source: src/linux/universal/linux-4.9/kernel/trace/ring_buffer.c @ 31859

Last change on this file since 31859 was 31859, checked in by brainslayer, 3 months ago

kernel update

File size: 130.5 KB
Line 
1/*
2 * Generic ring buffer
3 *
4 * Copyright (C) 2008 Steven Rostedt <srostedt@redhat.com>
5 */
6#include <linux/trace_events.h>
7#include <linux/ring_buffer.h>
8#include <linux/trace_clock.h>
9#include <linux/trace_seq.h>
10#include <linux/spinlock.h>
11#include <linux/irq_work.h>
12#include <linux/uaccess.h>
13#include <linux/hardirq.h>
14#include <linux/kthread.h>      /* for self test */
15#include <linux/kmemcheck.h>
16#include <linux/module.h>
17#include <linux/percpu.h>
18#include <linux/mutex.h>
19#include <linux/delay.h>
20#include <linux/slab.h>
21#include <linux/init.h>
22#include <linux/hash.h>
23#include <linux/list.h>
24#include <linux/cpu.h>
25
26#include <asm/local.h>
27
28static void update_pages_handler(struct work_struct *work);
29
30/*
31 * The ring buffer header is special. We must manually up keep it.
32 */
33int ring_buffer_print_entry_header(struct trace_seq *s)
34{
35        trace_seq_puts(s, "# compressed entry header\n");
36        trace_seq_puts(s, "\ttype_len    :    5 bits\n");
37        trace_seq_puts(s, "\ttime_delta  :   27 bits\n");
38        trace_seq_puts(s, "\tarray       :   32 bits\n");
39        trace_seq_putc(s, '\n');
40        trace_seq_printf(s, "\tpadding     : type == %d\n",
41                         RINGBUF_TYPE_PADDING);
42        trace_seq_printf(s, "\ttime_extend : type == %d\n",
43                         RINGBUF_TYPE_TIME_EXTEND);
44        trace_seq_printf(s, "\tdata max type_len  == %d\n",
45                         RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
46
47        return !trace_seq_has_overflowed(s);
48}
49
50/*
51 * The ring buffer is made up of a list of pages. A separate list of pages is
52 * allocated for each CPU. A writer may only write to a buffer that is
53 * associated with the CPU it is currently executing on.  A reader may read
54 * from any per cpu buffer.
55 *
56 * The reader is special. For each per cpu buffer, the reader has its own
57 * reader page. When a reader has read the entire reader page, this reader
58 * page is swapped with another page in the ring buffer.
59 *
60 * Now, as long as the writer is off the reader page, the reader can do what
61 * ever it wants with that page. The writer will never write to that page
62 * again (as long as it is out of the ring buffer).
63 *
64 * Here's some silly ASCII art.
65 *
66 *   +------+
67 *   |reader|          RING BUFFER
68 *   |page  |
69 *   +------+        +---+   +---+   +---+
70 *                   |   |-->|   |-->|   |
71 *                   +---+   +---+   +---+
72 *                     ^               |
73 *                     |               |
74 *                     +---------------+
75 *
76 *
77 *   +------+
78 *   |reader|          RING BUFFER
79 *   |page  |------------------v
80 *   +------+        +---+   +---+   +---+
81 *                   |   |-->|   |-->|   |
82 *                   +---+   +---+   +---+
83 *                     ^               |
84 *                     |               |
85 *                     +---------------+
86 *
87 *
88 *   +------+
89 *   |reader|          RING BUFFER
90 *   |page  |------------------v
91 *   +------+        +---+   +---+   +---+
92 *      ^            |   |-->|   |-->|   |
93 *      |            +---+   +---+   +---+
94 *      |                              |
95 *      |                              |
96 *      +------------------------------+
97 *
98 *
99 *   +------+
100 *   |buffer|          RING BUFFER
101 *   |page  |------------------v
102 *   +------+        +---+   +---+   +---+
103 *      ^            |   |   |   |-->|   |
104 *      |   New      +---+   +---+   +---+
105 *      |  Reader------^               |
106 *      |   page                       |
107 *      +------------------------------+
108 *
109 *
110 * After we make this swap, the reader can hand this page off to the splice
111 * code and be done with it. It can even allocate a new page if it needs to
112 * and swap that into the ring buffer.
113 *
114 * We will be using cmpxchg soon to make all this lockless.
115 *
116 */
117
118/* Used for individual buffers (after the counter) */
119#define RB_BUFFER_OFF           (1 << 20)
120
121#define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
122
123#define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
124#define RB_ALIGNMENT            4U
125#define RB_MAX_SMALL_DATA       (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
126#define RB_EVNT_MIN_SIZE        8U      /* two 32bit words */
127
128#ifndef CONFIG_HAVE_64BIT_ALIGNED_ACCESS
129# define RB_FORCE_8BYTE_ALIGNMENT       0
130# define RB_ARCH_ALIGNMENT              RB_ALIGNMENT
131#else
132# define RB_FORCE_8BYTE_ALIGNMENT       1
133# define RB_ARCH_ALIGNMENT              8U
134#endif
135
136#define RB_ALIGN_DATA           __aligned(RB_ARCH_ALIGNMENT)
137
138/* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
139#define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
140
141enum {
142        RB_LEN_TIME_EXTEND = 8,
143        RB_LEN_TIME_STAMP = 16,
144};
145
146#define skip_time_extend(event) \
147        ((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
148
149static inline int rb_null_event(struct ring_buffer_event *event)
150{
151        return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
152}
153
154static void rb_event_set_padding(struct ring_buffer_event *event)
155{
156        /* padding has a NULL time_delta */
157        event->type_len = RINGBUF_TYPE_PADDING;
158        event->time_delta = 0;
159}
160
161static unsigned
162rb_event_data_length(struct ring_buffer_event *event)
163{
164        unsigned length;
165
166        if (event->type_len)
167                length = event->type_len * RB_ALIGNMENT;
168        else
169                length = event->array[0];
170        return length + RB_EVNT_HDR_SIZE;
171}
172
173/*
174 * Return the length of the given event. Will return
175 * the length of the time extend if the event is a
176 * time extend.
177 */
178static inline unsigned
179rb_event_length(struct ring_buffer_event *event)
180{
181        switch (event->type_len) {
182        case RINGBUF_TYPE_PADDING:
183                if (rb_null_event(event))
184                        /* undefined */
185                        return -1;
186                return  event->array[0] + RB_EVNT_HDR_SIZE;
187
188        case RINGBUF_TYPE_TIME_EXTEND:
189                return RB_LEN_TIME_EXTEND;
190
191        case RINGBUF_TYPE_TIME_STAMP:
192                return RB_LEN_TIME_STAMP;
193
194        case RINGBUF_TYPE_DATA:
195                return rb_event_data_length(event);
196        default:
197                BUG();
198        }
199        /* not hit */
200        return 0;
201}
202
203/*
204 * Return total length of time extend and data,
205 *   or just the event length for all other events.
206 */
207static inline unsigned
208rb_event_ts_length(struct ring_buffer_event *event)
209{
210        unsigned len = 0;
211
212        if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
213                /* time extends include the data event after it */
214                len = RB_LEN_TIME_EXTEND;
215                event = skip_time_extend(event);
216        }
217        return len + rb_event_length(event);
218}
219
220/**
221 * ring_buffer_event_length - return the length of the event
222 * @event: the event to get the length of
223 *
224 * Returns the size of the data load of a data event.
225 * If the event is something other than a data event, it
226 * returns the size of the event itself. With the exception
227 * of a TIME EXTEND, where it still returns the size of the
228 * data load of the data event after it.
229 */
230unsigned ring_buffer_event_length(struct ring_buffer_event *event)
231{
232        unsigned length;
233
234        if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
235                event = skip_time_extend(event);
236
237        length = rb_event_length(event);
238        if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
239                return length;
240        length -= RB_EVNT_HDR_SIZE;
241        if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
242                length -= sizeof(event->array[0]);
243        return length;
244}
245EXPORT_SYMBOL_GPL(ring_buffer_event_length);
246
247/* inline for ring buffer fast paths */
248static void *
249rb_event_data(struct ring_buffer_event *event)
250{
251        if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
252                event = skip_time_extend(event);
253        BUG_ON(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
254        /* If length is in len field, then array[0] has the data */
255        if (event->type_len)
256                return (void *)&event->array[0];
257        /* Otherwise length is in array[0] and array[1] has the data */
258        return (void *)&event->array[1];
259}
260
261/**
262 * ring_buffer_event_data - return the data of the event
263 * @event: the event to get the data from
264 */
265void *ring_buffer_event_data(struct ring_buffer_event *event)
266{
267        return rb_event_data(event);
268}
269EXPORT_SYMBOL_GPL(ring_buffer_event_data);
270
271#define for_each_buffer_cpu(buffer, cpu)                \
272        for_each_cpu(cpu, buffer->cpumask)
273
274#define TS_SHIFT        27
275#define TS_MASK         ((1ULL << TS_SHIFT) - 1)
276#define TS_DELTA_TEST   (~TS_MASK)
277
278/* Flag when events were overwritten */
279#define RB_MISSED_EVENTS        (1 << 31)
280/* Missed count stored at end */
281#define RB_MISSED_STORED        (1 << 30)
282
283struct buffer_data_page {
284        u64              time_stamp;    /* page time stamp */
285        local_t          commit;        /* write committed index */
286        unsigned char    data[] RB_ALIGN_DATA;  /* data of buffer page */
287};
288
289/*
290 * Note, the buffer_page list must be first. The buffer pages
291 * are allocated in cache lines, which means that each buffer
292 * page will be at the beginning of a cache line, and thus
293 * the least significant bits will be zero. We use this to
294 * add flags in the list struct pointers, to make the ring buffer
295 * lockless.
296 */
297struct buffer_page {
298        struct list_head list;          /* list of buffer pages */
299        local_t          write;         /* index for next write */
300        unsigned         read;          /* index for next read */
301        local_t          entries;       /* entries on this page */
302        unsigned long    real_end;      /* real end of data */
303        struct buffer_data_page *page;  /* Actual data page */
304};
305
306/*
307 * The buffer page counters, write and entries, must be reset
308 * atomically when crossing page boundaries. To synchronize this
309 * update, two counters are inserted into the number. One is
310 * the actual counter for the write position or count on the page.
311 *
312 * The other is a counter of updaters. Before an update happens
313 * the update partition of the counter is incremented. This will
314 * allow the updater to update the counter atomically.
315 *
316 * The counter is 20 bits, and the state data is 12.
317 */
318#define RB_WRITE_MASK           0xfffff
319#define RB_WRITE_INTCNT         (1 << 20)
320
321static void rb_init_page(struct buffer_data_page *bpage)
322{
323        local_set(&bpage->commit, 0);
324}
325
326/**
327 * ring_buffer_page_len - the size of data on the page.
328 * @page: The page to read
329 *
330 * Returns the amount of data on the page, including buffer page header.
331 */
332size_t ring_buffer_page_len(void *page)
333{
334        return local_read(&((struct buffer_data_page *)page)->commit)
335                + BUF_PAGE_HDR_SIZE;
336}
337
338/*
339 * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
340 * this issue out.
341 */
342static void free_buffer_page(struct buffer_page *bpage)
343{
344        free_page((unsigned long)bpage->page);
345        kfree(bpage);
346}
347
348/*
349 * We need to fit the time_stamp delta into 27 bits.
350 */
351static inline int test_time_stamp(u64 delta)
352{
353        if (delta & TS_DELTA_TEST)
354                return 1;
355        return 0;
356}
357
358#define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
359
360/* Max payload is BUF_PAGE_SIZE - header (8bytes) */
361#define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
362
363int ring_buffer_print_page_header(struct trace_seq *s)
364{
365        struct buffer_data_page field;
366
367        trace_seq_printf(s, "\tfield: u64 timestamp;\t"
368                         "offset:0;\tsize:%u;\tsigned:%u;\n",
369                         (unsigned int)sizeof(field.time_stamp),
370                         (unsigned int)is_signed_type(u64));
371
372        trace_seq_printf(s, "\tfield: local_t commit;\t"
373                         "offset:%u;\tsize:%u;\tsigned:%u;\n",
374                         (unsigned int)offsetof(typeof(field), commit),
375                         (unsigned int)sizeof(field.commit),
376                         (unsigned int)is_signed_type(long));
377
378        trace_seq_printf(s, "\tfield: int overwrite;\t"
379                         "offset:%u;\tsize:%u;\tsigned:%u;\n",
380                         (unsigned int)offsetof(typeof(field), commit),
381                         1,
382                         (unsigned int)is_signed_type(long));
383
384        trace_seq_printf(s, "\tfield: char data;\t"
385                         "offset:%u;\tsize:%u;\tsigned:%u;\n",
386                         (unsigned int)offsetof(typeof(field), data),
387                         (unsigned int)BUF_PAGE_SIZE,
388                         (unsigned int)is_signed_type(char));
389
390        return !trace_seq_has_overflowed(s);
391}
392
393struct rb_irq_work {
394        struct irq_work                 work;
395        wait_queue_head_t               waiters;
396        wait_queue_head_t               full_waiters;
397        bool                            waiters_pending;
398        bool                            full_waiters_pending;
399        bool                            wakeup_full;
400};
401
402/*
403 * Structure to hold event state and handle nested events.
404 */
405struct rb_event_info {
406        u64                     ts;
407        u64                     delta;
408        unsigned long           length;
409        struct buffer_page      *tail_page;
410        int                     add_timestamp;
411};
412
413/*
414 * Used for which event context the event is in.
415 *  NMI     = 0
416 *  IRQ     = 1
417 *  SOFTIRQ = 2
418 *  NORMAL  = 3
419 *
420 * See trace_recursive_lock() comment below for more details.
421 */
422enum {
423        RB_CTX_NMI,
424        RB_CTX_IRQ,
425        RB_CTX_SOFTIRQ,
426        RB_CTX_NORMAL,
427        RB_CTX_MAX
428};
429
430/*
431 * head_page == tail_page && head == tail then buffer is empty.
432 */
433struct ring_buffer_per_cpu {
434        int                             cpu;
435        atomic_t                        record_disabled;
436        struct ring_buffer              *buffer;
437        raw_spinlock_t                  reader_lock;    /* serialize readers */
438        arch_spinlock_t                 lock;
439        struct lock_class_key           lock_key;
440        unsigned long                   nr_pages;
441        unsigned int                    current_context;
442        struct list_head                *pages;
443        struct buffer_page              *head_page;     /* read from head */
444        struct buffer_page              *tail_page;     /* write to tail */
445        struct buffer_page              *commit_page;   /* committed pages */
446        struct buffer_page              *reader_page;
447        unsigned long                   lost_events;
448        unsigned long                   last_overrun;
449        local_t                         entries_bytes;
450        local_t                         entries;
451        local_t                         overrun;
452        local_t                         commit_overrun;
453        local_t                         dropped_events;
454        local_t                         committing;
455        local_t                         commits;
456        unsigned long                   read;
457        unsigned long                   read_bytes;
458        u64                             write_stamp;
459        u64                             read_stamp;
460        /* ring buffer pages to update, > 0 to add, < 0 to remove */
461        long                            nr_pages_to_update;
462        struct list_head                new_pages; /* new pages to add */
463        struct work_struct              update_pages_work;
464        struct completion               update_done;
465
466        struct rb_irq_work              irq_work;
467};
468
469struct ring_buffer {
470        unsigned                        flags;
471        int                             cpus;
472        atomic_t                        record_disabled;
473        atomic_t                        resize_disabled;
474        cpumask_var_t                   cpumask;
475
476        struct lock_class_key           *reader_lock_key;
477
478        struct mutex                    mutex;
479
480        struct ring_buffer_per_cpu      **buffers;
481
482#ifdef CONFIG_HOTPLUG_CPU
483        struct notifier_block           cpu_notify;
484#endif
485        u64                             (*clock)(void);
486
487        struct rb_irq_work              irq_work;
488};
489
490struct ring_buffer_iter {
491        struct ring_buffer_per_cpu      *cpu_buffer;
492        unsigned long                   head;
493        struct buffer_page              *head_page;
494        struct buffer_page              *cache_reader_page;
495        unsigned long                   cache_read;
496        u64                             read_stamp;
497};
498
499/*
500 * rb_wake_up_waiters - wake up tasks waiting for ring buffer input
501 *
502 * Schedules a delayed work to wake up any task that is blocked on the
503 * ring buffer waiters queue.
504 */
505static void rb_wake_up_waiters(struct irq_work *work)
506{
507        struct rb_irq_work *rbwork = container_of(work, struct rb_irq_work, work);
508
509        wake_up_all(&rbwork->waiters);
510        if (rbwork->wakeup_full) {
511                rbwork->wakeup_full = false;
512                wake_up_all(&rbwork->full_waiters);
513        }
514}
515
516/**
517 * ring_buffer_wait - wait for input to the ring buffer
518 * @buffer: buffer to wait on
519 * @cpu: the cpu buffer to wait on
520 * @full: wait until a full page is available, if @cpu != RING_BUFFER_ALL_CPUS
521 *
522 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
523 * as data is added to any of the @buffer's cpu buffers. Otherwise
524 * it will wait for data to be added to a specific cpu buffer.
525 */
526int ring_buffer_wait(struct ring_buffer *buffer, int cpu, bool full)
527{
528        struct ring_buffer_per_cpu *uninitialized_var(cpu_buffer);
529        DEFINE_WAIT(wait);
530        struct rb_irq_work *work;
531        int ret = 0;
532
533        /*
534         * Depending on what the caller is waiting for, either any
535         * data in any cpu buffer, or a specific buffer, put the
536         * caller on the appropriate wait queue.
537         */
538        if (cpu == RING_BUFFER_ALL_CPUS) {
539                work = &buffer->irq_work;
540                /* Full only makes sense on per cpu reads */
541                full = false;
542        } else {
543                if (!cpumask_test_cpu(cpu, buffer->cpumask))
544                        return -ENODEV;
545                cpu_buffer = buffer->buffers[cpu];
546                work = &cpu_buffer->irq_work;
547        }
548
549
550        while (true) {
551                if (full)
552                        prepare_to_wait(&work->full_waiters, &wait, TASK_INTERRUPTIBLE);
553                else
554                        prepare_to_wait(&work->waiters, &wait, TASK_INTERRUPTIBLE);
555
556                /*
557                 * The events can happen in critical sections where
558                 * checking a work queue can cause deadlocks.
559                 * After adding a task to the queue, this flag is set
560                 * only to notify events to try to wake up the queue
561                 * using irq_work.
562                 *
563                 * We don't clear it even if the buffer is no longer
564                 * empty. The flag only causes the next event to run
565                 * irq_work to do the work queue wake up. The worse
566                 * that can happen if we race with !trace_empty() is that
567                 * an event will cause an irq_work to try to wake up
568                 * an empty queue.
569                 *
570                 * There's no reason to protect this flag either, as
571                 * the work queue and irq_work logic will do the necessary
572                 * synchronization for the wake ups. The only thing
573                 * that is necessary is that the wake up happens after
574                 * a task has been queued. It's OK for spurious wake ups.
575                 */
576                if (full)
577                        work->full_waiters_pending = true;
578                else
579                        work->waiters_pending = true;
580
581                if (signal_pending(current)) {
582                        ret = -EINTR;
583                        break;
584                }
585
586                if (cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer))
587                        break;
588
589                if (cpu != RING_BUFFER_ALL_CPUS &&
590                    !ring_buffer_empty_cpu(buffer, cpu)) {
591                        unsigned long flags;
592                        bool pagebusy;
593
594                        if (!full)
595                                break;
596
597                        raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
598                        pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
599                        raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
600
601                        if (!pagebusy)
602                                break;
603                }
604
605                schedule();
606        }
607
608        if (full)
609                finish_wait(&work->full_waiters, &wait);
610        else
611                finish_wait(&work->waiters, &wait);
612
613        return ret;
614}
615
616/**
617 * ring_buffer_poll_wait - poll on buffer input
618 * @buffer: buffer to wait on
619 * @cpu: the cpu buffer to wait on
620 * @filp: the file descriptor
621 * @poll_table: The poll descriptor
622 *
623 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
624 * as data is added to any of the @buffer's cpu buffers. Otherwise
625 * it will wait for data to be added to a specific cpu buffer.
626 *
627 * Returns POLLIN | POLLRDNORM if data exists in the buffers,
628 * zero otherwise.
629 */
630int ring_buffer_poll_wait(struct ring_buffer *buffer, int cpu,
631                          struct file *filp, poll_table *poll_table)
632{
633        struct ring_buffer_per_cpu *cpu_buffer;
634        struct rb_irq_work *work;
635
636        if (cpu == RING_BUFFER_ALL_CPUS)
637                work = &buffer->irq_work;
638        else {
639                if (!cpumask_test_cpu(cpu, buffer->cpumask))
640                        return -EINVAL;
641
642                cpu_buffer = buffer->buffers[cpu];
643                work = &cpu_buffer->irq_work;
644        }
645
646        poll_wait(filp, &work->waiters, poll_table);
647        work->waiters_pending = true;
648        /*
649         * There's a tight race between setting the waiters_pending and
650         * checking if the ring buffer is empty.  Once the waiters_pending bit
651         * is set, the next event will wake the task up, but we can get stuck
652         * if there's only a single event in.
653         *
654         * FIXME: Ideally, we need a memory barrier on the writer side as well,
655         * but adding a memory barrier to all events will cause too much of a
656         * performance hit in the fast path.  We only need a memory barrier when
657         * the buffer goes from empty to having content.  But as this race is
658         * extremely small, and it's not a problem if another event comes in, we
659         * will fix it later.
660         */
661        smp_mb();
662
663        if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
664            (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
665                return POLLIN | POLLRDNORM;
666        return 0;
667}
668
669/* buffer may be either ring_buffer or ring_buffer_per_cpu */
670#define RB_WARN_ON(b, cond)                                             \
671        ({                                                              \
672                int _____ret = unlikely(cond);                          \
673                if (_____ret) {                                         \
674                        if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
675                                struct ring_buffer_per_cpu *__b =       \
676                                        (void *)b;                      \
677                                atomic_inc(&__b->buffer->record_disabled); \
678                        } else                                          \
679                                atomic_inc(&b->record_disabled);        \
680                        WARN_ON(1);                                     \
681                }                                                       \
682                _____ret;                                               \
683        })
684
685/* Up this if you want to test the TIME_EXTENTS and normalization */
686#define DEBUG_SHIFT 0
687
688static inline u64 rb_time_stamp(struct ring_buffer *buffer)
689{
690        /* shift to debug/test normalization and TIME_EXTENTS */
691        return buffer->clock() << DEBUG_SHIFT;
692}
693
694u64 ring_buffer_time_stamp(struct ring_buffer *buffer, int cpu)
695{
696        u64 time;
697
698        preempt_disable_notrace();
699        time = rb_time_stamp(buffer);
700        preempt_enable_no_resched_notrace();
701
702        return time;
703}
704EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
705
706void ring_buffer_normalize_time_stamp(struct ring_buffer *buffer,
707                                      int cpu, u64 *ts)
708{
709        /* Just stupid testing the normalize function and deltas */
710        *ts >>= DEBUG_SHIFT;
711}
712EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
713
714/*
715 * Making the ring buffer lockless makes things tricky.
716 * Although writes only happen on the CPU that they are on,
717 * and they only need to worry about interrupts. Reads can
718 * happen on any CPU.
719 *
720 * The reader page is always off the ring buffer, but when the
721 * reader finishes with a page, it needs to swap its page with
722 * a new one from the buffer. The reader needs to take from
723 * the head (writes go to the tail). But if a writer is in overwrite
724 * mode and wraps, it must push the head page forward.
725 *
726 * Here lies the problem.
727 *
728 * The reader must be careful to replace only the head page, and
729 * not another one. As described at the top of the file in the
730 * ASCII art, the reader sets its old page to point to the next
731 * page after head. It then sets the page after head to point to
732 * the old reader page. But if the writer moves the head page
733 * during this operation, the reader could end up with the tail.
734 *
735 * We use cmpxchg to help prevent this race. We also do something
736 * special with the page before head. We set the LSB to 1.
737 *
738 * When the writer must push the page forward, it will clear the
739 * bit that points to the head page, move the head, and then set
740 * the bit that points to the new head page.
741 *
742 * We also don't want an interrupt coming in and moving the head
743 * page on another writer. Thus we use the second LSB to catch
744 * that too. Thus:
745 *
746 * head->list->prev->next        bit 1          bit 0
747 *                              -------        -------
748 * Normal page                     0              0
749 * Points to head page             0              1
750 * New head page                   1              0
751 *
752 * Note we can not trust the prev pointer of the head page, because:
753 *
754 * +----+       +-----+        +-----+
755 * |    |------>|  T  |---X--->|  N  |
756 * |    |<------|     |        |     |
757 * +----+       +-----+        +-----+
758 *   ^                           ^ |
759 *   |          +-----+          | |
760 *   +----------|  R  |----------+ |
761 *              |     |<-----------+
762 *              +-----+
763 *
764 * Key:  ---X-->  HEAD flag set in pointer
765 *         T      Tail page
766 *         R      Reader page
767 *         N      Next page
768 *
769 * (see __rb_reserve_next() to see where this happens)
770 *
771 *  What the above shows is that the reader just swapped out
772 *  the reader page with a page in the buffer, but before it
773 *  could make the new header point back to the new page added
774 *  it was preempted by a writer. The writer moved forward onto
775 *  the new page added by the reader and is about to move forward
776 *  again.
777 *
778 *  You can see, it is legitimate for the previous pointer of
779 *  the head (or any page) not to point back to itself. But only
780 *  temporarially.
781 */
782
783#define RB_PAGE_NORMAL          0UL
784#define RB_PAGE_HEAD            1UL
785#define RB_PAGE_UPDATE          2UL
786
787
788#define RB_FLAG_MASK            3UL
789
790/* PAGE_MOVED is not part of the mask */
791#define RB_PAGE_MOVED           4UL
792
793/*
794 * rb_list_head - remove any bit
795 */
796static struct list_head *rb_list_head(struct list_head *list)
797{
798        unsigned long val = (unsigned long)list;
799
800        return (struct list_head *)(val & ~RB_FLAG_MASK);
801}
802
803/*
804 * rb_is_head_page - test if the given page is the head page
805 *
806 * Because the reader may move the head_page pointer, we can
807 * not trust what the head page is (it may be pointing to
808 * the reader page). But if the next page is a header page,
809 * its flags will be non zero.
810 */
811static inline int
812rb_is_head_page(struct ring_buffer_per_cpu *cpu_buffer,
813                struct buffer_page *page, struct list_head *list)
814{
815        unsigned long val;
816
817        val = (unsigned long)list->next;
818
819        if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
820                return RB_PAGE_MOVED;
821
822        return val & RB_FLAG_MASK;
823}
824
825/*
826 * rb_is_reader_page
827 *
828 * The unique thing about the reader page, is that, if the
829 * writer is ever on it, the previous pointer never points
830 * back to the reader page.
831 */
832static bool rb_is_reader_page(struct buffer_page *page)
833{
834        struct list_head *list = page->list.prev;
835
836        return rb_list_head(list->next) != &page->list;
837}
838
839/*
840 * rb_set_list_to_head - set a list_head to be pointing to head.
841 */
842static void rb_set_list_to_head(struct ring_buffer_per_cpu *cpu_buffer,
843                                struct list_head *list)
844{
845        unsigned long *ptr;
846
847        ptr = (unsigned long *)&list->next;
848        *ptr |= RB_PAGE_HEAD;
849        *ptr &= ~RB_PAGE_UPDATE;
850}
851
852/*
853 * rb_head_page_activate - sets up head page
854 */
855static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
856{
857        struct buffer_page *head;
858
859        head = cpu_buffer->head_page;
860        if (!head)
861                return;
862
863        /*
864         * Set the previous list pointer to have the HEAD flag.
865         */
866        rb_set_list_to_head(cpu_buffer, head->list.prev);
867}
868
869static void rb_list_head_clear(struct list_head *list)
870{
871        unsigned long *ptr = (unsigned long *)&list->next;
872
873        *ptr &= ~RB_FLAG_MASK;
874}
875
876/*
877 * rb_head_page_dactivate - clears head page ptr (for free list)
878 */
879static void
880rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
881{
882        struct list_head *hd;
883
884        /* Go through the whole list and clear any pointers found. */
885        rb_list_head_clear(cpu_buffer->pages);
886
887        list_for_each(hd, cpu_buffer->pages)
888                rb_list_head_clear(hd);
889}
890
891static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
892                            struct buffer_page *head,
893                            struct buffer_page *prev,
894                            int old_flag, int new_flag)
895{
896        struct list_head *list;
897        unsigned long val = (unsigned long)&head->list;
898        unsigned long ret;
899
900        list = &prev->list;
901
902        val &= ~RB_FLAG_MASK;
903
904        ret = cmpxchg((unsigned long *)&list->next,
905                      val | old_flag, val | new_flag);
906
907        /* check if the reader took the page */
908        if ((ret & ~RB_FLAG_MASK) != val)
909                return RB_PAGE_MOVED;
910
911        return ret & RB_FLAG_MASK;
912}
913
914static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
915                                   struct buffer_page *head,
916                                   struct buffer_page *prev,
917                                   int old_flag)
918{
919        return rb_head_page_set(cpu_buffer, head, prev,
920                                old_flag, RB_PAGE_UPDATE);
921}
922
923static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
924                                 struct buffer_page *head,
925                                 struct buffer_page *prev,
926                                 int old_flag)
927{
928        return rb_head_page_set(cpu_buffer, head, prev,
929                                old_flag, RB_PAGE_HEAD);
930}
931
932static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
933                                   struct buffer_page *head,
934                                   struct buffer_page *prev,
935                                   int old_flag)
936{
937        return rb_head_page_set(cpu_buffer, head, prev,
938                                old_flag, RB_PAGE_NORMAL);
939}
940
941static inline void rb_inc_page(struct ring_buffer_per_cpu *cpu_buffer,
942                               struct buffer_page **bpage)
943{
944        struct list_head *p = rb_list_head((*bpage)->list.next);
945
946        *bpage = list_entry(p, struct buffer_page, list);
947}
948
949static struct buffer_page *
950rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
951{
952        struct buffer_page *head;
953        struct buffer_page *page;
954        struct list_head *list;
955        int i;
956
957        if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
958                return NULL;
959
960        /* sanity check */
961        list = cpu_buffer->pages;
962        if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
963                return NULL;
964
965        page = head = cpu_buffer->head_page;
966        /*
967         * It is possible that the writer moves the header behind
968         * where we started, and we miss in one loop.
969         * A second loop should grab the header, but we'll do
970         * three loops just because I'm paranoid.
971         */
972        for (i = 0; i < 3; i++) {
973                do {
974                        if (rb_is_head_page(cpu_buffer, page, page->list.prev)) {
975                                cpu_buffer->head_page = page;
976                                return page;
977                        }
978                        rb_inc_page(cpu_buffer, &page);
979                } while (page != head);
980        }
981
982        RB_WARN_ON(cpu_buffer, 1);
983
984        return NULL;
985}
986
987static int rb_head_page_replace(struct buffer_page *old,
988                                struct buffer_page *new)
989{
990        unsigned long *ptr = (unsigned long *)&old->list.prev->next;
991        unsigned long val;
992        unsigned long ret;
993
994        val = *ptr & ~RB_FLAG_MASK;
995        val |= RB_PAGE_HEAD;
996
997        ret = cmpxchg(ptr, val, (unsigned long)&new->list);
998
999        return ret == val;
1000}
1001
1002/*
1003 * rb_tail_page_update - move the tail page forward
1004 */
1005static void rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
1006                               struct buffer_page *tail_page,
1007                               struct buffer_page *next_page)
1008{
1009        unsigned long old_entries;
1010        unsigned long old_write;
1011
1012        /*
1013         * The tail page now needs to be moved forward.
1014         *
1015         * We need to reset the tail page, but without messing
1016         * with possible erasing of data brought in by interrupts
1017         * that have moved the tail page and are currently on it.
1018         *
1019         * We add a counter to the write field to denote this.
1020         */
1021        old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
1022        old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
1023
1024        /*
1025         * Just make sure we have seen our old_write and synchronize
1026         * with any interrupts that come in.
1027         */
1028        barrier();
1029
1030        /*
1031         * If the tail page is still the same as what we think
1032         * it is, then it is up to us to update the tail
1033         * pointer.
1034         */
1035        if (tail_page == READ_ONCE(cpu_buffer->tail_page)) {
1036                /* Zero the write counter */
1037                unsigned long val = old_write & ~RB_WRITE_MASK;
1038                unsigned long eval = old_entries & ~RB_WRITE_MASK;
1039
1040                /*
1041                 * This will only succeed if an interrupt did
1042                 * not come in and change it. In which case, we
1043                 * do not want to modify it.
1044                 *
1045                 * We add (void) to let the compiler know that we do not care
1046                 * about the return value of these functions. We use the
1047                 * cmpxchg to only update if an interrupt did not already
1048                 * do it for us. If the cmpxchg fails, we don't care.
1049                 */
1050                (void)local_cmpxchg(&next_page->write, old_write, val);
1051                (void)local_cmpxchg(&next_page->entries, old_entries, eval);
1052
1053                /*
1054                 * No need to worry about races with clearing out the commit.
1055                 * it only can increment when a commit takes place. But that
1056                 * only happens in the outer most nested commit.
1057                 */
1058                local_set(&next_page->page->commit, 0);
1059
1060                /* Again, either we update tail_page or an interrupt does */
1061                (void)cmpxchg(&cpu_buffer->tail_page, tail_page, next_page);
1062        }
1063}
1064
1065static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
1066                          struct buffer_page *bpage)
1067{
1068        unsigned long val = (unsigned long)bpage;
1069
1070        if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK))
1071                return 1;
1072
1073        return 0;
1074}
1075
1076/**
1077 * rb_check_list - make sure a pointer to a list has the last bits zero
1078 */
1079static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer,
1080                         struct list_head *list)
1081{
1082        if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev))
1083                return 1;
1084        if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next))
1085                return 1;
1086        return 0;
1087}
1088
1089/**
1090 * rb_check_pages - integrity check of buffer pages
1091 * @cpu_buffer: CPU buffer with pages to test
1092 *
1093 * As a safety measure we check to make sure the data pages have not
1094 * been corrupted.
1095 */
1096static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
1097{
1098        struct list_head *head = cpu_buffer->pages;
1099        struct buffer_page *bpage, *tmp;
1100
1101        /* Reset the head page if it exists */
1102        if (cpu_buffer->head_page)
1103                rb_set_head_page(cpu_buffer);
1104
1105        rb_head_page_deactivate(cpu_buffer);
1106
1107        if (RB_WARN_ON(cpu_buffer, head->next->prev != head))
1108                return -1;
1109        if (RB_WARN_ON(cpu_buffer, head->prev->next != head))
1110                return -1;
1111
1112        if (rb_check_list(cpu_buffer, head))
1113                return -1;
1114
1115        list_for_each_entry_safe(bpage, tmp, head, list) {
1116                if (RB_WARN_ON(cpu_buffer,
1117                               bpage->list.next->prev != &bpage->list))
1118                        return -1;
1119                if (RB_WARN_ON(cpu_buffer,
1120                               bpage->list.prev->next != &bpage->list))
1121                        return -1;
1122                if (rb_check_list(cpu_buffer, &bpage->list))
1123                        return -1;
1124        }
1125
1126        rb_head_page_activate(cpu_buffer);
1127
1128        return 0;
1129}
1130
1131static int __rb_allocate_pages(long nr_pages, struct list_head *pages, int cpu)
1132{
1133        struct buffer_page *bpage, *tmp;
1134        long i;
1135
1136        for (i = 0; i < nr_pages; i++) {
1137                struct page *page;
1138                /*
1139                 * __GFP_NORETRY flag makes sure that the allocation fails
1140                 * gracefully without invoking oom-killer and the system is
1141                 * not destabilized.
1142                 */
1143                bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1144                                    GFP_KERNEL | __GFP_NORETRY,
1145                                    cpu_to_node(cpu));
1146                if (!bpage)
1147                        goto free_pages;
1148
1149                list_add(&bpage->list, pages);
1150
1151                page = alloc_pages_node(cpu_to_node(cpu),
1152                                        GFP_KERNEL | __GFP_NORETRY, 0);
1153                if (!page)
1154                        goto free_pages;
1155                bpage->page = page_address(page);
1156                rb_init_page(bpage->page);
1157        }
1158
1159        return 0;
1160
1161free_pages:
1162        list_for_each_entry_safe(bpage, tmp, pages, list) {
1163                list_del_init(&bpage->list);
1164                free_buffer_page(bpage);
1165        }
1166
1167        return -ENOMEM;
1168}
1169
1170static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1171                             unsigned long nr_pages)
1172{
1173        LIST_HEAD(pages);
1174
1175        WARN_ON(!nr_pages);
1176
1177        if (__rb_allocate_pages(nr_pages, &pages, cpu_buffer->cpu))
1178                return -ENOMEM;
1179
1180        /*
1181         * The ring buffer page list is a circular list that does not
1182         * start and end with a list head. All page list items point to
1183         * other pages.
1184         */
1185        cpu_buffer->pages = pages.next;
1186        list_del(&pages);
1187
1188        cpu_buffer->nr_pages = nr_pages;
1189
1190        rb_check_pages(cpu_buffer);
1191
1192        return 0;
1193}
1194
1195static struct ring_buffer_per_cpu *
1196rb_allocate_cpu_buffer(struct ring_buffer *buffer, long nr_pages, int cpu)
1197{
1198        struct ring_buffer_per_cpu *cpu_buffer;
1199        struct buffer_page *bpage;
1200        struct page *page;
1201        int ret;
1202
1203        cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1204                                  GFP_KERNEL, cpu_to_node(cpu));
1205        if (!cpu_buffer)
1206                return NULL;
1207
1208        cpu_buffer->cpu = cpu;
1209        cpu_buffer->buffer = buffer;
1210        raw_spin_lock_init(&cpu_buffer->reader_lock);
1211        lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1212        cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1213        INIT_WORK(&cpu_buffer->update_pages_work, update_pages_handler);
1214        init_completion(&cpu_buffer->update_done);
1215        init_irq_work(&cpu_buffer->irq_work.work, rb_wake_up_waiters);
1216        init_waitqueue_head(&cpu_buffer->irq_work.waiters);
1217        init_waitqueue_head(&cpu_buffer->irq_work.full_waiters);
1218
1219        bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1220                            GFP_KERNEL, cpu_to_node(cpu));
1221        if (!bpage)
1222                goto fail_free_buffer;
1223
1224        rb_check_bpage(cpu_buffer, bpage);
1225
1226        cpu_buffer->reader_page = bpage;
1227        page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0);
1228        if (!page)
1229                goto fail_free_reader;
1230        bpage->page = page_address(page);
1231        rb_init_page(bpage->page);
1232
1233        INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1234        INIT_LIST_HEAD(&cpu_buffer->new_pages);
1235
1236        ret = rb_allocate_pages(cpu_buffer, nr_pages);
1237        if (ret < 0)
1238                goto fail_free_reader;
1239
1240        cpu_buffer->head_page
1241                = list_entry(cpu_buffer->pages, struct buffer_page, list);
1242        cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1243
1244        rb_head_page_activate(cpu_buffer);
1245
1246        return cpu_buffer;
1247
1248 fail_free_reader:
1249        free_buffer_page(cpu_buffer->reader_page);
1250
1251 fail_free_buffer:
1252        kfree(cpu_buffer);
1253        return NULL;
1254}
1255
1256static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1257{
1258        struct list_head *head = cpu_buffer->pages;
1259        struct buffer_page *bpage, *tmp;
1260
1261        free_buffer_page(cpu_buffer->reader_page);
1262
1263        rb_head_page_deactivate(cpu_buffer);
1264
1265        if (head) {
1266                list_for_each_entry_safe(bpage, tmp, head, list) {
1267                        list_del_init(&bpage->list);
1268                        free_buffer_page(bpage);
1269                }
1270                bpage = list_entry(head, struct buffer_page, list);
1271                free_buffer_page(bpage);
1272        }
1273
1274        kfree(cpu_buffer);
1275}
1276
1277#ifdef CONFIG_HOTPLUG_CPU
1278static int rb_cpu_notify(struct notifier_block *self,
1279                         unsigned long action, void *hcpu);
1280#endif
1281
1282/**
1283 * __ring_buffer_alloc - allocate a new ring_buffer
1284 * @size: the size in bytes per cpu that is needed.
1285 * @flags: attributes to set for the ring buffer.
1286 *
1287 * Currently the only flag that is available is the RB_FL_OVERWRITE
1288 * flag. This flag means that the buffer will overwrite old data
1289 * when the buffer wraps. If this flag is not set, the buffer will
1290 * drop data when the tail hits the head.
1291 */
1292struct ring_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1293                                        struct lock_class_key *key)
1294{
1295        struct ring_buffer *buffer;
1296        long nr_pages;
1297        int bsize;
1298        int cpu;
1299
1300        /* keep it in its own cache line */
1301        buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1302                         GFP_KERNEL);
1303        if (!buffer)
1304                return NULL;
1305
1306        if (!alloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1307                goto fail_free_buffer;
1308
1309        nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1310        buffer->flags = flags;
1311        buffer->clock = trace_clock_local;
1312        buffer->reader_lock_key = key;
1313
1314        init_irq_work(&buffer->irq_work.work, rb_wake_up_waiters);
1315        init_waitqueue_head(&buffer->irq_work.waiters);
1316
1317        /* need at least two pages */
1318        if (nr_pages < 2)
1319                nr_pages = 2;
1320
1321        /*
1322         * In case of non-hotplug cpu, if the ring-buffer is allocated
1323         * in early initcall, it will not be notified of secondary cpus.
1324         * In that off case, we need to allocate for all possible cpus.
1325         */
1326#ifdef CONFIG_HOTPLUG_CPU
1327        cpu_notifier_register_begin();
1328        cpumask_copy(buffer->cpumask, cpu_online_mask);
1329#else
1330        cpumask_copy(buffer->cpumask, cpu_possible_mask);
1331#endif
1332        buffer->cpus = nr_cpu_ids;
1333
1334        bsize = sizeof(void *) * nr_cpu_ids;
1335        buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1336                                  GFP_KERNEL);
1337        if (!buffer->buffers)
1338                goto fail_free_cpumask;
1339
1340        for_each_buffer_cpu(buffer, cpu) {
1341                buffer->buffers[cpu] =
1342                        rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
1343                if (!buffer->buffers[cpu])
1344                        goto fail_free_buffers;
1345        }
1346
1347#ifdef CONFIG_HOTPLUG_CPU
1348        buffer->cpu_notify.notifier_call = rb_cpu_notify;
1349        buffer->cpu_notify.priority = 0;
1350        __register_cpu_notifier(&buffer->cpu_notify);
1351        cpu_notifier_register_done();
1352#endif
1353
1354        mutex_init(&buffer->mutex);
1355
1356        return buffer;
1357
1358 fail_free_buffers:
1359        for_each_buffer_cpu(buffer, cpu) {
1360                if (buffer->buffers[cpu])
1361                        rb_free_cpu_buffer(buffer->buffers[cpu]);
1362        }
1363        kfree(buffer->buffers);
1364
1365 fail_free_cpumask:
1366        free_cpumask_var(buffer->cpumask);
1367#ifdef CONFIG_HOTPLUG_CPU
1368        cpu_notifier_register_done();
1369#endif
1370
1371 fail_free_buffer:
1372        kfree(buffer);
1373        return NULL;
1374}
1375EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1376
1377/**
1378 * ring_buffer_free - free a ring buffer.
1379 * @buffer: the buffer to free.
1380 */
1381void
1382ring_buffer_free(struct ring_buffer *buffer)
1383{
1384        int cpu;
1385
1386#ifdef CONFIG_HOTPLUG_CPU
1387        cpu_notifier_register_begin();
1388        __unregister_cpu_notifier(&buffer->cpu_notify);
1389#endif
1390
1391        for_each_buffer_cpu(buffer, cpu)
1392                rb_free_cpu_buffer(buffer->buffers[cpu]);
1393
1394#ifdef CONFIG_HOTPLUG_CPU
1395        cpu_notifier_register_done();
1396#endif
1397
1398        kfree(buffer->buffers);
1399        free_cpumask_var(buffer->cpumask);
1400
1401        kfree(buffer);
1402}
1403EXPORT_SYMBOL_GPL(ring_buffer_free);
1404
1405void ring_buffer_set_clock(struct ring_buffer *buffer,
1406                           u64 (*clock)(void))
1407{
1408        buffer->clock = clock;
1409}
1410
1411static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1412
1413static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1414{
1415        return local_read(&bpage->entries) & RB_WRITE_MASK;
1416}
1417
1418static inline unsigned long rb_page_write(struct buffer_page *bpage)
1419{
1420        return local_read(&bpage->write) & RB_WRITE_MASK;
1421}
1422
1423static int
1424rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned long nr_pages)
1425{
1426        struct list_head *tail_page, *to_remove, *next_page;
1427        struct buffer_page *to_remove_page, *tmp_iter_page;
1428        struct buffer_page *last_page, *first_page;
1429        unsigned long nr_removed;
1430        unsigned long head_bit;
1431        int page_entries;
1432
1433        head_bit = 0;
1434
1435        raw_spin_lock_irq(&cpu_buffer->reader_lock);
1436        atomic_inc(&cpu_buffer->record_disabled);
1437        /*
1438         * We don't race with the readers since we have acquired the reader
1439         * lock. We also don't race with writers after disabling recording.
1440         * This makes it easy to figure out the first and the last page to be
1441         * removed from the list. We unlink all the pages in between including
1442         * the first and last pages. This is done in a busy loop so that we
1443         * lose the least number of traces.
1444         * The pages are freed after we restart recording and unlock readers.
1445         */
1446        tail_page = &cpu_buffer->tail_page->list;
1447
1448        /*
1449         * tail page might be on reader page, we remove the next page
1450         * from the ring buffer
1451         */
1452        if (cpu_buffer->tail_page == cpu_buffer->reader_page)
1453                tail_page = rb_list_head(tail_page->next);
1454        to_remove = tail_page;
1455
1456        /* start of pages to remove */
1457        first_page = list_entry(rb_list_head(to_remove->next),
1458                                struct buffer_page, list);
1459
1460        for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
1461                to_remove = rb_list_head(to_remove)->next;
1462                head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
1463        }
1464
1465        next_page = rb_list_head(to_remove)->next;
1466
1467        /*
1468         * Now we remove all pages between tail_page and next_page.
1469         * Make sure that we have head_bit value preserved for the
1470         * next page
1471         */
1472        tail_page->next = (struct list_head *)((unsigned long)next_page |
1473                                                head_bit);
1474        next_page = rb_list_head(next_page);
1475        next_page->prev = tail_page;
1476
1477        /* make sure pages points to a valid page in the ring buffer */
1478        cpu_buffer->pages = next_page;
1479
1480        /* update head page */
1481        if (head_bit)
1482                cpu_buffer->head_page = list_entry(next_page,
1483                                                struct buffer_page, list);
1484
1485        /*
1486         * change read pointer to make sure any read iterators reset
1487         * themselves
1488         */
1489        cpu_buffer->read = 0;
1490
1491        /* pages are removed, resume tracing and then free the pages */
1492        atomic_dec(&cpu_buffer->record_disabled);
1493        raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1494
1495        RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
1496
1497        /* last buffer page to remove */
1498        last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
1499                                list);
1500        tmp_iter_page = first_page;
1501
1502        do {
1503                to_remove_page = tmp_iter_page;
1504                rb_inc_page(cpu_buffer, &tmp_iter_page);
1505
1506                /* update the counters */
1507                page_entries = rb_page_entries(to_remove_page);
1508                if (page_entries) {
1509                        /*
1510                         * If something was added to this page, it was full
1511                         * since it is not the tail page. So we deduct the
1512                         * bytes consumed in ring buffer from here.
1513                         * Increment overrun to account for the lost events.
1514                         */
1515                        local_add(page_entries, &cpu_buffer->overrun);
1516                        local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1517                }
1518
1519                /*
1520                 * We have already removed references to this list item, just
1521                 * free up the buffer_page and its page
1522                 */
1523                free_buffer_page(to_remove_page);
1524                nr_removed--;
1525
1526        } while (to_remove_page != last_page);
1527
1528        RB_WARN_ON(cpu_buffer, nr_removed);
1529
1530        return nr_removed == 0;
1531}
1532
1533static int
1534rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
1535{
1536        struct list_head *pages = &cpu_buffer->new_pages;
1537        int retries, success;
1538
1539        raw_spin_lock_irq(&cpu_buffer->reader_lock);
1540        /*
1541         * We are holding the reader lock, so the reader page won't be swapped
1542         * in the ring buffer. Now we are racing with the writer trying to
1543         * move head page and the tail page.
1544         * We are going to adapt the reader page update process where:
1545         * 1. We first splice the start and end of list of new pages between
1546         *    the head page and its previous page.
1547         * 2. We cmpxchg the prev_page->next to point from head page to the
1548         *    start of new pages list.
1549         * 3. Finally, we update the head->prev to the end of new list.
1550         *
1551         * We will try this process 10 times, to make sure that we don't keep
1552         * spinning.
1553         */
1554        retries = 10;
1555        success = 0;
1556        while (retries--) {
1557                struct list_head *head_page, *prev_page, *r;
1558                struct list_head *last_page, *first_page;
1559                struct list_head *head_page_with_bit;
1560
1561                head_page = &rb_set_head_page(cpu_buffer)->list;
1562                if (!head_page)
1563                        break;
1564                prev_page = head_page->prev;
1565
1566                first_page = pages->next;
1567                last_page  = pages->prev;
1568
1569                head_page_with_bit = (struct list_head *)
1570                                     ((unsigned long)head_page | RB_PAGE_HEAD);
1571
1572                last_page->next = head_page_with_bit;
1573                first_page->prev = prev_page;
1574
1575                r = cmpxchg(&prev_page->next, head_page_with_bit, first_page);
1576
1577                if (r == head_page_with_bit) {
1578                        /*
1579                         * yay, we replaced the page pointer to our new list,
1580                         * now, we just have to update to head page's prev
1581                         * pointer to point to end of list
1582                         */
1583                        head_page->prev = last_page;
1584                        success = 1;
1585                        break;
1586                }
1587        }
1588
1589        if (success)
1590                INIT_LIST_HEAD(pages);
1591        /*
1592         * If we weren't successful in adding in new pages, warn and stop
1593         * tracing
1594         */
1595        RB_WARN_ON(cpu_buffer, !success);
1596        raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1597
1598        /* free pages if they weren't inserted */
1599        if (!success) {
1600                struct buffer_page *bpage, *tmp;
1601                list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1602                                         list) {
1603                        list_del_init(&bpage->list);
1604                        free_buffer_page(bpage);
1605                }
1606        }
1607        return success;
1608}
1609
1610static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
1611{
1612        int success;
1613
1614        if (cpu_buffer->nr_pages_to_update > 0)
1615                success = rb_insert_pages(cpu_buffer);
1616        else
1617                success = rb_remove_pages(cpu_buffer,
1618                                        -cpu_buffer->nr_pages_to_update);
1619
1620        if (success)
1621                cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
1622}
1623
1624static void update_pages_handler(struct work_struct *work)
1625{
1626        struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
1627                        struct ring_buffer_per_cpu, update_pages_work);
1628        rb_update_pages(cpu_buffer);
1629        complete(&cpu_buffer->update_done);
1630}
1631
1632/**
1633 * ring_buffer_resize - resize the ring buffer
1634 * @buffer: the buffer to resize.
1635 * @size: the new size.
1636 * @cpu_id: the cpu buffer to resize
1637 *
1638 * Minimum size is 2 * BUF_PAGE_SIZE.
1639 *
1640 * Returns 0 on success and < 0 on failure.
1641 */
1642int ring_buffer_resize(struct ring_buffer *buffer, unsigned long size,
1643                        int cpu_id)
1644{
1645        struct ring_buffer_per_cpu *cpu_buffer;
1646        unsigned long nr_pages;
1647        int cpu, err = 0;
1648
1649        /*
1650         * Always succeed at resizing a non-existent buffer:
1651         */
1652        if (!buffer)
1653                return size;
1654
1655        /* Make sure the requested buffer exists */
1656        if (cpu_id != RING_BUFFER_ALL_CPUS &&
1657            !cpumask_test_cpu(cpu_id, buffer->cpumask))
1658                return size;
1659
1660        nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1661
1662        /* we need a minimum of two pages */
1663        if (nr_pages < 2)
1664                nr_pages = 2;
1665
1666        size = nr_pages * BUF_PAGE_SIZE;
1667
1668        /*
1669         * Don't succeed if resizing is disabled, as a reader might be
1670         * manipulating the ring buffer and is expecting a sane state while
1671         * this is true.
1672         */
1673        if (atomic_read(&buffer->resize_disabled))
1674                return -EBUSY;
1675
1676        /* prevent another thread from changing buffer sizes */
1677        mutex_lock(&buffer->mutex);
1678
1679        if (cpu_id == RING_BUFFER_ALL_CPUS) {
1680                /* calculate the pages to update */
1681                for_each_buffer_cpu(buffer, cpu) {
1682                        cpu_buffer = buffer->buffers[cpu];
1683
1684                        cpu_buffer->nr_pages_to_update = nr_pages -
1685                                                        cpu_buffer->nr_pages;
1686                        /*
1687                         * nothing more to do for removing pages or no update
1688                         */
1689                        if (cpu_buffer->nr_pages_to_update <= 0)
1690                                continue;
1691                        /*
1692                         * to add pages, make sure all new pages can be
1693                         * allocated without receiving ENOMEM
1694                         */
1695                        INIT_LIST_HEAD(&cpu_buffer->new_pages);
1696                        if (__rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1697                                                &cpu_buffer->new_pages, cpu)) {
1698                                /* not enough memory for new pages */
1699                                err = -ENOMEM;
1700                                goto out_err;
1701                        }
1702                }
1703
1704                get_online_cpus();
1705                /*
1706                 * Fire off all the required work handlers
1707                 * We can't schedule on offline CPUs, but it's not necessary
1708                 * since we can change their buffer sizes without any race.
1709                 */
1710                for_each_buffer_cpu(buffer, cpu) {
1711                        cpu_buffer = buffer->buffers[cpu];
1712                        if (!cpu_buffer->nr_pages_to_update)
1713                                continue;
1714
1715                        /* Can't run something on an offline CPU. */
1716                        if (!cpu_online(cpu)) {
1717                                rb_update_pages(cpu_buffer);
1718                                cpu_buffer->nr_pages_to_update = 0;
1719                        } else {
1720                                schedule_work_on(cpu,
1721                                                &cpu_buffer->update_pages_work);
1722                        }
1723                }
1724
1725                /* wait for all the updates to complete */
1726                for_each_buffer_cpu(buffer, cpu) {
1727                        cpu_buffer = buffer->buffers[cpu];
1728                        if (!cpu_buffer->nr_pages_to_update)
1729                                continue;
1730
1731                        if (cpu_online(cpu))
1732                                wait_for_completion(&cpu_buffer->update_done);
1733                        cpu_buffer->nr_pages_to_update = 0;
1734                }
1735
1736                put_online_cpus();
1737        } else {
1738                /* Make sure this CPU has been intitialized */
1739                if (!cpumask_test_cpu(cpu_id, buffer->cpumask))
1740                        goto out;
1741
1742                cpu_buffer = buffer->buffers[cpu_id];
1743
1744                if (nr_pages == cpu_buffer->nr_pages)
1745                        goto out;
1746
1747                cpu_buffer->nr_pages_to_update = nr_pages -
1748                                                cpu_buffer->nr_pages;
1749
1750                INIT_LIST_HEAD(&cpu_buffer->new_pages);
1751                if (cpu_buffer->nr_pages_to_update > 0 &&
1752                        __rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1753                                            &cpu_buffer->new_pages, cpu_id)) {
1754                        err = -ENOMEM;
1755                        goto out_err;
1756                }
1757
1758                get_online_cpus();
1759
1760                /* Can't run something on an offline CPU. */
1761                if (!cpu_online(cpu_id))
1762                        rb_update_pages(cpu_buffer);
1763                else {
1764                        schedule_work_on(cpu_id,
1765                                         &cpu_buffer->update_pages_work);
1766                        wait_for_completion(&cpu_buffer->update_done);
1767                }
1768
1769                cpu_buffer->nr_pages_to_update = 0;
1770                put_online_cpus();
1771        }
1772
1773 out:
1774        /*
1775         * The ring buffer resize can happen with the ring buffer
1776         * enabled, so that the update disturbs the tracing as little
1777         * as possible. But if the buffer is disabled, we do not need
1778         * to worry about that, and we can take the time to verify
1779         * that the buffer is not corrupt.
1780         */
1781        if (atomic_read(&buffer->record_disabled)) {
1782                atomic_inc(&buffer->record_disabled);
1783                /*
1784                 * Even though the buffer was disabled, we must make sure
1785                 * that it is truly disabled before calling rb_check_pages.
1786                 * There could have been a race between checking
1787                 * record_disable and incrementing it.
1788                 */
1789                synchronize_sched();
1790                for_each_buffer_cpu(buffer, cpu) {
1791                        cpu_buffer = buffer->buffers[cpu];
1792                        rb_check_pages(cpu_buffer);
1793                }
1794                atomic_dec(&buffer->record_disabled);
1795        }
1796
1797        mutex_unlock(&buffer->mutex);
1798        return size;
1799
1800 out_err:
1801        for_each_buffer_cpu(buffer, cpu) {
1802                struct buffer_page *bpage, *tmp;
1803
1804                cpu_buffer = buffer->buffers[cpu];
1805                cpu_buffer->nr_pages_to_update = 0;
1806
1807                if (list_empty(&cpu_buffer->new_pages))
1808                        continue;
1809
1810                list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1811                                        list) {
1812                        list_del_init(&bpage->list);
1813                        free_buffer_page(bpage);
1814                }
1815        }
1816        mutex_unlock(&buffer->mutex);
1817        return err;
1818}
1819EXPORT_SYMBOL_GPL(ring_buffer_resize);
1820
1821void ring_buffer_change_overwrite(struct ring_buffer *buffer, int val)
1822{
1823        mutex_lock(&buffer->mutex);
1824        if (val)
1825                buffer->flags |= RB_FL_OVERWRITE;
1826        else
1827                buffer->flags &= ~RB_FL_OVERWRITE;
1828        mutex_unlock(&buffer->mutex);
1829}
1830EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
1831
1832static inline void *
1833__rb_data_page_index(struct buffer_data_page *bpage, unsigned index)
1834{
1835        return bpage->data + index;
1836}
1837
1838static inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
1839{
1840        return bpage->page->data + index;
1841}
1842
1843static inline struct ring_buffer_event *
1844rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
1845{
1846        return __rb_page_index(cpu_buffer->reader_page,
1847                               cpu_buffer->reader_page->read);
1848}
1849
1850static inline struct ring_buffer_event *
1851rb_iter_head_event(struct ring_buffer_iter *iter)
1852{
1853        return __rb_page_index(iter->head_page, iter->head);
1854}
1855
1856static inline unsigned rb_page_commit(struct buffer_page *bpage)
1857{
1858        return local_read(&bpage->page->commit);
1859}
1860
1861/* Size is determined by what has been committed */
1862static inline unsigned rb_page_size(struct buffer_page *bpage)
1863{
1864        return rb_page_commit(bpage);
1865}
1866
1867static inline unsigned
1868rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
1869{
1870        return rb_page_commit(cpu_buffer->commit_page);
1871}
1872
1873static inline unsigned
1874rb_event_index(struct ring_buffer_event *event)
1875{
1876        unsigned long addr = (unsigned long)event;
1877
1878        return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
1879}
1880
1881static void rb_inc_iter(struct ring_buffer_iter *iter)
1882{
1883        struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
1884
1885        /*
1886         * The iterator could be on the reader page (it starts there).
1887         * But the head could have moved, since the reader was
1888         * found. Check for this case and assign the iterator
1889         * to the head page instead of next.
1890         */
1891        if (iter->head_page == cpu_buffer->reader_page)
1892                iter->head_page = rb_set_head_page(cpu_buffer);
1893        else
1894                rb_inc_page(cpu_buffer, &iter->head_page);
1895
1896        iter->read_stamp = iter->head_page->page->time_stamp;
1897        iter->head = 0;
1898}
1899
1900/*
1901 * rb_handle_head_page - writer hit the head page
1902 *
1903 * Returns: +1 to retry page
1904 *           0 to continue
1905 *          -1 on error
1906 */
1907static int
1908rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
1909                    struct buffer_page *tail_page,
1910                    struct buffer_page *next_page)
1911{
1912        struct buffer_page *new_head;
1913        int entries;
1914        int type;
1915        int ret;
1916
1917        entries = rb_page_entries(next_page);
1918
1919        /*
1920         * The hard part is here. We need to move the head
1921         * forward, and protect against both readers on
1922         * other CPUs and writers coming in via interrupts.
1923         */
1924        type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
1925                                       RB_PAGE_HEAD);
1926
1927        /*
1928         * type can be one of four:
1929         *  NORMAL - an interrupt already moved it for us
1930         *  HEAD   - we are the first to get here.
1931         *  UPDATE - we are the interrupt interrupting
1932         *           a current move.
1933         *  MOVED  - a reader on another CPU moved the next
1934         *           pointer to its reader page. Give up
1935         *           and try again.
1936         */
1937
1938        switch (type) {
1939        case RB_PAGE_HEAD:
1940                /*
1941                 * We changed the head to UPDATE, thus
1942                 * it is our responsibility to update
1943                 * the counters.
1944                 */
1945                local_add(entries, &cpu_buffer->overrun);
1946                local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1947
1948                /*
1949                 * The entries will be zeroed out when we move the
1950                 * tail page.
1951                 */
1952
1953                /* still more to do */
1954                break;
1955
1956        case RB_PAGE_UPDATE:
1957                /*
1958                 * This is an interrupt that interrupt the
1959                 * previous update. Still more to do.
1960                 */
1961                break;
1962        case RB_PAGE_NORMAL:
1963                /*
1964                 * An interrupt came in before the update
1965                 * and processed this for us.
1966                 * Nothing left to do.
1967                 */
1968                return 1;
1969        case RB_PAGE_MOVED:
1970                /*
1971                 * The reader is on another CPU and just did
1972                 * a swap with our next_page.
1973                 * Try again.
1974                 */
1975                return 1;
1976        default:
1977                RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
1978                return -1;
1979        }
1980
1981        /*
1982         * Now that we are here, the old head pointer is
1983         * set to UPDATE. This will keep the reader from
1984         * swapping the head page with the reader page.
1985         * The reader (on another CPU) will spin till
1986         * we are finished.
1987         *
1988         * We just need to protect against interrupts
1989         * doing the job. We will set the next pointer
1990         * to HEAD. After that, we set the old pointer
1991         * to NORMAL, but only if it was HEAD before.
1992         * otherwise we are an interrupt, and only
1993         * want the outer most commit to reset it.
1994         */
1995        new_head = next_page;
1996        rb_inc_page(cpu_buffer, &new_head);
1997
1998        ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
1999                                    RB_PAGE_NORMAL);
2000
2001        /*
2002         * Valid returns are:
2003         *  HEAD   - an interrupt came in and already set it.
2004         *  NORMAL - One of two things:
2005         *            1) We really set it.
2006         *            2) A bunch of interrupts came in and moved
2007         *               the page forward again.
2008         */
2009        switch (ret) {
2010        case RB_PAGE_HEAD:
2011        case RB_PAGE_NORMAL:
2012                /* OK */
2013                break;
2014        default:
2015                RB_WARN_ON(cpu_buffer, 1);
2016                return -1;
2017        }
2018
2019        /*
2020         * It is possible that an interrupt came in,
2021         * set the head up, then more interrupts came in
2022         * and moved it again. When we get back here,
2023         * the page would have been set to NORMAL but we
2024         * just set it back to HEAD.
2025         *
2026         * How do you detect this? Well, if that happened
2027         * the tail page would have moved.
2028         */
2029        if (ret == RB_PAGE_NORMAL) {
2030                struct buffer_page *buffer_tail_page;
2031
2032                buffer_tail_page = READ_ONCE(cpu_buffer->tail_page);
2033                /*
2034                 * If the tail had moved passed next, then we need
2035                 * to reset the pointer.
2036                 */
2037                if (buffer_tail_page != tail_page &&
2038                    buffer_tail_page != next_page)
2039                        rb_head_page_set_normal(cpu_buffer, new_head,
2040                                                next_page,
2041                                                RB_PAGE_HEAD);
2042        }
2043
2044        /*
2045         * If this was the outer most commit (the one that
2046         * changed the original pointer from HEAD to UPDATE),
2047         * then it is up to us to reset it to NORMAL.
2048         */
2049        if (type == RB_PAGE_HEAD) {
2050                ret = rb_head_page_set_normal(cpu_buffer, next_page,
2051                                              tail_page,
2052                                              RB_PAGE_UPDATE);
2053                if (RB_WARN_ON(cpu_buffer,
2054                               ret != RB_PAGE_UPDATE))
2055                        return -1;
2056        }
2057
2058        return 0;
2059}
2060
2061static inline void
2062rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2063              unsigned long tail, struct rb_event_info *info)
2064{
2065        struct buffer_page *tail_page = info->tail_page;
2066        struct ring_buffer_event *event;
2067        unsigned long length = info->length;
2068
2069        /*
2070         * Only the event that crossed the page boundary
2071         * must fill the old tail_page with padding.
2072         */
2073        if (tail >= BUF_PAGE_SIZE) {
2074                /*
2075                 * If the page was filled, then we still need
2076                 * to update the real_end. Reset it to zero
2077                 * and the reader will ignore it.
2078                 */
2079                if (tail == BUF_PAGE_SIZE)
2080                        tail_page->real_end = 0;
2081
2082                local_sub(length, &tail_page->write);
2083                return;
2084        }
2085
2086        event = __rb_page_index(tail_page, tail);
2087        kmemcheck_annotate_bitfield(event, bitfield);
2088
2089        /* account for padding bytes */
2090        local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes);
2091
2092        /*
2093         * Save the original length to the meta data.
2094         * This will be used by the reader to add lost event
2095         * counter.
2096         */
2097        tail_page->real_end = tail;
2098
2099        /*
2100         * If this event is bigger than the minimum size, then
2101         * we need to be careful that we don't subtract the
2102         * write counter enough to allow another writer to slip
2103         * in on this page.
2104         * We put in a discarded commit instead, to make sure
2105         * that this space is not used again.
2106         *
2107         * If we are less than the minimum size, we don't need to
2108         * worry about it.
2109         */
2110        if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
2111                /* No room for any events */
2112
2113                /* Mark the rest of the page with padding */
2114                rb_event_set_padding(event);
2115
2116                /* Set the write back to the previous setting */
2117                local_sub(length, &tail_page->write);
2118                return;
2119        }
2120
2121        /* Put in a discarded event */
2122        event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
2123        event->type_len = RINGBUF_TYPE_PADDING;
2124        /* time delta must be non zero */
2125        event->time_delta = 1;
2126
2127        /* Set write to end of buffer */
2128        length = (tail + length) - BUF_PAGE_SIZE;
2129        local_sub(length, &tail_page->write);
2130}
2131
2132static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer);
2133
2134/*
2135 * This is the slow path, force gcc not to inline it.
2136 */
2137static noinline struct ring_buffer_event *
2138rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2139             unsigned long tail, struct rb_event_info *info)
2140{
2141        struct buffer_page *tail_page = info->tail_page;
2142        struct buffer_page *commit_page = cpu_buffer->commit_page;
2143        struct ring_buffer *buffer = cpu_buffer->buffer;
2144        struct buffer_page *next_page;
2145        int ret;
2146
2147        next_page = tail_page;
2148
2149        rb_inc_page(cpu_buffer, &next_page);
2150
2151        /*
2152         * If for some reason, we had an interrupt storm that made
2153         * it all the way around the buffer, bail, and warn
2154         * about it.
2155         */
2156        if (unlikely(next_page == commit_page)) {
2157                local_inc(&cpu_buffer->commit_overrun);
2158                goto out_reset;
2159        }
2160
2161        /*
2162         * This is where the fun begins!
2163         *
2164         * We are fighting against races between a reader that
2165         * could be on another CPU trying to swap its reader
2166         * page with the buffer head.
2167         *
2168         * We are also fighting against interrupts coming in and
2169         * moving the head or tail on us as well.
2170         *
2171         * If the next page is the head page then we have filled
2172         * the buffer, unless the commit page is still on the
2173         * reader page.
2174         */
2175        if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) {
2176
2177                /*
2178                 * If the commit is not on the reader page, then
2179                 * move the header page.
2180                 */
2181                if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2182                        /*
2183                         * If we are not in overwrite mode,
2184                         * this is easy, just stop here.
2185                         */
2186                        if (!(buffer->flags & RB_FL_OVERWRITE)) {
2187                                local_inc(&cpu_buffer->dropped_events);
2188                                goto out_reset;
2189                        }
2190
2191                        ret = rb_handle_head_page(cpu_buffer,
2192                                                  tail_page,
2193                                                  next_page);
2194                        if (ret < 0)
2195                                goto out_reset;
2196                        if (ret)
2197                                goto out_again;
2198                } else {
2199                        /*
2200                         * We need to be careful here too. The
2201                         * commit page could still be on the reader
2202                         * page. We could have a small buffer, and
2203                         * have filled up the buffer with events
2204                         * from interrupts and such, and wrapped.
2205                         *
2206                         * Note, if the tail page is also the on the
2207                         * reader_page, we let it move out.
2208                         */
2209                        if (unlikely((cpu_buffer->commit_page !=
2210                                      cpu_buffer->tail_page) &&
2211                                     (cpu_buffer->commit_page ==
2212                                      cpu_buffer->reader_page))) {
2213                                local_inc(&cpu_buffer->commit_overrun);
2214                                goto out_reset;
2215                        }
2216                }
2217        }
2218
2219        rb_tail_page_update(cpu_buffer, tail_page, next_page);
2220
2221 out_again:
2222
2223        rb_reset_tail(cpu_buffer, tail, info);
2224
2225        /* Commit what we have for now. */
2226        rb_end_commit(cpu_buffer);
2227        /* rb_end_commit() decs committing */
2228        local_inc(&cpu_buffer->committing);
2229
2230        /* fail and let the caller try again */
2231        return ERR_PTR(-EAGAIN);
2232
2233 out_reset:
2234        /* reset write */
2235        rb_reset_tail(cpu_buffer, tail, info);
2236
2237        return NULL;
2238}
2239
2240/* Slow path, do not inline */
2241static noinline struct ring_buffer_event *
2242rb_add_time_stamp(struct ring_buffer_event *event, u64 delta)
2243{
2244        event->type_len = RINGBUF_TYPE_TIME_EXTEND;
2245
2246        /* Not the first event on the page? */
2247        if (rb_event_index(event)) {
2248                event->time_delta = delta & TS_MASK;
2249                event->array[0] = delta >> TS_SHIFT;
2250        } else {
2251                /* nope, just zero it */
2252                event->time_delta = 0;
2253                event->array[0] = 0;
2254        }
2255
2256        return skip_time_extend(event);
2257}
2258
2259static inline bool rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2260                                     struct ring_buffer_event *event);
2261
2262/**
2263 * rb_update_event - update event type and data
2264 * @event: the event to update
2265 * @type: the type of event
2266 * @length: the size of the event field in the ring buffer
2267 *
2268 * Update the type and data fields of the event. The length
2269 * is the actual size that is written to the ring buffer,
2270 * and with this, we can determine what to place into the
2271 * data field.
2272 */
2273static void
2274rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
2275                struct ring_buffer_event *event,
2276                struct rb_event_info *info)
2277{
2278        unsigned length = info->length;
2279        u64 delta = info->delta;
2280
2281        /* Only a commit updates the timestamp */
2282        if (unlikely(!rb_event_is_commit(cpu_buffer, event)))
2283                delta = 0;
2284
2285        /*
2286         * If we need to add a timestamp, then we
2287         * add it to the start of the resevered space.
2288         */
2289        if (unlikely(info->add_timestamp)) {
2290                event = rb_add_time_stamp(event, delta);
2291                length -= RB_LEN_TIME_EXTEND;
2292                delta = 0;
2293        }
2294
2295        event->time_delta = delta;
2296        length -= RB_EVNT_HDR_SIZE;
2297        if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
2298                event->type_len = 0;
2299                event->array[0] = length;
2300        } else
2301                event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
2302}
2303
2304static unsigned rb_calculate_event_length(unsigned length)
2305{
2306        struct ring_buffer_event event; /* Used only for sizeof array */
2307
2308        /* zero length can cause confusions */
2309        if (!length)
2310                length++;
2311
2312        if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
2313                length += sizeof(event.array[0]);
2314
2315        length += RB_EVNT_HDR_SIZE;
2316        length = ALIGN(length, RB_ARCH_ALIGNMENT);
2317
2318        /*
2319         * In case the time delta is larger than the 27 bits for it
2320         * in the header, we need to add a timestamp. If another
2321         * event comes in when trying to discard this one to increase
2322         * the length, then the timestamp will be added in the allocated
2323         * space of this event. If length is bigger than the size needed
2324         * for the TIME_EXTEND, then padding has to be used. The events
2325         * length must be either RB_LEN_TIME_EXTEND, or greater than or equal
2326         * to RB_LEN_TIME_EXTEND + 8, as 8 is the minimum size for padding.
2327         * As length is a multiple of 4, we only need to worry if it
2328         * is 12 (RB_LEN_TIME_EXTEND + 4).
2329         */
2330        if (length == RB_LEN_TIME_EXTEND + RB_ALIGNMENT)
2331                length += RB_ALIGNMENT;
2332
2333        return length;
2334}
2335
2336#ifndef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2337static inline bool sched_clock_stable(void)
2338{
2339        return true;
2340}
2341#endif
2342
2343static inline int
2344rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2345                  struct ring_buffer_event *event)
2346{
2347        unsigned long new_index, old_index;
2348        struct buffer_page *bpage;
2349        unsigned long index;
2350        unsigned long addr;
2351
2352        new_index = rb_event_index(event);
2353        old_index = new_index + rb_event_ts_length(event);
2354        addr = (unsigned long)event;
2355        addr &= PAGE_MASK;
2356
2357        bpage = READ_ONCE(cpu_buffer->tail_page);
2358
2359        if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2360                unsigned long write_mask =
2361                        local_read(&bpage->write) & ~RB_WRITE_MASK;
2362                unsigned long event_length = rb_event_length(event);
2363                /*
2364                 * This is on the tail page. It is possible that
2365                 * a write could come in and move the tail page
2366                 * and write to the next page. That is fine
2367                 * because we just shorten what is on this page.
2368                 */
2369                old_index += write_mask;
2370                new_index += write_mask;
2371                index = local_cmpxchg(&bpage->write, old_index, new_index);
2372                if (index == old_index) {
2373                        /* update counters */
2374                        local_sub(event_length, &cpu_buffer->entries_bytes);
2375                        return 1;
2376                }
2377        }
2378
2379        /* could not discard */
2380        return 0;
2381}
2382
2383static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2384{
2385        local_inc(&cpu_buffer->committing);
2386        local_inc(&cpu_buffer->commits);
2387}
2388
2389static void
2390rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
2391{
2392        unsigned long max_count;
2393
2394        /*
2395         * We only race with interrupts and NMIs on this CPU.
2396         * If we own the commit event, then we can commit
2397         * all others that interrupted us, since the interruptions
2398         * are in stack format (they finish before they come
2399         * back to us). This allows us to do a simple loop to
2400         * assign the commit to the tail.
2401         */
2402 again:
2403        max_count = cpu_buffer->nr_pages * 100;
2404
2405        while (cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)) {
2406                if (RB_WARN_ON(cpu_buffer, !(--max_count)))
2407                        return;
2408                if (RB_WARN_ON(cpu_buffer,
2409                               rb_is_reader_page(cpu_buffer->tail_page)))
2410                        return;
2411                local_set(&cpu_buffer->commit_page->page->commit,
2412                          rb_page_write(cpu_buffer->commit_page));
2413                rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
2414                /* Only update the write stamp if the page has an event */
2415                if (rb_page_write(cpu_buffer->commit_page))
2416                        cpu_buffer->write_stamp =
2417                                cpu_buffer->commit_page->page->time_stamp;
2418                /* add barrier to keep gcc from optimizing too much */
2419                barrier();
2420        }
2421        while (rb_commit_index(cpu_buffer) !=
2422               rb_page_write(cpu_buffer->commit_page)) {
2423
2424                local_set(&cpu_buffer->commit_page->page->commit,
2425                          rb_page_write(cpu_buffer->commit_page));
2426                RB_WARN_ON(cpu_buffer,
2427                           local_read(&cpu_buffer->commit_page->page->commit) &
2428                           ~RB_WRITE_MASK);
2429                barrier();
2430        }
2431
2432        /* again, keep gcc from optimizing */
2433        barrier();
2434
2435        /*
2436         * If an interrupt came in just after the first while loop
2437         * and pushed the tail page forward, we will be left with
2438         * a dangling commit that will never go forward.
2439         */
2440        if (unlikely(cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)))
2441                goto again;
2442}
2443
2444static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
2445{
2446        unsigned long commits;
2447
2448        if (RB_WARN_ON(cpu_buffer,
2449                       !local_read(&cpu_buffer->committing)))
2450                return;
2451
2452 again:
2453        commits = local_read(&cpu_buffer->commits);
2454        /* synchronize with interrupts */
2455        barrier();
2456        if (local_read(&cpu_buffer->committing) == 1)
2457                rb_set_commit_to_write(cpu_buffer);
2458
2459        local_dec(&cpu_buffer->committing);
2460
2461        /* synchronize with interrupts */
2462        barrier();
2463
2464        /*
2465         * Need to account for interrupts coming in between the
2466         * updating of the commit page and the clearing of the
2467         * committing counter.
2468         */
2469        if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
2470            !local_read(&cpu_buffer->committing)) {
2471                local_inc(&cpu_buffer->committing);
2472                goto again;
2473        }
2474}
2475
2476static inline void rb_event_discard(struct ring_buffer_event *event)
2477{
2478        if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
2479                event = skip_time_extend(event);
2480
2481        /* array[0] holds the actual length for the discarded event */
2482        event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
2483        event->type_len = RINGBUF_TYPE_PADDING;
2484        /* time delta must be non zero */
2485        if (!event->time_delta)
2486                event->time_delta = 1;
2487}
2488
2489static inline bool
2490rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2491                   struct ring_buffer_event *event)
2492{
2493        unsigned long addr = (unsigned long)event;
2494        unsigned long index;
2495
2496        index = rb_event_index(event);
2497        addr &= PAGE_MASK;
2498
2499        return cpu_buffer->commit_page->page == (void *)addr &&
2500                rb_commit_index(cpu_buffer) == index;
2501}
2502
2503static void
2504rb_update_write_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2505                      struct ring_buffer_event *event)
2506{
2507        u64 delta;
2508
2509        /*
2510         * The event first in the commit queue updates the
2511         * time stamp.
2512         */
2513        if (rb_event_is_commit(cpu_buffer, event)) {
2514                /*
2515                 * A commit event that is first on a page
2516                 * updates the write timestamp with the page stamp
2517                 */
2518                if (!rb_event_index(event))
2519                        cpu_buffer->write_stamp =
2520                                cpu_buffer->commit_page->page->time_stamp;
2521                else if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
2522                        delta = event->array[0];
2523                        delta <<= TS_SHIFT;
2524                        delta += event->time_delta;
2525                        cpu_buffer->write_stamp += delta;
2526                } else
2527                        cpu_buffer->write_stamp += event->time_delta;
2528        }
2529}
2530
2531static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
2532                      struct ring_buffer_event *event)
2533{
2534        local_inc(&cpu_buffer->entries);
2535        rb_update_write_stamp(cpu_buffer, event);
2536        rb_end_commit(cpu_buffer);
2537}
2538
2539static __always_inline void
2540rb_wakeups(struct ring_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer)
2541{
2542        bool pagebusy;
2543
2544        if (buffer->irq_work.waiters_pending) {
2545                buffer->irq_work.waiters_pending = false;
2546                /* irq_work_queue() supplies it's own memory barriers */
2547                irq_work_queue(&buffer->irq_work.work);
2548        }
2549
2550        if (cpu_buffer->irq_work.waiters_pending) {
2551                cpu_buffer->irq_work.waiters_pending = false;
2552                /* irq_work_queue() supplies it's own memory barriers */
2553                irq_work_queue(&cpu_buffer->irq_work.work);
2554        }
2555
2556        pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
2557
2558        if (!pagebusy && cpu_buffer->irq_work.full_waiters_pending) {
2559                cpu_buffer->irq_work.wakeup_full = true;
2560                cpu_buffer->irq_work.full_waiters_pending = false;
2561                /* irq_work_queue() supplies it's own memory barriers */
2562                irq_work_queue(&cpu_buffer->irq_work.work);
2563        }
2564}
2565
2566/*
2567 * The lock and unlock are done within a preempt disable section.
2568 * The current_context per_cpu variable can only be modified
2569 * by the current task between lock and unlock. But it can
2570 * be modified more than once via an interrupt. To pass this
2571 * information from the lock to the unlock without having to
2572 * access the 'in_interrupt()' functions again (which do show
2573 * a bit of overhead in something as critical as function tracing,
2574 * we use a bitmask trick.
2575 *
2576 *  bit 0 =  NMI context
2577 *  bit 1 =  IRQ context
2578 *  bit 2 =  SoftIRQ context
2579 *  bit 3 =  normal context.
2580 *
2581 * This works because this is the order of contexts that can
2582 * preempt other contexts. A SoftIRQ never preempts an IRQ
2583 * context.
2584 *
2585 * When the context is determined, the corresponding bit is
2586 * checked and set (if it was set, then a recursion of that context
2587 * happened).
2588 *
2589 * On unlock, we need to clear this bit. To do so, just subtract
2590 * 1 from the current_context and AND it to itself.
2591 *
2592 * (binary)
2593 *  101 - 1 = 100
2594 *  101 & 100 = 100 (clearing bit zero)
2595 *
2596 *  1010 - 1 = 1001
2597 *  1010 & 1001 = 1000 (clearing bit 1)
2598 *
2599 * The least significant bit can be cleared this way, and it
2600 * just so happens that it is the same bit corresponding to
2601 * the current context.
2602 */
2603
2604static __always_inline int
2605trace_recursive_lock(struct ring_buffer_per_cpu *cpu_buffer)
2606{
2607        unsigned int val = cpu_buffer->current_context;
2608        int bit;
2609
2610        if (in_interrupt()) {
2611                if (in_nmi())
2612                        bit = RB_CTX_NMI;
2613                else if (in_irq())
2614                        bit = RB_CTX_IRQ;
2615                else
2616                        bit = RB_CTX_SOFTIRQ;
2617        } else
2618                bit = RB_CTX_NORMAL;
2619
2620        if (unlikely(val & (1 << bit)))
2621                return 1;
2622
2623        val |= (1 << bit);
2624        cpu_buffer->current_context = val;
2625
2626        return 0;
2627}
2628
2629static __always_inline void
2630trace_recursive_unlock(struct ring_buffer_per_cpu *cpu_buffer)
2631{
2632        cpu_buffer->current_context &= cpu_buffer->current_context - 1;
2633}
2634
2635/**
2636 * ring_buffer_unlock_commit - commit a reserved
2637 * @buffer: The buffer to commit to
2638 * @event: The event pointer to commit.
2639 *
2640 * This commits the data to the ring buffer, and releases any locks held.
2641 *
2642 * Must be paired with ring_buffer_lock_reserve.
2643 */
2644int ring_buffer_unlock_commit(struct ring_buffer *buffer,
2645                              struct ring_buffer_event *event)
2646{
2647        struct ring_buffer_per_cpu *cpu_buffer;
2648        int cpu = raw_smp_processor_id();
2649
2650        cpu_buffer = buffer->buffers[cpu];
2651
2652        rb_commit(cpu_buffer, event);
2653
2654        rb_wakeups(buffer, cpu_buffer);
2655
2656        trace_recursive_unlock(cpu_buffer);
2657
2658        preempt_enable_notrace();
2659
2660        return 0;
2661}
2662EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
2663
2664static noinline void
2665rb_handle_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2666                    struct rb_event_info *info)
2667{
2668        WARN_ONCE(info->delta > (1ULL << 59),
2669                  KERN_WARNING "Delta way too big! %llu ts=%llu write stamp = %llu\n%s",
2670                  (unsigned long long)info->delta,
2671                  (unsigned long long)info->ts,
2672                  (unsigned long long)cpu_buffer->write_stamp,
2673                  sched_clock_stable() ? "" :
2674                  "If you just came from a suspend/resume,\n"
2675                  "please switch to the trace global clock:\n"
2676                  "  echo global > /sys/kernel/debug/tracing/trace_clock\n");
2677        info->add_timestamp = 1;
2678}
2679
2680static struct ring_buffer_event *
2681__rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
2682                  struct rb_event_info *info)
2683{
2684        struct ring_buffer_event *event;
2685        struct buffer_page *tail_page;
2686        unsigned long tail, write;
2687
2688        /*
2689         * If the time delta since the last event is too big to
2690         * hold in the time field of the event, then we append a
2691         * TIME EXTEND event ahead of the data event.
2692         */
2693        if (unlikely(info->add_timestamp))
2694                info->length += RB_LEN_TIME_EXTEND;
2695
2696        /* Don't let the compiler play games with cpu_buffer->tail_page */
2697        tail_page = info->tail_page = READ_ONCE(cpu_buffer->tail_page);
2698        write = local_add_return(info->length, &tail_page->write);
2699
2700        /* set write to only the index of the write */
2701        write &= RB_WRITE_MASK;
2702        tail = write - info->length;
2703
2704        /*
2705         * If this is the first commit on the page, then it has the same
2706         * timestamp as the page itself.
2707         */
2708        if (!tail)
2709                info->delta = 0;
2710
2711        /* See if we shot pass the end of this buffer page */
2712        if (unlikely(write > BUF_PAGE_SIZE))
2713                return rb_move_tail(cpu_buffer, tail, info);
2714
2715        /* We reserved something on the buffer */
2716
2717        event = __rb_page_index(tail_page, tail);
2718        kmemcheck_annotate_bitfield(event, bitfield);
2719        rb_update_event(cpu_buffer, event, info);
2720
2721        local_inc(&tail_page->entries);
2722
2723        /*
2724         * If this is the first commit on the page, then update
2725         * its timestamp.
2726         */
2727        if (!tail)
2728                tail_page->page->time_stamp = info->ts;
2729
2730        /* account for these added bytes */
2731        local_add(info->length, &cpu_buffer->entries_bytes);
2732
2733        return event;
2734}
2735
2736static struct ring_buffer_event *
2737rb_reserve_next_event(struct ring_buffer *buffer,
2738                      struct ring_buffer_per_cpu *cpu_buffer,
2739                      unsigned long length)
2740{
2741        struct ring_buffer_event *event;
2742        struct rb_event_info info;
2743        int nr_loops = 0;
2744        u64 diff;
2745
2746        rb_start_commit(cpu_buffer);
2747
2748#ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2749        /*
2750         * Due to the ability to swap a cpu buffer from a buffer
2751         * it is possible it was swapped before we committed.
2752         * (committing stops a swap). We check for it here and
2753         * if it happened, we have to fail the write.
2754         */
2755        barrier();
2756        if (unlikely(ACCESS_ONCE(cpu_buffer->buffer) != buffer)) {
2757                local_dec(&cpu_buffer->committing);
2758                local_dec(&cpu_buffer->commits);
2759                return NULL;
2760        }
2761#endif
2762
2763        info.length = rb_calculate_event_length(length);
2764 again:
2765        info.add_timestamp = 0;
2766        info.delta = 0;
2767
2768        /*
2769         * We allow for interrupts to reenter here and do a trace.
2770         * If one does, it will cause this original code to loop
2771         * back here. Even with heavy interrupts happening, this
2772         * should only happen a few times in a row. If this happens
2773         * 1000 times in a row, there must be either an interrupt
2774         * storm or we have something buggy.
2775         * Bail!
2776         */
2777        if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
2778                goto out_fail;
2779
2780        info.ts = rb_time_stamp(cpu_buffer->buffer);
2781        diff = info.ts - cpu_buffer->write_stamp;
2782
2783        /* make sure this diff is calculated here */
2784        barrier();
2785
2786        /* Did the write stamp get updated already? */
2787        if (likely(info.ts >= cpu_buffer->write_stamp)) {
2788                info.delta = diff;
2789                if (unlikely(test_time_stamp(info.delta)))
2790                        rb_handle_timestamp(cpu_buffer, &info);
2791        }
2792
2793        event = __rb_reserve_next(cpu_buffer, &info);
2794
2795        if (unlikely(PTR_ERR(event) == -EAGAIN)) {
2796                if (info.add_timestamp)
2797                        info.length -= RB_LEN_TIME_EXTEND;
2798                goto again;
2799        }
2800
2801        if (!event)
2802                goto out_fail;
2803
2804        return event;
2805
2806 out_fail:
2807        rb_end_commit(cpu_buffer);
2808        return NULL;
2809}
2810
2811/**
2812 * ring_buffer_lock_reserve - reserve a part of the buffer
2813 * @buffer: the ring buffer to reserve from
2814 * @length: the length of the data to reserve (excluding event header)
2815 *
2816 * Returns a reseverd event on the ring buffer to copy directly to.
2817 * The user of this interface will need to get the body to write into
2818 * and can use the ring_buffer_event_data() interface.
2819 *
2820 * The length is the length of the data needed, not the event length
2821 * which also includes the event header.
2822 *
2823 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2824 * If NULL is returned, then nothing has been allocated or locked.
2825 */
2826struct ring_buffer_event *
2827ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length)
2828{
2829        struct ring_buffer_per_cpu *cpu_buffer;
2830        struct ring_buffer_event *event;
2831        int cpu;
2832
2833        /* If we are tracing schedule, we don't want to recurse */
2834        preempt_disable_notrace();
2835
2836        if (unlikely(atomic_read(&buffer->record_disabled)))
2837                goto out;
2838
2839        cpu = raw_smp_processor_id();
2840
2841        if (unlikely(!cpumask_test_cpu(cpu, buffer->cpumask)))
2842                goto out;
2843
2844        cpu_buffer = buffer->buffers[cpu];
2845
2846        if (unlikely(atomic_read(&cpu_buffer->record_disabled)))
2847                goto out;
2848
2849        if (unlikely(length > BUF_MAX_DATA_SIZE))
2850                goto out;
2851
2852        if (unlikely(trace_recursive_lock(cpu_buffer)))
2853                goto out;
2854
2855        event = rb_reserve_next_event(buffer, cpu_buffer, length);
2856        if (!event)
2857                goto out_unlock;
2858
2859        return event;
2860
2861 out_unlock:
2862        trace_recursive_unlock(cpu_buffer);
2863 out:
2864        preempt_enable_notrace();
2865        return NULL;
2866}
2867EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
2868
2869/*
2870 * Decrement the entries to the page that an event is on.
2871 * The event does not even need to exist, only the pointer
2872 * to the page it is on. This may only be called before the commit
2873 * takes place.
2874 */
2875static inline void
2876rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
2877                   struct ring_buffer_event *event)
2878{
2879        unsigned long addr = (unsigned long)event;
2880        struct buffer_page *bpage = cpu_buffer->commit_page;
2881        struct buffer_page *start;
2882
2883        addr &= PAGE_MASK;
2884
2885        /* Do the likely case first */
2886        if (likely(bpage->page == (void *)addr)) {
2887                local_dec(&bpage->entries);
2888                return;
2889        }
2890
2891        /*
2892         * Because the commit page may be on the reader page we
2893         * start with the next page and check the end loop there.
2894         */
2895        rb_inc_page(cpu_buffer, &bpage);
2896        start = bpage;
2897        do {
2898                if (bpage->page == (void *)addr) {
2899                        local_dec(&bpage->entries);
2900                        return;
2901                }
2902                rb_inc_page(cpu_buffer, &bpage);
2903        } while (bpage != start);
2904
2905        /* commit not part of this buffer?? */
2906        RB_WARN_ON(cpu_buffer, 1);
2907}
2908
2909/**
2910 * ring_buffer_commit_discard - discard an event that has not been committed
2911 * @buffer: the ring buffer
2912 * @event: non committed event to discard
2913 *
2914 * Sometimes an event that is in the ring buffer needs to be ignored.
2915 * This function lets the user discard an event in the ring buffer
2916 * and then that event will not be read later.
2917 *
2918 * This function only works if it is called before the the item has been
2919 * committed. It will try to free the event from the ring buffer
2920 * if another event has not been added behind it.
2921 *
2922 * If another event has been added behind it, it will set the event
2923 * up as discarded, and perform the commit.
2924 *
2925 * If this function is called, do not call ring_buffer_unlock_commit on
2926 * the event.
2927 */
2928void ring_buffer_discard_commit(struct ring_buffer *buffer,
2929                                struct ring_buffer_event *event)
2930{
2931        struct ring_buffer_per_cpu *cpu_buffer;
2932        int cpu;
2933
2934        /* The event is discarded regardless */
2935        rb_event_discard(event);
2936
2937        cpu = smp_processor_id();
2938        cpu_buffer = buffer->buffers[cpu];
2939
2940        /*
2941         * This must only be called if the event has not been
2942         * committed yet. Thus we can assume that preemption
2943         * is still disabled.
2944         */
2945        RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
2946
2947        rb_decrement_entry(cpu_buffer, event);
2948        if (rb_try_to_discard(cpu_buffer, event))
2949                goto out;
2950
2951        /*
2952         * The commit is still visible by the reader, so we
2953         * must still update the timestamp.
2954         */
2955        rb_update_write_stamp(cpu_buffer, event);
2956 out:
2957        rb_end_commit(cpu_buffer);
2958
2959        trace_recursive_unlock(cpu_buffer);
2960
2961        preempt_enable_notrace();
2962
2963}
2964EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
2965
2966/**
2967 * ring_buffer_write - write data to the buffer without reserving
2968 * @buffer: The ring buffer to write to.
2969 * @length: The length of the data being written (excluding the event header)
2970 * @data: The data to write to the buffer.
2971 *
2972 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
2973 * one function. If you already have the data to write to the buffer, it
2974 * may be easier to simply call this function.
2975 *
2976 * Note, like ring_buffer_lock_reserve, the length is the length of the data
2977 * and not the length of the event which would hold the header.
2978 */
2979int ring_buffer_write(struct ring_buffer *buffer,
2980                      unsigned long length,
2981                      void *data)
2982{
2983        struct ring_buffer_per_cpu *cpu_buffer;
2984        struct ring_buffer_event *event;
2985        void *body;
2986        int ret = -EBUSY;
2987        int cpu;
2988
2989        preempt_disable_notrace();
2990
2991        if (atomic_read(&buffer->record_disabled))
2992                goto out;
2993
2994        cpu = raw_smp_processor_id();
2995
2996        if (!cpumask_test_cpu(cpu, buffer->cpumask))
2997                goto out;
2998
2999        cpu_buffer = buffer->buffers[cpu];
3000
3001        if (atomic_read(&cpu_buffer->record_disabled))
3002                goto out;
3003
3004        if (length > BUF_MAX_DATA_SIZE)
3005                goto out;
3006
3007        if (unlikely(trace_recursive_lock(cpu_buffer)))
3008                goto out;
3009
3010        event = rb_reserve_next_event(buffer, cpu_buffer, length);
3011        if (!event)
3012                goto out_unlock;
3013
3014        body = rb_event_data(event);
3015
3016        memcpy(body, data, length);
3017
3018        rb_commit(cpu_buffer, event);
3019
3020        rb_wakeups(buffer, cpu_buffer);
3021
3022        ret = 0;
3023
3024 out_unlock:
3025        trace_recursive_unlock(cpu_buffer);
3026
3027 out:
3028        preempt_enable_notrace();
3029
3030        return ret;
3031}
3032EXPORT_SYMBOL_GPL(ring_buffer_write);
3033
3034static bool rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
3035{
3036        struct buffer_page *reader = cpu_buffer->reader_page;
3037        struct buffer_page *head = rb_set_head_page(cpu_buffer);
3038        struct buffer_page *commit = cpu_buffer->commit_page;
3039
3040        /* In case of error, head will be NULL */
3041        if (unlikely(!head))
3042                return true;
3043
3044        return reader->read == rb_page_commit(reader) &&
3045                (commit == reader ||
3046                 (commit == head &&
3047                  head->read == rb_page_commit(commit)));
3048}
3049
3050/**
3051 * ring_buffer_record_disable - stop all writes into the buffer
3052 * @buffer: The ring buffer to stop writes to.
3053 *
3054 * This prevents all writes to the buffer. Any attempt to write
3055 * to the buffer after this will fail and return NULL.
3056 *
3057 * The caller should call synchronize_sched() after this.
3058 */
3059void ring_buffer_record_disable(struct ring_buffer *buffer)
3060{
3061        atomic_inc(&buffer->record_disabled);
3062}
3063EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
3064
3065/**
3066 * ring_buffer_record_enable - enable writes to the buffer
3067 * @buffer: The ring buffer to enable writes
3068 *
3069 * Note, multiple disables will need the same number of enables
3070 * to truly enable the writing (much like preempt_disable).
3071 */
3072void ring_buffer_record_enable(struct ring_buffer *buffer)
3073{
3074        atomic_dec(&buffer->record_disabled);
3075}
3076EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
3077
3078/**
3079 * ring_buffer_record_off - stop all writes into the buffer
3080 * @buffer: The ring buffer to stop writes to.
3081 *
3082 * This prevents all writes to the buffer. Any attempt to write
3083 * to the buffer after this will fail and return NULL.
3084 *
3085 * This is different than ring_buffer_record_disable() as
3086 * it works like an on/off switch, where as the disable() version
3087 * must be paired with a enable().
3088 */
3089void ring_buffer_record_off(struct ring_buffer *buffer)
3090{
3091        unsigned int rd;
3092        unsigned int new_rd;
3093
3094        do {
3095                rd = atomic_read(&buffer->record_disabled);
3096                new_rd = rd | RB_BUFFER_OFF;
3097        } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3098}
3099EXPORT_SYMBOL_GPL(ring_buffer_record_off);
3100
3101/**
3102 * ring_buffer_record_on - restart writes into the buffer
3103 * @buffer: The ring buffer to start writes to.
3104 *
3105 * This enables all writes to the buffer that was disabled by
3106 * ring_buffer_record_off().
3107 *
3108 * This is different than ring_buffer_record_enable() as
3109 * it works like an on/off switch, where as the enable() version
3110 * must be paired with a disable().
3111 */
3112void ring_buffer_record_on(struct ring_buffer *buffer)
3113{
3114        unsigned int rd;
3115        unsigned int new_rd;
3116
3117        do {
3118                rd = atomic_read(&buffer->record_disabled);
3119                new_rd = rd & ~RB_BUFFER_OFF;
3120        } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3121}
3122EXPORT_SYMBOL_GPL(ring_buffer_record_on);
3123
3124/**
3125 * ring_buffer_record_is_on - return true if the ring buffer can write
3126 * @buffer: The ring buffer to see if write is enabled
3127 *
3128 * Returns true if the ring buffer is in a state that it accepts writes.
3129 */
3130int ring_buffer_record_is_on(struct ring_buffer *buffer)
3131{
3132        return !atomic_read(&buffer->record_disabled);
3133}
3134
3135/**
3136 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
3137 * @buffer: The ring buffer to stop writes to.
3138 * @cpu: The CPU buffer to stop
3139 *
3140 * This prevents all writes to the buffer. Any attempt to write
3141 * to the buffer after this will fail and return NULL.
3142 *
3143 * The caller should call synchronize_sched() after this.
3144 */
3145void ring_buffer_record_disable_cpu(struct ring_buffer *buffer, int cpu)
3146{
3147        struct ring_buffer_per_cpu *cpu_buffer;
3148
3149        if (!cpumask_test_cpu(cpu, buffer->cpumask))
3150                return;
3151
3152        cpu_buffer = buffer->buffers[cpu];
3153        atomic_inc(&cpu_buffer->record_disabled);
3154}
3155EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
3156
3157/**
3158 * ring_buffer_record_enable_cpu - enable writes to the buffer
3159 * @buffer: The ring buffer to enable writes
3160 * @cpu: The CPU to enable.
3161 *
3162 * Note, multiple disables will need the same number of enables
3163 * to truly enable the writing (much like preempt_disable).
3164 */
3165void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu)
3166{
3167        struct ring_buffer_per_cpu *cpu_buffer;
3168
3169        if (!cpumask_test_cpu(cpu, buffer->cpumask))
3170                return;
3171
3172        cpu_buffer = buffer->buffers[cpu];
3173        atomic_dec(&cpu_buffer->record_disabled);
3174}
3175EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
3176
3177/*
3178 * The total entries in the ring buffer is the running counter
3179 * of entries entered into the ring buffer, minus the sum of
3180 * the entries read from the ring buffer and the number of
3181 * entries that were overwritten.
3182 */
3183static inline unsigned long
3184rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
3185{
3186        return local_read(&cpu_buffer->entries) -
3187                (local_read(&cpu_buffer->overrun) + cpu_buffer->read);
3188}
3189
3190/**
3191 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
3192 * @buffer: The ring buffer
3193 * @cpu: The per CPU buffer to read from.
3194 */
3195u64 ring_buffer_oldest_event_ts(struct ring_buffer *buffer, int cpu)
3196{
3197        unsigned long flags;
3198        struct ring_buffer_per_cpu *cpu_buffer;
3199        struct buffer_page *bpage;
3200        u64 ret = 0;
3201
3202        if (!cpumask_test_cpu(cpu, buffer->cpumask))
3203                return 0;
3204
3205        cpu_buffer = buffer->buffers[cpu];
3206        raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3207        /*
3208         * if the tail is on reader_page, oldest time stamp is on the reader
3209         * page
3210         */
3211        if (cpu_buffer->tail_page == cpu_buffer->reader_page)
3212                bpage = cpu_buffer->reader_page;
3213        else
3214                bpage = rb_set_head_page(cpu_buffer);
3215        if (bpage)
3216                ret = bpage->page->time_stamp;
3217        raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3218
3219        return ret;
3220}
3221EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
3222
3223/**
3224 * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
3225 * @buffer: The ring buffer
3226 * @cpu: The per CPU buffer to read from.
3227 */
3228unsigned long ring_buffer_bytes_cpu(struct ring_buffer *buffer, int cpu)
3229{
3230        struct ring_buffer_per_cpu *cpu_buffer;
3231        unsigned long ret;
3232
3233        if (!cpumask_test_cpu(cpu, buffer->cpumask))
3234                return 0;
3235
3236        cpu_buffer = buffer->buffers[cpu];
3237        ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
3238
3239        return ret;
3240}
3241EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
3242
3243/**
3244 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
3245 * @buffer: The ring buffer
3246 * @cpu: The per CPU buffer to get the entries from.
3247 */
3248unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu)
3249{
3250        struct ring_buffer_per_cpu *cpu_buffer;
3251
3252        if (!cpumask_test_cpu(cpu, buffer->cpumask))
3253                return 0;
3254
3255        cpu_buffer = buffer->buffers[cpu];
3256
3257        return rb_num_of_entries(cpu_buffer);
3258}
3259EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
3260
3261/**
3262 * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
3263 * buffer wrapping around (only if RB_FL_OVERWRITE is on).
3264 * @buffer: The ring buffer
3265 * @cpu: The per CPU buffer to get the number of overruns from
3266 */
3267unsigned long ring_buffer_overrun_cpu(struct ring_buffer *buffer, int cpu)
3268{
3269        struct ring_buffer_per_cpu *cpu_buffer;
3270        unsigned long ret;
3271
3272        if (!cpumask_test_cpu(cpu, buffer->cpumask))
3273                return 0;
3274
3275        cpu_buffer = buffer->buffers[cpu];
3276        ret = local_read(&cpu_buffer->overrun);
3277
3278        return ret;
3279}
3280EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
3281
3282/**
3283 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
3284 * commits failing due to the buffer wrapping around while there are uncommitted
3285 * events, such as during an interrupt storm.
3286 * @buffer: The ring buffer
3287 * @cpu: The per CPU buffer to get the number of overruns from
3288 */
3289unsigned long
3290ring_buffer_commit_overrun_cpu(struct ring_buffer *buffer, int cpu)
3291{
3292        struct ring_buffer_per_cpu *cpu_buffer;
3293        unsigned long ret;
3294
3295        if (!cpumask_test_cpu(cpu, buffer->cpumask))
3296                return 0;
3297
3298        cpu_buffer = buffer->buffers[cpu];
3299        ret = local_read(&cpu_buffer->commit_overrun);
3300
3301        return ret;
3302}
3303EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
3304
3305/**
3306 * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
3307 * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
3308 * @buffer: The ring buffer
3309 * @cpu: The per CPU buffer to get the number of overruns from
3310 */
3311unsigned long
3312ring_buffer_dropped_events_cpu(struct ring_buffer *buffer, int cpu)
3313{
3314        struct ring_buffer_per_cpu *cpu_buffer;
3315        unsigned long ret;
3316
3317        if (!cpumask_test_cpu(cpu, buffer->cpumask))
3318                return 0;
3319
3320        cpu_buffer = buffer->buffers[cpu];
3321        ret = local_read(&cpu_buffer->dropped_events);
3322
3323        return ret;
3324}
3325EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
3326
3327/**
3328 * ring_buffer_read_events_cpu - get the number of events successfully read
3329 * @buffer: The ring buffer
3330 * @cpu: The per CPU buffer to get the number of events read
3331 */
3332unsigned long
3333ring_buffer_read_events_cpu(struct ring_buffer *buffer, int cpu)
3334{
3335        struct ring_buffer_per_cpu *cpu_buffer;
3336
3337        if (!cpumask_test_cpu(cpu, buffer->cpumask))
3338                return 0;
3339
3340        cpu_buffer = buffer->buffers[cpu];
3341        return cpu_buffer->read;
3342}
3343EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu);
3344
3345/**
3346 * ring_buffer_entries - get the number of entries in a buffer
3347 * @buffer: The ring buffer
3348 *
3349 * Returns the total number of entries in the ring buffer
3350 * (all CPU entries)
3351 */
3352unsigned long ring_buffer_entries(struct ring_buffer *buffer)
3353{
3354        struct ring_buffer_per_cpu *cpu_buffer;
3355        unsigned long entries = 0;
3356        int cpu;
3357
3358        /* if you care about this being correct, lock the buffer */
3359        for_each_buffer_cpu(buffer, cpu) {
3360                cpu_buffer = buffer->buffers[cpu];
3361                entries += rb_num_of_entries(cpu_buffer);
3362        }
3363
3364        return entries;
3365}
3366EXPORT_SYMBOL_GPL(ring_buffer_entries);
3367
3368/**
3369 * ring_buffer_overruns - get the number of overruns in buffer
3370 * @buffer: The ring buffer
3371 *
3372 * Returns the total number of overruns in the ring buffer
3373 * (all CPU entries)
3374 */
3375unsigned long ring_buffer_overruns(struct ring_buffer *buffer)
3376{
3377        struct ring_buffer_per_cpu *cpu_buffer;
3378        unsigned long overruns = 0;
3379        int cpu;
3380
3381        /* if you care about this being correct, lock the buffer */
3382        for_each_buffer_cpu(buffer, cpu) {
3383                cpu_buffer = buffer->buffers[cpu];
3384                overruns += local_read(&cpu_buffer->overrun);
3385        }
3386
3387        return overruns;
3388}
3389EXPORT_SYMBOL_GPL(ring_buffer_overruns);
3390
3391static void rb_iter_reset(struct ring_buffer_iter *iter)
3392{
3393        struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3394
3395        /* Iterator usage is expected to have record disabled */
3396        iter->head_page = cpu_buffer->reader_page;
3397        iter->head = cpu_buffer->reader_page->read;
3398
3399        iter->cache_reader_page = iter->head_page;
3400        iter->cache_read = cpu_buffer->read;
3401
3402        if (iter->head)
3403                iter->read_stamp = cpu_buffer->read_stamp;
3404        else
3405                iter->read_stamp = iter->head_page->page->time_stamp;
3406}
3407
3408/**
3409 * ring_buffer_iter_reset - reset an iterator
3410 * @iter: The iterator to reset
3411 *
3412 * Resets the iterator, so that it will start from the beginning
3413 * again.
3414 */
3415void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
3416{
3417        struct ring_buffer_per_cpu *cpu_buffer;
3418        unsigned long flags;
3419
3420        if (!iter)
3421                return;
3422
3423        cpu_buffer = iter->cpu_buffer;
3424
3425        raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3426        rb_iter_reset(iter);
3427        raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3428}
3429EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
3430
3431/**
3432 * ring_buffer_iter_empty - check if an iterator has no more to read
3433 * @iter: The iterator to check
3434 */
3435int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
3436{
3437        struct ring_buffer_per_cpu *cpu_buffer;
3438
3439        cpu_buffer = iter->cpu_buffer;
3440
3441        return iter->head_page == cpu_buffer->commit_page &&
3442                iter->head == rb_commit_index(cpu_buffer);
3443}
3444EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
3445
3446static void
3447rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
3448                     struct ring_buffer_event *event)
3449{
3450        u64 delta;
3451
3452        switch (event->type_len) {
3453        case RINGBUF_TYPE_PADDING:
3454                return;
3455
3456        case RINGBUF_TYPE_TIME_EXTEND:
3457                delta = event->array[0];
3458                delta <<= TS_SHIFT;
3459                delta += event->time_delta;
3460                cpu_buffer->read_stamp += delta;
3461                return;
3462
3463        case RINGBUF_TYPE_TIME_STAMP:
3464                /* FIXME: not implemented */
3465                return;
3466
3467        case RINGBUF_TYPE_DATA:
3468                cpu_buffer->read_stamp += event->time_delta;
3469                return;
3470
3471        default:
3472                BUG();
3473        }
3474        return;
3475}
3476
3477static void
3478rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
3479                          struct ring_buffer_event *event)
3480{
3481        u64 delta;
3482
3483        switch (event->type_len) {
3484        case RINGBUF_TYPE_PADDING:
3485                return;
3486
3487        case RINGBUF_TYPE_TIME_EXTEND:
3488                delta = event->array[0];
3489                delta <<= TS_SHIFT;
3490                delta += event->time_delta;
3491                iter->read_stamp += delta;
3492                return;
3493
3494        case RINGBUF_TYPE_TIME_STAMP:
3495                /* FIXME: not implemented */
3496                return;
3497
3498        case RINGBUF_TYPE_DATA:
3499                iter->read_stamp += event->time_delta;
3500                return;
3501
3502        default:
3503                BUG();
3504        }
3505        return;
3506}
3507
3508static struct buffer_page *
3509rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
3510{
3511        struct buffer_page *reader = NULL;
3512        unsigned long overwrite;
3513        unsigned long flags;
3514        int nr_loops = 0;
3515        int ret;
3516
3517        local_irq_save(flags);
3518        arch_spin_lock(&cpu_buffer->lock);
3519
3520 again:
3521        /*
3522         * This should normally only loop twice. But because the
3523         * start of the reader inserts an empty page, it causes
3524         * a case where we will loop three times. There should be no
3525         * reason to loop four times (that I know of).
3526         */
3527        if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
3528                reader = NULL;
3529                goto out;
3530        }
3531
3532        reader = cpu_buffer->reader_page;
3533
3534        /* If there's more to read, return this page */
3535        if (cpu_buffer->reader_page->read < rb_page_size(reader))
3536                goto out;
3537
3538        /* Never should we have an index greater than the size */
3539        if (RB_WARN_ON(cpu_buffer,
3540                       cpu_buffer->reader_page->read > rb_page_size(reader)))
3541                goto out;
3542
3543        /* check if we caught up to the tail */
3544        reader = NULL;
3545        if (cpu_buffer->commit_page == cpu_buffer->reader_page)
3546                goto out;
3547
3548        /* Don't bother swapping if the ring buffer is empty */
3549        if (rb_num_of_entries(cpu_buffer) == 0)
3550                goto out;
3551
3552        /*
3553         * Reset the reader page to size zero.
3554         */
3555        local_set(&cpu_buffer->reader_page->write, 0);
3556        local_set(&cpu_buffer->reader_page->entries, 0);
3557        local_set(&cpu_buffer->reader_page->page->commit, 0);
3558        cpu_buffer->reader_page->real_end = 0;
3559
3560 spin:
3561        /*
3562         * Splice the empty reader page into the list around the head.
3563         */
3564        reader = rb_set_head_page(cpu_buffer);
3565        if (!reader)
3566                goto out;
3567        cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
3568        cpu_buffer->reader_page->list.prev = reader->list.prev;
3569
3570        /*
3571         * cpu_buffer->pages just needs to point to the buffer, it
3572         *  has no specific buffer page to point to. Lets move it out
3573         *  of our way so we don't accidentally swap it.
3574         */
3575        cpu_buffer->pages = reader->list.prev;
3576
3577        /* The reader page will be pointing to the new head */
3578        rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list);
3579
3580        /*
3581         * We want to make sure we read the overruns after we set up our
3582         * pointers to the next object. The writer side does a
3583         * cmpxchg to cross pages which acts as the mb on the writer
3584         * side. Note, the reader will constantly fail the swap
3585         * while the writer is updating the pointers, so this
3586         * guarantees that the overwrite recorded here is the one we
3587         * want to compare with the last_overrun.
3588         */
3589        smp_mb();
3590        overwrite = local_read(&(cpu_buffer->overrun));
3591
3592        /*
3593         * Here's the tricky part.
3594         *
3595         * We need to move the pointer past the header page.
3596         * But we can only do that if a writer is not currently
3597         * moving it. The page before the header page has the
3598         * flag bit '1' set if it is pointing to the page we want.
3599         * but if the writer is in the process of moving it
3600         * than it will be '2' or already moved '0'.
3601         */
3602
3603        ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
3604
3605        /*
3606         * If we did not convert it, then we must try again.
3607         */
3608        if (!ret)
3609                goto spin;
3610
3611        /*
3612         * Yeah! We succeeded in replacing the page.
3613         *
3614         * Now make the new head point back to the reader page.
3615         */
3616        rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
3617        rb_inc_page(cpu_buffer, &cpu_buffer->head_page);
3618
3619        /* Finally update the reader page to the new head */
3620        cpu_buffer->reader_page = reader;
3621        cpu_buffer->reader_page->read = 0;
3622
3623        if (overwrite != cpu_buffer->last_overrun) {
3624                cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
3625                cpu_buffer->last_overrun = overwrite;
3626        }
3627
3628        goto again;
3629
3630 out:
3631        /* Update the read_stamp on the first event */
3632        if (reader && reader->read == 0)
3633                cpu_buffer->read_stamp = reader->page->time_stamp;
3634
3635        arch_spin_unlock(&cpu_buffer->lock);
3636        local_irq_restore(flags);
3637
3638        return reader;
3639}
3640
3641static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
3642{
3643        struct ring_buffer_event *event;
3644        struct buffer_page *reader;
3645        unsigned length;
3646
3647        reader = rb_get_reader_page(cpu_buffer);
3648
3649        /* This function should not be called when buffer is empty */
3650        if (RB_WARN_ON(cpu_buffer, !reader))
3651                return;
3652
3653        event = rb_reader_event(cpu_buffer);
3654
3655        if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
3656                cpu_buffer->read++;
3657
3658        rb_update_read_stamp(cpu_buffer, event);
3659
3660        length = rb_event_length(event);
3661        cpu_buffer->reader_page->read += length;
3662}
3663
3664static void rb_advance_iter(struct ring_buffer_iter *iter)
3665{
3666        struct ring_buffer_per_cpu *cpu_buffer;
3667        struct ring_buffer_event *event;
3668        unsigned length;
3669
3670        cpu_buffer = iter->cpu_buffer;
3671
3672        /*
3673         * Check if we are at the end of the buffer.
3674         */
3675        if (iter->head >= rb_page_size(iter->head_page)) {
3676                /* discarded commits can make the page empty */
3677                if (iter->head_page == cpu_buffer->commit_page)
3678                        return;
3679                rb_inc_iter(iter);
3680                return;
3681        }
3682
3683        event = rb_iter_head_event(iter);
3684
3685        length = rb_event_length(event);
3686
3687        /*
3688         * This should not be called to advance the header if we are
3689         * at the tail of the buffer.
3690         */
3691        if (RB_WARN_ON(cpu_buffer,
3692                       (iter->head_page == cpu_buffer->commit_page) &&
3693                       (iter->head + length > rb_commit_index(cpu_buffer))))
3694                return;
3695
3696        rb_update_iter_read_stamp(iter, event);
3697
3698        iter->head += length;
3699
3700        /* check for end of page padding */
3701        if ((iter->head >= rb_page_size(iter->head_page)) &&
3702            (iter->head_page != cpu_buffer->commit_page))
3703                rb_inc_iter(iter);
3704}
3705
3706static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
3707{
3708        return cpu_buffer->lost_events;
3709}
3710
3711static struct ring_buffer_event *
3712rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
3713               unsigned long *lost_events)
3714{
3715        struct ring_buffer_event *event;
3716        struct buffer_page *reader;
3717        int nr_loops = 0;
3718
3719 again:
3720        /*
3721         * We repeat when a time extend is encountered.
3722         * Since the time extend is always attached to a data event,
3723         * we should never loop more than once.
3724         * (We never hit the following condition more than twice).
3725         */
3726        if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3727                return NULL;
3728
3729        reader = rb_get_reader_page(cpu_buffer);
3730        if (!reader)
3731                return NULL;
3732
3733        event = rb_reader_event(cpu_buffer);
3734
3735        switch (event->type_len) {
3736        case RINGBUF_TYPE_PADDING:
3737                if (rb_null_event(event))
3738                        RB_WARN_ON(cpu_buffer, 1);
3739                /*
3740                 * Because the writer could be discarding every
3741                 * event it creates (which would probably be bad)
3742                 * if we were to go back to "again" then we may never
3743                 * catch up, and will trigger the warn on, or lock
3744                 * the box. Return the padding, and we will release
3745                 * the current locks, and try again.
3746                 */
3747                return event;
3748
3749        case RINGBUF_TYPE_TIME_EXTEND:
3750                /* Internal data, OK to advance */
3751                rb_advance_reader(cpu_buffer);
3752                goto again;
3753
3754        case RINGBUF_TYPE_TIME_STAMP:
3755                /* FIXME: not implemented */
3756                rb_advance_reader(cpu_buffer);
3757                goto again;
3758
3759        case RINGBUF_TYPE_DATA:
3760                if (ts) {
3761                        *ts = cpu_buffer->read_stamp + event->time_delta;
3762                        ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3763                                                         cpu_buffer->cpu, ts);
3764                }
3765                if (lost_events)
3766                        *lost_events = rb_lost_events(cpu_buffer);
3767                return event;
3768
3769        default:
3770                BUG();
3771        }
3772
3773        return NULL;
3774}
3775EXPORT_SYMBOL_GPL(ring_buffer_peek);
3776
3777static struct ring_buffer_event *
3778rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3779{
3780        struct ring_buffer *buffer;
3781        struct ring_buffer_per_cpu *cpu_buffer;
3782        struct ring_buffer_event *event;
3783        int nr_loops = 0;
3784
3785        cpu_buffer = iter->cpu_buffer;
3786        buffer = cpu_buffer->buffer;
3787
3788        /*
3789         * Check if someone performed a consuming read to
3790         * the buffer. A consuming read invalidates the iterator
3791         * and we need to reset the iterator in this case.
3792         */
3793        if (unlikely(iter->cache_read != cpu_buffer->read ||
3794                     iter->cache_reader_page != cpu_buffer->reader_page))
3795                rb_iter_reset(iter);
3796
3797 again:
3798        if (ring_buffer_iter_empty(iter))
3799                return NULL;
3800
3801        /*
3802         * We repeat when a time extend is encountered or we hit
3803         * the end of the page. Since the time extend is always attached
3804         * to a data event, we should never loop more than three times.
3805         * Once for going to next page, once on time extend, and
3806         * finally once to get the event.
3807         * (We never hit the following condition more than thrice).
3808         */
3809        if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3))
3810                return NULL;
3811
3812        if (rb_per_cpu_empty(cpu_buffer))
3813                return NULL;
3814
3815        if (iter->head >= rb_page_size(iter->head_page)) {
3816                rb_inc_iter(iter);
3817                goto again;
3818        }
3819
3820        event = rb_iter_head_event(iter);
3821
3822        switch (event->type_len) {
3823        case RINGBUF_TYPE_PADDING:
3824                if (rb_null_event(event)) {
3825                        rb_inc_iter(iter);
3826                        goto again;
3827                }
3828                rb_advance_iter(iter);
3829                return event;
3830
3831        case RINGBUF_TYPE_TIME_EXTEND:
3832                /* Internal data, OK to advance */
3833                rb_advance_iter(iter);
3834                goto again;
3835
3836        case RINGBUF_TYPE_TIME_STAMP:
3837                /* FIXME: not implemented */
3838                rb_advance_iter(iter);
3839                goto again;
3840
3841        case RINGBUF_TYPE_DATA:
3842                if (ts) {
3843                        *ts = iter->read_stamp + event->time_delta;
3844                        ring_buffer_normalize_time_stamp(buffer,
3845                                                         cpu_buffer->cpu, ts);
3846                }
3847                return event;
3848
3849        default:
3850                BUG();
3851        }
3852
3853        return NULL;
3854}
3855EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
3856
3857static inline bool rb_reader_lock(struct ring_buffer_per_cpu *cpu_buffer)
3858{
3859        if (likely(!in_nmi())) {
3860                raw_spin_lock(&cpu_buffer->reader_lock);
3861                return true;
3862        }
3863
3864        /*
3865         * If an NMI die dumps out the content of the ring buffer
3866         * trylock must be used to prevent a deadlock if the NMI
3867         * preempted a task that holds the ring buffer locks. If
3868         * we get the lock then all is fine, if not, then continue
3869         * to do the read, but this can corrupt the ring buffer,
3870         * so it must be permanently disabled from future writes.
3871         * Reading from NMI is a oneshot deal.
3872         */
3873        if (raw_spin_trylock(&cpu_buffer->reader_lock))
3874                return true;
3875
3876        /* Continue without locking, but disable the ring buffer */
3877        atomic_inc(&cpu_buffer->record_disabled);
3878        return false;
3879}
3880
3881static inline void
3882rb_reader_unlock(struct ring_buffer_per_cpu *cpu_buffer, bool locked)
3883{
3884        if (likely(locked))
3885                raw_spin_unlock(&cpu_buffer->reader_lock);
3886        return;
3887}
3888
3889/**
3890 * ring_buffer_peek - peek at the next event to be read
3891 * @buffer: The ring buffer to read
3892 * @cpu: The cpu to peak at
3893 * @ts: The timestamp counter of this event.
3894 * @lost_events: a variable to store if events were lost (may be NULL)
3895 *
3896 * This will return the event that will be read next, but does
3897 * not consume the data.
3898 */
3899struct ring_buffer_event *
3900ring_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts,
3901                 unsigned long *lost_events)
3902{
3903        struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3904        struct ring_buffer_event *event;
3905        unsigned long flags;
3906        bool dolock;
3907
3908        if (!cpumask_test_cpu(cpu, buffer->cpumask))
3909                return NULL;
3910
3911 again:
3912        local_irq_save(flags);
3913        dolock = rb_reader_lock(cpu_buffer);
3914        event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3915        if (event && event->type_len == RINGBUF_TYPE_PADDING)
3916                rb_advance_reader(cpu_buffer);
3917        rb_reader_unlock(cpu_buffer, dolock);
3918        local_irq_restore(flags);
3919
3920        if (event && event->type_len == RINGBUF_TYPE_PADDING)
3921                goto again;
3922
3923        return event;
3924}
3925
3926/**
3927 * ring_buffer_iter_peek - peek at the next event to be read
3928 * @iter: The ring buffer iterator
3929 * @ts: The timestamp counter of this event.
3930 *
3931 * This will return the event that will be read next, but does
3932 * not increment the iterator.
3933 */
3934struct ring_buffer_event *
3935ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3936{
3937        struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3938        struct ring_buffer_event *event;
3939        unsigned long flags;
3940
3941 again:
3942        raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3943        event = rb_iter_peek(iter, ts);
3944        raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3945
3946        if (event && event->type_len == RINGBUF_TYPE_PADDING)
3947                goto again;
3948
3949        return event;
3950}
3951
3952/**
3953 * ring_buffer_consume - return an event and consume it
3954 * @buffer: The ring buffer to get the next event from
3955 * @cpu: the cpu to read the buffer from
3956 * @ts: a variable to store the timestamp (may be NULL)
3957 * @lost_events: a variable to store if events were lost (may be NULL)
3958 *
3959 * Returns the next event in the ring buffer, and that event is consumed.
3960 * Meaning, that sequential reads will keep returning a different event,
3961 * and eventually empty the ring buffer if the producer is slower.
3962 */
3963struct ring_buffer_event *
3964ring_buffer_consume(struct ring_buffer *buffer, int cpu, u64 *ts,
3965                    unsigned long *lost_events)
3966{
3967        struct ring_buffer_per_cpu *cpu_buffer;
3968        struct ring_buffer_event *event = NULL;
3969        unsigned long flags;
3970        bool dolock;
3971
3972 again:
3973        /* might be called in atomic */
3974        preempt_disable();
3975
3976        if (!cpumask_test_cpu(cpu, buffer->cpumask))
3977                goto out;
3978
3979        cpu_buffer = buffer->buffers[cpu];
3980        local_irq_save(flags);
3981        dolock = rb_reader_lock(cpu_buffer);
3982
3983        event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3984        if (event) {
3985                cpu_buffer->lost_events = 0;
3986                rb_advance_reader(cpu_buffer);
3987        }
3988
3989        rb_reader_unlock(cpu_buffer, dolock);
3990        local_irq_restore(flags);
3991
3992 out:
3993        preempt_enable();
3994
3995        if (event && event->type_len == RINGBUF_TYPE_PADDING)
3996                goto again;
3997
3998        return event;
3999}
4000EXPORT_SYMBOL_GPL(ring_buffer_consume);
4001
4002/**
4003 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
4004 * @buffer: The ring buffer to read from
4005 * @cpu: The cpu buffer to iterate over
4006 *
4007 * This performs the initial preparations necessary to iterate
4008 * through the buffer.  Memory is allocated, buffer recording
4009 * is disabled, and the iterator pointer is returned to the caller.
4010 *
4011 * Disabling buffer recordng prevents the reading from being
4012 * corrupted. This is not a consuming read, so a producer is not
4013 * expected.
4014 *
4015 * After a sequence of ring_buffer_read_prepare calls, the user is
4016 * expected to make at least one call to ring_buffer_read_prepare_sync.
4017 * Afterwards, ring_buffer_read_start is invoked to get things going
4018 * for real.
4019 *
4020 * This overall must be paired with ring_buffer_read_finish.
4021 */
4022struct ring_buffer_iter *
4023ring_buffer_read_prepare(struct ring_buffer *buffer, int cpu)
4024{
4025        struct ring_buffer_per_cpu *cpu_buffer;
4026        struct ring_buffer_iter *iter;
4027
4028        if (!cpumask_test_cpu(cpu, buffer->cpumask))
4029                return NULL;
4030
4031        iter = kmalloc(sizeof(*iter), GFP_KERNEL);
4032        if (!iter)
4033                return NULL;
4034
4035        cpu_buffer = buffer->buffers[cpu];
4036
4037        iter->cpu_buffer = cpu_buffer;
4038
4039        atomic_inc(&buffer->resize_disabled);
4040        atomic_inc(&cpu_buffer->record_disabled);
4041
4042        return iter;
4043}
4044EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
4045
4046/**
4047 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
4048 *
4049 * All previously invoked ring_buffer_read_prepare calls to prepare
4050 * iterators will be synchronized.  Afterwards, read_buffer_read_start
4051 * calls on those iterators are allowed.
4052 */
4053void
4054ring_buffer_read_prepare_sync(void)
4055{
4056        synchronize_sched();
4057}
4058EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
4059
4060/**
4061 * ring_buffer_read_start - start a non consuming read of the buffer
4062 * @iter: The iterator returned by ring_buffer_read_prepare
4063 *
4064 * This finalizes the startup of an iteration through the buffer.
4065 * The iterator comes from a call to ring_buffer_read_prepare and
4066 * an intervening ring_buffer_read_prepare_sync must have been
4067 * performed.
4068 *
4069 * Must be paired with ring_buffer_read_finish.
4070 */
4071void
4072ring_buffer_read_start(struct ring_buffer_iter *iter)
4073{
4074        struct ring_buffer_per_cpu *cpu_buffer;
4075        unsigned long flags;
4076
4077        if (!iter)
4078                return;
4079
4080        cpu_buffer = iter->cpu_buffer;
4081
4082        raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4083        arch_spin_lock(&cpu_buffer->lock);
4084        rb_iter_reset(iter);
4085        arch_spin_unlock(&cpu_buffer->lock);
4086        raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4087}
4088EXPORT_SYMBOL_GPL(ring_buffer_read_start);
4089
4090/**
4091 * ring_buffer_read_finish - finish reading the iterator of the buffer
4092 * @iter: The iterator retrieved by ring_buffer_start
4093 *
4094 * This re-enables the recording to the buffer, and frees the
4095 * iterator.
4096 */
4097void
4098ring_buffer_read_finish(struct ring_buffer_iter *iter)
4099{
4100        struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4101        unsigned long flags;
4102
4103        /*
4104         * Ring buffer is disabled from recording, here's a good place
4105         * to check the integrity of the ring buffer.
4106         * Must prevent readers from trying to read, as the check
4107         * clears the HEAD page and readers require it.
4108         */
4109        raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4110        rb_check_pages(cpu_buffer);
4111        raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4112
4113        atomic_dec(&cpu_buffer->record_disabled);
4114        atomic_dec(&cpu_buffer->buffer->resize_disabled);
4115        kfree(iter);
4116}
4117EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
4118
4119/**
4120 * ring_buffer_read - read the next item in the ring buffer by the iterator
4121 * @iter: The ring buffer iterator
4122 * @ts: The time stamp of the event read.
4123 *
4124 * This reads the next event in the ring buffer and increments the iterator.
4125 */
4126struct ring_buffer_event *
4127ring_buffer_read(struct ring_buffer_iter *iter, u64 *ts)
4128{
4129        struct ring_buffer_event *event;
4130        struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4131        unsigned long flags;
4132
4133        raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4134 again:
4135        event = rb_iter_peek(iter, ts);
4136        if (!event)
4137                goto out;
4138
4139        if (event->type_len == RINGBUF_TYPE_PADDING)
4140                goto again;
4141
4142        rb_advance_iter(iter);
4143 out:
4144        raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4145
4146        return event;
4147}
4148EXPORT_SYMBOL_GPL(ring_buffer_read);
4149
4150/**
4151 * ring_buffer_size - return the size of the ring buffer (in bytes)
4152 * @buffer: The ring buffer.
4153 */
4154unsigned long ring_buffer_size(struct ring_buffer *buffer, int cpu)
4155{
4156        /*
4157         * Earlier, this method returned
4158         *      BUF_PAGE_SIZE * buffer->nr_pages
4159         * Since the nr_pages field is now removed, we have converted this to
4160         * return the per cpu buffer value.
4161         */
4162        if (!cpumask_test_cpu(cpu, buffer->cpumask))
4163                return 0;
4164
4165        return BUF_PAGE_SIZE * buffer->buffers[cpu]->nr_pages;
4166}
4167EXPORT_SYMBOL_GPL(ring_buffer_size);
4168
4169static void
4170rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
4171{
4172        rb_head_page_deactivate(cpu_buffer);
4173
4174        cpu_buffer->head_page
4175                = list_entry(cpu_buffer->pages, struct buffer_page, list);
4176        local_set(&cpu_buffer->head_page->write, 0);
4177        local_set(&cpu_buffer->head_page->entries, 0);
4178        local_set(&cpu_buffer->head_page->page->commit, 0);
4179
4180        cpu_buffer->head_page->read = 0;
4181
4182        cpu_buffer->tail_page = cpu_buffer->head_page;
4183        cpu_buffer->commit_page = cpu_buffer->head_page;
4184
4185        INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
4186        INIT_LIST_HEAD(&cpu_buffer->new_pages);
4187        local_set(&cpu_buffer->reader_page->write, 0);
4188        local_set(&cpu_buffer->reader_page->entries, 0);
4189        local_set(&cpu_buffer->reader_page->page->commit, 0);
4190        cpu_buffer->reader_page->read = 0;
4191
4192        local_set(&cpu_buffer->entries_bytes, 0);
4193        local_set(&cpu_buffer->overrun, 0);
4194        local_set(&cpu_buffer->commit_overrun, 0);
4195        local_set(&cpu_buffer->dropped_events, 0);
4196        local_set(&cpu_buffer->entries, 0);
4197        local_set(&cpu_buffer->committing, 0);
4198        local_set(&cpu_buffer->commits, 0);
4199        cpu_buffer->read = 0;
4200        cpu_buffer->read_bytes = 0;
4201
4202        cpu_buffer->write_stamp = 0;
4203        cpu_buffer->read_stamp = 0;
4204
4205        cpu_buffer->lost_events = 0;
4206        cpu_buffer->last_overrun = 0;
4207
4208        rb_head_page_activate(cpu_buffer);
4209}
4210
4211/**
4212 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
4213 * @buffer: The ring buffer to reset a per cpu buffer of
4214 * @cpu: The CPU buffer to be reset
4215 */
4216void ring_buffer_reset_cpu(struct ring_buffer *buffer, int cpu)
4217{
4218        struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4219        unsigned long flags;
4220
4221        if (!cpumask_test_cpu(cpu, buffer->cpumask))
4222                return;
4223
4224        atomic_inc(&buffer->resize_disabled);
4225        atomic_inc(&cpu_buffer->record_disabled);
4226
4227        /* Make sure all commits have finished */
4228        synchronize_sched();
4229
4230        raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4231
4232        if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
4233                goto out;
4234
4235        arch_spin_lock(&cpu_buffer->lock);
4236
4237        rb_reset_cpu(cpu_buffer);
4238
4239        arch_spin_unlock(&cpu_buffer->lock);
4240
4241 out:
4242        raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4243
4244        atomic_dec(&cpu_buffer->record_disabled);
4245        atomic_dec(&buffer->resize_disabled);
4246}
4247EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
4248
4249/**
4250 * ring_buffer_reset - reset a ring buffer
4251 * @buffer: The ring buffer to reset all cpu buffers
4252 */
4253void ring_buffer_reset(struct ring_buffer *buffer)
4254{
4255        int cpu;
4256
4257        for_each_buffer_cpu(buffer, cpu)
4258                ring_buffer_reset_cpu(buffer, cpu);
4259}
4260EXPORT_SYMBOL_GPL(ring_buffer_reset);
4261
4262/**
4263 * rind_buffer_empty - is the ring buffer empty?
4264 * @buffer: The ring buffer to test
4265 */
4266bool ring_buffer_empty(struct ring_buffer *buffer)
4267{
4268        struct ring_buffer_per_cpu *cpu_buffer;
4269        unsigned long flags;
4270        bool dolock;
4271        int cpu;
4272        int ret;
4273
4274        /* yes this is racy, but if you don't like the race, lock the buffer */
4275        for_each_buffer_cpu(buffer, cpu) {
4276                cpu_buffer = buffer->buffers[cpu];
4277                local_irq_save(flags);
4278                dolock = rb_reader_lock(cpu_buffer);
4279                ret = rb_per_cpu_empty(cpu_buffer);
4280                rb_reader_unlock(cpu_buffer, dolock);
4281                local_irq_restore(flags);
4282
4283                if (!ret)
4284                        return false;
4285        }
4286
4287        return true;
4288}
4289EXPORT_SYMBOL_GPL(ring_buffer_empty);
4290
4291/**
4292 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
4293 * @buffer: The ring buffer
4294 * @cpu: The CPU buffer to test
4295 */
4296bool ring_buffer_empty_cpu(struct ring_buffer *buffer, int cpu)
4297{
4298        struct ring_buffer_per_cpu *cpu_buffer;
4299        unsigned long flags;
4300        bool dolock;
4301        int ret;
4302
4303        if (!cpumask_test_cpu(cpu, buffer->cpumask))
4304                return true;
4305
4306        cpu_buffer = buffer->buffers[cpu];
4307        local_irq_save(flags);
4308        dolock = rb_reader_lock(cpu_buffer);
4309        ret = rb_per_cpu_empty(cpu_buffer);
4310        rb_reader_unlock(cpu_buffer, dolock);
4311        local_irq_restore(flags);
4312
4313        return ret;
4314}
4315EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
4316
4317#ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
4318/**
4319 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
4320 * @buffer_a: One buffer to swap with
4321 * @buffer_b: The other buffer to swap with
4322 *
4323 * This function is useful for tracers that want to take a "snapshot"
4324 * of a CPU buffer and has another back up buffer lying around.
4325 * it is expected that the tracer handles the cpu buffer not being
4326 * used at the moment.
4327 */
4328int ring_buffer_swap_cpu(struct ring_buffer *buffer_a,
4329                         struct ring_buffer *buffer_b, int cpu)
4330{
4331        struct ring_buffer_per_cpu *cpu_buffer_a;
4332        struct ring_buffer_per_cpu *cpu_buffer_b;
4333        int ret = -EINVAL;
4334
4335        if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
4336            !cpumask_test_cpu(cpu, buffer_b->cpumask))
4337                goto out;
4338
4339        cpu_buffer_a = buffer_a->buffers[cpu];
4340        cpu_buffer_b = buffer_b->buffers[cpu];
4341
4342        /* At least make sure the two buffers are somewhat the same */
4343        if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
4344                goto out;
4345
4346        ret = -EAGAIN;
4347
4348        if (atomic_read(&buffer_a->record_disabled))
4349                goto out;
4350
4351        if (atomic_read(&buffer_b->record_disabled))
4352                goto out;
4353
4354        if (atomic_read(&cpu_buffer_a->record_disabled))
4355                goto out;
4356
4357        if (atomic_read(&cpu_buffer_b->record_disabled))
4358                goto out;
4359
4360        /*
4361         * We can't do a synchronize_sched here because this
4362         * function can be called in atomic context.
4363         * Normally this will be called from the same CPU as cpu.
4364         * If not it's up to the caller to protect this.
4365         */
4366        atomic_inc(&cpu_buffer_a->record_disabled);
4367        atomic_inc(&cpu_buffer_b->record_disabled);
4368
4369        ret = -EBUSY;
4370        if (local_read(&cpu_buffer_a->committing))
4371                goto out_dec;
4372        if (local_read(&cpu_buffer_b->committing))
4373                goto out_dec;
4374
4375        buffer_a->buffers[cpu] = cpu_buffer_b;
4376        buffer_b->buffers[cpu] = cpu_buffer_a;
4377
4378        cpu_buffer_b->buffer = buffer_a;
4379        cpu_buffer_a->buffer = buffer_b;
4380
4381        ret = 0;
4382
4383out_dec:
4384        atomic_dec(&cpu_buffer_a->record_disabled);
4385        atomic_dec(&cpu_buffer_b->record_disabled);
4386out:
4387        return ret;
4388}
4389EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
4390#endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
4391
4392/**
4393 * ring_buffer_alloc_read_page - allocate a page to read from buffer
4394 * @buffer: the buffer to allocate for.
4395 * @cpu: the cpu buffer to allocate.
4396 *
4397 * This function is used in conjunction with ring_buffer_read_page.
4398 * When reading a full page from the ring buffer, these functions
4399 * can be used to speed up the process. The calling function should
4400 * allocate a few pages first with this function. Then when it
4401 * needs to get pages from the ring buffer, it passes the result
4402 * of this function into ring_buffer_read_page, which will swap
4403 * the page that was allocated, with the read page of the buffer.
4404 *
4405 * Returns:
4406 *  The page allocated, or NULL on error.
4407 */
4408void *ring_buffer_alloc_read_page(struct ring_buffer *buffer, int cpu)
4409{
4410        struct buffer_data_page *bpage;
4411        struct page *page;
4412
4413        page = alloc_pages_node(cpu_to_node(cpu),
4414                                GFP_KERNEL | __GFP_NORETRY, 0);
4415        if (!page)
4416                return NULL;
4417
4418        bpage = page_address(page);
4419
4420        rb_init_page(bpage);
4421
4422        return bpage;
4423}
4424EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
4425
4426/**
4427 * ring_buffer_free_read_page - free an allocated read page
4428 * @buffer: the buffer the page was allocate for
4429 * @data: the page to free
4430 *
4431 * Free a page allocated from ring_buffer_alloc_read_page.
4432 */
4433void ring_buffer_free_read_page(struct ring_buffer *buffer, void *data)
4434{
4435        free_page((unsigned long)data);
4436}
4437EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
4438
4439/**
4440 * ring_buffer_read_page - extract a page from the ring buffer
4441 * @buffer: buffer to extract from
4442 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
4443 * @len: amount to extract
4444 * @cpu: the cpu of the buffer to extract
4445 * @full: should the extraction only happen when the page is full.
4446 *
4447 * This function will pull out a page from the ring buffer and consume it.
4448 * @data_page must be the address of the variable that was returned
4449 * from ring_buffer_alloc_read_page. This is because the page might be used
4450 * to swap with a page in the ring buffer.
4451 *
4452 * for example:
4453 *      rpage = ring_buffer_alloc_read_page(buffer, cpu);
4454 *      if (!rpage)
4455 *              return error;
4456 *      ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
4457 *      if (ret >= 0)
4458 *              process_page(rpage, ret);
4459 *
4460 * When @full is set, the function will not return true unless
4461 * the writer is off the reader page.
4462 *
4463 * Note: it is up to the calling functions to handle sleeps and wakeups.
4464 *  The ring buffer can be used anywhere in the kernel and can not
4465 *  blindly call wake_up. The layer that uses the ring buffer must be
4466 *  responsible for that.
4467 *
4468 * Returns:
4469 *  >=0 if data has been transferred, returns the offset of consumed data.
4470 *  <0 if no data has been transferred.
4471 */
4472int ring_buffer_read_page(struct ring_buffer *buffer,
4473                          void **data_page, size_t len, int cpu, int full)
4474{
4475        struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4476        struct ring_buffer_event *event;
4477        struct buffer_data_page *bpage;
4478        struct buffer_page *reader;
4479        unsigned long missed_events;
4480        unsigned long flags;
4481        unsigned int commit;
4482        unsigned int read;
4483        u64 save_timestamp;
4484        int ret = -1;
4485
4486        if (!cpumask_test_cpu(cpu, buffer->cpumask))
4487                goto out;
4488
4489        /*
4490         * If len is not big enough to hold the page header, then
4491         * we can not copy anything.
4492         */
4493        if (len <= BUF_PAGE_HDR_SIZE)
4494                goto out;
4495
4496        len -= BUF_PAGE_HDR_SIZE;
4497
4498        if (!data_page)
4499                goto out;
4500
4501        bpage = *data_page;
4502        if (!bpage)
4503                goto out;
4504
4505        raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4506
4507        reader = rb_get_reader_page(cpu_buffer);
4508        if (!reader)
4509                goto out_unlock;
4510
4511        event = rb_reader_event(cpu_buffer);
4512
4513        read = reader->read;
4514        commit = rb_page_commit(reader);
4515
4516        /* Check if any events were dropped */
4517        missed_events = cpu_buffer->lost_events;
4518
4519        /*
4520         * If this page has been partially read or
4521         * if len is not big enough to read the rest of the page or
4522         * a writer is still on the page, then
4523         * we must copy the data from the page to the buffer.
4524         * Otherwise, we can simply swap the page with the one passed in.
4525         */
4526        if (read || (len < (commit - read)) ||
4527            cpu_buffer->reader_page == cpu_buffer->commit_page) {
4528                struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
4529                unsigned int rpos = read;
4530                unsigned int pos = 0;
4531                unsigned int size;
4532
4533                if (full)
4534                        goto out_unlock;
4535
4536                if (len > (commit - read))
4537                        len = (commit - read);
4538
4539                /* Always keep the time extend and data together */
4540                size = rb_event_ts_length(event);
4541
4542                if (len < size)
4543                        goto out_unlock;
4544
4545                /* save the current timestamp, since the user will need it */
4546                save_timestamp = cpu_buffer->read_stamp;
4547
4548                /* Need to copy one event at a time */
4549                do {
4550                        /* We need the size of one event, because
4551                         * rb_advance_reader only advances by one event,
4552                         * whereas rb_event_ts_length may include the size of
4553                         * one or two events.
4554                         * We have already ensured there's enough space if this
4555                         * is a time extend. */
4556                        size = rb_event_length(event);
4557                        memcpy(bpage->data + pos, rpage->data + rpos, size);
4558
4559                        len -= size;
4560
4561                        rb_advance_reader(cpu_buffer);
4562                        rpos = reader->read;
4563                        pos += size;
4564
4565                        if (rpos >= commit)
4566                                break;
4567
4568                        event = rb_reader_event(cpu_buffer);
4569                        /* Always keep the time extend and data together */
4570                        size = rb_event_ts_length(event);
4571                } while (len >= size);
4572
4573                /* update bpage */
4574                local_set(&bpage->commit, pos);
4575                bpage->time_stamp = save_timestamp;
4576
4577                /* we copied everything to the beginning */
4578                read = 0;
4579        } else {
4580                /* update the entry counter */
4581                cpu_buffer->read += rb_page_entries(reader);
4582                cpu_buffer->read_bytes += BUF_PAGE_SIZE;
4583
4584                /* swap the pages */
4585                rb_init_page(bpage);
4586                bpage = reader->page;
4587                reader->page = *data_page;
4588                local_set(&reader->write, 0);
4589                local_set(&reader->entries, 0);
4590                reader->read = 0;
4591                *data_page = bpage;
4592
4593                /*
4594                 * Use the real_end for the data size,
4595                 * This gives us a chance to store the lost events
4596                 * on the page.
4597                 */
4598                if (reader->real_end)
4599                        local_set(&bpage->commit, reader->real_end);
4600        }
4601        ret = read;
4602
4603        cpu_buffer->lost_events = 0;
4604
4605        commit = local_read(&bpage->commit);
4606        /*
4607         * Set a flag in the commit field if we lost events
4608         */
4609        if (missed_events) {
4610                /* If there is room at the end of the page to save the
4611                 * missed events, then record it there.
4612                 */
4613                if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
4614                        memcpy(&bpage->data[commit], &missed_events,
4615                               sizeof(missed_events));
4616                        local_add(RB_MISSED_STORED, &bpage->commit);
4617                        commit += sizeof(missed_events);
4618                }
4619                local_add(RB_MISSED_EVENTS, &bpage->commit);
4620        }
4621
4622        /*
4623         * This page may be off to user land. Zero it out here.
4624         */
4625        if (commit < BUF_PAGE_SIZE)
4626                memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
4627
4628 out_unlock:
4629        raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4630
4631 out:
4632        return ret;
4633}
4634EXPORT_SYMBOL_GPL(ring_buffer_read_page);
4635
4636#ifdef CONFIG_HOTPLUG_CPU
4637static int rb_cpu_notify(struct notifier_block *self,
4638                         unsigned long action, void *hcpu)
4639{
4640        struct ring_buffer *buffer =
4641                container_of(self, struct ring_buffer, cpu_notify);
4642        long cpu = (long)hcpu;
4643        long nr_pages_same;
4644        int cpu_i;
4645        unsigned long nr_pages;
4646
4647        switch (action) {
4648        case CPU_UP_PREPARE:
4649        case CPU_UP_PREPARE_FROZEN:
4650                if (cpumask_test_cpu(cpu, buffer->cpumask))
4651                        return NOTIFY_OK;
4652
4653                nr_pages = 0;
4654                nr_pages_same = 1;
4655                /* check if all cpu sizes are same */
4656                for_each_buffer_cpu(buffer, cpu_i) {
4657                        /* fill in the size from first enabled cpu */
4658                        if (nr_pages == 0)
4659                                nr_pages = buffer->buffers[cpu_i]->nr_pages;
4660                        if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
4661                                nr_pages_same = 0;
4662                                break;
4663                        }
4664                }
4665                /* allocate minimum pages, user can later expand it */
4666                if (!nr_pages_same)
4667                        nr_pages = 2;
4668                buffer->buffers[cpu] =
4669                        rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
4670                if (!buffer->buffers[cpu]) {
4671                        WARN(1, "failed to allocate ring buffer on CPU %ld\n",
4672                             cpu);
4673                        return NOTIFY_OK;
4674                }
4675                smp_wmb();
4676                cpumask_set_cpu(cpu, buffer->cpumask);
4677                break;
4678        case CPU_DOWN_PREPARE:
4679        case CPU_DOWN_PREPARE_FROZEN:
4680                /*
4681                 * Do nothing.
4682                 *  If we were to free the buffer, then the user would
4683                 *  lose any trace that was in the buffer.
4684                 */
4685                break;
4686        default:
4687                break;
4688        }
4689        return NOTIFY_OK;
4690}
4691#endif
4692
4693#ifdef CONFIG_RING_BUFFER_STARTUP_TEST
4694/*
4695 * This is a basic integrity check of the ring buffer.
4696 * Late in the boot cycle this test will run when configured in.
4697 * It will kick off a thread per CPU that will go into a loop
4698 * writing to the per cpu ring buffer various sizes of data.
4699 * Some of the data will be large items, some small.
4700 *
4701 * Another thread is created that goes into a spin, sending out
4702 * IPIs to the other CPUs to also write into the ring buffer.
4703 * this is to test the nesting ability of the buffer.
4704 *
4705 * Basic stats are recorded and reported. If something in the
4706 * ring buffer should happen that's not expected, a big warning
4707 * is displayed and all ring buffers are disabled.
4708 */
4709static struct task_struct *rb_threads[NR_CPUS] __initdata;
4710
4711struct rb_test_data {
4712        struct ring_buffer      *buffer;
4713        unsigned long           events;
4714        unsigned long           bytes_written;
4715        unsigned long           bytes_alloc;
4716        unsigned long           bytes_dropped;
4717        unsigned long           events_nested;
4718        unsigned long           bytes_written_nested;
4719        unsigned long           bytes_alloc_nested;
4720        unsigned long           bytes_dropped_nested;
4721        int                     min_size_nested;
4722        int                     max_size_nested;
4723        int                     max_size;
4724        int                     min_size;
4725        int                     cpu;
4726        int                     cnt;
4727};
4728
4729static struct rb_test_data rb_data[NR_CPUS] __initdata;
4730
4731/* 1 meg per cpu */
4732#define RB_TEST_BUFFER_SIZE     1048576
4733
4734static char rb_string[] __initdata =
4735        "abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
4736        "?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
4737        "!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
4738
4739static bool rb_test_started __initdata;
4740
4741struct rb_item {
4742        int size;
4743        char str[];
4744};
4745
4746static __init int rb_write_something(struct rb_test_data *data, bool nested)
4747{
4748        struct ring_buffer_event *event;
4749        struct rb_item *item;
4750        bool started;
4751        int event_len;
4752        int size;
4753        int len;
4754        int cnt;
4755
4756        /* Have nested writes different that what is written */
4757        cnt = data->cnt + (nested ? 27 : 0);
4758
4759        /* Multiply cnt by ~e, to make some unique increment */
4760        size = (data->cnt * 68 / 25) % (sizeof(rb_string) - 1);
4761
4762        len = size + sizeof(struct rb_item);
4763
4764        started = rb_test_started;
4765        /* read rb_test_started before checking buffer enabled */
4766        smp_rmb();
4767
4768        event = ring_buffer_lock_reserve(data->buffer, len);
4769        if (!event) {
4770                /* Ignore dropped events before test starts. */
4771                if (started) {
4772                        if (nested)
4773                                data->bytes_dropped += len;
4774                        else
4775                                data->bytes_dropped_nested += len;
4776                }
4777                return len;
4778        }
4779
4780        event_len = ring_buffer_event_length(event);
4781
4782        if (RB_WARN_ON(data->buffer, event_len < len))
4783                goto out;
4784
4785        item = ring_buffer_event_data(event);
4786        item->size = size;
4787        memcpy(item->str, rb_string, size);
4788
4789        if (nested) {
4790                data->bytes_alloc_nested += event_len;
4791                data->bytes_written_nested += len;
4792                data->events_nested++;
4793                if (!data->min_size_nested || len < data->min_size_nested)
4794                        data->min_size_nested = len;
4795                if (len > data->max_size_nested)
4796                        data->max_size_nested = len;
4797        } else {
4798                data->bytes_alloc += event_len;
4799                data->bytes_written += len;
4800                data->events++;
4801                if (!data->min_size || len < data->min_size)
4802                        data->max_size = len;
4803                if (len > data->max_size)
4804                        data->max_size = len;
4805        }
4806
4807 out:
4808        ring_buffer_unlock_commit(data->buffer, event);
4809
4810        return 0;
4811}
4812
4813static __init int rb_test(void *arg)
4814{
4815        struct rb_test_data *data = arg;
4816
4817        while (!kthread_should_stop()) {
4818                rb_write_something(data, false);
4819                data->cnt++;
4820
4821                set_current_state(TASK_INTERRUPTIBLE);
4822                /* Now sleep between a min of 100-300us and a max of 1ms */
4823                usleep_range(((data->cnt % 3) + 1) * 100, 1000);
4824        }
4825
4826        return 0;
4827}
4828
4829static __init void rb_ipi(void *ignore)
4830{
4831        struct rb_test_data *data;
4832        int cpu = smp_processor_id();
4833
4834        data = &rb_data[cpu];
4835        rb_write_something(data, true);
4836}
4837
4838static __init int rb_hammer_test(void *arg)
4839{
4840        while (!kthread_should_stop()) {
4841
4842                /* Send an IPI to all cpus to write data! */
4843                smp_call_function(rb_ipi, NULL, 1);
4844                /* No sleep, but for non preempt, let others run */
4845                schedule();
4846        }
4847
4848        return 0;
4849}
4850
4851static __init int test_ringbuffer(void)
4852{
4853        struct task_struct *rb_hammer;
4854        struct ring_buffer *buffer;
4855        int cpu;
4856        int ret = 0;
4857
4858        pr_info("Running ring buffer tests...\n");
4859
4860        buffer = ring_buffer_alloc(RB_TEST_BUFFER_SIZE, RB_FL_OVERWRITE);
4861        if (WARN_ON(!buffer))
4862                return 0;
4863
4864        /* Disable buffer so that threads can't write to it yet */
4865        ring_buffer_record_off(buffer);
4866
4867        for_each_online_cpu(cpu) {
4868                rb_data[cpu].buffer = buffer;
4869                rb_data[cpu].cpu = cpu;
4870                rb_data[cpu].cnt = cpu;
4871                rb_threads[cpu] = kthread_create(rb_test, &rb_data[cpu],
4872                                                 "rbtester/%d", cpu);
4873                if (WARN_ON(IS_ERR(rb_threads[cpu]))) {
4874                        pr_cont("FAILED\n");
4875                        ret = PTR_ERR(rb_threads[cpu]);
4876                        goto out_free;
4877                }
4878
4879                kthread_bind(rb_threads[cpu], cpu);
4880                wake_up_process(rb_threads[cpu]);
4881        }
4882
4883        /* Now create the rb hammer! */
4884        rb_hammer = kthread_run(rb_hammer_test, NULL, "rbhammer");
4885        if (WARN_ON(IS_ERR(rb_hammer))) {
4886                pr_cont("FAILED\n");
4887                ret = PTR_ERR(rb_hammer);
4888                goto out_free;
4889        }
4890
4891        ring_buffer_record_on(buffer);
4892        /*
4893         * Show buffer is enabled before setting rb_test_started.
4894         * Yes there's a small race window where events could be
4895         * dropped and the thread wont catch it. But when a ring
4896         * buffer gets enabled, there will always be some kind of
4897         * delay before other CPUs see it. Thus, we don't care about
4898         * those dropped events. We care about events dropped after
4899         * the threads see that the buffer is active.
4900         */
4901        smp_wmb();
4902        rb_test_started = true;
4903
4904        set_current_state(TASK_INTERRUPTIBLE);
4905        /* Just run for 10 seconds */;
4906        schedule_timeout(10 * HZ);
4907
4908        kthread_stop(rb_hammer);
4909
4910 out_free:
4911        for_each_online_cpu(cpu) {
4912                if (!rb_threads[cpu])
4913                        break;
4914                kthread_stop(rb_threads[cpu]);
4915        }
4916        if (ret) {
4917                ring_buffer_free(buffer);
4918                return ret;
4919        }
4920
4921        /* Report! */
4922        pr_info("finished\n");
4923        for_each_online_cpu(cpu) {
4924                struct ring_buffer_event *event;
4925                struct rb_test_data *data = &rb_data[cpu];
4926                struct rb_item *item;
4927                unsigned long total_events;
4928                unsigned long total_dropped;
4929                unsigned long total_written;
4930                unsigned long total_alloc;
4931                unsigned long total_read = 0;
4932                unsigned long total_size = 0;
4933                unsigned long total_len = 0;
4934                unsigned long total_lost = 0;
4935                unsigned long lost;
4936                int big_event_size;
4937                int small_event_size;
4938
4939                ret = -1;
4940
4941                total_events = data->events + data->events_nested;
4942                total_written = data->bytes_written + data->bytes_written_nested;
4943                total_alloc = data->bytes_alloc + data->bytes_alloc_nested;
4944                total_dropped = data->bytes_dropped + data->bytes_dropped_nested;
4945
4946                big_event_size = data->max_size + data->max_size_nested;
4947                small_event_size = data->min_size + data->min_size_nested;
4948
4949                pr_info("CPU %d:\n", cpu);
4950                pr_info("              events:    %ld\n", total_events);
4951                pr_info("       dropped bytes:    %ld\n", total_dropped);
4952                pr_info("       alloced bytes:    %ld\n", total_alloc);
4953                pr_info("       written bytes:    %ld\n", total_written);
4954                pr_info("       biggest event:    %d\n", big_event_size);
4955                pr_info("      smallest event:    %d\n", small_event_size);
4956
4957                if (RB_WARN_ON(buffer, total_dropped))
4958                        break;
4959
4960                ret = 0;
4961
4962                while ((event = ring_buffer_consume(buffer, cpu, NULL, &lost))) {
4963                        total_lost += lost;
4964                        item = ring_buffer_event_data(event);
4965                        total_len += ring_buffer_event_length(event);
4966                        total_size += item->size + sizeof(struct rb_item);
4967                        if (memcmp(&item->str[0], rb_string, item->size) != 0) {
4968                                pr_info("FAILED!\n");
4969                                pr_info("buffer had: %.*s\n", item->size, item->str);
4970                                pr_info("expected:   %.*s\n", item->size, rb_string);
4971                                RB_WARN_ON(buffer, 1);
4972                                ret = -1;
4973                                break;
4974                        }
4975                        total_read++;
4976                }
4977                if (ret)
4978                        break;
4979
4980                ret = -1;
4981
4982                pr_info("         read events:   %ld\n", total_read);
4983                pr_info("         lost events:   %ld\n", total_lost);
4984                pr_info("        total events:   %ld\n", total_lost + total_read);
4985                pr_info("  recorded len bytes:   %ld\n", total_len);
4986                pr_info(" recorded size bytes:   %ld\n", total_size);
4987                if (total_lost)
4988                        pr_info(" With dropped events, record len and size may not match\n"
4989                                " alloced and written from above\n");
4990                if (!total_lost) {
4991                        if (RB_WARN_ON(buffer, total_len != total_alloc ||
4992                                       total_size != total_written))
4993                                break;
4994                }
4995                if (RB_WARN_ON(buffer, total_lost + total_read != total_events))
4996                        break;
4997
4998                ret = 0;
4999        }
5000        if (!ret)
5001                pr_info("Ring buffer PASSED!\n");
5002
5003        ring_buffer_free(buffer);
5004        return 0;
5005}
5006
5007late_initcall(test_ringbuffer);
5008#endif /* CONFIG_RING_BUFFER_STARTUP_TEST */
Note: See TracBrowser for help on using the repository browser.