length;
2640
2641         /*
2642          * If this is the first commit on the page, then it has the same
2643          * timestamp as the page itself.
2644          */
2645         if (!tail)
2646                 info->delta = 0;
2647
2648         /* See if we shot pass the end of this buffer page */
2649         if (unlikely(write > BUF_PAGE_SIZE))
2650                 return rb_move_tail(cpu_buffer, tail, info);
2651
2652         /* We reserved something on the buffer */
2653
2654         event = __rb_page_index(tail_page, tail);
2655         rb_update_event(cpu_buffer, event, info);
2656
2657         local_inc(&tail_page->entries);
2658
2659         /*
2660          * If this is the first commit on the page, then update
2661          * its timestamp.
2662          */
2663         if (!tail)
2664                 tail_page->page->time_stamp = info->ts;
2665
2666         /* account for these added bytes */
2667         local_add(info->length, &cpu_buffer->entries_bytes);
2668
2669         return event;
2670 }
2671
2672 static __always_inline struct ring_buffer_event *
2673 rb_reserve_next_event(struct ring_buffer *buffer,
2674                       struct ring_buffer_per_cpu *cpu_buffer,
2675                       unsigned long length)
2676 {
2677         struct ring_buffer_event *event;
2678         struct rb_event_info info;
2679         int nr_loops = 0;
2680         u64 diff;
2681
2682         rb_start_commit(cpu_buffer);
2683
2684 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2685         /*
2686          * Due to the ability to swap a cpu buffer from a buffer
2687          * it is possible it was swapped before we committed.
2688          * (committing stops a swap). We check for it here and
2689          * if it happened, we have to fail the write.
2690          */
2691         barrier();
2692         if (unlikely(READ_ONCE(cpu_buffer->buffer) != buffer)) {
2693                 local_dec(&cpu_buffer->committing);
2694                 local_dec(&cpu_buffer->commits);
2695                 return NULL;
2696         }
2697 #endif
2698
2699         info.length = rb_calculate_event_length(length);
2700  again:
2701         info.add_timestamp = 0;
2702         info.delta = 0;
2703
2704         /*
2705          * We allow for interrupts to reenter here and do a trace.
2706          * If one does, it will cause this original code to loop
2707          * back here. Even with heavy interrupts happening, this
2708          * should only happen a few times in a row. If this happens
2709          * 1000 times in a row, there must be either an interrupt
2710          * storm or we have something buggy.
2711          * Bail!
2712          */
2713         if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
2714                 goto out_fail;
2715
2716         info.ts = rb_time_stamp(cpu_buffer->buffer);
2717         diff = info.ts - cpu_buffer->write_stamp;
2718
2719         /* make sure this diff is calculated here */
2720         barrier();
2721
2722         /* Did the write stamp get updated already? */
2723         if (likely(info.ts >= cpu_buffer->write_stamp)) {
2724                 info.delta = diff;
2725                 if (unlikely(test_time_stamp(info.delta)))
2726                         rb_handle_timestamp(cpu_buffer, &info);
2727         }
2728
2729         event = __rb_reserve_next(cpu_buffer, &info);
2730
2731         if (unlikely(PTR_ERR(event) == -EAGAIN)) {
2732                 if (info.add_timestamp)
2733                         info.length -= RB_LEN_TIME_EXTEND;
2734                 goto again;
2735         }
2736
2737         if (!event)
2738                 goto out_fail;
2739
2740         return event;
2741
2742  out_fail:
2743         rb_end_commit(cpu_buffer);
2744         return NULL;
2745 }
2746
2747 /**
2748  * ring_buffer_lock_reserve - reserve a part of the buffer
2749  * @buffer: the ring buffer to reserve from
2750  * @length: the length of the data to reserve (excluding event header)
2751  *
2752  * Returns a reseverd event on the ring buffer to copy directly to.
2753  * The user of this interface will need to get the body to write into
2754  * and can use the ring_buffer_event_data() interface.
2755  *
2756  * The length is the length of the data needed, not the event length
2757  * which also includes the event header.
2758  *
2759  * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2760  * If NULL is returned, then nothing has been allocated or locked.
2761  */
2762 struct ring_buffer_event *
2763 ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length)
2764 {
2765         struct ring_buffer_per_cpu *cpu_buffer;
2766         struct ring_buffer_event *event;
2767         int cpu;
2768
2769         /* If we are tracing schedule, we don't want to recurse */
2770         preempt_disable_notrace();
2771
2772         if (unlikely(atomic_read(&buffer->record_disabled)))
2773                 goto out;
2774
2775         cpu = raw_smp_processor_id();
2776
2777         if (unlikely(!cpumask_test_cpu(cpu, buffer->cpumask)))
2778                 goto out;
2779
2780         cpu_buffer = buffer->buffers[cpu];
2781
2782         if (unlikely(atomic_read(&cpu_buffer->record_disabled)))
2783                 goto out;
2784
2785         if (unlikely(length > BUF_MAX_DATA_SIZE))
2786                 goto out;
2787
2788         if (unlikely(trace_recursive_lock(cpu_buffer)))
2789                 goto out;
2790
2791         event = rb_reserve_next_event(buffer, cpu_buffer, length);
2792         if (!event)
2793                 goto out_unlock;
2794
2795         return event;
2796
2797  out_unlock:
2798         trace_recursive_unlock(cpu_buffer);
2799  out:
2800         preempt_enable_notrace();
2801         return NULL;
2802 }
2803 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
2804
2805 /*
2806  * Decrement the entries to the page that an event is on.
2807  * The event does not even need to exist, only the pointer
2808  * to the page it is on. This may only be called before the commit
2809  * takes place.
2810  */
2811 static inline void
2812 rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
2813                    struct ring_buffer_event *event)
2814 {
2815         unsigned long addr = (unsigned long)event;
2816         struct buffer_page *bpage = cpu_buffer->commit_page;
2817         struct buffer_page *start;
2818
2819         addr &= PAGE_MASK;
2820
2821         /* Do the likely case first */
2822         if (likely(bpage->page == (void *)addr)) {
2823                 local_dec(&bpage->entries);
2824                 return;
2825         }
2826
2827         /*
2828          * Because the commit page may be on the reader page we
2829          * start with the next page and check the end loop there.
2830          */
2831         rb_inc_page(cpu_buffer, &bpage);
2832         start = bpage;
2833         do {
2834                 if (bpage->page == (void *)addr) {
2835                         local_dec(&bpage->entries);
2836                         return;
2837                 }
2838                 rb_inc_page(cpu_buffer, &bpage);
2839         } while (bpage != start);
2840
2841         /* commit not part of this buffer?? */
2842         RB_WARN_ON(cpu_buffer, 1);
2843 }
2844
2845 /**
2846  * ring_buffer_commit_discard - discard an event that has not been committed
2847  * @buffer: the ring buffer
2848  * @event: non committed event to discard
2849  *
2850  * Sometimes an event that is in the ring buffer needs to be ignored.
2851  * This function lets the user discard an event in the ring buffer
2852  * and then that event will not be read later.
2853  *
2854  * This function only works if it is called before the the item has been
2855  * committed. It will try to free the event from the ring buffer
2856  * if another event has not been added behind it.
2857  *
2858  * If another event has been added behind it, it will set the event
2859  * up as discarded, and perform the commit.
2860  *
2861  * If this function is called, do not call ring_buffer_unlock_commit on
2862  * the event.
2863  */
2864 void ring_buffer_discard_commit(struct ring_buffer *buffer,
2865                                 struct ring_buffer_event *event)
2866 {
2867         struct ring_buffer_per_cpu *cpu_buffer;
2868         int cpu;
2869
2870         /* The event is discarded regardless */
2871         rb_event_discard(event);
2872
2873         cpu = smp_processor_id();
2874         cpu_buffer = buffer->buffers[cpu];
2875
2876         /*
2877          * This must only be called if the event has not been
2878          * committed yet. Thus we can assume that preemption
2879          * is still disabled.
2880          */
2881         RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
2882
2883         rb_decrement_entry(cpu_buffer, event);
2884         if (rb_try_to_discard(cpu_buffer, event))
2885                 goto out;
2886
2887         /*
2888          * The commit is still visible by the reader, so we
2889          * must still update the timestamp.
2890          */
2891         rb_update_write_stamp(cpu_buffer, event);
2892  out:
2893         rb_end_commit(cpu_buffer);
2894
2895         trace_recursive_unlock(cpu_buffer);
2896
2897         preempt_enable_notrace();
2898
2899 }
2900 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
2901
2902 /**
2903  * ring_buffer_write - write data to the buffer without reserving
2904  * @buffer: The ring buffer to write to.
2905  * @length: The length of the data being written (excluding the event header)
2906  * @data: The data to write to the buffer.
2907  *
2908  * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
2909  * one function. If you already have the data to write to the buffer, it
2910  * may be easier to simply call this function.
2911  *
2912  * Note, like ring_buffer_lock_reserve, the length is the length of the data
2913  * and not the length of the event which would hold the header.
2914  */
2915 int ring_buffer_write(struct ring_buffer *buffer,
2916                       unsigned long length,
2917                       void *data)
2918 {
2919         struct ring_buffer_per_cpu *cpu_buffer;
2920         struct ring_buffer_event *event;
2921         void *body;
2922         int ret = -EBUSY;
2923         int cpu;
2924
2925         preempt_disable_notrace();
2926
2927         if (atomic_read(&buffer->record_disabled))
2928                 goto out;
2929
2930         cpu = raw_smp_processor_id();
2931
2932         if (!cpumask_test_cpu(cpu, buffer->cpumask))
2933                 goto out;
2934
2935         cpu_buffer = buffer->buffers[cpu];
2936
2937         if (atomic_read(&cpu_buffer->record_disabled))
2938                 goto out;
2939
2940         if (length > BUF_MAX_DATA_SIZE)
2941                 goto out;
2942
2943         if (unlikely(trace_recursive_lock(cpu_buffer)))
2944                 goto out;
2945
2946         event = rb_reserve_next_event(buffer, cpu_buffer, length);
2947         if (!event)
2948                 goto out_unlock;
2949
2950         body = rb_event_data(event);
2951
2952         memcpy(body, data, length);
2953
2954         rb_commit(cpu_buffer, event);
2955
2956         rb_wakeups(buffer, cpu_buffer);
2957
2958         ret = 0;
2959
2960  out_unlock:
2961         trace_recursive_unlock(cpu_buffer);
2962
2963  out:
2964         preempt_enable_notrace();
2965
2966         return ret;
2967 }
2968 EXPORT_SYMBOL_GPL(ring_buffer_write);
2969
2970 static bool rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
2971 {
2972         struct buffer_page *reader = cpu_buffer->reader_page;
2973         struct buffer_page *head = rb_set_head_page(cpu_buffer);
2974         struct buffer_page *commit = cpu_buffer->commit_page;
2975
2976         /* In case of error, head will be NULL */
2977         if (unlikely(!head))
2978                 return true;
2979
2980         return reader->read == rb_page_commit(reader) &&
2981                 (commit == reader ||
2982                  (commit == head &&
2983                   head->read == rb_page_commit(commit)));
2984 }
2985
2986 /**
2987  * ring_buffer_record_disable - stop all writes into the buffer
2988  * @buffer: The ring buffer to stop writes to.
2989  *
2990  * This prevents all writes to the buffer. Any attempt to write
2991  * to the buffer after this will fail and return NULL.
2992  *
2993  * The caller should call synchronize_sched() after this.
2994  */
2995 void ring_buffer_record_disable(struct ring_buffer *buffer)
2996 {
2997         atomic_inc(&buffer->record_disabled);
2998 }
2999 EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
3000
3001 /**
3002  * ring_buffer_record_enable - enable writes to the buffer
3003  * @buffer: The ring buffer to enable writes
3004  *
3005  * Note, multiple disables will need the same number of enables
3006  * to truly enable the writing (much like preempt_disable).
3007  */
3008 void ring_buffer_record_enable(struct ring_buffer *buffer)
3009 {
3010         atomic_dec(&buffer->record_disabled);
3011 }
3012 EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
3013
3014 /**
3015  * ring_buffer_record_off - stop all writes into the buffer
3016  * @buffer: The ring buffer to stop writes to.
3017  *
3018  * This prevents all writes to the buffer. Any attempt to write
3019  * to the buffer after this will fail and return NULL.
3020  *
3021  * This is different than ring_buffer_record_disable() as
3022  * it works like an on/off switch, where as the disable() version
3023  * must be paired with a enable().
3024  */
3025 void ring_buffer_record_off(struct ring_buffer *buffer)
3026 {
3027         unsigned int rd;
3028         unsigned int new_rd;
3029
3030         do {
3031                 rd = atomic_read(&buffer->record_disabled);
3032                 new_rd = rd | RB_BUFFER_OFF;
3033         } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3034 }
3035 EXPORT_SYMBOL_GPL(ring_buffer_record_off);
3036
3037 /**
3038  * ring_buffer_record_on - restart writes into the buffer
3039  * @buffer: The ring buffer to start writes to.
3040  *
3041  * This enables all writes to the buffer that was disabled by
3042  * ring_buffer_record_off().
3043  *
3044  * This is different than ring_buffer_record_enable() as
3045  * it works like an on/off switch, where as the enable() version
3046  * must be paired with a disable().
3047  */
3048 void ring_buffer_record_on(struct ring_buffer *buffer)
3049 {
3050         unsigned int rd;
3051         unsigned int new_rd;
3052
3053         do {
3054                 rd = atomic_read(&buffer->record_disabled);
3055                 new_rd = rd & ~RB_BUFFER_OFF;
3056         } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3057 }
3058 EXPORT_SYMBOL_GPL(ring_buffer_record_on);
3059
3060 /**
3061  * ring_buffer_record_is_on - return true if the ring buffer can write
3062  * @buffer: The ring buffer to see if write is enabled
3063  *
3064  * Returns true if the ring buffer is in a state that it accepts writes.
3065  */
3066 int ring_buffer_record_is_on(struct ring_buffer *buffer)
3067 {
3068         return !atomic_read(&buffer->record_disabled);
3069 }
3070
3071 /**
3072  * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
3073  * @buffer: The ring buffer to stop writes to.
3074  * @cpu: The CPU buffer to stop
3075  *
3076  * This prevents all writes to the buffer. Any attempt to write
3077  * to the buffer after this will fail and return NULL.
3078  *
3079  * The caller should call synchronize_sched() after this.
3080  */
3081 void ring_buffer_record_disable_cpu(struct ring_buffer *buffer, int cpu)
3082 {
3083         struct ring_buffer_per_cpu *cpu_buffer;
3084
3085         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3086                 return;
3087
3088         cpu_buffer = buffer->buffers[cpu];
3089         atomic_inc(&cpu_buffer->record_disabled);
3090 }
3091 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
3092
3093 /**
3094  * ring_buffer_record_enable_cpu - enable writes to the buffer
3095  * @buffer: The ring buffer to enable writes
3096  * @cpu: The CPU to enable.
3097  *
3098  * Note, multiple disables will need the same number of enables
3099  * to truly enable the writing (much like preempt_disable).
3100  */
3101 void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu)
3102 {
3103         struct ring_buffer_per_cpu *cpu_buffer;
3104
3105         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3106                 return;
3107
3108         cpu_buffer = buffer->buffers[cpu];
3109         atomic_dec(&cpu_buffer->record_disabled);
3110 }
3111 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
3112
3113 /*
3114  * The total entries in the ring buffer is the running counter
3115  * of entries entered into the ring buffer, minus the sum of
3116  * the entries read from the ring buffer and the number of
3117  * entries that were overwritten.
3118  */
3119 static inline unsigned long
3120 rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
3121 {
3122         return local_read(&cpu_buffer->entries) -
3123                 (local_read(&cpu_buffer->overrun) + cpu_buffer->read);
3124 }
3125
3126 /**
3127  * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
3128  * @buffer: The ring buffer
3129  * @cpu: The per CPU buffer to read from.
3130  */
3131 u64 ring_buffer_oldest_event_ts(struct ring_buffer *buffer, int cpu)
3132 {
3133         unsigned long flags;
3134         struct ring_buffer_per_cpu *cpu_buffer;
3135         struct buffer_page *bpage;
3136         u64 ret = 0;
3137
3138         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3139                 return 0;
3140
3141         cpu_buffer = buffer->buffers[cpu];
3142         raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3143         /*
3144          * if the tail is on reader_page, oldest time stamp is on the reader
3145          * page
3146          */
3147         if (cpu_buffer->tail_page == cpu_buffer->reader_page)
3148                 bpage = cpu_buffer->reader_page;
3149         else
3150                 bpage = rb_set_head_page(cpu_buffer);
3151         if (bpage)
3152                 ret = bpage->page->time_stamp;
3153         raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3154
3155         return ret;
3156 }
3157 EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
3158
3159 /**
3160  * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
3161  * @buffer: The ring buffer
3162  * @cpu: The per CPU buffer to read from.
3163  */
3164 unsigned long ring_buffer_bytes_cpu(struct ring_buffer *buffer, int cpu)
3165 {
3166         struct ring_buffer_per_cpu *cpu_buffer;
3167         unsigned long ret;
3168
3169         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3170                 return 0;
3171
3172         cpu_buffer = buffer->buffers[cpu];
3173         ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
3174
3175         return ret;
3176 }
3177 EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
3178
3179 /**
3180  * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
3181  * @buffer: The ring buffer
3182  * @cpu: The per CPU buffer to get the entries from.
3183  */
3184 unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu)
3185 {
3186         struct ring_buffer_per_cpu *cpu_buffer;
3187
3188         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3189                 return 0;
3190
3191         cpu_buffer = buffer->buffers[cpu];
3192
3193         return rb_num_of_entries(cpu_buffer);
3194 }
3195 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
3196
3197 /**
3198  * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
3199  * buffer wrapping around (only if RB_FL_OVERWRITE is on).
3200  * @buffer: The ring buffer
3201  * @cpu: The per CPU buffer to get the number of overruns from
3202  */
3203 unsigned long ring_buffer_overrun_cpu(struct ring_buffer *buffer, int cpu)
3204 {
3205         struct ring_buffer_per_cpu *cpu_buffer;
3206         unsigned long ret;
3207
3208         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3209                 return 0;
3210
3211         cpu_buffer = buffer->buffers[cpu];
3212         ret = local_read(&cpu_buffer->overrun);
3213
3214         return ret;
3215 }
3216 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
3217
3218 /**
3219  * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
3220  * commits failing due to the buffer wrapping around while there are uncommitted
3221  * events, such as during an interrupt storm.
3222  * @buffer: The ring buffer
3223  * @cpu: The per CPU buffer to get the number of overruns from
3224  */
3225 unsigned long
3226 ring_buffer_commit_overrun_cpu(struct ring_buffer *buffer, int cpu)
3227 {
3228         struct ring_buffer_per_cpu *cpu_buffer;
3229         unsigned long ret;
3230
3231         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3232                 return 0;
3233
3234         cpu_buffer = buffer->buffers[cpu];
3235         ret = local_read(&cpu_buffer->commit_overrun);
3236
3237         return ret;
3238 }
3239 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
3240
3241 /**
3242  * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
3243  * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
3244  * @buffer: The ring buffer
3245  * @cpu: The per CPU buffer to get the number of overruns from
3246  */
3247 unsigned long
3248 ring_buffer_dropped_events_cpu(struct ring_buffer *buffer, int cpu)
3249 {
3250         struct ring_buffer_per_cpu *cpu_buffer;
3251         unsigned long ret;
3252
3253         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3254                 return 0;
3255
3256         cpu_buffer = buffer->buffers[cpu];
3257         ret = local_read(&cpu_buffer->dropped_events);
3258
3259         return ret;
3260 }
3261 EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
3262
3263 /**
3264  * ring_buffer_read_events_cpu - get the number of events successfully read
3265  * @buffer: The ring buffer
3266  * @cpu: The per CPU buffer to get the number of events read
3267  */
3268 unsigned long
3269 ring_buffer_read_events_cpu(struct ring_buffer *buffer, int cpu)
3270 {
3271         struct ring_buffer_per_cpu *cpu_buffer;
3272
3273         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3274                 return 0;
3275
3276         cpu_buffer = buffer->buffers[cpu];
3277         return cpu_buffer->read;
3278 }
3279 EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu);
3280
3281 /**
3282  * ring_buffer_entries - get the number of entries in a buffer
3283  * @buffer: The ring buffer
3284  *
3285  * Returns the total number of entries in the ring buffer
3286  * (all CPU entries)
3287  */
3288 unsigned long ring_buffer_entries(struct ring_buffer *buffer)
3289 {
3290         struct ring_buffer_per_cpu *cpu_buffer;
3291         unsigned long entries = 0;
3292         int cpu;
3293
3294         /* if you care about this being correct, lock the buffer */
3295         for_each_buffer_cpu(buffer, cpu) {
3296                 cpu_buffer = buffer->buffers[cpu];
3297                 entries += rb_num_of_entries(cpu_buffer);
3298         }
3299
3300         return entries;
3301 }
3302 EXPORT_SYMBOL_GPL(ring_buffer_entries);
3303
3304 /**
3305  * ring_buffer_overruns - get the number of overruns in buffer
3306  * @buffer: The ring buffer
3307  *
3308  * Returns the total number of overruns in the ring buffer
3309  * (all CPU entries)
3310  */
3311 unsigned long ring_buffer_overruns(struct ring_buffer *buffer)
3312 {
3313         struct ring_buffer_per_cpu *cpu_buffer;
3314         unsigned long overruns = 0;
3315         int cpu;
3316
3317         /* if you care about this being correct, lock the buffer */
3318         for_each_buffer_cpu(buffer, cpu) {
3319                 cpu_buffer = buffer->buffers[cpu];
3320                 overruns += local_read(&cpu_buffer->overrun);
3321         }
3322
3323         return overruns;
3324 }
3325 EXPORT_SYMBOL_GPL(ring_buffer_overruns);
3326
3327 static void rb_iter_reset(struct ring_buffer_iter *iter)
3328 {
3329         struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3330
3331         /* Iterator usage is expected to have record disabled */
3332         iter->head_page = cpu_buffer->reader_page;
3333         iter->head = cpu_buffer->reader_page->read;
3334
3335         iter->cache_reader_page = iter->head_page;
3336         iter->cache_read = cpu_buffer->read;
3337
3338         if (iter->head)
3339                 iter->read_stamp = cpu_buffer->read_stamp;
3340         else
3341                 iter->read_stamp = iter->head_page->page->time_stamp;
3342 }
3343
3344 /**
3345  * ring_buffer_iter_reset - reset an iterator
3346  * @iter: The iterator to reset
3347  *
3348  * Resets the iterator, so that it will start from the beginning
3349  * again.
3350  */
3351 void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
3352 {
3353         struct ring_buffer_per_cpu *cpu_buffer;
3354         unsigned long flags;
3355
3356         if (!iter)
3357                 return;
3358
3359         cpu_buffer = iter->cpu_buffer;
3360
3361         raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3362         rb_iter_reset(iter);
3363         raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3364 }
3365 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
3366
3367 /**
3368  * ring_buffer_iter_empty - check if an iterator has no more to read
3369  * @iter: The iterator to check
3370  */
3371 int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
3372 {
3373         struct ring_buffer_per_cpu *cpu_buffer;
3374         struct buffer_page *reader;
3375         struct buffer_page *head_page;
3376         struct buffer_page *commit_page;
3377         unsigned commit;
3378
3379         cpu_buffer = iter->cpu_buffer;
3380
3381         /* Remember, trace recording is off when iterator is in use */
3382         reader = cpu_buffer->reader_page;
3383         head_page = cpu_buffer->head_page;
3384         commit_page = cpu_buffer->commit_page;
3385         commit = rb_page_commit(commit_page);
3386
3387         return ((iter->head_page == commit_page && iter->head == commit) ||
3388                 (iter->head_page == reader && commit_page == head_page &&
3389                  head_page->read == commit &&
3390                  iter->head == rb_page_commit(cpu_buffer->reader_page)));
3391 }
3392 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
3393
3394 static void
3395 rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
3396                      struct ring_buffer_event *event)
3397 {
3398         u64 delta;
3399
3400         switch (event->type_len) {
3401         case RINGBUF_TYPE_PADDING:
3402                 return;
3403
3404         case RINGBUF_TYPE_TIME_EXTEND:
3405                 delta = event->array[0];
3406                 delta <<= TS_SHIFT;
3407                 delta += event->time_delta;
3408                 cpu_buffer->read_stamp += delta;
3409                 return;
3410
3411         case RINGBUF_TYPE_TIME_STAMP:
3412                 /* FIXME: not implemented */
3413                 return;
3414
3415         case RINGBUF_TYPE_DATA:
3416                 cpu_buffer->read_stamp += event->time_delta;
3417                 return;
3418
3419         default:
3420                 BUG();
3421         }
3422         return;
3423 }
3424
3425 static void
3426 rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
3427                           struct ring_buffer_event *event)
3428 {
3429         u64 delta;
3430
3431         switch (event->type_len) {
3432         case RINGBUF_TYPE_PADDING:
3433                 return;
3434
3435         case RINGBUF_TYPE_TIME_EXTEND:
3436                 delta = event->array[0];
3437                 delta <<= TS_SHIFT;
3438                 delta += event->time_delta;
3439                 iter->read_stamp += delta;
3440                 return;
3441
3442         case RINGBUF_TYPE_TIME_STAMP:
3443                 /* FIXME: not implemented */
3444                 return;
3445
3446         case RINGBUF_TYPE_DATA:
3447                 iter->read_stamp += event->time_delta;
3448                 return;
3449
3450         default:
3451                 BUG();
3452         }</