~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

TOMOYO Linux Cross Reference
Linux/kernel/trace/ring_buffer.c

Version: ~ [ linux-5.14-rc3 ] ~ [ linux-5.13.5 ] ~ [ linux-5.12.19 ] ~ [ linux-5.11.22 ] ~ [ linux-5.10.53 ] ~ [ linux-5.9.16 ] ~ [ linux-5.8.18 ] ~ [ linux-5.7.19 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.135 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.198 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.240 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.276 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.276 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.18.140 ] ~ [ linux-3.16.85 ] ~ [ linux-3.14.79 ] ~ [ linux-3.12.74 ] ~ [ linux-3.10.108 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

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

~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

kernel.org | git.kernel.org | LWN.net | Project Home | Wiki (Japanese) | Wiki (English) | SVN repository | Mail admin

Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.

osdn.jp