Linux 5.2
[sfrench/cifs-2.6.git] / kernel / trace / ring_buffer.c
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Generic ring buffer
4  *
5  * Copyright (C) 2008 Steven Rostedt <srostedt@redhat.com>
6  */
7 #include <linux/trace_events.h>
8 #include <linux/ring_buffer.h>
9 #include <linux/trace_clock.h>
10 #include <linux/sched/clock.h>
11 #include <linux/trace_seq.h>
12 #include <linux/spinlock.h>
13 #include <linux/irq_work.h>
14 #include <linux/uaccess.h>
15 #include <linux/hardirq.h>
16 #include <linux/kthread.h>      /* for self test */
17 #include <linux/module.h>
18 #include <linux/percpu.h>
19 #include <linux/mutex.h>
20 #include <linux/delay.h>
21 #include <linux/slab.h>
22 #include <linux/init.h>
23 #include <linux/hash.h>
24 #include <linux/list.h>
25 #include <linux/cpu.h>
26 #include <linux/oom.h>
27
28 #include <asm/local.h>
29
30 static void update_pages_handler(struct work_struct *work);
31
32 /*
33  * The ring buffer header is special. We must manually up keep it.
34  */
35 int ring_buffer_print_entry_header(struct trace_seq *s)
36 {
37         trace_seq_puts(s, "# compressed entry header\n");
38         trace_seq_puts(s, "\ttype_len    :    5 bits\n");
39         trace_seq_puts(s, "\ttime_delta  :   27 bits\n");
40         trace_seq_puts(s, "\tarray       :   32 bits\n");
41         trace_seq_putc(s, '\n');
42         trace_seq_printf(s, "\tpadding     : type == %d\n",
43                          RINGBUF_TYPE_PADDING);
44         trace_seq_printf(s, "\ttime_extend : type == %d\n",
45                          RINGBUF_TYPE_TIME_EXTEND);
46         trace_seq_printf(s, "\ttime_stamp : type == %d\n",
47                          RINGBUF_TYPE_TIME_STAMP);
48         trace_seq_printf(s, "\tdata max type_len  == %d\n",
49                          RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
50
51         return !trace_seq_has_overflowed(s);
52 }
53
54 /*
55  * The ring buffer is made up of a list of pages. A separate list of pages is
56  * allocated for each CPU. A writer may only write to a buffer that is
57  * associated with the CPU it is currently executing on.  A reader may read
58  * from any per cpu buffer.
59  *
60  * The reader is special. For each per cpu buffer, the reader has its own
61  * reader page. When a reader has read the entire reader page, this reader
62  * page is swapped with another page in the ring buffer.
63  *
64  * Now, as long as the writer is off the reader page, the reader can do what
65  * ever it wants with that page. The writer will never write to that page
66  * again (as long as it is out of the ring buffer).
67  *
68  * Here's some silly ASCII art.
69  *
70  *   +------+
71  *   |reader|          RING BUFFER
72  *   |page  |
73  *   +------+        +---+   +---+   +---+
74  *                   |   |-->|   |-->|   |
75  *                   +---+   +---+   +---+
76  *                     ^               |
77  *                     |               |
78  *                     +---------------+
79  *
80  *
81  *   +------+
82  *   |reader|          RING BUFFER
83  *   |page  |------------------v
84  *   +------+        +---+   +---+   +---+
85  *                   |   |-->|   |-->|   |
86  *                   +---+   +---+   +---+
87  *                     ^               |
88  *                     |               |
89  *                     +---------------+
90  *
91  *
92  *   +------+
93  *   |reader|          RING BUFFER
94  *   |page  |------------------v
95  *   +------+        +---+   +---+   +---+
96  *      ^            |   |-->|   |-->|   |
97  *      |            +---+   +---+   +---+
98  *      |                              |
99  *      |                              |
100  *      +------------------------------+
101  *
102  *
103  *   +------+
104  *   |buffer|          RING BUFFER
105  *   |page  |------------------v
106  *   +------+        +---+   +---+   +---+
107  *      ^            |   |   |   |-->|   |
108  *      |   New      +---+   +---+   +---+
109  *      |  Reader------^               |
110  *      |   page                       |
111  *      +------------------------------+
112  *
113  *
114  * After we make this swap, the reader can hand this page off to the splice
115  * code and be done with it. It can even allocate a new page if it needs to
116  * and swap that into the ring buffer.
117  *
118  * We will be using cmpxchg soon to make all this lockless.
119  *
120  */
121
122 /* Used for individual buffers (after the counter) */
123 #define RB_BUFFER_OFF           (1 << 20)
124
125 #define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
126
127 #define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
128 #define RB_ALIGNMENT            4U
129 #define RB_MAX_SMALL_DATA       (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
130 #define RB_EVNT_MIN_SIZE        8U      /* two 32bit words */
131
132 #ifndef CONFIG_HAVE_64BIT_ALIGNED_ACCESS
133 # define RB_FORCE_8BYTE_ALIGNMENT       0
134 # define RB_ARCH_ALIGNMENT              RB_ALIGNMENT
135 #else
136 # define RB_FORCE_8BYTE_ALIGNMENT       1
137 # define RB_ARCH_ALIGNMENT              8U
138 #endif
139
140 #define RB_ALIGN_DATA           __aligned(RB_ARCH_ALIGNMENT)
141
142 /* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
143 #define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
144
145 enum {
146         RB_LEN_TIME_EXTEND = 8,
147         RB_LEN_TIME_STAMP =  8,
148 };
149
150 #define skip_time_extend(event) \
151         ((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
152
153 #define extended_time(event) \
154         (event->type_len >= RINGBUF_TYPE_TIME_EXTEND)
155
156 static inline int rb_null_event(struct ring_buffer_event *event)
157 {
158         return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
159 }
160
161 static void rb_event_set_padding(struct ring_buffer_event *event)
162 {
163         /* padding has a NULL time_delta */
164         event->type_len = RINGBUF_TYPE_PADDING;
165         event->time_delta = 0;
166 }
167
168 static unsigned
169 rb_event_data_length(struct ring_buffer_event *event)
170 {
171         unsigned length;
172
173         if (event->type_len)
174                 length = event->type_len * RB_ALIGNMENT;
175         else
176                 length = event->array[0];
177         return length + RB_EVNT_HDR_SIZE;
178 }
179
180 /*
181  * Return the length of the given event. Will return
182  * the length of the time extend if the event is a
183  * time extend.
184  */
185 static inline unsigned
186 rb_event_length(struct ring_buffer_event *event)
187 {
188         switch (event->type_len) {
189         case RINGBUF_TYPE_PADDING:
190                 if (rb_null_event(event))
191                         /* undefined */
192                         return -1;
193                 return  event->array[0] + RB_EVNT_HDR_SIZE;
194
195         case RINGBUF_TYPE_TIME_EXTEND:
196                 return RB_LEN_TIME_EXTEND;
197
198         case RINGBUF_TYPE_TIME_STAMP:
199                 return RB_LEN_TIME_STAMP;
200
201         case RINGBUF_TYPE_DATA:
202                 return rb_event_data_length(event);
203         default:
204                 BUG();
205         }
206         /* not hit */
207         return 0;
208 }
209
210 /*
211  * Return total length of time extend and data,
212  *   or just the event length for all other events.
213  */
214 static inline unsigned
215 rb_event_ts_length(struct ring_buffer_event *event)
216 {
217         unsigned len = 0;
218
219         if (extended_time(event)) {
220                 /* time extends include the data event after it */
221                 len = RB_LEN_TIME_EXTEND;
222                 event = skip_time_extend(event);
223         }
224         return len + rb_event_length(event);
225 }
226
227 /**
228  * ring_buffer_event_length - return the length of the event
229  * @event: the event to get the length of
230  *
231  * Returns the size of the data load of a data event.
232  * If the event is something other than a data event, it
233  * returns the size of the event itself. With the exception
234  * of a TIME EXTEND, where it still returns the size of the
235  * data load of the data event after it.
236  */
237 unsigned ring_buffer_event_length(struct ring_buffer_event *event)
238 {
239         unsigned length;
240
241         if (extended_time(event))
242                 event = skip_time_extend(event);
243
244         length = rb_event_length(event);
245         if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
246                 return length;
247         length -= RB_EVNT_HDR_SIZE;
248         if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
249                 length -= sizeof(event->array[0]);
250         return length;
251 }
252 EXPORT_SYMBOL_GPL(ring_buffer_event_length);
253
254 /* inline for ring buffer fast paths */
255 static __always_inline void *
256 rb_event_data(struct ring_buffer_event *event)
257 {
258         if (extended_time(event))
259                 event = skip_time_extend(event);
260         BUG_ON(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
261         /* If length is in len field, then array[0] has the data */
262         if (event->type_len)
263                 return (void *)&event->array[0];
264         /* Otherwise length is in array[0] and array[1] has the data */
265         return (void *)&event->array[1];
266 }
267
268 /**
269  * ring_buffer_event_data - return the data of the event
270  * @event: the event to get the data from
271  */
272 void *ring_buffer_event_data(struct ring_buffer_event *event)
273 {
274         return rb_event_data(event);
275 }
276 EXPORT_SYMBOL_GPL(ring_buffer_event_data);
277
278 #define for_each_buffer_cpu(buffer, cpu)                \
279         for_each_cpu(cpu, buffer->cpumask)
280
281 #define TS_SHIFT        27
282 #define TS_MASK         ((1ULL << TS_SHIFT) - 1)
283 #define TS_DELTA_TEST   (~TS_MASK)
284
285 /**
286  * ring_buffer_event_time_stamp - return the event's extended timestamp
287  * @event: the event to get the timestamp of
288  *
289  * Returns the extended timestamp associated with a data event.
290  * An extended time_stamp is a 64-bit timestamp represented
291  * internally in a special way that makes the best use of space
292  * contained within a ring buffer event.  This function decodes
293  * it and maps it to a straight u64 value.
294  */
295 u64 ring_buffer_event_time_stamp(struct ring_buffer_event *event)
296 {
297         u64 ts;
298
299         ts = event->array[0];
300         ts <<= TS_SHIFT;
301         ts += event->time_delta;
302
303         return ts;
304 }
305
306 /* Flag when events were overwritten */
307 #define RB_MISSED_EVENTS        (1 << 31)
308 /* Missed count stored at end */
309 #define RB_MISSED_STORED        (1 << 30)
310
311 #define RB_MISSED_FLAGS         (RB_MISSED_EVENTS|RB_MISSED_STORED)
312
313 struct buffer_data_page {
314         u64              time_stamp;    /* page time stamp */
315         local_t          commit;        /* write committed index */
316         unsigned char    data[] RB_ALIGN_DATA;  /* data of buffer page */
317 };
318
319 /*
320  * Note, the buffer_page list must be first. The buffer pages
321  * are allocated in cache lines, which means that each buffer
322  * page will be at the beginning of a cache line, and thus
323  * the least significant bits will be zero. We use this to
324  * add flags in the list struct pointers, to make the ring buffer
325  * lockless.
326  */
327 struct buffer_page {
328         struct list_head list;          /* list of buffer pages */
329         local_t          write;         /* index for next write */
330         unsigned         read;          /* index for next read */
331         local_t          entries;       /* entries on this page */
332         unsigned long    real_end;      /* real end of data */
333         struct buffer_data_page *page;  /* Actual data page */
334 };
335
336 /*
337  * The buffer page counters, write and entries, must be reset
338  * atomically when crossing page boundaries. To synchronize this
339  * update, two counters are inserted into the number. One is
340  * the actual counter for the write position or count on the page.
341  *
342  * The other is a counter of updaters. Before an update happens
343  * the update partition of the counter is incremented. This will
344  * allow the updater to update the counter atomically.
345  *
346  * The counter is 20 bits, and the state data is 12.
347  */
348 #define RB_WRITE_MASK           0xfffff
349 #define RB_WRITE_INTCNT         (1 << 20)
350
351 static void rb_init_page(struct buffer_data_page *bpage)
352 {
353         local_set(&bpage->commit, 0);
354 }
355
356 /*
357  * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
358  * this issue out.
359  */
360 static void free_buffer_page(struct buffer_page *bpage)
361 {
362         free_page((unsigned long)bpage->page);
363         kfree(bpage);
364 }
365
366 /*
367  * We need to fit the time_stamp delta into 27 bits.
368  */
369 static inline int test_time_stamp(u64 delta)
370 {
371         if (delta & TS_DELTA_TEST)
372                 return 1;
373         return 0;
374 }
375
376 #define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
377
378 /* Max payload is BUF_PAGE_SIZE - header (8bytes) */
379 #define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
380
381 int ring_buffer_print_page_header(struct trace_seq *s)
382 {
383         struct buffer_data_page field;
384
385         trace_seq_printf(s, "\tfield: u64 timestamp;\t"
386                          "offset:0;\tsize:%u;\tsigned:%u;\n",
387                          (unsigned int)sizeof(field.time_stamp),
388                          (unsigned int)is_signed_type(u64));
389
390         trace_seq_printf(s, "\tfield: local_t commit;\t"
391                          "offset:%u;\tsize:%u;\tsigned:%u;\n",
392                          (unsigned int)offsetof(typeof(field), commit),
393                          (unsigned int)sizeof(field.commit),
394                          (unsigned int)is_signed_type(long));
395
396         trace_seq_printf(s, "\tfield: int overwrite;\t"
397                          "offset:%u;\tsize:%u;\tsigned:%u;\n",
398                          (unsigned int)offsetof(typeof(field), commit),
399                          1,
400                          (unsigned int)is_signed_type(long));
401
402         trace_seq_printf(s, "\tfield: char data;\t"
403                          "offset:%u;\tsize:%u;\tsigned:%u;\n",
404                          (unsigned int)offsetof(typeof(field), data),
405                          (unsigned int)BUF_PAGE_SIZE,
406                          (unsigned int)is_signed_type(char));
407
408         return !trace_seq_has_overflowed(s);
409 }
410
411 struct rb_irq_work {
412         struct irq_work                 work;
413         wait_queue_head_t               waiters;
414         wait_queue_head_t               full_waiters;
415         bool                            waiters_pending;
416         bool                            full_waiters_pending;
417         bool                            wakeup_full;
418 };
419
420 /*
421  * Structure to hold event state and handle nested events.
422  */
423 struct rb_event_info {
424         u64                     ts;
425         u64                     delta;
426         unsigned long           length;
427         struct buffer_page      *tail_page;
428         int                     add_timestamp;
429 };
430
431 /*
432  * Used for which event context the event is in.
433  *  NMI     = 0
434  *  IRQ     = 1
435  *  SOFTIRQ = 2
436  *  NORMAL  = 3
437  *
438  * See trace_recursive_lock() comment below for more details.
439  */
440 enum {
441         RB_CTX_NMI,
442         RB_CTX_IRQ,
443         RB_CTX_SOFTIRQ,
444         RB_CTX_NORMAL,
445         RB_CTX_MAX
446 };
447
448 /*
449  * head_page == tail_page && head == tail then buffer is empty.
450  */
451 struct ring_buffer_per_cpu {
452         int                             cpu;
453         atomic_t                        record_disabled;
454         struct ring_buffer              *buffer;
455         raw_spinlock_t                  reader_lock;    /* serialize readers */
456         arch_spinlock_t                 lock;
457         struct lock_class_key           lock_key;
458         struct buffer_data_page         *free_page;
459         unsigned long                   nr_pages;
460         unsigned int                    current_context;
461         struct list_head                *pages;
462         struct buffer_page              *head_page;     /* read from head */
463         struct buffer_page              *tail_page;     /* write to tail */
464         struct buffer_page              *commit_page;   /* committed pages */
465         struct buffer_page              *reader_page;
466         unsigned long                   lost_events;
467         unsigned long                   last_overrun;
468         unsigned long                   nest;
469         local_t                         entries_bytes;
470         local_t                         entries;
471         local_t                         overrun;
472         local_t                         commit_overrun;
473         local_t                         dropped_events;
474         local_t                         committing;
475         local_t                         commits;
476         local_t                         pages_touched;
477         local_t                         pages_read;
478         long                            last_pages_touch;
479         size_t                          shortest_full;
480         unsigned long                   read;
481         unsigned long                   read_bytes;
482         u64                             write_stamp;
483         u64                             read_stamp;
484         /* ring buffer pages to update, > 0 to add, < 0 to remove */
485         long                            nr_pages_to_update;
486         struct list_head                new_pages; /* new pages to add */
487         struct work_struct              update_pages_work;
488         struct completion               update_done;
489
490         struct rb_irq_work              irq_work;
491 };
492
493 struct ring_buffer {
494         unsigned                        flags;
495         int                             cpus;
496         atomic_t                        record_disabled;
497         atomic_t                        resize_disabled;
498         cpumask_var_t                   cpumask;
499
500         struct lock_class_key           *reader_lock_key;
501
502         struct mutex                    mutex;
503
504         struct ring_buffer_per_cpu      **buffers;
505
506         struct hlist_node               node;
507         u64                             (*clock)(void);
508
509         struct rb_irq_work              irq_work;
510         bool                            time_stamp_abs;
511 };
512
513 struct ring_buffer_iter {
514         struct ring_buffer_per_cpu      *cpu_buffer;
515         unsigned long                   head;
516         struct buffer_page              *head_page;
517         struct buffer_page              *cache_reader_page;
518         unsigned long                   cache_read;
519         u64                             read_stamp;
520 };
521
522 /**
523  * ring_buffer_nr_pages - get the number of buffer pages in the ring buffer
524  * @buffer: The ring_buffer to get the number of pages from
525  * @cpu: The cpu of the ring_buffer to get the number of pages from
526  *
527  * Returns the number of pages used by a per_cpu buffer of the ring buffer.
528  */
529 size_t ring_buffer_nr_pages(struct ring_buffer *buffer, int cpu)
530 {
531         return buffer->buffers[cpu]->nr_pages;
532 }
533
534 /**
535  * ring_buffer_nr_pages_dirty - get the number of used pages in the ring buffer
536  * @buffer: The ring_buffer to get the number of pages from
537  * @cpu: The cpu of the ring_buffer to get the number of pages from
538  *
539  * Returns the number of pages that have content in the ring buffer.
540  */
541 size_t ring_buffer_nr_dirty_pages(struct ring_buffer *buffer, int cpu)
542 {
543         size_t read;
544         size_t cnt;
545
546         read = local_read(&buffer->buffers[cpu]->pages_read);
547         cnt = local_read(&buffer->buffers[cpu]->pages_touched);
548         /* The reader can read an empty page, but not more than that */
549         if (cnt < read) {
550                 WARN_ON_ONCE(read > cnt + 1);
551                 return 0;
552         }
553
554         return cnt - read;
555 }
556
557 /*
558  * rb_wake_up_waiters - wake up tasks waiting for ring buffer input
559  *
560  * Schedules a delayed work to wake up any task that is blocked on the
561  * ring buffer waiters queue.
562  */
563 static void rb_wake_up_waiters(struct irq_work *work)
564 {
565         struct rb_irq_work *rbwork = container_of(work, struct rb_irq_work, work);
566
567         wake_up_all(&rbwork->waiters);
568         if (rbwork->wakeup_full) {
569                 rbwork->wakeup_full = false;
570                 wake_up_all(&rbwork->full_waiters);
571         }
572 }
573
574 /**
575  * ring_buffer_wait - wait for input to the ring buffer
576  * @buffer: buffer to wait on
577  * @cpu: the cpu buffer to wait on
578  * @full: wait until a full page is available, if @cpu != RING_BUFFER_ALL_CPUS
579  *
580  * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
581  * as data is added to any of the @buffer's cpu buffers. Otherwise
582  * it will wait for data to be added to a specific cpu buffer.
583  */
584 int ring_buffer_wait(struct ring_buffer *buffer, int cpu, int full)
585 {
586         struct ring_buffer_per_cpu *uninitialized_var(cpu_buffer);
587         DEFINE_WAIT(wait);
588         struct rb_irq_work *work;
589         int ret = 0;
590
591         /*
592          * Depending on what the caller is waiting for, either any
593          * data in any cpu buffer, or a specific buffer, put the
594          * caller on the appropriate wait queue.
595          */
596         if (cpu == RING_BUFFER_ALL_CPUS) {
597                 work = &buffer->irq_work;
598                 /* Full only makes sense on per cpu reads */
599                 full = 0;
600         } else {
601                 if (!cpumask_test_cpu(cpu, buffer->cpumask))
602                         return -ENODEV;
603                 cpu_buffer = buffer->buffers[cpu];
604                 work = &cpu_buffer->irq_work;
605         }
606
607
608         while (true) {
609                 if (full)
610                         prepare_to_wait(&work->full_waiters, &wait, TASK_INTERRUPTIBLE);
611                 else
612                         prepare_to_wait(&work->waiters, &wait, TASK_INTERRUPTIBLE);
613
614                 /*
615                  * The events can happen in critical sections where
616                  * checking a work queue can cause deadlocks.
617                  * After adding a task to the queue, this flag is set
618                  * only to notify events to try to wake up the queue
619                  * using irq_work.
620                  *
621                  * We don't clear it even if the buffer is no longer
622                  * empty. The flag only causes the next event to run
623                  * irq_work to do the work queue wake up. The worse
624                  * that can happen if we race with !trace_empty() is that
625                  * an event will cause an irq_work to try to wake up
626                  * an empty queue.
627                  *
628                  * There's no reason to protect this flag either, as
629                  * the work queue and irq_work logic will do the necessary
630                  * synchronization for the wake ups. The only thing
631                  * that is necessary is that the wake up happens after
632                  * a task has been queued. It's OK for spurious wake ups.
633                  */
634                 if (full)
635                         work->full_waiters_pending = true;
636                 else
637                         work->waiters_pending = true;
638
639                 if (signal_pending(current)) {
640                         ret = -EINTR;
641                         break;
642                 }
643
644                 if (cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer))
645                         break;
646
647                 if (cpu != RING_BUFFER_ALL_CPUS &&
648                     !ring_buffer_empty_cpu(buffer, cpu)) {
649                         unsigned long flags;
650                         bool pagebusy;
651                         size_t nr_pages;
652                         size_t dirty;
653
654                         if (!full)
655                                 break;
656
657                         raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
658                         pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
659                         nr_pages = cpu_buffer->nr_pages;
660                         dirty = ring_buffer_nr_dirty_pages(buffer, cpu);
661                         if (!cpu_buffer->shortest_full ||
662                             cpu_buffer->shortest_full < full)
663                                 cpu_buffer->shortest_full = full;
664                         raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
665                         if (!pagebusy &&
666                             (!nr_pages || (dirty * 100) > full * nr_pages))
667                                 break;
668                 }
669
670                 schedule();
671         }
672
673         if (full)
674                 finish_wait(&work->full_waiters, &wait);
675         else
676                 finish_wait(&work->waiters, &wait);
677
678         return ret;
679 }
680
681 /**
682  * ring_buffer_poll_wait - poll on buffer input
683  * @buffer: buffer to wait on
684  * @cpu: the cpu buffer to wait on
685  * @filp: the file descriptor
686  * @poll_table: The poll descriptor
687  *
688  * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
689  * as data is added to any of the @buffer's cpu buffers. Otherwise
690  * it will wait for data to be added to a specific cpu buffer.
691  *
692  * Returns EPOLLIN | EPOLLRDNORM if data exists in the buffers,
693  * zero otherwise.
694  */
695 __poll_t ring_buffer_poll_wait(struct ring_buffer *buffer, int cpu,
696                           struct file *filp, poll_table *poll_table)
697 {
698         struct ring_buffer_per_cpu *cpu_buffer;
699         struct rb_irq_work *work;
700
701         if (cpu == RING_BUFFER_ALL_CPUS)
702                 work = &buffer->irq_work;
703         else {
704                 if (!cpumask_test_cpu(cpu, buffer->cpumask))
705                         return -EINVAL;
706
707                 cpu_buffer = buffer->buffers[cpu];
708                 work = &cpu_buffer->irq_work;
709         }
710
711         poll_wait(filp, &work->waiters, poll_table);
712         work->waiters_pending = true;
713         /*
714          * There's a tight race between setting the waiters_pending and
715          * checking if the ring buffer is empty.  Once the waiters_pending bit
716          * is set, the next event will wake the task up, but we can get stuck
717          * if there's only a single event in.
718          *
719          * FIXME: Ideally, we need a memory barrier on the writer side as well,
720          * but adding a memory barrier to all events will cause too much of a
721          * performance hit in the fast path.  We only need a memory barrier when
722          * the buffer goes from empty to having content.  But as this race is
723          * extremely small, and it's not a problem if another event comes in, we
724          * will fix it later.
725          */
726         smp_mb();
727
728         if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
729             (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
730                 return EPOLLIN | EPOLLRDNORM;
731         return 0;
732 }
733
734 /* buffer may be either ring_buffer or ring_buffer_per_cpu */
735 #define RB_WARN_ON(b, cond)                                             \
736         ({                                                              \
737                 int _____ret = unlikely(cond);                          \
738                 if (_____ret) {                                         \
739                         if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
740                                 struct ring_buffer_per_cpu *__b =       \
741                                         (void *)b;                      \
742                                 atomic_inc(&__b->buffer->record_disabled); \
743                         } else                                          \
744                                 atomic_inc(&b->record_disabled);        \
745                         WARN_ON(1);                                     \
746                 }                                                       \
747                 _____ret;                                               \
748         })
749
750 /* Up this if you want to test the TIME_EXTENTS and normalization */
751 #define DEBUG_SHIFT 0
752
753 static inline u64 rb_time_stamp(struct ring_buffer *buffer)
754 {
755         /* shift to debug/test normalization and TIME_EXTENTS */
756         return buffer->clock() << DEBUG_SHIFT;
757 }
758
759 u64 ring_buffer_time_stamp(struct ring_buffer *buffer, int cpu)
760 {
761         u64 time;
762
763         preempt_disable_notrace();
764         time = rb_time_stamp(buffer);
765         preempt_enable_no_resched_notrace();
766
767         return time;
768 }
769 EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
770
771 void ring_buffer_normalize_time_stamp(struct ring_buffer *buffer,
772                                       int cpu, u64 *ts)
773 {
774         /* Just stupid testing the normalize function and deltas */
775         *ts >>= DEBUG_SHIFT;
776 }
777 EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
778
779 /*
780  * Making the ring buffer lockless makes things tricky.
781  * Although writes only happen on the CPU that they are on,
782  * and they only need to worry about interrupts. Reads can
783  * happen on any CPU.
784  *
785  * The reader page is always off the ring buffer, but when the
786  * reader finishes with a page, it needs to swap its page with
787  * a new one from the buffer. The reader needs to take from
788  * the head (writes go to the tail). But if a writer is in overwrite
789  * mode and wraps, it must push the head page forward.
790  *
791  * Here lies the problem.
792  *
793  * The reader must be careful to replace only the head page, and
794  * not another one. As described at the top of the file in the
795  * ASCII art, the reader sets its old page to point to the next
796  * page after head. It then sets the page after head to point to
797  * the old reader page. But if the writer moves the head page
798  * during this operation, the reader could end up with the tail.
799  *
800  * We use cmpxchg to help prevent this race. We also do something
801  * special with the page before head. We set the LSB to 1.
802  *
803  * When the writer must push the page forward, it will clear the
804  * bit that points to the head page, move the head, and then set
805  * the bit that points to the new head page.
806  *
807  * We also don't want an interrupt coming in and moving the head
808  * page on another writer. Thus we use the second LSB to catch
809  * that too. Thus:
810  *
811  * head->list->prev->next        bit 1          bit 0
812  *                              -------        -------
813  * Normal page                     0              0
814  * Points to head page             0              1
815  * New head page                   1              0
816  *
817  * Note we can not trust the prev pointer of the head page, because:
818  *
819  * +----+       +-----+        +-----+
820  * |    |------>|  T  |---X--->|  N  |
821  * |    |<------|     |        |     |
822  * +----+       +-----+        +-----+
823  *   ^                           ^ |
824  *   |          +-----+          | |
825  *   +----------|  R  |----------+ |
826  *              |     |<-----------+
827  *              +-----+
828  *
829  * Key:  ---X-->  HEAD flag set in pointer
830  *         T      Tail page
831  *         R      Reader page
832  *         N      Next page
833  *
834  * (see __rb_reserve_next() to see where this happens)
835  *
836  *  What the above shows is that the reader just swapped out
837  *  the reader page with a page in the buffer, but before it
838  *  could make the new header point back to the new page added
839  *  it was preempted by a writer. The writer moved forward onto
840  *  the new page added by the reader and is about to move forward
841  *  again.
842  *
843  *  You can see, it is legitimate for the previous pointer of
844  *  the head (or any page) not to point back to itself. But only
845  *  temporarily.
846  */
847
848 #define RB_PAGE_NORMAL          0UL
849 #define RB_PAGE_HEAD            1UL
850 #define RB_PAGE_UPDATE          2UL
851
852
853 #define RB_FLAG_MASK            3UL
854
855 /* PAGE_MOVED is not part of the mask */
856 #define RB_PAGE_MOVED           4UL
857
858 /*
859  * rb_list_head - remove any bit
860  */
861 static struct list_head *rb_list_head(struct list_head *list)
862 {
863         unsigned long val = (unsigned long)list;
864
865         return (struct list_head *)(val & ~RB_FLAG_MASK);
866 }
867
868 /*
869  * rb_is_head_page - test if the given page is the head page
870  *
871  * Because the reader may move the head_page pointer, we can
872  * not trust what the head page is (it may be pointing to
873  * the reader page). But if the next page is a header page,
874  * its flags will be non zero.
875  */
876 static inline int
877 rb_is_head_page(struct ring_buffer_per_cpu *cpu_buffer,
878                 struct buffer_page *page, struct list_head *list)
879 {
880         unsigned long val;
881
882         val = (unsigned long)list->next;
883
884         if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
885                 return RB_PAGE_MOVED;
886
887         return val & RB_FLAG_MASK;
888 }
889
890 /*
891  * rb_is_reader_page
892  *
893  * The unique thing about the reader page, is that, if the
894  * writer is ever on it, the previous pointer never points
895  * back to the reader page.
896  */
897 static bool rb_is_reader_page(struct buffer_page *page)
898 {
899         struct list_head *list = page->list.prev;
900
901         return rb_list_head(list->next) != &page->list;
902 }
903
904 /*
905  * rb_set_list_to_head - set a list_head to be pointing to head.
906  */
907 static void rb_set_list_to_head(struct ring_buffer_per_cpu *cpu_buffer,
908                                 struct list_head *list)
909 {
910         unsigned long *ptr;
911
912         ptr = (unsigned long *)&list->next;
913         *ptr |= RB_PAGE_HEAD;
914         *ptr &= ~RB_PAGE_UPDATE;
915 }
916
917 /*
918  * rb_head_page_activate - sets up head page
919  */
920 static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
921 {
922         struct buffer_page *head;
923
924         head = cpu_buffer->head_page;
925         if (!head)
926                 return;
927
928         /*
929          * Set the previous list pointer to have the HEAD flag.
930          */
931         rb_set_list_to_head(cpu_buffer, head->list.prev);
932 }
933
934 static void rb_list_head_clear(struct list_head *list)
935 {
936         unsigned long *ptr = (unsigned long *)&list->next;
937
938         *ptr &= ~RB_FLAG_MASK;
939 }
940
941 /*
942  * rb_head_page_deactivate - clears head page ptr (for free list)
943  */
944 static void
945 rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
946 {
947         struct list_head *hd;
948
949         /* Go through the whole list and clear any pointers found. */
950         rb_list_head_clear(cpu_buffer->pages);
951
952         list_for_each(hd, cpu_buffer->pages)
953                 rb_list_head_clear(hd);
954 }
955
956 static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
957                             struct buffer_page *head,
958                             struct buffer_page *prev,
959                             int old_flag, int new_flag)
960 {
961         struct list_head *list;
962         unsigned long val = (unsigned long)&head->list;
963         unsigned long ret;
964
965         list = &prev->list;
966
967         val &= ~RB_FLAG_MASK;
968
969         ret = cmpxchg((unsigned long *)&list->next,
970                       val | old_flag, val | new_flag);
971
972         /* check if the reader took the page */
973         if ((ret & ~RB_FLAG_MASK) != val)
974                 return RB_PAGE_MOVED;
975
976         return ret & RB_FLAG_MASK;
977 }
978
979 static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
980                                    struct buffer_page *head,
981                                    struct buffer_page *prev,
982                                    int old_flag)
983 {
984         return rb_head_page_set(cpu_buffer, head, prev,
985                                 old_flag, RB_PAGE_UPDATE);
986 }
987
988 static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
989                                  struct buffer_page *head,
990                                  struct buffer_page *prev,
991                                  int old_flag)
992 {
993         return rb_head_page_set(cpu_buffer, head, prev,
994                                 old_flag, RB_PAGE_HEAD);
995 }
996
997 static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
998                                    struct buffer_page *head,
999                                    struct buffer_page *prev,
1000                                    int old_flag)
1001 {
1002         return rb_head_page_set(cpu_buffer, head, prev,
1003                                 old_flag, RB_PAGE_NORMAL);
1004 }
1005
1006 static inline void rb_inc_page(struct ring_buffer_per_cpu *cpu_buffer,
1007                                struct buffer_page **bpage)
1008 {
1009         struct list_head *p = rb_list_head((*bpage)->list.next);
1010
1011         *bpage = list_entry(p, struct buffer_page, list);
1012 }
1013
1014 static struct buffer_page *
1015 rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
1016 {
1017         struct buffer_page *head;
1018         struct buffer_page *page;
1019         struct list_head *list;
1020         int i;
1021
1022         if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
1023                 return NULL;
1024
1025         /* sanity check */
1026         list = cpu_buffer->pages;
1027         if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
1028                 return NULL;
1029
1030         page = head = cpu_buffer->head_page;
1031         /*
1032          * It is possible that the writer moves the header behind
1033          * where we started, and we miss in one loop.
1034          * A second loop should grab the header, but we'll do
1035          * three loops just because I'm paranoid.
1036          */
1037         for (i = 0; i < 3; i++) {
1038                 do {
1039                         if (rb_is_head_page(cpu_buffer, page, page->list.prev)) {
1040                                 cpu_buffer->head_page = page;
1041                                 return page;
1042                         }
1043                         rb_inc_page(cpu_buffer, &page);
1044                 } while (page != head);
1045         }
1046
1047         RB_WARN_ON(cpu_buffer, 1);
1048
1049         return NULL;
1050 }
1051
1052 static int rb_head_page_replace(struct buffer_page *old,
1053                                 struct buffer_page *new)
1054 {
1055         unsigned long *ptr = (unsigned long *)&old->list.prev->next;
1056         unsigned long val;
1057         unsigned long ret;
1058
1059         val = *ptr & ~RB_FLAG_MASK;
1060         val |= RB_PAGE_HEAD;
1061
1062         ret = cmpxchg(ptr, val, (unsigned long)&new->list);
1063
1064         return ret == val;
1065 }
1066
1067 /*
1068  * rb_tail_page_update - move the tail page forward
1069  */
1070 static void rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
1071                                struct buffer_page *tail_page,
1072                                struct buffer_page *next_page)
1073 {
1074         unsigned long old_entries;
1075         unsigned long old_write;
1076
1077         /*
1078          * The tail page now needs to be moved forward.
1079          *
1080          * We need to reset the tail page, but without messing
1081          * with possible erasing of data brought in by interrupts
1082          * that have moved the tail page and are currently on it.
1083          *
1084          * We add a counter to the write field to denote this.
1085          */
1086         old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
1087         old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
1088
1089         local_inc(&cpu_buffer->pages_touched);
1090         /*
1091          * Just make sure we have seen our old_write and synchronize
1092          * with any interrupts that come in.
1093          */
1094         barrier();
1095
1096         /*
1097          * If the tail page is still the same as what we think
1098          * it is, then it is up to us to update the tail
1099          * pointer.
1100          */
1101         if (tail_page == READ_ONCE(cpu_buffer->tail_page)) {
1102                 /* Zero the write counter */
1103                 unsigned long val = old_write & ~RB_WRITE_MASK;
1104                 unsigned long eval = old_entries & ~RB_WRITE_MASK;
1105
1106                 /*
1107                  * This will only succeed if an interrupt did
1108                  * not come in and change it. In which case, we
1109                  * do not want to modify it.
1110                  *
1111                  * We add (void) to let the compiler know that we do not care
1112                  * about the return value of these functions. We use the
1113                  * cmpxchg to only update if an interrupt did not already
1114                  * do it for us. If the cmpxchg fails, we don't care.
1115                  */
1116                 (void)local_cmpxchg(&next_page->write, old_write, val);
1117                 (void)local_cmpxchg(&next_page->entries, old_entries, eval);
1118
1119                 /*
1120                  * No need to worry about races with clearing out the commit.
1121                  * it only can increment when a commit takes place. But that
1122                  * only happens in the outer most nested commit.
1123                  */
1124                 local_set(&next_page->page->commit, 0);
1125
1126                 /* Again, either we update tail_page or an interrupt does */
1127                 (void)cmpxchg(&cpu_buffer->tail_page, tail_page, next_page);
1128         }
1129 }
1130
1131 static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
1132                           struct buffer_page *bpage)
1133 {
1134         unsigned long val = (unsigned long)bpage;
1135
1136         if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK))
1137                 return 1;
1138
1139         return 0;
1140 }
1141
1142 /**
1143  * rb_check_list - make sure a pointer to a list has the last bits zero
1144  */
1145 static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer,
1146                          struct list_head *list)
1147 {
1148         if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev))
1149                 return 1;
1150         if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next))
1151                 return 1;
1152         return 0;
1153 }
1154
1155 /**
1156  * rb_check_pages - integrity check of buffer pages
1157  * @cpu_buffer: CPU buffer with pages to test
1158  *
1159  * As a safety measure we check to make sure the data pages have not
1160  * been corrupted.
1161  */
1162 static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
1163 {
1164         struct list_head *head = cpu_buffer->pages;
1165         struct buffer_page *bpage, *tmp;
1166
1167         /* Reset the head page if it exists */
1168         if (cpu_buffer->head_page)
1169                 rb_set_head_page(cpu_buffer);
1170
1171         rb_head_page_deactivate(cpu_buffer);
1172
1173         if (RB_WARN_ON(cpu_buffer, head->next->prev != head))
1174                 return -1;
1175         if (RB_WARN_ON(cpu_buffer, head->prev->next != head))
1176                 return -1;
1177
1178         if (rb_check_list(cpu_buffer, head))
1179                 return -1;
1180
1181         list_for_each_entry_safe(bpage, tmp, head, list) {
1182                 if (RB_WARN_ON(cpu_buffer,
1183                                bpage->list.next->prev != &bpage->list))
1184                         return -1;
1185                 if (RB_WARN_ON(cpu_buffer,
1186                                bpage->list.prev->next != &bpage->list))
1187                         return -1;
1188                 if (rb_check_list(cpu_buffer, &bpage->list))
1189                         return -1;
1190         }
1191
1192         rb_head_page_activate(cpu_buffer);
1193
1194         return 0;
1195 }
1196
1197 static int __rb_allocate_pages(long nr_pages, struct list_head *pages, int cpu)
1198 {
1199         struct buffer_page *bpage, *tmp;
1200         bool user_thread = current->mm != NULL;
1201         gfp_t mflags;
1202         long i;
1203
1204         /*
1205          * Check if the available memory is there first.
1206          * Note, si_mem_available() only gives us a rough estimate of available
1207          * memory. It may not be accurate. But we don't care, we just want
1208          * to prevent doing any allocation when it is obvious that it is
1209          * not going to succeed.
1210          */
1211         i = si_mem_available();
1212         if (i < nr_pages)
1213                 return -ENOMEM;
1214
1215         /*
1216          * __GFP_RETRY_MAYFAIL flag makes sure that the allocation fails
1217          * gracefully without invoking oom-killer and the system is not
1218          * destabilized.
1219          */
1220         mflags = GFP_KERNEL | __GFP_RETRY_MAYFAIL;
1221
1222         /*
1223          * If a user thread allocates too much, and si_mem_available()
1224          * reports there's enough memory, even though there is not.
1225          * Make sure the OOM killer kills this thread. This can happen
1226          * even with RETRY_MAYFAIL because another task may be doing
1227          * an allocation after this task has taken all memory.
1228          * This is the task the OOM killer needs to take out during this
1229          * loop, even if it was triggered by an allocation somewhere else.
1230          */
1231         if (user_thread)
1232                 set_current_oom_origin();
1233         for (i = 0; i < nr_pages; i++) {
1234                 struct page *page;
1235
1236                 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1237                                     mflags, cpu_to_node(cpu));
1238                 if (!bpage)
1239                         goto free_pages;
1240
1241                 list_add(&bpage->list, pages);
1242
1243                 page = alloc_pages_node(cpu_to_node(cpu), mflags, 0);
1244                 if (!page)
1245                         goto free_pages;
1246                 bpage->page = page_address(page);
1247                 rb_init_page(bpage->page);
1248
1249                 if (user_thread && fatal_signal_pending(current))
1250                         goto free_pages;
1251         }
1252         if (user_thread)
1253                 clear_current_oom_origin();
1254
1255         return 0;
1256
1257 free_pages:
1258         list_for_each_entry_safe(bpage, tmp, pages, list) {
1259                 list_del_init(&bpage->list);
1260                 free_buffer_page(bpage);
1261         }
1262         if (user_thread)
1263                 clear_current_oom_origin();
1264
1265         return -ENOMEM;
1266 }
1267
1268 static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1269                              unsigned long nr_pages)
1270 {
1271         LIST_HEAD(pages);
1272
1273         WARN_ON(!nr_pages);
1274
1275         if (__rb_allocate_pages(nr_pages, &pages, cpu_buffer->cpu))
1276                 return -ENOMEM;
1277
1278         /*
1279          * The ring buffer page list is a circular list that does not
1280          * start and end with a list head. All page list items point to
1281          * other pages.
1282          */
1283         cpu_buffer->pages = pages.next;
1284         list_del(&pages);
1285
1286         cpu_buffer->nr_pages = nr_pages;
1287
1288         rb_check_pages(cpu_buffer);
1289
1290         return 0;
1291 }
1292
1293 static struct ring_buffer_per_cpu *
1294 rb_allocate_cpu_buffer(struct ring_buffer *buffer, long nr_pages, int cpu)
1295 {
1296         struct ring_buffer_per_cpu *cpu_buffer;
1297         struct buffer_page *bpage;
1298         struct page *page;
1299         int ret;
1300
1301         cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1302                                   GFP_KERNEL, cpu_to_node(cpu));
1303         if (!cpu_buffer)
1304                 return NULL;
1305
1306         cpu_buffer->cpu = cpu;
1307         cpu_buffer->buffer = buffer;
1308         raw_spin_lock_init(&cpu_buffer->reader_lock);
1309         lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1310         cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1311         INIT_WORK(&cpu_buffer->update_pages_work, update_pages_handler);
1312         init_completion(&cpu_buffer->update_done);
1313         init_irq_work(&cpu_buffer->irq_work.work, rb_wake_up_waiters);
1314         init_waitqueue_head(&cpu_buffer->irq_work.waiters);
1315         init_waitqueue_head(&cpu_buffer->irq_work.full_waiters);
1316
1317         bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1318                             GFP_KERNEL, cpu_to_node(cpu));
1319         if (!bpage)
1320                 goto fail_free_buffer;
1321
1322         rb_check_bpage(cpu_buffer, bpage);
1323
1324         cpu_buffer->reader_page = bpage;
1325         page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0);
1326         if (!page)
1327                 goto fail_free_reader;
1328         bpage->page = page_address(page);
1329         rb_init_page(bpage->page);
1330
1331         INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1332         INIT_LIST_HEAD(&cpu_buffer->new_pages);
1333
1334         ret = rb_allocate_pages(cpu_buffer, nr_pages);
1335         if (ret < 0)
1336                 goto fail_free_reader;
1337
1338         cpu_buffer->head_page
1339                 = list_entry(cpu_buffer->pages, struct buffer_page, list);
1340         cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1341
1342         rb_head_page_activate(cpu_buffer);
1343
1344         return cpu_buffer;
1345
1346  fail_free_reader:
1347         free_buffer_page(cpu_buffer->reader_page);
1348
1349  fail_free_buffer:
1350         kfree(cpu_buffer);
1351         return NULL;
1352 }
1353
1354 static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1355 {
1356         struct list_head *head = cpu_buffer->pages;
1357         struct buffer_page *bpage, *tmp;
1358
1359         free_buffer_page(cpu_buffer->reader_page);
1360
1361         rb_head_page_deactivate(cpu_buffer);
1362
1363         if (head) {
1364                 list_for_each_entry_safe(bpage, tmp, head, list) {
1365                         list_del_init(&bpage->list);
1366                         free_buffer_page(bpage);
1367                 }
1368                 bpage = list_entry(head, struct buffer_page, list);
1369                 free_buffer_page(bpage);
1370         }
1371
1372         kfree(cpu_buffer);
1373 }
1374
1375 /**
1376  * __ring_buffer_alloc - allocate a new ring_buffer
1377  * @size: the size in bytes per cpu that is needed.
1378  * @flags: attributes to set for the ring buffer.
1379  *
1380  * Currently the only flag that is available is the RB_FL_OVERWRITE
1381  * flag. This flag means that the buffer will overwrite old data
1382  * when the buffer wraps. If this flag is not set, the buffer will
1383  * drop data when the tail hits the head.
1384  */
1385 struct ring_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1386                                         struct lock_class_key *key)
1387 {
1388         struct ring_buffer *buffer;
1389         long nr_pages;
1390         int bsize;
1391         int cpu;
1392         int ret;
1393
1394         /* keep it in its own cache line */
1395         buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1396                          GFP_KERNEL);
1397         if (!buffer)
1398                 return NULL;
1399
1400         if (!zalloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1401                 goto fail_free_buffer;
1402
1403         nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1404         buffer->flags = flags;
1405         buffer->clock = trace_clock_local;
1406         buffer->reader_lock_key = key;
1407
1408         init_irq_work(&buffer->irq_work.work, rb_wake_up_waiters);
1409         init_waitqueue_head(&buffer->irq_work.waiters);
1410
1411         /* need at least two pages */
1412         if (nr_pages < 2)
1413                 nr_pages = 2;
1414
1415         buffer->cpus = nr_cpu_ids;
1416
1417         bsize = sizeof(void *) * nr_cpu_ids;
1418         buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1419                                   GFP_KERNEL);
1420         if (!buffer->buffers)
1421                 goto fail_free_cpumask;
1422
1423         cpu = raw_smp_processor_id();
1424         cpumask_set_cpu(cpu, buffer->cpumask);
1425         buffer->buffers[cpu] = rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
1426         if (!buffer->buffers[cpu])
1427                 goto fail_free_buffers;
1428
1429         ret = cpuhp_state_add_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1430         if (ret < 0)
1431                 goto fail_free_buffers;
1432
1433         mutex_init(&buffer->mutex);
1434
1435         return buffer;
1436
1437  fail_free_buffers:
1438         for_each_buffer_cpu(buffer, cpu) {
1439                 if (buffer->buffers[cpu])
1440                         rb_free_cpu_buffer(buffer->buffers[cpu]);
1441         }
1442         kfree(buffer->buffers);
1443
1444  fail_free_cpumask:
1445         free_cpumask_var(buffer->cpumask);
1446
1447  fail_free_buffer:
1448         kfree(buffer);
1449         return NULL;
1450 }
1451 EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1452
1453 /**
1454  * ring_buffer_free - free a ring buffer.
1455  * @buffer: the buffer to free.
1456  */
1457 void
1458 ring_buffer_free(struct ring_buffer *buffer)
1459 {
1460         int cpu;
1461
1462         cpuhp_state_remove_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1463
1464         for_each_buffer_cpu(buffer, cpu)
1465                 rb_free_cpu_buffer(buffer->buffers[cpu]);
1466
1467         kfree(buffer->buffers);
1468         free_cpumask_var(buffer->cpumask);
1469
1470         kfree(buffer);
1471 }
1472 EXPORT_SYMBOL_GPL(ring_buffer_free);
1473
1474 void ring_buffer_set_clock(struct ring_buffer *buffer,
1475                            u64 (*clock)(void))
1476 {
1477         buffer->clock = clock;
1478 }
1479
1480 void ring_buffer_set_time_stamp_abs(struct ring_buffer *buffer, bool abs)
1481 {
1482         buffer->time_stamp_abs = abs;
1483 }
1484
1485 bool ring_buffer_time_stamp_abs(struct ring_buffer *buffer)
1486 {
1487         return buffer->time_stamp_abs;
1488 }
1489
1490 static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1491
1492 static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1493 {
1494         return local_read(&bpage->entries) & RB_WRITE_MASK;
1495 }
1496
1497 static inline unsigned long rb_page_write(struct buffer_page *bpage)
1498 {
1499         return local_read(&bpage->write) & RB_WRITE_MASK;
1500 }
1501
1502 static int
1503 rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned long nr_pages)
1504 {
1505         struct list_head *tail_page, *to_remove, *next_page;
1506         struct buffer_page *to_remove_page, *tmp_iter_page;
1507         struct buffer_page *last_page, *first_page;
1508         unsigned long nr_removed;
1509         unsigned long head_bit;
1510         int page_entries;
1511
1512         head_bit = 0;
1513
1514         raw_spin_lock_irq(&cpu_buffer->reader_lock);
1515         atomic_inc(&cpu_buffer->record_disabled);
1516         /*
1517          * We don't race with the readers since we have acquired the reader
1518          * lock. We also don't race with writers after disabling recording.
1519          * This makes it easy to figure out the first and the last page to be
1520          * removed from the list. We unlink all the pages in between including
1521          * the first and last pages. This is done in a busy loop so that we
1522          * lose the least number of traces.
1523          * The pages are freed after we restart recording and unlock readers.
1524          */
1525         tail_page = &cpu_buffer->tail_page->list;
1526
1527         /*
1528          * tail page might be on reader page, we remove the next page
1529          * from the ring buffer
1530          */
1531         if (cpu_buffer->tail_page == cpu_buffer->reader_page)
1532                 tail_page = rb_list_head(tail_page->next);
1533         to_remove = tail_page;
1534
1535         /* start of pages to remove */
1536         first_page = list_entry(rb_list_head(to_remove->next),
1537                                 struct buffer_page, list);
1538
1539         for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
1540                 to_remove = rb_list_head(to_remove)->next;
1541                 head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
1542         }
1543
1544         next_page = rb_list_head(to_remove)->next;
1545
1546         /*
1547          * Now we remove all pages between tail_page and next_page.
1548          * Make sure that we have head_bit value preserved for the
1549          * next page
1550          */
1551         tail_page->next = (struct list_head *)((unsigned long)next_page |
1552                                                 head_bit);
1553         next_page = rb_list_head(next_page);
1554         next_page->prev = tail_page;
1555
1556         /* make sure pages points to a valid page in the ring buffer */
1557         cpu_buffer->pages = next_page;
1558
1559         /* update head page */
1560         if (head_bit)
1561                 cpu_buffer->head_page = list_entry(next_page,
1562                                                 struct buffer_page, list);
1563
1564         /*
1565          * change read pointer to make sure any read iterators reset
1566          * themselves
1567          */
1568         cpu_buffer->read = 0;
1569
1570         /* pages are removed, resume tracing and then free the pages */
1571         atomic_dec(&cpu_buffer->record_disabled);
1572         raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1573
1574         RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
1575
1576         /* last buffer page to remove */
1577         last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
1578                                 list);
1579         tmp_iter_page = first_page;
1580
1581         do {
1582                 cond_resched();
1583
1584                 to_remove_page = tmp_iter_page;
1585                 rb_inc_page(cpu_buffer, &tmp_iter_page);
1586
1587                 /* update the counters */
1588                 page_entries = rb_page_entries(to_remove_page);
1589                 if (page_entries) {
1590                         /*
1591                          * If something was added to this page, it was full
1592                          * since it is not the tail page. So we deduct the
1593                          * bytes consumed in ring buffer from here.
1594                          * Increment overrun to account for the lost events.
1595                          */
1596                         local_add(page_entries, &cpu_buffer->overrun);
1597                         local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1598                 }
1599
1600                 /*
1601                  * We have already removed references to this list item, just
1602                  * free up the buffer_page and its page
1603                  */
1604                 free_buffer_page(to_remove_page);
1605                 nr_removed--;
1606
1607         } while (to_remove_page != last_page);
1608
1609         RB_WARN_ON(cpu_buffer, nr_removed);
1610
1611         return nr_removed == 0;
1612 }
1613
1614 static int
1615 rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
1616 {
1617         struct list_head *pages = &cpu_buffer->new_pages;
1618         int retries, success;
1619
1620         raw_spin_lock_irq(&cpu_buffer->reader_lock);
1621         /*
1622          * We are holding the reader lock, so the reader page won't be swapped
1623          * in the ring buffer. Now we are racing with the writer trying to
1624          * move head page and the tail page.
1625          * We are going to adapt the reader page update process where:
1626          * 1. We first splice the start and end of list of new pages between
1627          *    the head page and its previous page.
1628          * 2. We cmpxchg the prev_page->next to point from head page to the
1629          *    start of new pages list.
1630          * 3. Finally, we update the head->prev to the end of new list.
1631          *
1632          * We will try this process 10 times, to make sure that we don't keep
1633          * spinning.
1634          */
1635         retries = 10;
1636         success = 0;
1637         while (retries--) {
1638                 struct list_head *head_page, *prev_page, *r;
1639                 struct list_head *last_page, *first_page;
1640                 struct list_head *head_page_with_bit;
1641
1642                 head_page = &rb_set_head_page(cpu_buffer)->list;
1643                 if (!head_page)
1644                         break;
1645                 prev_page = head_page->prev;
1646
1647                 first_page = pages->next;
1648                 last_page  = pages->prev;
1649
1650                 head_page_with_bit = (struct list_head *)
1651                                      ((unsigned long)head_page | RB_PAGE_HEAD);
1652
1653                 last_page->next = head_page_with_bit;
1654                 first_page->prev = prev_page;
1655
1656                 r = cmpxchg(&prev_page->next, head_page_with_bit, first_page);
1657
1658                 if (r == head_page_with_bit) {
1659                         /*
1660                          * yay, we replaced the page pointer to our new list,
1661                          * now, we just have to update to head page's prev
1662                          * pointer to point to end of list
1663                          */
1664                         head_page->prev = last_page;
1665                         success = 1;
1666                         break;
1667                 }
1668         }
1669
1670         if (success)
1671                 INIT_LIST_HEAD(pages);
1672         /*
1673          * If we weren't successful in adding in new pages, warn and stop
1674          * tracing
1675          */
1676         RB_WARN_ON(cpu_buffer, !success);
1677         raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1678
1679         /* free pages if they weren't inserted */
1680         if (!success) {
1681                 struct buffer_page *bpage, *tmp;
1682                 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1683                                          list) {
1684                         list_del_init(&bpage->list);
1685                         free_buffer_page(bpage);
1686                 }
1687         }
1688         return success;
1689 }
1690
1691 static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
1692 {
1693         int success;
1694
1695         if (cpu_buffer->nr_pages_to_update > 0)
1696                 success = rb_insert_pages(cpu_buffer);
1697         else
1698                 success = rb_remove_pages(cpu_buffer,
1699                                         -cpu_buffer->nr_pages_to_update);
1700
1701         if (success)
1702                 cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
1703 }
1704
1705 static void update_pages_handler(struct work_struct *work)
1706 {
1707         struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
1708                         struct ring_buffer_per_cpu, update_pages_work);
1709         rb_update_pages(cpu_buffer);
1710         complete(&cpu_buffer->update_done);
1711 }
1712
1713 /**
1714  * ring_buffer_resize - resize the ring buffer
1715  * @buffer: the buffer to resize.
1716  * @size: the new size.
1717  * @cpu_id: the cpu buffer to resize
1718  *
1719  * Minimum size is 2 * BUF_PAGE_SIZE.
1720  *
1721  * Returns 0 on success and < 0 on failure.
1722  */
1723 int ring_buffer_resize(struct ring_buffer *buffer, unsigned long size,
1724                         int cpu_id)
1725 {
1726         struct ring_buffer_per_cpu *cpu_buffer;
1727         unsigned long nr_pages;
1728         int cpu, err = 0;
1729
1730         /*
1731          * Always succeed at resizing a non-existent buffer:
1732          */
1733         if (!buffer)
1734                 return size;
1735
1736         /* Make sure the requested buffer exists */
1737         if (cpu_id != RING_BUFFER_ALL_CPUS &&
1738             !cpumask_test_cpu(cpu_id, buffer->cpumask))
1739                 return size;
1740
1741         nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1742
1743         /* we need a minimum of two pages */
1744         if (nr_pages < 2)
1745                 nr_pages = 2;
1746
1747         size = nr_pages * BUF_PAGE_SIZE;
1748
1749         /*
1750          * Don't succeed if resizing is disabled, as a reader might be
1751          * manipulating the ring buffer and is expecting a sane state while
1752          * this is true.
1753          */
1754         if (atomic_read(&buffer->resize_disabled))
1755                 return -EBUSY;
1756
1757         /* prevent another thread from changing buffer sizes */
1758         mutex_lock(&buffer->mutex);
1759
1760         if (cpu_id == RING_BUFFER_ALL_CPUS) {
1761                 /* calculate the pages to update */
1762                 for_each_buffer_cpu(buffer, cpu) {
1763                         cpu_buffer = buffer->buffers[cpu];
1764
1765                         cpu_buffer->nr_pages_to_update = nr_pages -
1766                                                         cpu_buffer->nr_pages;
1767                         /*
1768                          * nothing more to do for removing pages or no update
1769                          */
1770                         if (cpu_buffer->nr_pages_to_update <= 0)
1771                                 continue;
1772                         /*
1773                          * to add pages, make sure all new pages can be
1774                          * allocated without receiving ENOMEM
1775                          */
1776                         INIT_LIST_HEAD(&cpu_buffer->new_pages);
1777                         if (__rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1778                                                 &cpu_buffer->new_pages, cpu)) {
1779                                 /* not enough memory for new pages */
1780                                 err = -ENOMEM;
1781                                 goto out_err;
1782                         }
1783                 }
1784
1785                 get_online_cpus();
1786                 /*
1787                  * Fire off all the required work handlers
1788                  * We can't schedule on offline CPUs, but it's not necessary
1789                  * since we can change their buffer sizes without any race.
1790                  */
1791                 for_each_buffer_cpu(buffer, cpu) {
1792                         cpu_buffer = buffer->buffers[cpu];
1793                         if (!cpu_buffer->nr_pages_to_update)
1794                                 continue;
1795
1796                         /* Can't run something on an offline CPU. */
1797                         if (!cpu_online(cpu)) {
1798                                 rb_update_pages(cpu_buffer);
1799                                 cpu_buffer->nr_pages_to_update = 0;
1800                         } else {
1801                                 schedule_work_on(cpu,
1802                                                 &cpu_buffer->update_pages_work);
1803                         }
1804                 }
1805
1806                 /* wait for all the updates to complete */
1807                 for_each_buffer_cpu(buffer, cpu) {
1808                         cpu_buffer = buffer->buffers[cpu];
1809                         if (!cpu_buffer->nr_pages_to_update)
1810                                 continue;
1811
1812                         if (cpu_online(cpu))
1813                                 wait_for_completion(&cpu_buffer->update_done);
1814                         cpu_buffer->nr_pages_to_update = 0;
1815                 }
1816
1817                 put_online_cpus();
1818         } else {
1819                 /* Make sure this CPU has been initialized */
1820                 if (!cpumask_test_cpu(cpu_id, buffer->cpumask))
1821                         goto out;
1822
1823                 cpu_buffer = buffer->buffers[cpu_id];
1824
1825                 if (nr_pages == cpu_buffer->nr_pages)
1826                         goto out;
1827
1828                 cpu_buffer->nr_pages_to_update = nr_pages -
1829                                                 cpu_buffer->nr_pages;
1830
1831                 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1832                 if (cpu_buffer->nr_pages_to_update > 0 &&
1833                         __rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1834                                             &cpu_buffer->new_pages, cpu_id)) {
1835                         err = -ENOMEM;
1836                         goto out_err;
1837                 }
1838
1839                 get_online_cpus();
1840
1841                 /* Can't run something on an offline CPU. */
1842                 if (!cpu_online(cpu_id))
1843                         rb_update_pages(cpu_buffer);
1844                 else {
1845                         schedule_work_on(cpu_id,
1846                                          &cpu_buffer->update_pages_work);
1847                         wait_for_completion(&cpu_buffer->update_done);
1848                 }
1849
1850                 cpu_buffer->nr_pages_to_update = 0;
1851                 put_online_cpus();
1852         }
1853
1854  out:
1855         /*
1856          * The ring buffer resize can happen with the ring buffer
1857          * enabled, so that the update disturbs the tracing as little
1858          * as possible. But if the buffer is disabled, we do not need
1859          * to worry about that, and we can take the time to verify
1860          * that the buffer is not corrupt.
1861          */
1862         if (atomic_read(&buffer->record_disabled)) {
1863                 atomic_inc(&buffer->record_disabled);
1864                 /*
1865                  * Even though the buffer was disabled, we must make sure
1866                  * that it is truly disabled before calling rb_check_pages.
1867                  * There could have been a race between checking
1868                  * record_disable and incrementing it.
1869                  */
1870                 synchronize_rcu();
1871                 for_each_buffer_cpu(buffer, cpu) {
1872                         cpu_buffer = buffer->buffers[cpu];
1873                         rb_check_pages(cpu_buffer);
1874                 }
1875                 atomic_dec(&buffer->record_disabled);
1876         }
1877
1878         mutex_unlock(&buffer->mutex);
1879         return size;
1880
1881  out_err:
1882         for_each_buffer_cpu(buffer, cpu) {
1883                 struct buffer_page *bpage, *tmp;
1884
1885                 cpu_buffer = buffer->buffers[cpu];
1886                 cpu_buffer->nr_pages_to_update = 0;
1887
1888                 if (list_empty(&cpu_buffer->new_pages))
1889                         continue;
1890
1891                 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1892                                         list) {
1893                         list_del_init(&bpage->list);
1894                         free_buffer_page(bpage);
1895                 }
1896         }
1897         mutex_unlock(&buffer->mutex);
1898         return err;
1899 }
1900 EXPORT_SYMBOL_GPL(ring_buffer_resize);
1901
1902 void ring_buffer_change_overwrite(struct ring_buffer *buffer, int val)
1903 {
1904         mutex_lock(&buffer->mutex);
1905         if (val)
1906                 buffer->flags |= RB_FL_OVERWRITE;
1907         else
1908                 buffer->flags &= ~RB_FL_OVERWRITE;
1909         mutex_unlock(&buffer->mutex);
1910 }
1911 EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
1912
1913 static __always_inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
1914 {
1915         return bpage->page->data + index;
1916 }
1917
1918 static __always_inline struct ring_buffer_event *
1919 rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
1920 {
1921         return __rb_page_index(cpu_buffer->reader_page,
1922                                cpu_buffer->reader_page->read);
1923 }
1924
1925 static __always_inline struct ring_buffer_event *
1926 rb_iter_head_event(struct ring_buffer_iter *iter)
1927 {
1928         return __rb_page_index(iter->head_page, iter->head);
1929 }
1930
1931 static __always_inline unsigned rb_page_commit(struct buffer_page *bpage)
1932 {
1933         return local_read(&bpage->page->commit);
1934 }
1935
1936 /* Size is determined by what has been committed */
1937 static __always_inline unsigned rb_page_size(struct buffer_page *bpage)
1938 {
1939         return rb_page_commit(bpage);
1940 }
1941
1942 static __always_inline unsigned
1943 rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
1944 {
1945         return rb_page_commit(cpu_buffer->commit_page);
1946 }
1947
1948 static __always_inline unsigned
1949 rb_event_index(struct ring_buffer_event *event)
1950 {
1951         unsigned long addr = (unsigned long)event;
1952
1953         return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
1954 }
1955
1956 static void rb_inc_iter(struct ring_buffer_iter *iter)
1957 {
1958         struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
1959
1960         /*
1961          * The iterator could be on the reader page (it starts there).
1962          * But the head could have moved, since the reader was
1963          * found. Check for this case and assign the iterator
1964          * to the head page instead of next.
1965          */
1966         if (iter->head_page == cpu_buffer->reader_page)
1967                 iter->head_page = rb_set_head_page(cpu_buffer);
1968         else
1969                 rb_inc_page(cpu_buffer, &iter->head_page);
1970
1971         iter->read_stamp = iter->head_page->page->time_stamp;
1972         iter->head = 0;
1973 }
1974
1975 /*
1976  * rb_handle_head_page - writer hit the head page
1977  *
1978  * Returns: +1 to retry page
1979  *           0 to continue
1980  *          -1 on error
1981  */
1982 static int
1983 rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
1984                     struct buffer_page *tail_page,
1985                     struct buffer_page *next_page)
1986 {
1987         struct buffer_page *new_head;
1988         int entries;
1989         int type;
1990         int ret;
1991
1992         entries = rb_page_entries(next_page);
1993
1994         /*
1995          * The hard part is here. We need to move the head
1996          * forward, and protect against both readers on
1997          * other CPUs and writers coming in via interrupts.
1998          */
1999         type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
2000                                        RB_PAGE_HEAD);
2001
2002         /*
2003          * type can be one of four:
2004          *  NORMAL - an interrupt already moved it for us
2005          *  HEAD   - we are the first to get here.
2006          *  UPDATE - we are the interrupt interrupting
2007          *           a current move.
2008          *  MOVED  - a reader on another CPU moved the next
2009          *           pointer to its reader page. Give up
2010          *           and try again.
2011          */
2012
2013         switch (type) {
2014         case RB_PAGE_HEAD:
2015                 /*
2016                  * We changed the head to UPDATE, thus
2017                  * it is our responsibility to update
2018                  * the counters.
2019                  */
2020                 local_add(entries, &cpu_buffer->overrun);
2021                 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
2022
2023                 /*
2024                  * The entries will be zeroed out when we move the
2025                  * tail page.
2026                  */
2027
2028                 /* still more to do */
2029                 break;
2030
2031         case RB_PAGE_UPDATE:
2032                 /*
2033                  * This is an interrupt that interrupt the
2034                  * previous update. Still more to do.
2035                  */
2036                 break;
2037         case RB_PAGE_NORMAL:
2038                 /*
2039                  * An interrupt came in before the update
2040                  * and processed this for us.
2041                  * Nothing left to do.
2042                  */
2043                 return 1;
2044         case RB_PAGE_MOVED:
2045                 /*
2046                  * The reader is on another CPU and just did
2047                  * a swap with our next_page.
2048                  * Try again.
2049                  */
2050                 return 1;
2051         default:
2052                 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
2053                 return -1;
2054         }
2055
2056         /*
2057          * Now that we are here, the old head pointer is
2058          * set to UPDATE. This will keep the reader from
2059          * swapping the head page with the reader page.
2060          * The reader (on another CPU) will spin till
2061          * we are finished.
2062          *
2063          * We just need to protect against interrupts
2064          * doing the job. We will set the next pointer
2065          * to HEAD. After that, we set the old pointer
2066          * to NORMAL, but only if it was HEAD before.
2067          * otherwise we are an interrupt, and only
2068          * want the outer most commit to reset it.
2069          */
2070         new_head = next_page;
2071         rb_inc_page(cpu_buffer, &new_head);
2072
2073         ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
2074                                     RB_PAGE_NORMAL);
2075
2076         /*
2077          * Valid returns are:
2078          *  HEAD   - an interrupt came in and already set it.
2079          *  NORMAL - One of two things:
2080          *            1) We really set it.
2081          *            2) A bunch of interrupts came in and moved
2082          *               the page forward again.
2083          */
2084         switch (ret) {
2085         case RB_PAGE_HEAD:
2086         case RB_PAGE_NORMAL:
2087                 /* OK */
2088                 break;
2089         default:
2090                 RB_WARN_ON(cpu_buffer, 1);
2091                 return -1;
2092         }
2093
2094         /*
2095          * It is possible that an interrupt came in,
2096          * set the head up, then more interrupts came in
2097          * and moved it again. When we get back here,
2098          * the page would have been set to NORMAL but we
2099          * just set it back to HEAD.
2100          *
2101          * How do you detect this? Well, if that happened
2102          * the tail page would have moved.
2103          */
2104         if (ret == RB_PAGE_NORMAL) {
2105                 struct buffer_page *buffer_tail_page;
2106
2107                 buffer_tail_page = READ_ONCE(cpu_buffer->tail_page);
2108                 /*
2109                  * If the tail had moved passed next, then we need
2110                  * to reset the pointer.
2111                  */
2112                 if (buffer_tail_page != tail_page &&
2113                     buffer_tail_page != next_page)
2114                         rb_head_page_set_normal(cpu_buffer, new_head,
2115                                                 next_page,
2116                                                 RB_PAGE_HEAD);
2117         }
2118
2119         /*
2120          * If this was the outer most commit (the one that
2121          * changed the original pointer from HEAD to UPDATE),
2122          * then it is up to us to reset it to NORMAL.
2123          */
2124         if (type == RB_PAGE_HEAD) {
2125                 ret = rb_head_page_set_normal(cpu_buffer, next_page,
2126                                               tail_page,
2127                                               RB_PAGE_UPDATE);
2128                 if (RB_WARN_ON(cpu_buffer,
2129                                ret != RB_PAGE_UPDATE))
2130                         return -1;
2131         }
2132
2133         return 0;
2134 }
2135
2136 static inline void
2137 rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2138               unsigned long tail, struct rb_event_info *info)
2139 {
2140         struct buffer_page *tail_page = info->tail_page;
2141         struct ring_buffer_event *event;
2142         unsigned long length = info->length;
2143
2144         /*
2145          * Only the event that crossed the page boundary
2146          * must fill the old tail_page with padding.
2147          */
2148         if (tail >= BUF_PAGE_SIZE) {
2149                 /*
2150                  * If the page was filled, then we still need
2151                  * to update the real_end. Reset it to zero
2152                  * and the reader will ignore it.
2153                  */
2154                 if (tail == BUF_PAGE_SIZE)
2155                         tail_page->real_end = 0;
2156
2157                 local_sub(length, &tail_page->write);
2158                 return;
2159         }
2160
2161         event = __rb_page_index(tail_page, tail);
2162
2163         /* account for padding bytes */
2164         local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes);
2165
2166         /*
2167          * Save the original length to the meta data.
2168          * This will be used by the reader to add lost event
2169          * counter.
2170          */
2171         tail_page->real_end = tail;
2172
2173         /*
2174          * If this event is bigger than the minimum size, then
2175          * we need to be careful that we don't subtract the
2176          * write counter enough to allow another writer to slip
2177          * in on this page.
2178          * We put in a discarded commit instead, to make sure
2179          * that this space is not used again.
2180          *
2181          * If we are less than the minimum size, we don't need to
2182          * worry about it.
2183          */
2184         if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
2185                 /* No room for any events */
2186
2187                 /* Mark the rest of the page with padding */
2188                 rb_event_set_padding(event);
2189
2190                 /* Set the write back to the previous setting */
2191                 local_sub(length, &tail_page->write);
2192                 return;
2193         }
2194
2195         /* Put in a discarded event */
2196         event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
2197         event->type_len = RINGBUF_TYPE_PADDING;
2198         /* time delta must be non zero */
2199         event->time_delta = 1;
2200
2201         /* Set write to end of buffer */
2202         length = (tail + length) - BUF_PAGE_SIZE;
2203         local_sub(length, &tail_page->write);
2204 }
2205
2206 static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer);
2207
2208 /*
2209  * This is the slow path, force gcc not to inline it.
2210  */
2211 static noinline struct ring_buffer_event *
2212 rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2213              unsigned long tail, struct rb_event_info *info)
2214 {
2215         struct buffer_page *tail_page = info->tail_page;
2216         struct buffer_page *commit_page = cpu_buffer->commit_page;
2217         struct ring_buffer *buffer = cpu_buffer->buffer;
2218         struct buffer_page *next_page;
2219         int ret;
2220
2221         next_page = tail_page;
2222
2223         rb_inc_page(cpu_buffer, &next_page);
2224
2225         /*
2226          * If for some reason, we had an interrupt storm that made
2227          * it all the way around the buffer, bail, and warn
2228          * about it.
2229          */
2230         if (unlikely(next_page == commit_page)) {
2231                 local_inc(&cpu_buffer->commit_overrun);
2232                 goto out_reset;
2233         }
2234
2235         /*
2236          * This is where the fun begins!
2237          *
2238          * We are fighting against races between a reader that
2239          * could be on another CPU trying to swap its reader
2240          * page with the buffer head.
2241          *
2242          * We are also fighting against interrupts coming in and
2243          * moving the head or tail on us as well.
2244          *
2245          * If the next page is the head page then we have filled
2246          * the buffer, unless the commit page is still on the
2247          * reader page.
2248          */
2249         if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) {
2250
2251                 /*
2252                  * If the commit is not on the reader page, then
2253                  * move the header page.
2254                  */
2255                 if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2256                         /*
2257                          * If we are not in overwrite mode,
2258                          * this is easy, just stop here.
2259                          */
2260                         if (!(buffer->flags & RB_FL_OVERWRITE)) {
2261                                 local_inc(&cpu_buffer->dropped_events);
2262                                 goto out_reset;
2263                         }
2264
2265                         ret = rb_handle_head_page(cpu_buffer,
2266                                                   tail_page,
2267                                                   next_page);
2268                         if (ret < 0)
2269                                 goto out_reset;
2270                         if (ret)
2271                                 goto out_again;
2272                 } else {
2273                         /*
2274                          * We need to be careful here too. The
2275                          * commit page could still be on the reader
2276                          * page. We could have a small buffer, and
2277                          * have filled up the buffer with events
2278                          * from interrupts and such, and wrapped.
2279                          *
2280                          * Note, if the tail page is also the on the
2281                          * reader_page, we let it move out.
2282                          */
2283                         if (unlikely((cpu_buffer->commit_page !=
2284                                       cpu_buffer->tail_page) &&
2285                                      (cpu_buffer->commit_page ==
2286                                       cpu_buffer->reader_page))) {
2287                                 local_inc(&cpu_buffer->commit_overrun);
2288                                 goto out_reset;
2289                         }
2290                 }
2291         }
2292
2293         rb_tail_page_update(cpu_buffer, tail_page, next_page);
2294
2295  out_again:
2296
2297         rb_reset_tail(cpu_buffer, tail, info);
2298
2299         /* Commit what we have for now. */
2300         rb_end_commit(cpu_buffer);
2301         /* rb_end_commit() decs committing */
2302         local_inc(&cpu_buffer->committing);
2303
2304         /* fail and let the caller try again */
2305         return ERR_PTR(-EAGAIN);
2306
2307  out_reset:
2308         /* reset write */
2309         rb_reset_tail(cpu_buffer, tail, info);
2310
2311         return NULL;
2312 }
2313
2314 /* Slow path, do not inline */
2315 static noinline struct ring_buffer_event *
2316 rb_add_time_stamp(struct ring_buffer_event *event, u64 delta, bool abs)
2317 {
2318         if (abs)
2319                 event->type_len = RINGBUF_TYPE_TIME_STAMP;
2320         else
2321                 event->type_len = RINGBUF_TYPE_TIME_EXTEND;
2322
2323         /* Not the first event on the page, or not delta? */
2324         if (abs || rb_event_index(event)) {
2325                 event->time_delta = delta & TS_MASK;
2326                 event->array[0] = delta >> TS_SHIFT;
2327         } else {
2328                 /* nope, just zero it */
2329                 event->time_delta = 0;
2330                 event->array[0] = 0;
2331         }
2332
2333         return skip_time_extend(event);
2334 }
2335
2336 static inline bool rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2337                                      struct ring_buffer_event *event);
2338
2339 /**
2340  * rb_update_event - update event type and data
2341  * @event: the event to update
2342  * @type: the type of event
2343  * @length: the size of the event field in the ring buffer
2344  *
2345  * Update the type and data fields of the event. The length
2346  * is the actual size that is written to the ring buffer,
2347  * and with this, we can determine what to place into the
2348  * data field.
2349  */
2350 static void
2351 rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
2352                 struct ring_buffer_event *event,
2353                 struct rb_event_info *info)
2354 {
2355         unsigned length = info->length;
2356         u64 delta = info->delta;
2357
2358         /* Only a commit updates the timestamp */
2359         if (unlikely(!rb_event_is_commit(cpu_buffer, event)))
2360                 delta = 0;
2361
2362         /*
2363          * If we need to add a timestamp, then we
2364          * add it to the start of the reserved space.
2365          */
2366         if (unlikely(info->add_timestamp)) {
2367                 bool abs = ring_buffer_time_stamp_abs(cpu_buffer->buffer);
2368
2369                 event = rb_add_time_stamp(event, info->delta, abs);
2370                 length -= RB_LEN_TIME_EXTEND;
2371                 delta = 0;
2372         }
2373
2374         event->time_delta = delta;
2375         length -= RB_EVNT_HDR_SIZE;
2376         if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
2377                 event->type_len = 0;
2378                 event->array[0] = length;
2379         } else
2380                 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
2381 }
2382
2383 static unsigned rb_calculate_event_length(unsigned length)
2384 {
2385         struct ring_buffer_event event; /* Used only for sizeof array */
2386
2387         /* zero length can cause confusions */
2388         if (!length)
2389                 length++;
2390
2391         if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
2392                 length += sizeof(event.array[0]);
2393
2394         length += RB_EVNT_HDR_SIZE;
2395         length = ALIGN(length, RB_ARCH_ALIGNMENT);
2396
2397         /*
2398          * In case the time delta is larger than the 27 bits for it
2399          * in the header, we need to add a timestamp. If another
2400          * event comes in when trying to discard this one to increase
2401          * the length, then the timestamp will be added in the allocated
2402          * space of this event. If length is bigger than the size needed
2403          * for the TIME_EXTEND, then padding has to be used. The events
2404          * length must be either RB_LEN_TIME_EXTEND, or greater than or equal
2405          * to RB_LEN_TIME_EXTEND + 8, as 8 is the minimum size for padding.
2406          * As length is a multiple of 4, we only need to worry if it
2407          * is 12 (RB_LEN_TIME_EXTEND + 4).
2408          */
2409         if (length == RB_LEN_TIME_EXTEND + RB_ALIGNMENT)
2410                 length += RB_ALIGNMENT;
2411
2412         return length;
2413 }
2414
2415 #ifndef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2416 static inline bool sched_clock_stable(void)
2417 {
2418         return true;
2419 }
2420 #endif
2421
2422 static inline int
2423 rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2424                   struct ring_buffer_event *event)
2425 {
2426         unsigned long new_index, old_index;
2427         struct buffer_page *bpage;
2428         unsigned long index;
2429         unsigned long addr;
2430
2431         new_index = rb_event_index(event);
2432         old_index = new_index + rb_event_ts_length(event);
2433         addr = (unsigned long)event;
2434         addr &= PAGE_MASK;
2435
2436         bpage = READ_ONCE(cpu_buffer->tail_page);
2437
2438         if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2439                 unsigned long write_mask =
2440                         local_read(&bpage->write) & ~RB_WRITE_MASK;
2441                 unsigned long event_length = rb_event_length(event);
2442                 /*
2443                  * This is on the tail page. It is possible that
2444                  * a write could come in and move the tail page
2445                  * and write to the next page. That is fine
2446                  * because we just shorten what is on this page.
2447                  */
2448                 old_index += write_mask;
2449                 new_index += write_mask;
2450                 index = local_cmpxchg(&bpage->write, old_index, new_index);
2451                 if (index == old_index) {
2452                         /* update counters */
2453                         local_sub(event_length, &cpu_buffer->entries_bytes);
2454                         return 1;
2455                 }
2456         }
2457
2458         /* could not discard */
2459         return 0;
2460 }
2461
2462 static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2463 {
2464         local_inc(&cpu_buffer->committing);
2465         local_inc(&cpu_buffer->commits);
2466 }
2467
2468 static __always_inline void
2469 rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
2470 {
2471         unsigned long max_count;
2472
2473         /*
2474          * We only race with interrupts and NMIs on this CPU.
2475          * If we own the commit event, then we can commit
2476          * all others that interrupted us, since the interruptions
2477          * are in stack format (they finish before they come
2478          * back to us). This allows us to do a simple loop to
2479          * assign the commit to the tail.
2480          */
2481  again:
2482         max_count = cpu_buffer->nr_pages * 100;
2483
2484         while (cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)) {
2485                 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
2486                         return;
2487                 if (RB_WARN_ON(cpu_buffer,
2488                                rb_is_reader_page(cpu_buffer->tail_page)))
2489                         return;
2490                 local_set(&cpu_buffer->commit_page->page->commit,
2491                           rb_page_write(cpu_buffer->commit_page));
2492                 rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
2493                 /* Only update the write stamp if the page has an event */
2494                 if (rb_page_write(cpu_buffer->commit_page))
2495                         cpu_buffer->write_stamp =
2496                                 cpu_buffer->commit_page->page->time_stamp;
2497                 /* add barrier to keep gcc from optimizing too much */
2498                 barrier();
2499         }
2500         while (rb_commit_index(cpu_buffer) !=
2501                rb_page_write(cpu_buffer->commit_page)) {
2502
2503                 local_set(&cpu_buffer->commit_page->page->commit,
2504                           rb_page_write(cpu_buffer->commit_page));
2505                 RB_WARN_ON(cpu_buffer,
2506                            local_read(&cpu_buffer->commit_page->page->commit) &
2507                            ~RB_WRITE_MASK);
2508                 barrier();
2509         }
2510
2511         /* again, keep gcc from optimizing */
2512         barrier();
2513
2514         /*
2515          * If an interrupt came in just after the first while loop
2516          * and pushed the tail page forward, we will be left with
2517          * a dangling commit that will never go forward.
2518          */
2519         if (unlikely(cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)))
2520                 goto again;
2521 }
2522
2523 static __always_inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
2524 {
2525         unsigned long commits;
2526
2527         if (RB_WARN_ON(cpu_buffer,
2528                        !local_read(&cpu_buffer->committing)))
2529                 return;
2530
2531  again:
2532         commits = local_read(&cpu_buffer->commits);
2533         /* synchronize with interrupts */
2534         barrier();
2535         if (local_read(&cpu_buffer->committing) == 1)
2536                 rb_set_commit_to_write(cpu_buffer);
2537
2538         local_dec(&cpu_buffer->committing);
2539
2540         /* synchronize with interrupts */
2541         barrier();
2542
2543         /*
2544          * Need to account for interrupts coming in between the
2545          * updating of the commit page and the clearing of the
2546          * committing counter.
2547          */
2548         if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
2549             !local_read(&cpu_buffer->committing)) {
2550                 local_inc(&cpu_buffer->committing);
2551                 goto again;
2552         }
2553 }
2554
2555 static inline void rb_event_discard(struct ring_buffer_event *event)
2556 {
2557         if (extended_time(event))
2558                 event = skip_time_extend(event);
2559
2560         /* array[0] holds the actual length for the discarded event */
2561         event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
2562         event->type_len = RINGBUF_TYPE_PADDING;
2563         /* time delta must be non zero */
2564         if (!event->time_delta)
2565                 event->time_delta = 1;
2566 }
2567
2568 static __always_inline bool
2569 rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2570                    struct ring_buffer_event *event)
2571 {
2572         unsigned long addr = (unsigned long)event;
2573         unsigned long index;
2574
2575         index = rb_event_index(event);
2576         addr &= PAGE_MASK;
2577
2578         return cpu_buffer->commit_page->page == (void *)addr &&
2579                 rb_commit_index(cpu_buffer) == index;
2580 }
2581
2582 static __always_inline void
2583 rb_update_write_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2584                       struct ring_buffer_event *event)
2585 {
2586         u64 delta;
2587
2588         /*
2589          * The event first in the commit queue updates the
2590          * time stamp.
2591          */
2592         if (rb_event_is_commit(cpu_buffer, event)) {
2593                 /*
2594                  * A commit event that is first on a page
2595                  * updates the write timestamp with the page stamp
2596                  */
2597                 if (!rb_event_index(event))
2598                         cpu_buffer->write_stamp =
2599                                 cpu_buffer->commit_page->page->time_stamp;
2600                 else if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
2601                         delta = ring_buffer_event_time_stamp(event);
2602                         cpu_buffer->write_stamp += delta;
2603                 } else if (event->type_len == RINGBUF_TYPE_TIME_STAMP) {
2604                         delta = ring_buffer_event_time_stamp(event);
2605                         cpu_buffer->write_stamp = delta;
2606                 } else
2607                         cpu_buffer->write_stamp += event->time_delta;
2608         }
2609 }
2610
2611 static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
2612                       struct ring_buffer_event *event)
2613 {
2614         local_inc(&cpu_buffer->entries);
2615         rb_update_write_stamp(cpu_buffer, event);
2616         rb_end_commit(cpu_buffer);
2617 }
2618
2619 static __always_inline void
2620 rb_wakeups(struct ring_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer)
2621 {
2622         size_t nr_pages;
2623         size_t dirty;
2624         size_t full;
2625
2626         if (buffer->irq_work.waiters_pending) {
2627                 buffer->irq_work.waiters_pending = false;
2628                 /* irq_work_queue() supplies it's own memory barriers */
2629                 irq_work_queue(&buffer->irq_work.work);
2630         }
2631
2632         if (cpu_buffer->irq_work.waiters_pending) {
2633                 cpu_buffer->irq_work.waiters_pending = false;
2634                 /* irq_work_queue() supplies it's own memory barriers */
2635                 irq_work_queue(&cpu_buffer->irq_work.work);
2636         }
2637
2638         if (cpu_buffer->last_pages_touch == local_read(&cpu_buffer->pages_touched))
2639                 return;
2640
2641         if (cpu_buffer->reader_page == cpu_buffer->commit_page)
2642                 return;
2643
2644         if (!cpu_buffer->irq_work.full_waiters_pending)
2645                 return;
2646
2647         cpu_buffer->last_pages_touch = local_read(&cpu_buffer->pages_touched);
2648
2649         full = cpu_buffer->shortest_full;
2650         nr_pages = cpu_buffer->nr_pages;
2651         dirty = ring_buffer_nr_dirty_pages(buffer, cpu_buffer->cpu);
2652         if (full && nr_pages && (dirty * 100) <= full * nr_pages)
2653                 return;
2654
2655         cpu_buffer->irq_work.wakeup_full = true;
2656         cpu_buffer->irq_work.full_waiters_pending = false;
2657         /* irq_work_queue() supplies it's own memory barriers */
2658         irq_work_queue(&cpu_buffer->irq_work.work);
2659 }
2660
2661 /*
2662  * The lock and unlock are done within a preempt disable section.
2663  * The current_context per_cpu variable can only be modified
2664  * by the current task between lock and unlock. But it can
2665  * be modified more than once via an interrupt. To pass this
2666  * information from the lock to the unlock without having to
2667  * access the 'in_interrupt()' functions again (which do show
2668  * a bit of overhead in something as critical as function tracing,
2669  * we use a bitmask trick.
2670  *
2671  *  bit 0 =  NMI context
2672  *  bit 1 =  IRQ context
2673  *  bit 2 =  SoftIRQ context
2674  *  bit 3 =  normal context.
2675  *
2676  * This works because this is the order of contexts that can
2677  * preempt other contexts. A SoftIRQ never preempts an IRQ
2678  * context.
2679  *
2680  * When the context is determined, the corresponding bit is
2681  * checked and set (if it was set, then a recursion of that context
2682  * happened).
2683  *
2684  * On unlock, we need to clear this bit. To do so, just subtract
2685  * 1 from the current_context and AND it to itself.
2686  *
2687  * (binary)
2688  *  101 - 1 = 100
2689  *  101 & 100 = 100 (clearing bit zero)
2690  *
2691  *  1010 - 1 = 1001
2692  *  1010 & 1001 = 1000 (clearing bit 1)
2693  *
2694  * The least significant bit can be cleared this way, and it
2695  * just so happens that it is the same bit corresponding to
2696  * the current context.
2697  */
2698
2699 static __always_inline int
2700 trace_recursive_lock(struct ring_buffer_per_cpu *cpu_buffer)
2701 {
2702         unsigned int val = cpu_buffer->current_context;
2703         unsigned long pc = preempt_count();
2704         int bit;
2705
2706         if (!(pc & (NMI_MASK | HARDIRQ_MASK | SOFTIRQ_OFFSET)))
2707                 bit = RB_CTX_NORMAL;
2708         else
2709                 bit = pc & NMI_MASK ? RB_CTX_NMI :
2710                         pc & HARDIRQ_MASK ? RB_CTX_IRQ : RB_CTX_SOFTIRQ;
2711
2712         if (unlikely(val & (1 << (bit + cpu_buffer->nest))))
2713                 return 1;
2714
2715         val |= (1 << (bit + cpu_buffer->nest));
2716         cpu_buffer->current_context = val;
2717
2718         return 0;
2719 }
2720
2721 static __always_inline void
2722 trace_recursive_unlock(struct ring_buffer_per_cpu *cpu_buffer)
2723 {
2724         cpu_buffer->current_context &=
2725                 cpu_buffer->current_context - (1 << cpu_buffer->nest);
2726 }
2727
2728 /* The recursive locking above uses 4 bits */
2729 #define NESTED_BITS 4
2730
2731 /**
2732  * ring_buffer_nest_start - Allow to trace while nested
2733  * @buffer: The ring buffer to modify
2734  *
2735  * The ring buffer has a safety mechanism to prevent recursion.
2736  * But there may be a case where a trace needs to be done while
2737  * tracing something else. In this case, calling this function
2738  * will allow this function to nest within a currently active
2739  * ring_buffer_lock_reserve().
2740  *
2741  * Call this function before calling another ring_buffer_lock_reserve() and
2742  * call ring_buffer_nest_end() after the nested ring_buffer_unlock_commit().
2743  */
2744 void ring_buffer_nest_start(struct ring_buffer *buffer)
2745 {
2746         struct ring_buffer_per_cpu *cpu_buffer;
2747         int cpu;
2748
2749         /* Enabled by ring_buffer_nest_end() */
2750         preempt_disable_notrace();
2751         cpu = raw_smp_processor_id();
2752         cpu_buffer = buffer->buffers[cpu];
2753         /* This is the shift value for the above recursive locking */
2754         cpu_buffer->nest += NESTED_BITS;
2755 }
2756
2757 /**
2758  * ring_buffer_nest_end - Allow to trace while nested
2759  * @buffer: The ring buffer to modify
2760  *
2761  * Must be called after ring_buffer_nest_start() and after the
2762  * ring_buffer_unlock_commit().
2763  */
2764 void ring_buffer_nest_end(struct ring_buffer *buffer)
2765 {
2766         struct ring_buffer_per_cpu *cpu_buffer;
2767         int cpu;
2768
2769         /* disabled by ring_buffer_nest_start() */
2770         cpu = raw_smp_processor_id();
2771         cpu_buffer = buffer->buffers[cpu];
2772         /* This is the shift value for the above recursive locking */
2773         cpu_buffer->nest -= NESTED_BITS;
2774         preempt_enable_notrace();
2775 }
2776
2777 /**
2778  * ring_buffer_unlock_commit - commit a reserved
2779  * @buffer: The buffer to commit to
2780  * @event: The event pointer to commit.
2781  *
2782  * This commits the data to the ring buffer, and releases any locks held.
2783  *
2784  * Must be paired with ring_buffer_lock_reserve.
2785  */
2786 int ring_buffer_unlock_commit(struct ring_buffer *buffer,
2787                               struct ring_buffer_event *event)
2788 {
2789         struct ring_buffer_per_cpu *cpu_buffer;
2790         int cpu = raw_smp_processor_id();
2791
2792         cpu_buffer = buffer->buffers[cpu];
2793
2794         rb_commit(cpu_buffer, event);
2795
2796         rb_wakeups(buffer, cpu_buffer);
2797
2798         trace_recursive_unlock(cpu_buffer);
2799
2800         preempt_enable_notrace();
2801
2802         return 0;
2803 }
2804 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
2805
2806 static noinline void
2807 rb_handle_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2808                     struct rb_event_info *info)
2809 {
2810         WARN_ONCE(info->delta > (1ULL << 59),
2811                   KERN_WARNING "Delta way too big! %llu ts=%llu write stamp = %llu\n%s",
2812                   (unsigned long long)info->delta,
2813                   (unsigned long long)info->ts,
2814                   (unsigned long long)cpu_buffer->write_stamp,
2815                   sched_clock_stable() ? "" :
2816                   "If you just came from a suspend/resume,\n"
2817                   "please switch to the trace global clock:\n"
2818                   "  echo global > /sys/kernel/debug/tracing/trace_clock\n"
2819                   "or add trace_clock=global to the kernel command line\n");
2820         info->add_timestamp = 1;
2821 }
2822
2823 static struct ring_buffer_event *
2824 __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
2825                   struct rb_event_info *info)
2826 {
2827         struct ring_buffer_event *event;
2828         struct buffer_page *tail_page;
2829         unsigned long tail, write;
2830
2831         /*
2832          * If the time delta since the last event is too big to
2833          * hold in the time field of the event, then we append a
2834          * TIME EXTEND event ahead of the data event.
2835          */
2836         if (unlikely(info->add_timestamp))
2837                 info->length += RB_LEN_TIME_EXTEND;
2838
2839         /* Don't let the compiler play games with cpu_buffer->tail_page */
2840         tail_page = info->tail_page = READ_ONCE(cpu_buffer->tail_page);
2841         write = local_add_return(info->length, &tail_page->write);
2842
2843         /* set write to only the index of the write */
2844         write &= RB_WRITE_MASK;
2845         tail = write - info->length;
2846
2847         /*
2848          * If this is the first commit on the page, then it has the same
2849          * timestamp as the page itself.
2850          */
2851         if (!tail && !ring_buffer_time_stamp_abs(cpu_buffer->buffer))
2852                 info->delta = 0;
2853
2854         /* See if we shot pass the end of this buffer page */
2855         if (unlikely(write > BUF_PAGE_SIZE))
2856                 return rb_move_tail(cpu_buffer, tail, info);
2857
2858         /* We reserved something on the buffer */
2859
2860         event = __rb_page_index(tail_page, tail);
2861         rb_update_event(cpu_buffer, event, info);
2862
2863         local_inc(&tail_page->entries);
2864
2865         /*
2866          * If this is the first commit on the page, then update
2867          * its timestamp.
2868          */
2869         if (!tail)
2870                 tail_page->page->time_stamp = info->ts;
2871
2872         /* account for these added bytes */
2873         local_add(info->length, &cpu_buffer->entries_bytes);
2874
2875         return event;
2876 }
2877
2878 static __always_inline struct ring_buffer_event *
2879 rb_reserve_next_event(struct ring_buffer *buffer,
2880                       struct ring_buffer_per_cpu *cpu_buffer,
2881                       unsigned long length)
2882 {
2883         struct ring_buffer_event *event;
2884         struct rb_event_info info;
2885         int nr_loops = 0;
2886         u64 diff;
2887
2888         rb_start_commit(cpu_buffer);
2889
2890 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2891         /*
2892          * Due to the ability to swap a cpu buffer from a buffer
2893          * it is possible it was swapped before we committed.
2894          * (committing stops a swap). We check for it here and
2895          * if it happened, we have to fail the write.
2896          */
2897         barrier();
2898         if (unlikely(READ_ONCE(cpu_buffer->buffer) != buffer)) {
2899                 local_dec(&cpu_buffer->committing);
2900                 local_dec(&cpu_buffer->commits);
2901                 return NULL;
2902         }
2903 #endif
2904
2905         info.length = rb_calculate_event_length(length);
2906  again:
2907         info.add_timestamp = 0;
2908         info.delta = 0;
2909
2910         /*
2911          * We allow for interrupts to reenter here and do a trace.
2912          * If one does, it will cause this original code to loop
2913          * back here. Even with heavy interrupts happening, this
2914          * should only happen a few times in a row. If this happens
2915          * 1000 times in a row, there must be either an interrupt
2916          * storm or we have something buggy.
2917          * Bail!
2918          */
2919         if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
2920                 goto out_fail;
2921
2922         info.ts = rb_time_stamp(cpu_buffer->buffer);
2923         diff = info.ts - cpu_buffer->write_stamp;
2924
2925         /* make sure this diff is calculated here */
2926         barrier();
2927
2928         if (ring_buffer_time_stamp_abs(buffer)) {
2929                 info.delta = info.ts;
2930                 rb_handle_timestamp(cpu_buffer, &info);
2931         } else /* Did the write stamp get updated already? */
2932                 if (likely(info.ts >= cpu_buffer->write_stamp)) {
2933                 info.delta = diff;
2934                 if (unlikely(test_time_stamp(info.delta)))
2935                         rb_handle_timestamp(cpu_buffer, &info);
2936         }
2937
2938         event = __rb_reserve_next(cpu_buffer, &info);
2939
2940         if (unlikely(PTR_ERR(event) == -EAGAIN)) {
2941                 if (info.add_timestamp)
2942                         info.length -= RB_LEN_TIME_EXTEND;
2943                 goto again;
2944         }
2945
2946         if (!event)
2947                 goto out_fail;
2948
2949         return event;
2950
2951  out_fail:
2952         rb_end_commit(cpu_buffer);
2953         return NULL;
2954 }
2955
2956 /**
2957  * ring_buffer_lock_reserve - reserve a part of the buffer
2958  * @buffer: the ring buffer to reserve from
2959  * @length: the length of the data to reserve (excluding event header)
2960  *
2961  * Returns a reserved event on the ring buffer to copy directly to.
2962  * The user of this interface will need to get the body to write into
2963  * and can use the ring_buffer_event_data() interface.
2964  *
2965  * The length is the length of the data needed, not the event length
2966  * which also includes the event header.
2967  *
2968  * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2969  * If NULL is returned, then nothing has been allocated or locked.
2970  */
2971 struct ring_buffer_event *
2972 ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length)
2973 {
2974         struct ring_buffer_per_cpu *cpu_buffer;
2975         struct ring_buffer_event *event;
2976         int cpu;
2977
2978         /* If we are tracing schedule, we don't want to recurse */
2979         preempt_disable_notrace();
2980
2981         if (unlikely(atomic_read(&buffer->record_disabled)))
2982                 goto out;
2983
2984         cpu = raw_smp_processor_id();
2985
2986         if (unlikely(!cpumask_test_cpu(cpu, buffer->cpumask)))
2987                 goto out;
2988
2989         cpu_buffer = buffer->buffers[cpu];
2990
2991         if (unlikely(atomic_read(&cpu_buffer->record_disabled)))
2992                 goto out;
2993
2994         if (unlikely(length > BUF_MAX_DATA_SIZE))
2995                 goto out;
2996
2997         if (unlikely(trace_recursive_lock(cpu_buffer)))
2998                 goto out;
2999
3000         event = rb_reserve_next_event(buffer, cpu_buffer, length);
3001         if (!event)
3002                 goto out_unlock;
3003
3004         return event;
3005
3006  out_unlock:
3007         trace_recursive_unlock(cpu_buffer);
3008  out:
3009         preempt_enable_notrace();
3010         return NULL;
3011 }
3012 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
3013
3014 /*
3015  * Decrement the entries to the page that an event is on.
3016  * The event does not even need to exist, only the pointer
3017  * to the page it is on. This may only be called before the commit
3018  * takes place.
3019  */
3020 static inline void
3021 rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
3022                    struct ring_buffer_event *event)
3023 {
3024         unsigned long addr = (unsigned long)event;
3025         struct buffer_page *bpage = cpu_buffer->commit_page;
3026         struct buffer_page *start;
3027
3028         addr &= PAGE_MASK;
3029
3030         /* Do the likely case first */
3031         if (likely(bpage->page == (void *)addr)) {
3032                 local_dec(&bpage->entries);
3033                 return;
3034         }
3035
3036         /*
3037          * Because the commit page may be on the reader page we
3038          * start with the next page and check the end loop there.
3039          */
3040         rb_inc_page(cpu_buffer, &bpage);
3041         start = bpage;
3042         do {
3043                 if (bpage->page == (void *)addr) {
3044                         local_dec(&bpage->entries);
3045                         return;
3046                 }
3047                 rb_inc_page(cpu_buffer, &bpage);
3048         } while (bpage != start);
3049
3050         /* commit not part of this buffer?? */
3051         RB_WARN_ON(cpu_buffer, 1);
3052 }
3053
3054 /**
3055  * ring_buffer_commit_discard - discard an event that has not been committed
3056  * @buffer: the ring buffer
3057  * @event: non committed event to discard
3058  *
3059  * Sometimes an event that is in the ring buffer needs to be ignored.
3060  * This function lets the user discard an event in the ring buffer
3061  * and then that event will not be read later.
3062  *
3063  * This function only works if it is called before the item has been
3064  * committed. It will try to free the event from the ring buffer
3065  * if another event has not been added behind it.
3066  *
3067  * If another event has been added behind it, it will set the event
3068  * up as discarded, and perform the commit.
3069  *
3070  * If this function is called, do not call ring_buffer_unlock_commit on
3071  * the event.
3072  */
3073 void ring_buffer_discard_commit(struct ring_buffer *buffer,
3074                                 struct ring_buffer_event *event)
3075 {
3076         struct ring_buffer_per_cpu *cpu_buffer;
3077         int cpu;
3078
3079         /* The event is discarded regardless */
3080         rb_event_discard(event);
3081
3082         cpu = smp_processor_id();
3083         cpu_buffer = buffer->buffers[cpu];
3084
3085         /*
3086          * This must only be called if the event has not been
3087          * committed yet. Thus we can assume that preemption
3088          * is still disabled.
3089          */
3090         RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
3091
3092         rb_decrement_entry(cpu_buffer, event);
3093         if (rb_try_to_discard(cpu_buffer, event))
3094                 goto out;
3095
3096         /*
3097          * The commit is still visible by the reader, so we
3098          * must still update the timestamp.
3099          */
3100         rb_update_write_stamp(cpu_buffer, event);
3101  out:
3102         rb_end_commit(cpu_buffer);
3103
3104         trace_recursive_unlock(cpu_buffer);
3105
3106         preempt_enable_notrace();
3107
3108 }
3109 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
3110
3111 /**
3112  * ring_buffer_write - write data to the buffer without reserving
3113  * @buffer: The ring buffer to write to.
3114  * @length: The length of the data being written (excluding the event header)
3115  * @data: The data to write to the buffer.
3116  *
3117  * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
3118  * one function. If you already have the data to write to the buffer, it
3119  * may be easier to simply call this function.
3120  *
3121  * Note, like ring_buffer_lock_reserve, the length is the length of the data
3122  * and not the length of the event which would hold the header.
3123  */
3124 int ring_buffer_write(struct ring_buffer *buffer,
3125                       unsigned long length,
3126                       void *data)
3127 {
3128         struct ring_buffer_per_cpu *cpu_buffer;
3129         struct ring_buffer_event *event;
3130         void *body;
3131         int ret = -EBUSY;
3132         int cpu;
3133
3134         preempt_disable_notrace();
3135
3136         if (atomic_read(&buffer->record_disabled))
3137                 goto out;
3138
3139         cpu = raw_smp_processor_id();
3140
3141         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3142                 goto out;
3143
3144         cpu_buffer = buffer->buffers[cpu];
3145
3146         if (atomic_read(&cpu_buffer->record_disabled))
3147                 goto out;
3148
3149         if (length > BUF_MAX_DATA_SIZE)
3150                 goto out;
3151
3152         if (unlikely(trace_recursive_lock(cpu_buffer)))
3153                 goto out;
3154
3155         event = rb_reserve_next_event(buffer, cpu_buffer, length);
3156         if (!event)
3157                 goto out_unlock;
3158
3159         body = rb_event_data(event);
3160
3161         memcpy(body, data, length);
3162
3163         rb_commit(cpu_buffer, event);
3164
3165         rb_wakeups(buffer, cpu_buffer);
3166
3167         ret = 0;
3168
3169  out_unlock:
3170         trace_recursive_unlock(cpu_buffer);
3171
3172  out:
3173         preempt_enable_notrace();
3174
3175         return ret;
3176 }
3177 EXPORT_SYMBOL_GPL(ring_buffer_write);
3178
3179 static bool rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
3180 {
3181         struct buffer_page *reader = cpu_buffer->reader_page;
3182         struct buffer_page *head = rb_set_head_page(cpu_buffer);
3183         struct buffer_page *commit = cpu_buffer->commit_page;
3184
3185         /* In case of error, head will be NULL */
3186         if (unlikely(!head))
3187                 return true;
3188
3189         return reader->read == rb_page_commit(reader) &&
3190                 (commit == reader ||
3191                  (commit == head &&
3192                   head->read == rb_page_commit(commit)));
3193 }
3194
3195 /**
3196  * ring_buffer_record_disable - stop all writes into the buffer
3197  * @buffer: The ring buffer to stop writes to.
3198  *
3199  * This prevents all writes to the buffer. Any attempt to write
3200  * to the buffer after this will fail and return NULL.
3201  *
3202  * The caller should call synchronize_rcu() after this.
3203  */
3204 void ring_buffer_record_disable(struct ring_buffer *buffer)
3205 {
3206         atomic_inc(&buffer->record_disabled);
3207 }
3208 EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
3209
3210 /**
3211  * ring_buffer_record_enable - enable writes to the buffer
3212  * @buffer: The ring buffer to enable writes
3213  *
3214  * Note, multiple disables will need the same number of enables
3215  * to truly enable the writing (much like preempt_disable).
3216  */
3217 void ring_buffer_record_enable(struct ring_buffer *buffer)
3218 {
3219         atomic_dec(&buffer->record_disabled);
3220 }
3221 EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
3222
3223 /**
3224  * ring_buffer_record_off - stop all writes into the buffer
3225  * @buffer: The ring buffer to stop writes to.
3226  *
3227  * This prevents all writes to the buffer. Any attempt to write
3228  * to the buffer after this will fail and return NULL.
3229  *
3230  * This is different than ring_buffer_record_disable() as
3231  * it works like an on/off switch, where as the disable() version
3232  * must be paired with a enable().
3233  */
3234 void ring_buffer_record_off(struct ring_buffer *buffer)
3235 {
3236         unsigned int rd;
3237         unsigned int new_rd;
3238
3239         do {
3240                 rd = atomic_read(&buffer->record_disabled);
3241                 new_rd = rd | RB_BUFFER_OFF;
3242         } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3243 }
3244 EXPORT_SYMBOL_GPL(ring_buffer_record_off);
3245
3246 /**
3247  * ring_buffer_record_on - restart writes into the buffer
3248  * @buffer: The ring buffer to start writes to.
3249  *
3250  * This enables all writes to the buffer that was disabled by
3251  * ring_buffer_record_off().
3252  *
3253  * This is different than ring_buffer_record_enable() as
3254  * it works like an on/off switch, where as the enable() version
3255  * must be paired with a disable().
3256  */
3257 void ring_buffer_record_on(struct ring_buffer *buffer)
3258 {
3259         unsigned int rd;
3260         unsigned int new_rd;
3261
3262         do {
3263                 rd = atomic_read(&buffer->record_disabled);
3264                 new_rd = rd & ~RB_BUFFER_OFF;
3265         } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3266 }
3267 EXPORT_SYMBOL_GPL(ring_buffer_record_on);
3268
3269 /**
3270  * ring_buffer_record_is_on - return true if the ring buffer can write
3271  * @buffer: The ring buffer to see if write is enabled
3272  *
3273  * Returns true if the ring buffer is in a state that it accepts writes.
3274  */
3275 bool ring_buffer_record_is_on(struct ring_buffer *buffer)
3276 {
3277         return !atomic_read(&buffer->record_disabled);
3278 }
3279
3280 /**
3281  * ring_buffer_record_is_set_on - return true if the ring buffer is set writable
3282  * @buffer: The ring buffer to see if write is set enabled
3283  *
3284  * Returns true if the ring buffer is set writable by ring_buffer_record_on().
3285  * Note that this does NOT mean it is in a writable state.
3286  *
3287  * It may return true when the ring buffer has been disabled by
3288  * ring_buffer_record_disable(), as that is a temporary disabling of
3289  * the ring buffer.
3290  */
3291 bool ring_buffer_record_is_set_on(struct ring_buffer *buffer)
3292 {
3293         return !(atomic_read(&buffer->record_disabled) & RB_BUFFER_OFF);
3294 }
3295
3296 /**
3297  * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
3298  * @buffer: The ring buffer to stop writes to.
3299  * @cpu: The CPU buffer to stop
3300  *
3301  * This prevents all writes to the buffer. Any attempt to write
3302  * to the buffer after this will fail and return NULL.
3303  *
3304  * The caller should call synchronize_rcu() after this.
3305  */
3306 void ring_buffer_record_disable_cpu(struct ring_buffer *buffer, int cpu)
3307 {
3308         struct ring_buffer_per_cpu *cpu_buffer;
3309
3310         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3311                 return;
3312
3313         cpu_buffer = buffer->buffers[cpu];
3314         atomic_inc(&cpu_buffer->record_disabled);
3315 }
3316 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
3317
3318 /**
3319  * ring_buffer_record_enable_cpu - enable writes to the buffer
3320  * @buffer: The ring buffer to enable writes
3321  * @cpu: The CPU to enable.
3322  *
3323  * Note, multiple disables will need the same number of enables
3324  * to truly enable the writing (much like preempt_disable).
3325  */
3326 void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu)
3327 {
3328         struct ring_buffer_per_cpu *cpu_buffer;
3329
3330         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3331                 return;
3332
3333         cpu_buffer = buffer->buffers[cpu];
3334         atomic_dec(&cpu_buffer->record_disabled);
3335 }
3336 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
3337
3338 /*
3339  * The total entries in the ring buffer is the running counter
3340  * of entries entered into the ring buffer, minus the sum of
3341  * the entries read from the ring buffer and the number of
3342  * entries that were overwritten.
3343  */
3344 static inline unsigned long
3345 rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
3346 {
3347         return local_read(&cpu_buffer->entries) -
3348                 (local_read(&cpu_buffer->overrun) + cpu_buffer->read);
3349 }
3350
3351 /**
3352  * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
3353  * @buffer: The ring buffer
3354  * @cpu: The per CPU buffer to read from.
3355  */
3356 u64 ring_buffer_oldest_event_ts(struct ring_buffer *buffer, int cpu)
3357 {
3358         unsigned long flags;
3359         struct ring_buffer_per_cpu *cpu_buffer;
3360         struct buffer_page *bpage;
3361         u64 ret = 0;
3362
3363         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3364                 return 0;
3365
3366         cpu_buffer = buffer->buffers[cpu];
3367         raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3368         /*
3369          * if the tail is on reader_page, oldest time stamp is on the reader
3370          * page
3371          */
3372         if (cpu_buffer->tail_page == cpu_buffer->reader_page)
3373                 bpage = cpu_buffer->reader_page;
3374         else
3375                 bpage = rb_set_head_page(cpu_buffer);
3376         if (bpage)
3377                 ret = bpage->page->time_stamp;
3378         raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3379
3380         return ret;
3381 }
3382 EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
3383
3384 /**
3385  * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
3386  * @buffer: The ring buffer
3387  * @cpu: The per CPU buffer to read from.
3388  */
3389 unsigned long ring_buffer_bytes_cpu(struct ring_buffer *buffer, int cpu)
3390 {
3391         struct ring_buffer_per_cpu *cpu_buffer;
3392         unsigned long ret;
3393
3394         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3395                 return 0;
3396
3397         cpu_buffer = buffer->buffers[cpu];
3398         ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
3399
3400         return ret;
3401 }
3402 EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
3403
3404 /**
3405  * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
3406  * @buffer: The ring buffer
3407  * @cpu: The per CPU buffer to get the entries from.
3408  */
3409 unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu)
3410 {
3411         struct ring_buffer_per_cpu *cpu_buffer;
3412
3413         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3414                 return 0;
3415
3416         cpu_buffer = buffer->buffers[cpu];
3417
3418         return rb_num_of_entries(cpu_buffer);
3419 }
3420 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
3421
3422 /**
3423  * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
3424  * buffer wrapping around (only if RB_FL_OVERWRITE is on).
3425  * @buffer: The ring buffer
3426  * @cpu: The per CPU buffer to get the number of overruns from
3427  */
3428 unsigned long ring_buffer_overrun_cpu(struct ring_buffer *buffer, int cpu)
3429 {
3430         struct ring_buffer_per_cpu *cpu_buffer;
3431         unsigned long ret;
3432
3433         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3434                 return 0;
3435
3436         cpu_buffer = buffer->buffers[cpu];
3437         ret = local_read(&cpu_buffer->overrun);
3438
3439         return ret;
3440 }
3441 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
3442
3443 /**
3444  * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
3445  * commits failing due to the buffer wrapping around while there are uncommitted
3446  * events, such as during an interrupt storm.
3447  * @buffer: The ring buffer
3448  * @cpu: The per CPU buffer to get the number of overruns from
3449  */
3450 unsigned long
3451 ring_buffer_commit_overrun_cpu(struct ring_buffer *buffer, int cpu)
3452 {
3453         struct ring_buffer_per_cpu *cpu_buffer;
3454         unsigned long ret;
3455
3456         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3457                 return 0;
3458
3459         cpu_buffer = buffer->buffers[cpu];
3460         ret = local_read(&cpu_buffer->commit_overrun);
3461
3462         return ret;
3463 }
3464 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
3465
3466 /**
3467  * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
3468  * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
3469  * @buffer: The ring buffer
3470  * @cpu: The per CPU buffer to get the number of overruns from
3471  */
3472 unsigned long
3473 ring_buffer_dropped_events_cpu(struct ring_buffer *buffer, int cpu)
3474 {
3475         struct ring_buffer_per_cpu *cpu_buffer;
3476         unsigned long ret;
3477
3478         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3479                 return 0;
3480
3481         cpu_buffer = buffer->buffers[cpu];
3482         ret = local_read(&cpu_buffer->dropped_events);
3483
3484         return ret;
3485 }
3486 EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
3487
3488 /**
3489  * ring_buffer_read_events_cpu - get the number of events successfully read
3490  * @buffer: The ring buffer
3491  * @cpu: The per CPU buffer to get the number of events read
3492  */
3493 unsigned long
3494 ring_buffer_read_events_cpu(struct ring_buffer *buffer, int cpu)
3495 {
3496         struct ring_buffer_per_cpu *cpu_buffer;
3497
3498         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3499                 return 0;
3500
3501         cpu_buffer = buffer->buffers[cpu];
3502         return cpu_buffer->read;
3503 }
3504 EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu);
3505
3506 /**
3507  * ring_buffer_entries - get the number of entries in a buffer
3508  * @buffer: The ring buffer
3509  *
3510  * Returns the total number of entries in the ring buffer
3511  * (all CPU entries)
3512  */
3513 unsigned long ring_buffer_entries(struct ring_buffer *buffer)
3514 {
3515         struct ring_buffer_per_cpu *cpu_buffer;
3516         unsigned long entries = 0;
3517         int cpu;
3518
3519         /* if you care about this being correct, lock the buffer */
3520         for_each_buffer_cpu(buffer, cpu) {
3521                 cpu_buffer = buffer->buffers[cpu];
3522                 entries += rb_num_of_entries(cpu_buffer);
3523         }
3524
3525         return entries;
3526 }
3527 EXPORT_SYMBOL_GPL(ring_buffer_entries);
3528
3529 /**
3530  * ring_buffer_overruns - get the number of overruns in buffer
3531  * @buffer: The ring buffer
3532  *
3533  * Returns the total number of overruns in the ring buffer
3534  * (all CPU entries)
3535  */
3536 unsigned long ring_buffer_overruns(struct ring_buffer *buffer)
3537 {
3538         struct ring_buffer_per_cpu *cpu_buffer;
3539         unsigned long overruns = 0;
3540         int cpu;
3541
3542         /* if you care about this being correct, lock the buffer */
3543         for_each_buffer_cpu(buffer, cpu) {
3544                 cpu_buffer = buffer->buffers[cpu];
3545                 overruns += local_read(&cpu_buffer->overrun);
3546         }
3547
3548         return overruns;
3549 }
3550 EXPORT_SYMBOL_GPL(ring_buffer_overruns);
3551
3552 static void rb_iter_reset(struct ring_buffer_iter *iter)
3553 {
3554         struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3555
3556         /* Iterator usage is expected to have record disabled */
3557         iter->head_page = cpu_buffer->reader_page;
3558         iter->head = cpu_buffer->reader_page->read;
3559
3560         iter->cache_reader_page = iter->head_page;
3561         iter->cache_read = cpu_buffer->read;
3562
3563         if (iter->head)
3564                 iter->read_stamp = cpu_buffer->read_stamp;
3565         else
3566                 iter->read_stamp = iter->head_page->page->time_stamp;
3567 }
3568
3569 /**
3570  * ring_buffer_iter_reset - reset an iterator
3571  * @iter: The iterator to reset
3572  *
3573  * Resets the iterator, so that it will start from the beginning
3574  * again.
3575  */
3576 void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
3577 {
3578         struct ring_buffer_per_cpu *cpu_buffer;
3579         unsigned long flags;
3580
3581         if (!iter)
3582                 return;
3583
3584         cpu_buffer = iter->cpu_buffer;
3585
3586         raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3587         rb_iter_reset(iter);
3588         raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3589 }
3590 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
3591
3592 /**
3593  * ring_buffer_iter_empty - check if an iterator has no more to read
3594  * @iter: The iterator to check
3595  */
3596 int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
3597 {
3598         struct ring_buffer_per_cpu *cpu_buffer;
3599         struct buffer_page *reader;
3600         struct buffer_page *head_page;
3601         struct buffer_page *commit_page;
3602         unsigned commit;
3603
3604         cpu_buffer = iter->cpu_buffer;
3605
3606         /* Remember, trace recording is off when iterator is in use */
3607         reader = cpu_buffer->reader_page;
3608         head_page = cpu_buffer->head_page;
3609         commit_page = cpu_buffer->commit_page;
3610         commit = rb_page_commit(commit_page);
3611
3612         return ((iter->head_page == commit_page && iter->head == commit) ||
3613                 (iter->head_page == reader && commit_page == head_page &&
3614                  head_page->read == commit &&
3615                  iter->head == rb_page_commit(cpu_buffer->reader_page)));
3616 }
3617 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
3618
3619 static void
3620 rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
3621                      struct ring_buffer_event *event)
3622 {
3623         u64 delta;
3624
3625         switch (event->type_len) {
3626         case RINGBUF_TYPE_PADDING:
3627                 return;
3628
3629         case RINGBUF_TYPE_TIME_EXTEND:
3630                 delta = ring_buffer_event_time_stamp(event);
3631                 cpu_buffer->read_stamp += delta;
3632                 return;
3633
3634         case RINGBUF_TYPE_TIME_STAMP:
3635                 delta = ring_buffer_event_time_stamp(event);
3636                 cpu_buffer->read_stamp = delta;
3637                 return;
3638
3639         case RINGBUF_TYPE_DATA:
3640                 cpu_buffer->read_stamp += event->time_delta;
3641                 return;
3642
3643         default:
3644                 BUG();
3645         }
3646         return;
3647 }
3648
3649 static void
3650 rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
3651                           struct ring_buffer_event *event)
3652 {
3653         u64 delta;
3654
3655         switch (event->type_len) {
3656         case RINGBUF_TYPE_PADDING:
3657                 return;
3658
3659         case RINGBUF_TYPE_TIME_EXTEND:
3660                 delta = ring_buffer_event_time_stamp(event);
3661                 iter->read_stamp += delta;
3662                 return;
3663
3664         case RINGBUF_TYPE_TIME_STAMP:
3665                 delta = ring_buffer_event_time_stamp(event);
3666                 iter->read_stamp = delta;
3667                 return;
3668
3669         case RINGBUF_TYPE_DATA:
3670                 iter->read_stamp += event->time_delta;
3671                 return;
3672
3673         default:
3674                 BUG();
3675         }
3676         return;
3677 }
3678
3679 static struct buffer_page *
3680 rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
3681 {
3682         struct buffer_page *reader = NULL;
3683         unsigned long overwrite;
3684         unsigned long flags;
3685         int nr_loops = 0;
3686         int ret;
3687
3688         local_irq_save(flags);
3689         arch_spin_lock(&cpu_buffer->lock);
3690
3691  again:
3692         /*
3693          * This should normally only loop twice. But because the
3694          * start of the reader inserts an empty page, it causes
3695          * a case where we will loop three times. There should be no
3696          * reason to loop four times (that I know of).
3697          */
3698         if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
3699                 reader = NULL;