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

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

Version: ~ [ linux-6.1-rc7 ] ~ [ linux-6.0.10 ] ~ [ linux-5.19.17 ] ~ [ linux-5.18.19 ] ~ [ linux-5.17.15 ] ~ [ linux-5.16.20 ] ~ [ linux-5.15.80 ] ~ [ linux-5.14.21 ] ~ [ linux-5.13.19 ] ~ [ linux-5.12.19 ] ~ [ linux-5.11.22 ] ~ [ linux-5.10.156 ] ~ [ linux-5.9.16 ] ~ [ linux-5.8.18 ] ~ [ linux-5.7.19 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.225 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.267 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.300 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.334 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.302 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.10.108 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.9 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

  1 // SPDX-License-Identifier: GPL-2.0
  2 /*
  3  * Generic ring buffer
  4  *
  5  * Copyright (C) 2008 Steven Rostedt <srostedt@redhat.com>
  6  */
  7 #include <linux/trace_events.h>
  8 #include <linux/ring_buffer.h>
  9 #include <linux/trace_clock.h>
 10 #include <linux/sched/clock.h>
 11 #include <linux/trace_seq.h>
 12 #include <linux/spinlock.h>
 13 #include <linux/irq_work.h>
 14 #include <linux/security.h>
 15 #include <linux/uaccess.h>
 16 #include <linux/hardirq.h>
 17 #include <linux/kthread.h>      /* for self test */
 18 #include <linux/module.h>
 19 #include <linux/percpu.h>
 20 #include <linux/mutex.h>
 21 #include <linux/delay.h>
 22 #include <linux/slab.h>
 23 #include <linux/init.h>
 24 #include <linux/hash.h>
 25 #include <linux/list.h>
 26 #include <linux/cpu.h>
 27 #include <linux/oom.h>
 28 
 29 #include <asm/local.h>
 30 
 31 static void update_pages_handler(struct work_struct *work);
 32 
 33 /*
 34  * The ring buffer header is special. We must manually up keep it.
 35  */
 36 int ring_buffer_print_entry_header(struct trace_seq *s)
 37 {
 38         trace_seq_puts(s, "# compressed entry header\n");
 39         trace_seq_puts(s, "\ttype_len    :    5 bits\n");
 40         trace_seq_puts(s, "\ttime_delta  :   27 bits\n");
 41         trace_seq_puts(s, "\tarray       :   32 bits\n");
 42         trace_seq_putc(s, '\n');
 43         trace_seq_printf(s, "\tpadding     : type == %d\n",
 44                          RINGBUF_TYPE_PADDING);
 45         trace_seq_printf(s, "\ttime_extend : type == %d\n",
 46                          RINGBUF_TYPE_TIME_EXTEND);
 47         trace_seq_printf(s, "\ttime_stamp : type == %d\n",
 48                          RINGBUF_TYPE_TIME_STAMP);
 49         trace_seq_printf(s, "\tdata max type_len  == %d\n",
 50                          RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
 51 
 52         return !trace_seq_has_overflowed(s);
 53 }
 54 
 55 /*
 56  * The ring buffer is made up of a list of pages. A separate list of pages is
 57  * allocated for each CPU. A writer may only write to a buffer that is
 58  * associated with the CPU it is currently executing on.  A reader may read
 59  * from any per cpu buffer.
 60  *
 61  * The reader is special. For each per cpu buffer, the reader has its own
 62  * reader page. When a reader has read the entire reader page, this reader
 63  * page is swapped with another page in the ring buffer.
 64  *
 65  * Now, as long as the writer is off the reader page, the reader can do what
 66  * ever it wants with that page. The writer will never write to that page
 67  * again (as long as it is out of the ring buffer).
 68  *
 69  * Here's some silly ASCII art.
 70  *
 71  *   +------+
 72  *   |reader|          RING BUFFER
 73  *   |page  |
 74  *   +------+        +---+   +---+   +---+
 75  *                   |   |-->|   |-->|   |
 76  *                   +---+   +---+   +---+
 77  *                     ^               |
 78  *                     |               |
 79  *                     +---------------+
 80  *
 81  *
 82  *   +------+
 83  *   |reader|          RING BUFFER
 84  *   |page  |------------------v
 85  *   +------+        +---+   +---+   +---+
 86  *                   |   |-->|   |-->|   |
 87  *                   +---+   +---+   +---+
 88  *                     ^               |
 89  *                     |               |
 90  *                     +---------------+
 91  *
 92  *
 93  *   +------+
 94  *   |reader|          RING BUFFER
 95  *   |page  |------------------v
 96  *   +------+        +---+   +---+   +---+
 97  *      ^            |   |-->|   |-->|   |
 98  *      |            +---+   +---+   +---+
 99  *      |                              |
100  *      |                              |
101  *      +------------------------------+
102  *
103  *
104  *   +------+
105  *   |buffer|          RING BUFFER
106  *   |page  |------------------v
107  *   +------+        +---+   +---+   +---+
108  *      ^            |   |   |   |-->|   |
109  *      |   New      +---+   +---+   +---+
110  *      |  Reader------^               |
111  *      |   page                       |
112  *      +------------------------------+
113  *
114  *
115  * After we make this swap, the reader can hand this page off to the splice
116  * code and be done with it. It can even allocate a new page if it needs to
117  * and swap that into the ring buffer.
118  *
119  * We will be using cmpxchg soon to make all this lockless.
120  *
121  */
122 
123 /* Used for individual buffers (after the counter) */
124 #define RB_BUFFER_OFF           (1 << 20)
125 
126 #define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
127 
128 #define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
129 #define RB_ALIGNMENT            4U
130 #define RB_MAX_SMALL_DATA       (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
131 #define RB_EVNT_MIN_SIZE        8U      /* two 32bit words */
132 #define RB_ALIGN_DATA           __aligned(RB_ALIGNMENT)
133 
134 /* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
135 #define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
136 
137 enum {
138         RB_LEN_TIME_EXTEND = 8,
139         RB_LEN_TIME_STAMP =  8,
140 };
141 
142 #define skip_time_extend(event) \
143         ((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
144 
145 #define extended_time(event) \
146         (event->type_len >= RINGBUF_TYPE_TIME_EXTEND)
147 
148 static inline int rb_null_event(struct ring_buffer_event *event)
149 {
150         return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
151 }
152 
153 static void rb_event_set_padding(struct ring_buffer_event *event)
154 {
155         /* padding has a NULL time_delta */
156         event->type_len = RINGBUF_TYPE_PADDING;
157         event->time_delta = 0;
158 }
159 
160 static unsigned
161 rb_event_data_length(struct ring_buffer_event *event)
162 {
163         unsigned length;
164 
165         if (event->type_len)
166                 length = event->type_len * RB_ALIGNMENT;
167         else
168                 length = event->array[0];
169         return length + RB_EVNT_HDR_SIZE;
170 }
171 
172 /*
173  * Return the length of the given event. Will return
174  * the length of the time extend if the event is a
175  * time extend.
176  */
177 static inline unsigned
178 rb_event_length(struct ring_buffer_event *event)
179 {
180         switch (event->type_len) {
181         case RINGBUF_TYPE_PADDING:
182                 if (rb_null_event(event))
183                         /* undefined */
184                         return -1;
185                 return  event->array[0] + RB_EVNT_HDR_SIZE;
186 
187         case RINGBUF_TYPE_TIME_EXTEND:
188                 return RB_LEN_TIME_EXTEND;
189 
190         case RINGBUF_TYPE_TIME_STAMP:
191                 return RB_LEN_TIME_STAMP;
192 
193         case RINGBUF_TYPE_DATA:
194                 return rb_event_data_length(event);
195         default:
196                 BUG();
197         }
198         /* not hit */
199         return 0;
200 }
201 
202 /*
203  * Return total length of time extend and data,
204  *   or just the event length for all other events.
205  */
206 static inline unsigned
207 rb_event_ts_length(struct ring_buffer_event *event)
208 {
209         unsigned len = 0;
210 
211         if (extended_time(event)) {
212                 /* time extends include the data event after it */
213                 len = RB_LEN_TIME_EXTEND;
214                 event = skip_time_extend(event);
215         }
216         return len + rb_event_length(event);
217 }
218 
219 /**
220  * ring_buffer_event_length - return the length of the event
221  * @event: the event to get the length of
222  *
223  * Returns the size of the data load of a data event.
224  * If the event is something other than a data event, it
225  * returns the size of the event itself. With the exception
226  * of a TIME EXTEND, where it still returns the size of the
227  * data load of the data event after it.
228  */
229 unsigned ring_buffer_event_length(struct ring_buffer_event *event)
230 {
231         unsigned length;
232 
233         if (extended_time(event))
234                 event = skip_time_extend(event);
235 
236         length = rb_event_length(event);
237         if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
238                 return length;
239         length -= RB_EVNT_HDR_SIZE;
240         if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
241                 length -= sizeof(event->array[0]);
242         return length;
243 }
244 EXPORT_SYMBOL_GPL(ring_buffer_event_length);
245 
246 /* inline for ring buffer fast paths */
247 static __always_inline void *
248 rb_event_data(struct ring_buffer_event *event)
249 {
250         if (extended_time(event))
251                 event = skip_time_extend(event);
252         BUG_ON(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
253         /* If length is in len field, then array[0] has the data */
254         if (event->type_len)
255                 return (void *)&event->array[0];
256         /* Otherwise length is in array[0] and array[1] has the data */
257         return (void *)&event->array[1];
258 }
259 
260 /**
261  * ring_buffer_event_data - return the data of the event
262  * @event: the event to get the data from
263  */
264 void *ring_buffer_event_data(struct ring_buffer_event *event)
265 {
266         return rb_event_data(event);
267 }
268 EXPORT_SYMBOL_GPL(ring_buffer_event_data);
269 
270 #define for_each_buffer_cpu(buffer, cpu)                \
271         for_each_cpu(cpu, buffer->cpumask)
272 
273 #define TS_SHIFT        27
274 #define TS_MASK         ((1ULL << TS_SHIFT) - 1)
275 #define TS_DELTA_TEST   (~TS_MASK)
276 
277 /**
278  * ring_buffer_event_time_stamp - return the event's extended timestamp
279  * @event: the event to get the timestamp of
280  *
281  * Returns the extended timestamp associated with a data event.
282  * An extended time_stamp is a 64-bit timestamp represented
283  * internally in a special way that makes the best use of space
284  * contained within a ring buffer event.  This function decodes
285  * it and maps it to a straight u64 value.
286  */
287 u64 ring_buffer_event_time_stamp(struct ring_buffer_event *event)
288 {
289         u64 ts;
290 
291         ts = event->array[0];
292         ts <<= TS_SHIFT;
293         ts += event->time_delta;
294 
295         return ts;
296 }
297 
298 /* Flag when events were overwritten */
299 #define RB_MISSED_EVENTS        (1 << 31)
300 /* Missed count stored at end */
301 #define RB_MISSED_STORED        (1 << 30)
302 
303 struct buffer_data_page {
304         u64              time_stamp;    /* page time stamp */
305         local_t          commit;        /* write committed index */
306         unsigned char    data[] RB_ALIGN_DATA;  /* data of buffer page */
307 };
308 
309 /*
310  * Note, the buffer_page list must be first. The buffer pages
311  * are allocated in cache lines, which means that each buffer
312  * page will be at the beginning of a cache line, and thus
313  * the least significant bits will be zero. We use this to
314  * add flags in the list struct pointers, to make the ring buffer
315  * lockless.
316  */
317 struct buffer_page {
318         struct list_head list;          /* list of buffer pages */
319         local_t          write;         /* index for next write */
320         unsigned         read;          /* index for next read */
321         local_t          entries;       /* entries on this page */
322         unsigned long    real_end;      /* real end of data */
323         struct buffer_data_page *page;  /* Actual data page */
324 };
325 
326 /*
327  * The buffer page counters, write and entries, must be reset
328  * atomically when crossing page boundaries. To synchronize this
329  * update, two counters are inserted into the number. One is
330  * the actual counter for the write position or count on the page.
331  *
332  * The other is a counter of updaters. Before an update happens
333  * the update partition of the counter is incremented. This will
334  * allow the updater to update the counter atomically.
335  *
336  * The counter is 20 bits, and the state data is 12.
337  */
338 #define RB_WRITE_MASK           0xfffff
339 #define RB_WRITE_INTCNT         (1 << 20)
340 
341 static void rb_init_page(struct buffer_data_page *bpage)
342 {
343         local_set(&bpage->commit, 0);
344 }
345 
346 /*
347  * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
348  * this issue out.
349  */
350 static void free_buffer_page(struct buffer_page *bpage)
351 {
352         free_page((unsigned long)bpage->page);
353         kfree(bpage);
354 }
355 
356 /*
357  * We need to fit the time_stamp delta into 27 bits.
358  */
359 static inline int test_time_stamp(u64 delta)
360 {
361         if (delta & TS_DELTA_TEST)
362                 return 1;
363         return 0;
364 }
365 
366 #define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
367 
368 /* Max payload is BUF_PAGE_SIZE - header (8bytes) */
369 #define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
370 
371 int ring_buffer_print_page_header(struct trace_seq *s)
372 {
373         struct buffer_data_page field;
374 
375         trace_seq_printf(s, "\tfield: u64 timestamp;\t"
376                          "offset:0;\tsize:%u;\tsigned:%u;\n",
377                          (unsigned int)sizeof(field.time_stamp),
378                          (unsigned int)is_signed_type(u64));
379 
380         trace_seq_printf(s, "\tfield: local_t commit;\t"
381                          "offset:%u;\tsize:%u;\tsigned:%u;\n",
382                          (unsigned int)offsetof(typeof(field), commit),
383                          (unsigned int)sizeof(field.commit),
384                          (unsigned int)is_signed_type(long));
385 
386         trace_seq_printf(s, "\tfield: int overwrite;\t"
387                          "offset:%u;\tsize:%u;\tsigned:%u;\n",
388                          (unsigned int)offsetof(typeof(field), commit),
389                          1,
390                          (unsigned int)is_signed_type(long));
391 
392         trace_seq_printf(s, "\tfield: char data;\t"
393                          "offset:%u;\tsize:%u;\tsigned:%u;\n",
394                          (unsigned int)offsetof(typeof(field), data),
395                          (unsigned int)BUF_PAGE_SIZE,
396                          (unsigned int)is_signed_type(char));
397 
398         return !trace_seq_has_overflowed(s);
399 }
400 
401 struct rb_irq_work {
402         struct irq_work                 work;
403         wait_queue_head_t               waiters;
404         wait_queue_head_t               full_waiters;
405         bool                            waiters_pending;
406         bool                            full_waiters_pending;
407         bool                            wakeup_full;
408 };
409 
410 /*
411  * Structure to hold event state and handle nested events.
412  */
413 struct rb_event_info {
414         u64                     ts;
415         u64                     delta;
416         unsigned long           length;
417         struct buffer_page      *tail_page;
418         int                     add_timestamp;
419 };
420 
421 /*
422  * Used for which event context the event is in.
423  *  NMI     = 0
424  *  IRQ     = 1
425  *  SOFTIRQ = 2
426  *  NORMAL  = 3
427  *
428  * See trace_recursive_lock() comment below for more details.
429  */
430 enum {
431         RB_CTX_NMI,
432         RB_CTX_IRQ,
433         RB_CTX_SOFTIRQ,
434         RB_CTX_NORMAL,
435         RB_CTX_MAX
436 };
437 
438 /*
439  * head_page == tail_page && head == tail then buffer is empty.
440  */
441 struct ring_buffer_per_cpu {
442         int                             cpu;
443         atomic_t                        record_disabled;
444         struct trace_buffer     *buffer;
445         raw_spinlock_t                  reader_lock;    /* serialize readers */
446         arch_spinlock_t                 lock;
447         struct lock_class_key           lock_key;
448         struct buffer_data_page         *free_page;
449         unsigned long                   nr_pages;
450         unsigned int                    current_context;
451         struct list_head                *pages;
452         struct buffer_page              *head_page;     /* read from head */
453         struct buffer_page              *tail_page;     /* write to tail */
454         struct buffer_page              *commit_page;   /* committed pages */
455         struct buffer_page              *reader_page;
456         unsigned long                   lost_events;
457         unsigned long                   last_overrun;
458         unsigned long                   nest;
459         local_t                         entries_bytes;
460         local_t                         entries;
461         local_t                         overrun;
462         local_t                         commit_overrun;
463         local_t                         dropped_events;
464         local_t                         committing;
465         local_t                         commits;
466         local_t                         pages_touched;
467         local_t                         pages_read;
468         long                            last_pages_touch;
469         size_t                          shortest_full;
470         unsigned long                   read;
471         unsigned long                   read_bytes;
472         u64                             write_stamp;
473         u64                             read_stamp;
474         /* ring buffer pages to update, > 0 to add, < 0 to remove */
475         long                            nr_pages_to_update;
476         struct list_head                new_pages; /* new pages to add */
477         struct work_struct              update_pages_work;
478         struct completion               update_done;
479 
480         struct rb_irq_work              irq_work;
481 };
482 
483 struct trace_buffer {
484         unsigned                        flags;
485         int                             cpus;
486         atomic_t                        record_disabled;
487         atomic_t                        resize_disabled;
488         cpumask_var_t                   cpumask;
489 
490         struct lock_class_key           *reader_lock_key;
491 
492         struct mutex                    mutex;
493 
494         struct ring_buffer_per_cpu      **buffers;
495 
496         struct hlist_node               node;
497         u64                             (*clock)(void);
498 
499         struct rb_irq_work              irq_work;
500         bool                            time_stamp_abs;
501 };
502 
503 struct ring_buffer_iter {
504         struct ring_buffer_per_cpu      *cpu_buffer;
505         unsigned long                   head;
506         struct buffer_page              *head_page;
507         struct buffer_page              *cache_reader_page;
508         unsigned long                   cache_read;
509         u64                             read_stamp;
510 };
511 
512 /**
513  * ring_buffer_nr_pages - get the number of buffer pages in the ring buffer
514  * @buffer: The ring_buffer to get the number of pages from
515  * @cpu: The cpu of the ring_buffer to get the number of pages from
516  *
517  * Returns the number of pages used by a per_cpu buffer of the ring buffer.
518  */
519 size_t ring_buffer_nr_pages(struct trace_buffer *buffer, int cpu)
520 {
521         return buffer->buffers[cpu]->nr_pages;
522 }
523 
524 /**
525  * ring_buffer_nr_pages_dirty - get the number of used pages in the ring buffer
526  * @buffer: The ring_buffer to get the number of pages from
527  * @cpu: The cpu of the ring_buffer to get the number of pages from
528  *
529  * Returns the number of pages that have content in the ring buffer.
530  */
531 size_t ring_buffer_nr_dirty_pages(struct trace_buffer *buffer, int cpu)
532 {
533         size_t read;
534         size_t cnt;
535 
536         read = local_read(&buffer->buffers[cpu]->pages_read);
537         cnt = local_read(&buffer->buffers[cpu]->pages_touched);
538         /* The reader can read an empty page, but not more than that */
539         if (cnt < read) {
540                 WARN_ON_ONCE(read > cnt + 1);
541                 return 0;
542         }
543 
544         return cnt - read;
545 }
546 
547 /*
548  * rb_wake_up_waiters - wake up tasks waiting for ring buffer input
549  *
550  * Schedules a delayed work to wake up any task that is blocked on the
551  * ring buffer waiters queue.
552  */
553 static void rb_wake_up_waiters(struct irq_work *work)
554 {
555         struct rb_irq_work *rbwork = container_of(work, struct rb_irq_work, work);
556 
557         wake_up_all(&rbwork->waiters);
558         if (rbwork->wakeup_full) {
559                 rbwork->wakeup_full = false;
560                 wake_up_all(&rbwork->full_waiters);
561         }
562 }
563 
564 /**
565  * ring_buffer_wait - wait for input to the ring buffer
566  * @buffer: buffer to wait on
567  * @cpu: the cpu buffer to wait on
568  * @full: wait until a full page is available, if @cpu != RING_BUFFER_ALL_CPUS
569  *
570  * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
571  * as data is added to any of the @buffer's cpu buffers. Otherwise
572  * it will wait for data to be added to a specific cpu buffer.
573  */
574 int ring_buffer_wait(struct trace_buffer *buffer, int cpu, int full)
575 {
576         struct ring_buffer_per_cpu *uninitialized_var(cpu_buffer);
577         DEFINE_WAIT(wait);
578         struct rb_irq_work *work;
579         int ret = 0;
580 
581         /*
582          * Depending on what the caller is waiting for, either any
583          * data in any cpu buffer, or a specific buffer, put the
584          * caller on the appropriate wait queue.
585          */
586         if (cpu == RING_BUFFER_ALL_CPUS) {
587                 work = &buffer->irq_work;
588                 /* Full only makes sense on per cpu reads */
589                 full = 0;
590         } else {
591                 if (!cpumask_test_cpu(cpu, buffer->cpumask))
592                         return -ENODEV;
593                 cpu_buffer = buffer->buffers[cpu];
594                 work = &cpu_buffer->irq_work;
595         }
596 
597 
598         while (true) {
599                 if (full)
600                         prepare_to_wait(&work->full_waiters, &wait, TASK_INTERRUPTIBLE);
601                 else
602                         prepare_to_wait(&work->waiters, &wait, TASK_INTERRUPTIBLE);
603 
604                 /*
605                  * The events can happen in critical sections where
606                  * checking a work queue can cause deadlocks.
607                  * After adding a task to the queue, this flag is set
608                  * only to notify events to try to wake up the queue
609                  * using irq_work.
610                  *
611                  * We don't clear it even if the buffer is no longer
612                  * empty. The flag only causes the next event to run
613                  * irq_work to do the work queue wake up. The worse
614                  * that can happen if we race with !trace_empty() is that
615                  * an event will cause an irq_work to try to wake up
616                  * an empty queue.
617                  *
618                  * There's no reason to protect this flag either, as
619                  * the work queue and irq_work logic will do the necessary
620                  * synchronization for the wake ups. The only thing
621                  * that is necessary is that the wake up happens after
622                  * a task has been queued. It's OK for spurious wake ups.
623                  */
624                 if (full)
625                         work->full_waiters_pending = true;
626                 else
627                         work->waiters_pending = true;
628 
629                 if (signal_pending(current)) {
630                         ret = -EINTR;
631                         break;
632                 }
633 
634                 if (cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer))
635                         break;
636 
637                 if (cpu != RING_BUFFER_ALL_CPUS &&
638                     !ring_buffer_empty_cpu(buffer, cpu)) {
639                         unsigned long flags;
640                         bool pagebusy;
641                         size_t nr_pages;
642                         size_t dirty;
643 
644                         if (!full)
645                                 break;
646 
647                         raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
648                         pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
649                         nr_pages = cpu_buffer->nr_pages;
650                         dirty = ring_buffer_nr_dirty_pages(buffer, cpu);
651                         if (!cpu_buffer->shortest_full ||
652                             cpu_buffer->shortest_full < full)
653                                 cpu_buffer->shortest_full = full;
654                         raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
655                         if (!pagebusy &&
656                             (!nr_pages || (dirty * 100) > full * nr_pages))
657                                 break;
658                 }
659 
660                 schedule();
661         }
662 
663         if (full)
664                 finish_wait(&work->full_waiters, &wait);
665         else
666                 finish_wait(&work->waiters, &wait);
667 
668         return ret;
669 }
670 
671 /**
672  * ring_buffer_poll_wait - poll on buffer input
673  * @buffer: buffer to wait on
674  * @cpu: the cpu buffer to wait on
675  * @filp: the file descriptor
676  * @poll_table: The poll descriptor
677  *
678  * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
679  * as data is added to any of the @buffer's cpu buffers. Otherwise
680  * it will wait for data to be added to a specific cpu buffer.
681  *
682  * Returns EPOLLIN | EPOLLRDNORM if data exists in the buffers,
683  * zero otherwise.
684  */
685 __poll_t ring_buffer_poll_wait(struct trace_buffer *buffer, int cpu,
686                           struct file *filp, poll_table *poll_table)
687 {
688         struct ring_buffer_per_cpu *cpu_buffer;
689         struct rb_irq_work *work;
690 
691         if (cpu == RING_BUFFER_ALL_CPUS)
692                 work = &buffer->irq_work;
693         else {
694                 if (!cpumask_test_cpu(cpu, buffer->cpumask))
695                         return -EINVAL;
696 
697                 cpu_buffer = buffer->buffers[cpu];
698                 work = &cpu_buffer->irq_work;
699         }
700 
701         poll_wait(filp, &work->waiters, poll_table);
702         work->waiters_pending = true;
703         /*
704          * There's a tight race between setting the waiters_pending and
705          * checking if the ring buffer is empty.  Once the waiters_pending bit
706          * is set, the next event will wake the task up, but we can get stuck
707          * if there's only a single event in.
708          *
709          * FIXME: Ideally, we need a memory barrier on the writer side as well,
710          * but adding a memory barrier to all events will cause too much of a
711          * performance hit in the fast path.  We only need a memory barrier when
712          * the buffer goes from empty to having content.  But as this race is
713          * extremely small, and it's not a problem if another event comes in, we
714          * will fix it later.
715          */
716         smp_mb();
717 
718         if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
719             (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
720                 return EPOLLIN | EPOLLRDNORM;
721         return 0;
722 }
723 
724 /* buffer may be either ring_buffer or ring_buffer_per_cpu */
725 #define RB_WARN_ON(b, cond)                                             \
726         ({                                                              \
727                 int _____ret = unlikely(cond);                          \
728                 if (_____ret) {                                         \
729                         if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
730                                 struct ring_buffer_per_cpu *__b =       \
731                                         (void *)b;                      \
732                                 atomic_inc(&__b->buffer->record_disabled); \
733                         } else                                          \
734                                 atomic_inc(&b->record_disabled);        \
735                         WARN_ON(1);                                     \
736                 }                                                       \
737                 _____ret;                                               \
738         })
739 
740 /* Up this if you want to test the TIME_EXTENTS and normalization */
741 #define DEBUG_SHIFT 0
742 
743 static inline u64 rb_time_stamp(struct trace_buffer *buffer)
744 {
745         /* shift to debug/test normalization and TIME_EXTENTS */
746         return buffer->clock() << DEBUG_SHIFT;
747 }
748 
749 u64 ring_buffer_time_stamp(struct trace_buffer *buffer, int cpu)
750 {
751         u64 time;
752 
753         preempt_disable_notrace();
754         time = rb_time_stamp(buffer);
755         preempt_enable_notrace();
756 
757         return time;
758 }
759 EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
760 
761 void ring_buffer_normalize_time_stamp(struct trace_buffer *buffer,
762                                       int cpu, u64 *ts)
763 {
764         /* Just stupid testing the normalize function and deltas */
765         *ts >>= DEBUG_SHIFT;
766 }
767 EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
768 
769 /*
770  * Making the ring buffer lockless makes things tricky.
771  * Although writes only happen on the CPU that they are on,
772  * and they only need to worry about interrupts. Reads can
773  * happen on any CPU.
774  *
775  * The reader page is always off the ring buffer, but when the
776  * reader finishes with a page, it needs to swap its page with
777  * a new one from the buffer. The reader needs to take from
778  * the head (writes go to the tail). But if a writer is in overwrite
779  * mode and wraps, it must push the head page forward.
780  *
781  * Here lies the problem.
782  *
783  * The reader must be careful to replace only the head page, and
784  * not another one. As described at the top of the file in the
785  * ASCII art, the reader sets its old page to point to the next
786  * page after head. It then sets the page after head to point to
787  * the old reader page. But if the writer moves the head page
788  * during this operation, the reader could end up with the tail.
789  *
790  * We use cmpxchg to help prevent this race. We also do something
791  * special with the page before head. We set the LSB to 1.
792  *
793  * When the writer must push the page forward, it will clear the
794  * bit that points to the head page, move the head, and then set
795  * the bit that points to the new head page.
796  *
797  * We also don't want an interrupt coming in and moving the head
798  * page on another writer. Thus we use the second LSB to catch
799  * that too. Thus:
800  *
801  * head->list->prev->next        bit 1          bit 0
802  *                              -------        -------
803  * Normal page                     0              0
804  * Points to head page             0              1
805  * New head page                   1              0
806  *
807  * Note we can not trust the prev pointer of the head page, because:
808  *
809  * +----+       +-----+        +-----+
810  * |    |------>|  T  |---X--->|  N  |
811  * |    |<------|     |        |     |
812  * +----+       +-----+        +-----+
813  *   ^                           ^ |
814  *   |          +-----+          | |
815  *   +----------|  R  |----------+ |
816  *              |     |<-----------+
817  *              +-----+
818  *
819  * Key:  ---X-->  HEAD flag set in pointer
820  *         T      Tail page
821  *         R      Reader page
822  *         N      Next page
823  *
824  * (see __rb_reserve_next() to see where this happens)
825  *
826  *  What the above shows is that the reader just swapped out
827  *  the reader page with a page in the buffer, but before it
828  *  could make the new header point back to the new page added
829  *  it was preempted by a writer. The writer moved forward onto
830  *  the new page added by the reader and is about to move forward
831  *  again.
832  *
833  *  You can see, it is legitimate for the previous pointer of
834  *  the head (or any page) not to point back to itself. But only
835  *  temporarily.
836  */
837 
838 #define RB_PAGE_NORMAL          0UL
839 #define RB_PAGE_HEAD            1UL
840 #define RB_PAGE_UPDATE          2UL
841 
842 
843 #define RB_FLAG_MASK            3UL
844 
845 /* PAGE_MOVED is not part of the mask */
846 #define RB_PAGE_MOVED           4UL
847 
848 /*
849  * rb_list_head - remove any bit
850  */
851 static struct list_head *rb_list_head(struct list_head *list)
852 {
853         unsigned long val = (unsigned long)list;
854 
855         return (struct list_head *)(val & ~RB_FLAG_MASK);
856 }
857 
858 /*
859  * rb_is_head_page - test if the given page is the head page
860  *
861  * Because the reader may move the head_page pointer, we can
862  * not trust what the head page is (it may be pointing to
863  * the reader page). But if the next page is a header page,
864  * its flags will be non zero.
865  */
866 static inline int
867 rb_is_head_page(struct ring_buffer_per_cpu *cpu_buffer,
868                 struct buffer_page *page, struct list_head *list)
869 {
870         unsigned long val;
871 
872         val = (unsigned long)list->next;
873 
874         if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
875                 return RB_PAGE_MOVED;
876 
877         return val & RB_FLAG_MASK;
878 }
879 
880 /*
881  * rb_is_reader_page
882  *
883  * The unique thing about the reader page, is that, if the
884  * writer is ever on it, the previous pointer never points
885  * back to the reader page.
886  */
887 static bool rb_is_reader_page(struct buffer_page *page)
888 {
889         struct list_head *list = page->list.prev;
890 
891         return rb_list_head(list->next) != &page->list;
892 }
893 
894 /*
895  * rb_set_list_to_head - set a list_head to be pointing to head.
896  */
897 static void rb_set_list_to_head(struct ring_buffer_per_cpu *cpu_buffer,
898                                 struct list_head *list)
899 {
900         unsigned long *ptr;
901 
902         ptr = (unsigned long *)&list->next;
903         *ptr |= RB_PAGE_HEAD;
904         *ptr &= ~RB_PAGE_UPDATE;
905 }
906 
907 /*
908  * rb_head_page_activate - sets up head page
909  */
910 static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
911 {
912         struct buffer_page *head;
913 
914         head = cpu_buffer->head_page;
915         if (!head)
916                 return;
917 
918         /*
919          * Set the previous list pointer to have the HEAD flag.
920          */
921         rb_set_list_to_head(cpu_buffer, head->list.prev);
922 }
923 
924 static void rb_list_head_clear(struct list_head *list)
925 {
926         unsigned long *ptr = (unsigned long *)&list->next;
927 
928         *ptr &= ~RB_FLAG_MASK;
929 }
930 
931 /*
932  * rb_head_page_deactivate - clears head page ptr (for free list)
933  */
934 static void
935 rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
936 {
937         struct list_head *hd;
938 
939         /* Go through the whole list and clear any pointers found. */
940         rb_list_head_clear(cpu_buffer->pages);
941 
942         list_for_each(hd, cpu_buffer->pages)
943                 rb_list_head_clear(hd);
944 }
945 
946 static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
947                             struct buffer_page *head,
948                             struct buffer_page *prev,
949                             int old_flag, int new_flag)
950 {
951         struct list_head *list;
952         unsigned long val = (unsigned long)&head->list;
953         unsigned long ret;
954 
955         list = &prev->list;
956 
957         val &= ~RB_FLAG_MASK;
958 
959         ret = cmpxchg((unsigned long *)&list->next,
960                       val | old_flag, val | new_flag);
961 
962         /* check if the reader took the page */
963         if ((ret & ~RB_FLAG_MASK) != val)
964                 return RB_PAGE_MOVED;
965 
966         return ret & RB_FLAG_MASK;
967 }
968 
969 static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
970                                    struct buffer_page *head,
971                                    struct buffer_page *prev,
972                                    int old_flag)
973 {
974         return rb_head_page_set(cpu_buffer, head, prev,
975                                 old_flag, RB_PAGE_UPDATE);
976 }
977 
978 static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
979                                  struct buffer_page *head,
980                                  struct buffer_page *prev,
981                                  int old_flag)
982 {
983         return rb_head_page_set(cpu_buffer, head, prev,
984                                 old_flag, RB_PAGE_HEAD);
985 }
986 
987 static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
988                                    struct buffer_page *head,
989                                    struct buffer_page *prev,
990                                    int old_flag)
991 {
992         return rb_head_page_set(cpu_buffer, head, prev,
993                                 old_flag, RB_PAGE_NORMAL);
994 }
995 
996 static inline void rb_inc_page(struct ring_buffer_per_cpu *cpu_buffer,
997                                struct buffer_page **bpage)
998 {
999         struct list_head *p = rb_list_head((*bpage)->list.next);
1000 
1001         *bpage = list_entry(p, struct buffer_page, list);
1002 }
1003 
1004 static struct buffer_page *
1005 rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
1006 {
1007         struct buffer_page *head;
1008         struct buffer_page *page;
1009         struct list_head *list;
1010         int i;
1011 
1012         if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
1013                 return NULL;
1014 
1015         /* sanity check */
1016         list = cpu_buffer->pages;
1017         if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
1018                 return NULL;
1019 
1020         page = head = cpu_buffer->head_page;
1021         /*
1022          * It is possible that the writer moves the header behind
1023          * where we started, and we miss in one loop.
1024          * A second loop should grab the header, but we'll do
1025          * three loops just because I'm paranoid.
1026          */
1027         for (i = 0; i < 3; i++) {
1028                 do {
1029                         if (rb_is_head_page(cpu_buffer, page, page->list.prev)) {
1030                                 cpu_buffer->head_page = page;
1031                                 return page;
1032                         }
1033                         rb_inc_page(cpu_buffer, &page);
1034                 } while (page != head);
1035         }
1036 
1037         RB_WARN_ON(cpu_buffer, 1);
1038 
1039         return NULL;
1040 }
1041 
1042 static int rb_head_page_replace(struct buffer_page *old,
1043                                 struct buffer_page *new)
1044 {
1045         unsigned long *ptr = (unsigned long *)&old->list.prev->next;
1046         unsigned long val;
1047         unsigned long ret;
1048 
1049         val = *ptr & ~RB_FLAG_MASK;
1050         val |= RB_PAGE_HEAD;
1051 
1052         ret = cmpxchg(ptr, val, (unsigned long)&new->list);
1053 
1054         return ret == val;
1055 }
1056 
1057 /*
1058  * rb_tail_page_update - move the tail page forward
1059  */
1060 static void rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
1061                                struct buffer_page *tail_page,
1062                                struct buffer_page *next_page)
1063 {
1064         unsigned long old_entries;
1065         unsigned long old_write;
1066 
1067         /*
1068          * The tail page now needs to be moved forward.
1069          *
1070          * We need to reset the tail page, but without messing
1071          * with possible erasing of data brought in by interrupts
1072          * that have moved the tail page and are currently on it.
1073          *
1074          * We add a counter to the write field to denote this.
1075          */
1076         old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
1077         old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
1078 
1079         local_inc(&cpu_buffer->pages_touched);
1080         /*
1081          * Just make sure we have seen our old_write and synchronize
1082          * with any interrupts that come in.
1083          */
1084         barrier();
1085 
1086         /*
1087          * If the tail page is still the same as what we think
1088          * it is, then it is up to us to update the tail
1089          * pointer.
1090          */
1091         if (tail_page == READ_ONCE(cpu_buffer->tail_page)) {
1092                 /* Zero the write counter */
1093                 unsigned long val = old_write & ~RB_WRITE_MASK;
1094                 unsigned long eval = old_entries & ~RB_WRITE_MASK;
1095 
1096                 /*
1097                  * This will only succeed if an interrupt did
1098                  * not come in and change it. In which case, we
1099                  * do not want to modify it.
1100                  *
1101                  * We add (void) to let the compiler know that we do not care
1102                  * about the return value of these functions. We use the
1103                  * cmpxchg to only update if an interrupt did not already
1104                  * do it for us. If the cmpxchg fails, we don't care.
1105                  */
1106                 (void)local_cmpxchg(&next_page->write, old_write, val);
1107                 (void)local_cmpxchg(&next_page->entries, old_entries, eval);
1108 
1109                 /*
1110                  * No need to worry about races with clearing out the commit.
1111                  * it only can increment when a commit takes place. But that
1112                  * only happens in the outer most nested commit.
1113                  */
1114                 local_set(&next_page->page->commit, 0);
1115 
1116                 /* Again, either we update tail_page or an interrupt does */
1117                 (void)cmpxchg(&cpu_buffer->tail_page, tail_page, next_page);
1118         }
1119 }
1120 
1121 static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
1122                           struct buffer_page *bpage)
1123 {
1124         unsigned long val = (unsigned long)bpage;
1125 
1126         if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK))
1127                 return 1;
1128 
1129         return 0;
1130 }
1131 
1132 /**
1133  * rb_check_list - make sure a pointer to a list has the last bits zero
1134  */
1135 static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer,
1136                          struct list_head *list)
1137 {
1138         if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev))
1139                 return 1;
1140         if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next))
1141                 return 1;
1142         return 0;
1143 }
1144 
1145 /**
1146  * rb_check_pages - integrity check of buffer pages
1147  * @cpu_buffer: CPU buffer with pages to test
1148  *
1149  * As a safety measure we check to make sure the data pages have not
1150  * been corrupted.
1151  */
1152 static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
1153 {
1154         struct list_head *head = cpu_buffer->pages;
1155         struct buffer_page *bpage, *tmp;
1156 
1157         /* Reset the head page if it exists */
1158         if (cpu_buffer->head_page)
1159                 rb_set_head_page(cpu_buffer);
1160 
1161         rb_head_page_deactivate(cpu_buffer);
1162 
1163         if (RB_WARN_ON(cpu_buffer, head->next->prev != head))
1164                 return -1;
1165         if (RB_WARN_ON(cpu_buffer, head->prev->next != head))
1166                 return -1;
1167 
1168         if (rb_check_list(cpu_buffer, head))
1169                 return -1;
1170 
1171         list_for_each_entry_safe(bpage, tmp, head, list) {
1172                 if (RB_WARN_ON(cpu_buffer,
1173                                bpage->list.next->prev != &bpage->list))
1174                         return -1;
1175                 if (RB_WARN_ON(cpu_buffer,
1176                                bpage->list.prev->next != &bpage->list))
1177                         return -1;
1178                 if (rb_check_list(cpu_buffer, &bpage->list))
1179                         return -1;
1180         }
1181 
1182         rb_head_page_activate(cpu_buffer);
1183 
1184         return 0;
1185 }
1186 
1187 static int __rb_allocate_pages(long nr_pages, struct list_head *pages, int cpu)
1188 {
1189         struct buffer_page *bpage, *tmp;
1190         bool user_thread = current->mm != NULL;
1191         gfp_t mflags;
1192         long i;
1193 
1194         /*
1195          * Check if the available memory is there first.
1196          * Note, si_mem_available() only gives us a rough estimate of available
1197          * memory. It may not be accurate. But we don't care, we just want
1198          * to prevent doing any allocation when it is obvious that it is
1199          * not going to succeed.
1200          */
1201         i = si_mem_available();
1202         if (i < nr_pages)
1203                 return -ENOMEM;
1204 
1205         /*
1206          * __GFP_RETRY_MAYFAIL flag makes sure that the allocation fails
1207          * gracefully without invoking oom-killer and the system is not
1208          * destabilized.
1209          */
1210         mflags = GFP_KERNEL | __GFP_RETRY_MAYFAIL;
1211 
1212         /*
1213          * If a user thread allocates too much, and si_mem_available()
1214          * reports there's enough memory, even though there is not.
1215          * Make sure the OOM killer kills this thread. This can happen
1216          * even with RETRY_MAYFAIL because another task may be doing
1217          * an allocation after this task has taken all memory.
1218          * This is the task the OOM killer needs to take out during this
1219          * loop, even if it was triggered by an allocation somewhere else.
1220          */
1221         if (user_thread)
1222                 set_current_oom_origin();
1223         for (i = 0; i < nr_pages; i++) {
1224                 struct page *page;
1225 
1226                 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1227                                     mflags, cpu_to_node(cpu));
1228                 if (!bpage)
1229                         goto free_pages;
1230 
1231                 list_add(&bpage->list, pages);
1232 
1233                 page = alloc_pages_node(cpu_to_node(cpu), mflags, 0);
1234                 if (!page)
1235                         goto free_pages;
1236                 bpage->page = page_address(page);
1237                 rb_init_page(bpage->page);
1238 
1239                 if (user_thread && fatal_signal_pending(current))
1240                         goto free_pages;
1241         }
1242         if (user_thread)
1243                 clear_current_oom_origin();
1244 
1245         return 0;
1246 
1247 free_pages:
1248         list_for_each_entry_safe(bpage, tmp, pages, list) {
1249                 list_del_init(&bpage->list);
1250                 free_buffer_page(bpage);
1251         }
1252         if (user_thread)
1253                 clear_current_oom_origin();
1254 
1255         return -ENOMEM;
1256 }
1257 
1258 static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1259                              unsigned long nr_pages)
1260 {
1261         LIST_HEAD(pages);
1262 
1263         WARN_ON(!nr_pages);
1264 
1265         if (__rb_allocate_pages(nr_pages, &pages, cpu_buffer->cpu))
1266                 return -ENOMEM;
1267 
1268         /*
1269          * The ring buffer page list is a circular list that does not
1270          * start and end with a list head. All page list items point to
1271          * other pages.
1272          */
1273         cpu_buffer->pages = pages.next;
1274         list_del(&pages);
1275 
1276         cpu_buffer->nr_pages = nr_pages;
1277 
1278         rb_check_pages(cpu_buffer);
1279 
1280         return 0;
1281 }
1282 
1283 static struct ring_buffer_per_cpu *
1284 rb_allocate_cpu_buffer(struct trace_buffer *buffer, long nr_pages, int cpu)
1285 {
1286         struct ring_buffer_per_cpu *cpu_buffer;
1287         struct buffer_page *bpage;
1288         struct page *page;
1289         int ret;
1290 
1291         cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1292                                   GFP_KERNEL, cpu_to_node(cpu));
1293         if (!cpu_buffer)
1294                 return NULL;
1295 
1296         cpu_buffer->cpu = cpu;
1297         cpu_buffer->buffer = buffer;
1298         raw_spin_lock_init(&cpu_buffer->reader_lock);
1299         lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1300         cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1301         INIT_WORK(&cpu_buffer->update_pages_work, update_pages_handler);
1302         init_completion(&cpu_buffer->update_done);
1303         init_irq_work(&cpu_buffer->irq_work.work, rb_wake_up_waiters);
1304         init_waitqueue_head(&cpu_buffer->irq_work.waiters);
1305         init_waitqueue_head(&cpu_buffer->irq_work.full_waiters);
1306 
1307         bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1308                             GFP_KERNEL, cpu_to_node(cpu));
1309         if (!bpage)
1310                 goto fail_free_buffer;
1311 
1312         rb_check_bpage(cpu_buffer, bpage);
1313 
1314         cpu_buffer->reader_page = bpage;
1315         page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0);
1316         if (!page)
1317                 goto fail_free_reader;
1318         bpage->page = page_address(page);
1319         rb_init_page(bpage->page);
1320 
1321         INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1322         INIT_LIST_HEAD(&cpu_buffer->new_pages);
1323 
1324         ret = rb_allocate_pages(cpu_buffer, nr_pages);
1325         if (ret < 0)
1326                 goto fail_free_reader;
1327 
1328         cpu_buffer->head_page
1329                 = list_entry(cpu_buffer->pages, struct buffer_page, list);
1330         cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1331 
1332         rb_head_page_activate(cpu_buffer);
1333 
1334         return cpu_buffer;
1335 
1336  fail_free_reader:
1337         free_buffer_page(cpu_buffer->reader_page);
1338 
1339  fail_free_buffer:
1340         kfree(cpu_buffer);
1341         return NULL;
1342 }
1343 
1344 static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1345 {
1346         struct list_head *head = cpu_buffer->pages;
1347         struct buffer_page *bpage, *tmp;
1348 
1349         free_buffer_page(cpu_buffer->reader_page);
1350 
1351         rb_head_page_deactivate(cpu_buffer);
1352 
1353         if (head) {
1354                 list_for_each_entry_safe(bpage, tmp, head, list) {
1355                         list_del_init(&bpage->list);
1356                         free_buffer_page(bpage);
1357                 }
1358                 bpage = list_entry(head, struct buffer_page, list);
1359                 free_buffer_page(bpage);
1360         }
1361 
1362         kfree(cpu_buffer);
1363 }
1364 
1365 /**
1366  * __ring_buffer_alloc - allocate a new ring_buffer
1367  * @size: the size in bytes per cpu that is needed.
1368  * @flags: attributes to set for the ring buffer.
1369  * @key: ring buffer reader_lock_key.
1370  *
1371  * Currently the only flag that is available is the RB_FL_OVERWRITE
1372  * flag. This flag means that the buffer will overwrite old data
1373  * when the buffer wraps. If this flag is not set, the buffer will
1374  * drop data when the tail hits the head.
1375  */
1376 struct trace_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1377                                         struct lock_class_key *key)
1378 {
1379         struct trace_buffer *buffer;
1380         long nr_pages;
1381         int bsize;
1382         int cpu;
1383         int ret;
1384 
1385         /* keep it in its own cache line */
1386         buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1387                          GFP_KERNEL);
1388         if (!buffer)
1389                 return NULL;
1390 
1391         if (!zalloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1392                 goto fail_free_buffer;
1393 
1394         nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1395         buffer->flags = flags;
1396         buffer->clock = trace_clock_local;
1397         buffer->reader_lock_key = key;
1398 
1399         init_irq_work(&buffer->irq_work.work, rb_wake_up_waiters);
1400         init_waitqueue_head(&buffer->irq_work.waiters);
1401 
1402         /* need at least two pages */
1403         if (nr_pages < 2)
1404                 nr_pages = 2;
1405 
1406         buffer->cpus = nr_cpu_ids;
1407 
1408         bsize = sizeof(void *) * nr_cpu_ids;
1409         buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1410                                   GFP_KERNEL);
1411         if (!buffer->buffers)
1412                 goto fail_free_cpumask;
1413 
1414         cpu = raw_smp_processor_id();
1415         cpumask_set_cpu(cpu, buffer->cpumask);
1416         buffer->buffers[cpu] = rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
1417         if (!buffer->buffers[cpu])
1418                 goto fail_free_buffers;
1419 
1420         ret = cpuhp_state_add_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1421         if (ret < 0)
1422                 goto fail_free_buffers;
1423 
1424         mutex_init(&buffer->mutex);
1425 
1426         return buffer;
1427 
1428  fail_free_buffers:
1429         for_each_buffer_cpu(buffer, cpu) {
1430                 if (buffer->buffers[cpu])
1431                         rb_free_cpu_buffer(buffer->buffers[cpu]);
1432         }
1433         kfree(buffer->buffers);
1434 
1435  fail_free_cpumask:
1436         free_cpumask_var(buffer->cpumask);
1437 
1438  fail_free_buffer:
1439         kfree(buffer);
1440         return NULL;
1441 }
1442 EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1443 
1444 /**
1445  * ring_buffer_free - free a ring buffer.
1446  * @buffer: the buffer to free.
1447  */
1448 void
1449 ring_buffer_free(struct trace_buffer *buffer)
1450 {
1451         int cpu;
1452 
1453         cpuhp_state_remove_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1454 
1455         for_each_buffer_cpu(buffer, cpu)
1456                 rb_free_cpu_buffer(buffer->buffers[cpu]);
1457 
1458         kfree(buffer->buffers);
1459         free_cpumask_var(buffer->cpumask);
1460 
1461         kfree(buffer);
1462 }
1463 EXPORT_SYMBOL_GPL(ring_buffer_free);
1464 
1465 void ring_buffer_set_clock(struct trace_buffer *buffer,
1466                            u64 (*clock)(void))
1467 {
1468         buffer->clock = clock;
1469 }
1470 
1471 void ring_buffer_set_time_stamp_abs(struct trace_buffer *buffer, bool abs)
1472 {
1473         buffer->time_stamp_abs = abs;
1474 }
1475 
1476 bool ring_buffer_time_stamp_abs(struct trace_buffer *buffer)
1477 {
1478         return buffer->time_stamp_abs;
1479 }
1480 
1481 static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1482 
1483 static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1484 {
1485         return local_read(&bpage->entries) & RB_WRITE_MASK;
1486 }
1487 
1488 static inline unsigned long rb_page_write(struct buffer_page *bpage)
1489 {
1490         return local_read(&bpage->write) & RB_WRITE_MASK;
1491 }
1492 
1493 static int
1494 rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned long nr_pages)
1495 {
1496         struct list_head *tail_page, *to_remove, *next_page;
1497         struct buffer_page *to_remove_page, *tmp_iter_page;
1498         struct buffer_page *last_page, *first_page;
1499         unsigned long nr_removed;
1500         unsigned long head_bit;
1501         int page_entries;
1502 
1503         head_bit = 0;
1504 
1505         raw_spin_lock_irq(&cpu_buffer->reader_lock);
1506         atomic_inc(&cpu_buffer->record_disabled);
1507         /*
1508          * We don't race with the readers since we have acquired the reader
1509          * lock. We also don't race with writers after disabling recording.
1510          * This makes it easy to figure out the first and the last page to be
1511          * removed from the list. We unlink all the pages in between including
1512          * the first and last pages. This is done in a busy loop so that we
1513          * lose the least number of traces.
1514          * The pages are freed after we restart recording and unlock readers.
1515          */
1516         tail_page = &cpu_buffer->tail_page->list;
1517 
1518         /*
1519          * tail page might be on reader page, we remove the next page
1520          * from the ring buffer
1521          */
1522         if (cpu_buffer->tail_page == cpu_buffer->reader_page)
1523                 tail_page = rb_list_head(tail_page->next);
1524         to_remove = tail_page;
1525 
1526         /* start of pages to remove */
1527         first_page = list_entry(rb_list_head(to_remove->next),
1528                                 struct buffer_page, list);
1529 
1530         for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
1531                 to_remove = rb_list_head(to_remove)->next;
1532                 head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
1533         }
1534 
1535         next_page = rb_list_head(to_remove)->next;
1536 
1537         /*
1538          * Now we remove all pages between tail_page and next_page.
1539          * Make sure that we have head_bit value preserved for the
1540          * next page
1541          */
1542         tail_page->next = (struct list_head *)((unsigned long)next_page |
1543                                                 head_bit);
1544         next_page = rb_list_head(next_page);
1545         next_page->prev = tail_page;
1546 
1547         /* make sure pages points to a valid page in the ring buffer */
1548         cpu_buffer->pages = next_page;
1549 
1550         /* update head page */
1551         if (head_bit)
1552                 cpu_buffer->head_page = list_entry(next_page,
1553                                                 struct buffer_page, list);
1554 
1555         /*
1556          * change read pointer to make sure any read iterators reset
1557          * themselves
1558          */
1559         cpu_buffer->read = 0;
1560 
1561         /* pages are removed, resume tracing and then free the pages */
1562         atomic_dec(&cpu_buffer->record_disabled);
1563         raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1564 
1565         RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
1566 
1567         /* last buffer page to remove */
1568         last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
1569                                 list);
1570         tmp_iter_page = first_page;
1571 
1572         do {
1573                 cond_resched();
1574 
1575                 to_remove_page = tmp_iter_page;
1576                 rb_inc_page(cpu_buffer, &tmp_iter_page);
1577 
1578                 /* update the counters */
1579                 page_entries = rb_page_entries(to_remove_page);
1580                 if (page_entries) {
1581                         /*
1582                          * If something was added to this page, it was full
1583                          * since it is not the tail page. So we deduct the
1584                          * bytes consumed in ring buffer from here.
1585                          * Increment overrun to account for the lost events.
1586                          */
1587                         local_add(page_entries, &cpu_buffer->overrun);
1588                         local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1589                 }
1590 
1591                 /*
1592                  * We have already removed references to this list item, just
1593                  * free up the buffer_page and its page
1594                  */
1595                 free_buffer_page(to_remove_page);
1596                 nr_removed--;
1597 
1598         } while (to_remove_page != last_page);
1599 
1600         RB_WARN_ON(cpu_buffer, nr_removed);
1601 
1602         return nr_removed == 0;
1603 }
1604 
1605 static int
1606 rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
1607 {
1608         struct list_head *pages = &cpu_buffer->new_pages;
1609         int retries, success;
1610 
1611         raw_spin_lock_irq(&cpu_buffer->reader_lock);
1612         /*
1613          * We are holding the reader lock, so the reader page won't be swapped
1614          * in the ring buffer. Now we are racing with the writer trying to
1615          * move head page and the tail page.
1616          * We are going to adapt the reader page update process where:
1617          * 1. We first splice the start and end of list of new pages between
1618          *    the head page and its previous page.
1619          * 2. We cmpxchg the prev_page->next to point from head page to the
1620          *    start of new pages list.
1621          * 3. Finally, we update the head->prev to the end of new list.
1622          *
1623          * We will try this process 10 times, to make sure that we don't keep
1624          * spinning.
1625          */
1626         retries = 10;
1627         success = 0;
1628         while (retries--) {
1629                 struct list_head *head_page, *prev_page, *r;
1630                 struct list_head *last_page, *first_page;
1631                 struct list_head *head_page_with_bit;
1632 
1633                 head_page = &rb_set_head_page(cpu_buffer)->list;
1634                 if (!head_page)
1635                         break;
1636                 prev_page = head_page->prev;
1637 
1638                 first_page = pages->next;
1639                 last_page  = pages->prev;
1640 
1641                 head_page_with_bit = (struct list_head *)
1642                                      ((unsigned long)head_page | RB_PAGE_HEAD);
1643 
1644                 last_page->next = head_page_with_bit;
1645                 first_page->prev = prev_page;
1646 
1647                 r = cmpxchg(&prev_page->next, head_page_with_bit, first_page);
1648 
1649                 if (r == head_page_with_bit) {
1650                         /*
1651                          * yay, we replaced the page pointer to our new list,
1652                          * now, we just have to update to head page's prev
1653                          * pointer to point to end of list
1654                          */
1655                         head_page->prev = last_page;
1656                         success = 1;
1657                         break;
1658                 }
1659         }
1660 
1661         if (success)
1662                 INIT_LIST_HEAD(pages);
1663         /*
1664          * If we weren't successful in adding in new pages, warn and stop
1665          * tracing
1666          */
1667         RB_WARN_ON(cpu_buffer, !success);
1668         raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1669 
1670         /* free pages if they weren't inserted */
1671         if (!success) {
1672                 struct buffer_page *bpage, *tmp;
1673                 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1674                                          list) {
1675                         list_del_init(&bpage->list);
1676                         free_buffer_page(bpage);
1677                 }
1678         }
1679         return success;
1680 }
1681 
1682 static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
1683 {
1684         int success;
1685 
1686         if (cpu_buffer->nr_pages_to_update > 0)
1687                 success = rb_insert_pages(cpu_buffer);
1688         else
1689                 success = rb_remove_pages(cpu_buffer,
1690                                         -cpu_buffer->nr_pages_to_update);
1691 
1692         if (success)
1693                 cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
1694 }
1695 
1696 static void update_pages_handler(struct work_struct *work)
1697 {
1698         struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
1699                         struct ring_buffer_per_cpu, update_pages_work);
1700         rb_update_pages(cpu_buffer);
1701         complete(&cpu_buffer->update_done);
1702 }
1703 
1704 /**
1705  * ring_buffer_resize - resize the ring buffer
1706  * @buffer: the buffer to resize.
1707  * @size: the new size.
1708  * @cpu_id: the cpu buffer to resize
1709  *
1710  * Minimum size is 2 * BUF_PAGE_SIZE.
1711  *
1712  * Returns 0 on success and < 0 on failure.
1713  */
1714 int ring_buffer_resize(struct trace_buffer *buffer, unsigned long size,
1715                         int cpu_id)
1716 {
1717         struct ring_buffer_per_cpu *cpu_buffer;
1718         unsigned long nr_pages;
1719         int cpu, err = 0;
1720 
1721         /*
1722          * Always succeed at resizing a non-existent buffer:
1723          */
1724         if (!buffer)
1725                 return size;
1726 
1727         /* Make sure the requested buffer exists */
1728         if (cpu_id != RING_BUFFER_ALL_CPUS &&
1729             !cpumask_test_cpu(cpu_id, buffer->cpumask))
1730                 return size;
1731 
1732         nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1733 
1734         /* we need a minimum of two pages */
1735         if (nr_pages < 2)
1736                 nr_pages = 2;
1737 
1738         size = nr_pages * BUF_PAGE_SIZE;
1739 
1740         /*
1741          * Don't succeed if resizing is disabled, as a reader might be
1742          * manipulating the ring buffer and is expecting a sane state while
1743          * this is true.
1744          */
1745         if (atomic_read(&buffer->resize_disabled))
1746                 return -EBUSY;
1747 
1748         /* prevent another thread from changing buffer sizes */
1749         mutex_lock(&buffer->mutex);
1750 
1751         if (cpu_id == RING_BUFFER_ALL_CPUS) {
1752                 /* calculate the pages to update */
1753                 for_each_buffer_cpu(buffer, cpu) {
1754                         cpu_buffer = buffer->buffers[cpu];
1755 
1756                         cpu_buffer->nr_pages_to_update = nr_pages -
1757                                                         cpu_buffer->nr_pages;
1758                         /*
1759                          * nothing more to do for removing pages or no update
1760                          */
1761                         if (cpu_buffer->nr_pages_to_update <= 0)
1762                                 continue;
1763                         /*
1764                          * to add pages, make sure all new pages can be
1765                          * allocated without receiving ENOMEM
1766                          */
1767                         INIT_LIST_HEAD(&cpu_buffer->new_pages);
1768                         if (__rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1769                                                 &cpu_buffer->new_pages, cpu)) {
1770                                 /* not enough memory for new pages */
1771                                 err = -ENOMEM;
1772                                 goto out_err;
1773                         }
1774                 }
1775 
1776                 get_online_cpus();
1777                 /*
1778                  * Fire off all the required work handlers
1779                  * We can't schedule on offline CPUs, but it's not necessary
1780                  * since we can change their buffer sizes without any race.
1781                  */
1782                 for_each_buffer_cpu(buffer, cpu) {
1783                         cpu_buffer = buffer->buffers[cpu];
1784                         if (!cpu_buffer->nr_pages_to_update)
1785                                 continue;
1786 
1787                         /* Can't run something on an offline CPU. */
1788                         if (!cpu_online(cpu)) {
1789                                 rb_update_pages(cpu_buffer);
1790                                 cpu_buffer->nr_pages_to_update = 0;
1791                         } else {
1792                                 schedule_work_on(cpu,
1793                                                 &cpu_buffer->update_pages_work);
1794                         }
1795                 }
1796 
1797                 /* wait for all the updates to complete */
1798                 for_each_buffer_cpu(buffer, cpu) {
1799                         cpu_buffer = buffer->buffers[cpu];
1800                         if (!cpu_buffer->nr_pages_to_update)
1801                                 continue;
1802 
1803                         if (cpu_online(cpu))
1804                                 wait_for_completion(&cpu_buffer->update_done);
1805                         cpu_buffer->nr_pages_to_update = 0;
1806                 }
1807 
1808                 put_online_cpus();
1809         } else {
1810                 /* Make sure this CPU has been initialized */
1811                 if (!cpumask_test_cpu(cpu_id, buffer->cpumask))
1812                         goto out;
1813 
1814                 cpu_buffer = buffer->buffers[cpu_id];
1815 
1816                 if (nr_pages == cpu_buffer->nr_pages)
1817                         goto out;
1818 
1819                 cpu_buffer->nr_pages_to_update = nr_pages -
1820                                                 cpu_buffer->nr_pages;
1821 
1822                 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1823                 if (cpu_buffer->nr_pages_to_update > 0 &&
1824                         __rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1825                                             &cpu_buffer->new_pages, cpu_id)) {
1826                         err = -ENOMEM;
1827                         goto out_err;
1828                 }
1829 
1830                 get_online_cpus();
1831 
1832                 /* Can't run something on an offline CPU. */
1833                 if (!cpu_online(cpu_id))
1834                         rb_update_pages(cpu_buffer);
1835                 else {
1836                         schedule_work_on(cpu_id,
1837                                          &cpu_buffer->update_pages_work);
1838                         wait_for_completion(&cpu_buffer->update_done);
1839                 }
1840 
1841                 cpu_buffer->nr_pages_to_update = 0;
1842                 put_online_cpus();
1843         }
1844 
1845  out:
1846         /*
1847          * The ring buffer resize can happen with the ring buffer
1848          * enabled, so that the update disturbs the tracing as little
1849          * as possible. But if the buffer is disabled, we do not need
1850          * to worry about that, and we can take the time to verify
1851          * that the buffer is not corrupt.
1852          */
1853         if (atomic_read(&buffer->record_disabled)) {
1854                 atomic_inc(&buffer->record_disabled);
1855                 /*
1856                  * Even though the buffer was disabled, we must make sure
1857                  * that it is truly disabled before calling rb_check_pages.
1858                  * There could have been a race between checking
1859                  * record_disable and incrementing it.
1860                  */
1861                 synchronize_rcu();
1862                 for_each_buffer_cpu(buffer, cpu) {
1863                         cpu_buffer = buffer->buffers[cpu];
1864                         rb_check_pages(cpu_buffer);
1865                 }
1866                 atomic_dec(&buffer->record_disabled);
1867         }
1868 
1869         mutex_unlock(&buffer->mutex);
1870         return size;
1871 
1872  out_err:
1873         for_each_buffer_cpu(buffer, cpu) {
1874                 struct buffer_page *bpage, *tmp;
1875 
1876                 cpu_buffer = buffer->buffers[cpu];
1877                 cpu_buffer->nr_pages_to_update = 0;
1878 
1879                 if (list_empty(&cpu_buffer->new_pages))
1880                         continue;
1881 
1882                 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1883                                         list) {
1884                         list_del_init(&bpage->list);
1885                         free_buffer_page(bpage);
1886                 }
1887         }
1888         mutex_unlock(&buffer->mutex);
1889         return err;
1890 }
1891 EXPORT_SYMBOL_GPL(ring_buffer_resize);
1892 
1893 void ring_buffer_change_overwrite(struct trace_buffer *buffer, int val)
1894 {
1895         mutex_lock(&buffer->mutex);
1896         if (val)
1897                 buffer->flags |= RB_FL_OVERWRITE;
1898         else
1899                 buffer->flags &= ~RB_FL_OVERWRITE;
1900         mutex_unlock(&buffer->mutex);
1901 }
1902 EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
1903 
1904 static __always_inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
1905 {
1906         return bpage->page->data + index;
1907 }
1908 
1909 static __always_inline struct ring_buffer_event *
1910 rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
1911 {
1912         return __rb_page_index(cpu_buffer->reader_page,
1913                                cpu_buffer->reader_page->read);
1914 }
1915 
1916 static __always_inline struct ring_buffer_event *
1917 rb_iter_head_event(struct ring_buffer_iter *iter)
1918 {
1919         return __rb_page_index(iter->head_page, iter->head);
1920 }
1921 
1922 static __always_inline unsigned rb_page_commit(struct buffer_page *bpage)
1923 {
1924         return local_read(&bpage->page->commit);
1925 }
1926 
1927 /* Size is determined by what has been committed */
1928 static __always_inline unsigned rb_page_size(struct buffer_page *bpage)
1929 {
1930         return rb_page_commit(bpage);
1931 }
1932 
1933 static __always_inline unsigned
1934 rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
1935 {
1936         return rb_page_commit(cpu_buffer->commit_page);
1937 }
1938 
1939 static __always_inline unsigned
1940 rb_event_index(struct ring_buffer_event *event)
1941 {
1942         unsigned long addr = (unsigned long)event;
1943 
1944         return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
1945 }
1946 
1947 static void rb_inc_iter(struct ring_buffer_iter *iter)
1948 {
1949         struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
1950 
1951         /*
1952          * The iterator could be on the reader page (it starts there).
1953          * But the head could have moved, since the reader was
1954          * found. Check for this case and assign the iterator
1955          * to the head page instead of next.
1956          */
1957         if (iter->head_page == cpu_buffer->reader_page)
1958                 iter->head_page = rb_set_head_page(cpu_buffer);
1959         else
1960                 rb_inc_page(cpu_buffer, &iter->head_page);
1961 
1962         iter->read_stamp = iter->head_page->page->time_stamp;
1963         iter->head = 0;
1964 }
1965 
1966 /*
1967  * rb_handle_head_page - writer hit the head page
1968  *
1969  * Returns: +1 to retry page
1970  *           0 to continue
1971  *          -1 on error
1972  */
1973 static int
1974 rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
1975                     struct buffer_page *tail_page,
1976                     struct buffer_page *next_page)
1977 {
1978         struct buffer_page *new_head;
1979         int entries;
1980         int type;
1981         int ret;
1982 
1983         entries = rb_page_entries(next_page);
1984 
1985         /*
1986          * The hard part is here. We need to move the head
1987          * forward, and protect against both readers on
1988          * other CPUs and writers coming in via interrupts.
1989          */
1990         type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
1991                                        RB_PAGE_HEAD);
1992 
1993         /*
1994          * type can be one of four:
1995          *  NORMAL - an interrupt already moved it for us
1996          *  HEAD   - we are the first to get here.
1997          *  UPDATE - we are the interrupt interrupting
1998          *           a current move.
1999          *  MOVED  - a reader on another CPU moved the next
2000          *           pointer to its reader page. Give up
2001          *           and try again.
2002          */
2003 
2004         switch (type) {
2005         case RB_PAGE_HEAD:
2006                 /*
2007                  * We changed the head to UPDATE, thus
2008                  * it is our responsibility to update
2009                  * the counters.
2010                  */
2011                 local_add(entries, &cpu_buffer->overrun);
2012                 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
2013 
2014                 /*
2015                  * The entries will be zeroed out when we move the
2016                  * tail page.
2017                  */
2018 
2019                 /* still more to do */
2020                 break;
2021 
2022         case RB_PAGE_UPDATE:
2023                 /*
2024                  * This is an interrupt that interrupt the
2025                  * previous update. Still more to do.
2026                  */
2027                 break;
2028         case RB_PAGE_NORMAL:
2029                 /*
2030                  * An interrupt came in before the update
2031                  * and processed this for us.
2032                  * Nothing left to do.
2033                  */
2034                 return 1;
2035         case RB_PAGE_MOVED:
2036                 /*
2037                  * The reader is on another CPU and just did
2038                  * a swap with our next_page.
2039                  * Try again.
2040                  */
2041                 return 1;
2042         default:
2043                 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
2044                 return -1;
2045         }
2046 
2047         /*
2048          * Now that we are here, the old head pointer is
2049          * set to UPDATE. This will keep the reader from
2050          * swapping the head page with the reader page.
2051          * The reader (on another CPU) will spin till
2052          * we are finished.
2053          *
2054          * We just need to protect against interrupts
2055          * doing the job. We will set the next pointer
2056          * to HEAD. After that, we set the old pointer
2057          * to NORMAL, but only if it was HEAD before.
2058          * otherwise we are an interrupt, and only
2059          * want the outer most commit to reset it.
2060          */
2061         new_head = next_page;
2062         rb_inc_page(cpu_buffer, &new_head);
2063 
2064         ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
2065                                     RB_PAGE_NORMAL);
2066 
2067         /*
2068          * Valid returns are:
2069          *  HEAD   - an interrupt came in and already set it.
2070          *  NORMAL - One of two things:
2071          *            1) We really set it.
2072          *            2) A bunch of interrupts came in and moved
2073          *               the page forward again.
2074          */
2075         switch (ret) {
2076         case RB_PAGE_HEAD:
2077         case RB_PAGE_NORMAL:
2078                 /* OK */
2079                 break;
2080         default:
2081                 RB_WARN_ON(cpu_buffer, 1);
2082                 return -1;
2083         }
2084 
2085         /*
2086          * It is possible that an interrupt came in,
2087          * set the head up, then more interrupts came in
2088          * and moved it again. When we get back here,
2089          * the page would have been set to NORMAL but we
2090          * just set it back to HEAD.
2091          *
2092          * How do you detect this? Well, if that happened
2093          * the tail page would have moved.
2094          */
2095         if (ret == RB_PAGE_NORMAL) {
2096                 struct buffer_page *buffer_tail_page;
2097 
2098                 buffer_tail_page = READ_ONCE(cpu_buffer->tail_page);
2099                 /*
2100                  * If the tail had moved passed next, then we need
2101                  * to reset the pointer.
2102                  */
2103                 if (buffer_tail_page != tail_page &&
2104                     buffer_tail_page != next_page)
2105                         rb_head_page_set_normal(cpu_buffer, new_head,
2106                                                 next_page,
2107                                                 RB_PAGE_HEAD);
2108         }
2109 
2110         /*
2111          * If this was the outer most commit (the one that
2112          * changed the original pointer from HEAD to UPDATE),
2113          * then it is up to us to reset it to NORMAL.
2114          */
2115         if (type == RB_PAGE_HEAD) {
2116                 ret = rb_head_page_set_normal(cpu_buffer, next_page,
2117                                               tail_page,
2118                                               RB_PAGE_UPDATE);
2119                 if (RB_WARN_ON(cpu_buffer,
2120                                ret != RB_PAGE_UPDATE))
2121                         return -1;
2122         }
2123 
2124         return 0;
2125 }
2126 
2127 static inline void
2128 rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2129               unsigned long tail, struct rb_event_info *info)
2130 {
2131         struct buffer_page *tail_page = info->tail_page;
2132         struct ring_buffer_event *event;
2133         unsigned long length = info->length;
2134 
2135         /*
2136          * Only the event that crossed the page boundary
2137          * must fill the old tail_page with padding.
2138          */
2139         if (tail >= BUF_PAGE_SIZE) {
2140                 /*
2141                  * If the page was filled, then we still need
2142                  * to update the real_end. Reset it to zero
2143                  * and the reader will ignore it.
2144                  */
2145                 if (tail == BUF_PAGE_SIZE)
2146                         tail_page->real_end = 0;
2147 
2148                 local_sub(length, &tail_page->write);
2149                 return;
2150         }
2151 
2152         event = __rb_page_index(tail_page, tail);
2153 
2154         /* account for padding bytes */
2155         local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes);
2156 
2157         /*
2158          * Save the original length to the meta data.
2159          * This will be used by the reader to add lost event
2160          * counter.
2161          */
2162         tail_page->real_end = tail;
2163 
2164         /*
2165          * If this event is bigger than the minimum size, then
2166          * we need to be careful that we don't subtract the
2167          * write counter enough to allow another writer to slip
2168          * in on this page.
2169          * We put in a discarded commit instead, to make sure
2170          * that this space is not used again.
2171          *
2172          * If we are less than the minimum size, we don't need to
2173          * worry about it.
2174          */
2175         if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
2176                 /* No room for any events */
2177 
2178                 /* Mark the rest of the page with padding */
2179                 rb_event_set_padding(event);
2180 
2181                 /* Set the write back to the previous setting */
2182                 local_sub(length, &tail_page->write);
2183                 return;
2184         }
2185 
2186         /* Put in a discarded event */
2187         event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
2188         event->type_len = RINGBUF_TYPE_PADDING;
2189         /* time delta must be non zero */
2190         event->time_delta = 1;
2191 
2192         /* Set write to end of buffer */
2193         length = (tail + length) - BUF_PAGE_SIZE;
2194         local_sub(length, &tail_page->write);
2195 }
2196 
2197 static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer);
2198 
2199 /*
2200  * This is the slow path, force gcc not to inline it.
2201  */
2202 static noinline struct ring_buffer_event *
2203 rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2204              unsigned long tail, struct rb_event_info *info)
2205 {
2206         struct buffer_page *tail_page = info->tail_page;
2207         struct buffer_page *commit_page = cpu_buffer->commit_page;
2208         struct trace_buffer *buffer = cpu_buffer->buffer;
2209         struct buffer_page *next_page;
2210         int ret;
2211 
2212         next_page = tail_page;
2213 
2214         rb_inc_page(cpu_buffer, &next_page);
2215 
2216         /*
2217          * If for some reason, we had an interrupt storm that made
2218          * it all the way around the buffer, bail, and warn
2219          * about it.
2220          */
2221         if (unlikely(next_page == commit_page)) {
2222                 local_inc(&cpu_buffer->commit_overrun);
2223                 goto out_reset;
2224         }
2225 
2226         /*
2227          * This is where the fun begins!
2228          *
2229          * We are fighting against races between a reader that
2230          * could be on another CPU trying to swap its reader
2231          * page with the buffer head.
2232          *
2233          * We are also fighting against interrupts coming in and
2234          * moving the head or tail on us as well.
2235          *
2236          * If the next page is the head page then we have filled
2237          * the buffer, unless the commit page is still on the
2238          * reader page.
2239          */
2240         if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) {
2241 
2242                 /*
2243                  * If the commit is not on the reader page, then
2244                  * move the header page.
2245                  */
2246                 if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2247                         /*
2248                          * If we are not in overwrite mode,
2249                          * this is easy, just stop here.
2250                          */
2251                         if (!(buffer->flags & RB_FL_OVERWRITE)) {
2252                                 local_inc(&cpu_buffer->dropped_events);
2253                                 goto out_reset;
2254                         }
2255 
2256                         ret = rb_handle_head_page(cpu_buffer,
2257                                                   tail_page,
2258                                                   next_page);
2259                         if (ret < 0)
2260                                 goto out_reset;
2261                         if (ret)
2262                                 goto out_again;
2263                 } else {
2264                         /*
2265                          * We need to be careful here too. The
2266                          * commit page could still be on the reader
2267                          * page. We could have a small buffer, and
2268                          * have filled up the buffer with events
2269                          * from interrupts and such, and wrapped.
2270                          *
2271                          * Note, if the tail page is also the on the
2272                          * reader_page, we let it move out.
2273                          */
2274                         if (unlikely((cpu_buffer->commit_page !=
2275                                       cpu_buffer->tail_page) &&
2276                                      (cpu_buffer->commit_page ==
2277                                       cpu_buffer->reader_page))) {
2278                                 local_inc(&cpu_buffer->commit_overrun);
2279                                 goto out_reset;
2280                         }
2281                 }
2282         }
2283 
2284         rb_tail_page_update(cpu_buffer, tail_page, next_page);
2285 
2286  out_again:
2287 
2288         rb_reset_tail(cpu_buffer, tail, info);
2289 
2290         /* Commit what we have for now. */
2291         rb_end_commit(cpu_buffer);
2292         /* rb_end_commit() decs committing */
2293         local_inc(&cpu_buffer->committing);
2294 
2295         /* fail and let the caller try again */
2296         return ERR_PTR(-EAGAIN);
2297 
2298  out_reset:
2299         /* reset write */
2300         rb_reset_tail(cpu_buffer, tail, info);
2301 
2302         return NULL;
2303 }
2304 
2305 /* Slow path, do not inline */
2306 static noinline struct ring_buffer_event *
2307 rb_add_time_stamp(struct ring_buffer_event *event, u64 delta, bool abs)
2308 {
2309         if (abs)
2310                 event->type_len = RINGBUF_TYPE_TIME_STAMP;
2311         else
2312                 event->type_len = RINGBUF_TYPE_TIME_EXTEND;
2313 
2314         /* Not the first event on the page, or not delta? */
2315         if (abs || rb_event_index(event)) {
2316                 event->time_delta = delta & TS_MASK;
2317                 event->array[0] = delta >> TS_SHIFT;
2318         } else {
2319                 /* nope, just zero it */
2320                 event->time_delta = 0;
2321                 event->array[0] = 0;
2322         }
2323 
2324         return skip_time_extend(event);
2325 }
2326 
2327 static inline bool rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2328                                      struct ring_buffer_event *event);
2329 
2330 /**
2331  * rb_update_event - update event type and data
2332  * @cpu_buffer: The per cpu buffer of the @event
2333  * @event: the event to update
2334  * @info: The info to update the @event with (contains length and delta)
2335  *
2336  * Update the type and data fields of the @event. The length
2337  * is the actual size that is written to the ring buffer,
2338  * and with this, we can determine what to place into the
2339  * data field.
2340  */
2341 static void
2342 rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
2343                 struct ring_buffer_event *event,
2344                 struct rb_event_info *info)
2345 {
2346         unsigned length = info->length;
2347         u64 delta = info->delta;
2348 
2349         /* Only a commit updates the timestamp */
2350         if (unlikely(!rb_event_is_commit(cpu_buffer, event)))
2351                 delta = 0;
2352 
2353         /*
2354          * If we need to add a timestamp, then we
2355          * add it to the start of the reserved space.
2356          */
2357         if (unlikely(info->add_timestamp)) {
2358                 bool abs = ring_buffer_time_stamp_abs(cpu_buffer->buffer);
2359 
2360                 event = rb_add_time_stamp(event, info->delta, abs);
2361                 length -= RB_LEN_TIME_EXTEND;
2362                 delta = 0;
2363         }
2364 
2365         event->time_delta = delta;
2366         length -= RB_EVNT_HDR_SIZE;
2367         if (length > RB_MAX_SMALL_DATA) {
2368                 event->type_len = 0;
2369                 event->array[0] = length;
2370         } else
2371                 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
2372 }
2373 
2374 static unsigned rb_calculate_event_length(unsigned length)
2375 {
2376         struct ring_buffer_event event; /* Used only for sizeof array */
2377 
2378         /* zero length can cause confusions */
2379         if (!length)
2380                 length++;
2381 
2382         if (length > RB_MAX_SMALL_DATA)
2383                 length += sizeof(event.array[0]);
2384 
2385         length += RB_EVNT_HDR_SIZE;
2386         length = ALIGN(length, RB_ALIGNMENT);
2387 
2388         /*
2389          * In case the time delta is larger than the 27 bits for it
2390          * in the header, we need to add a timestamp. If another
2391          * event comes in when trying to discard this one to increase
2392          * the length, then the timestamp will be added in the allocated
2393          * space of this event. If length is bigger than the size needed
2394          * for the TIME_EXTEND, then padding has to be used. The events
2395          * length must be either RB_LEN_TIME_EXTEND, or greater than or equal
2396          * to RB_LEN_TIME_EXTEND + 8, as 8 is the minimum size for padding.
2397          * As length is a multiple of 4, we only need to worry if it
2398          * is 12 (RB_LEN_TIME_EXTEND + 4).
2399          */
2400         if (length == RB_LEN_TIME_EXTEND + RB_ALIGNMENT)
2401                 length += RB_ALIGNMENT;
2402 
2403         return length;
2404 }
2405 
2406 #ifndef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2407 static inline bool sched_clock_stable(void)
2408 {
2409         return true;
2410 }
2411 #endif
2412 
2413 static inline int
2414 rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2415                   struct ring_buffer_event *event)
2416 {
2417         unsigned long new_index, old_index;
2418         struct buffer_page *bpage;
2419         unsigned long index;
2420         unsigned long addr;
2421 
2422         new_index = rb_event_index(event);
2423         old_index = new_index + rb_event_ts_length(event);
2424         addr = (unsigned long)event;
2425         addr &= PAGE_MASK;
2426 
2427         bpage = READ_ONCE(cpu_buffer->tail_page);
2428 
2429         if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2430                 unsigned long write_mask =
2431                         local_read(&bpage->write) & ~RB_WRITE_MASK;
2432                 unsigned long event_length = rb_event_length(event);
2433                 /*
2434                  * This is on the tail page. It is possible that
2435                  * a write could come in and move the tail page
2436                  * and write to the next page. That is fine
2437                  * because we just shorten what is on this page.
2438                  */
2439                 old_index += write_mask;
2440                 new_index += write_mask;
2441                 index = local_cmpxchg(&bpage->write, old_index, new_index);
2442                 if (index == old_index) {
2443                         /* update counters */
2444                         local_sub(event_length, &cpu_buffer->entries_bytes);
2445                         return 1;
2446                 }
2447         }
2448 
2449         /* could not discard */
2450         return 0;
2451 }
2452 
2453 static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2454 {
2455         local_inc(&cpu_buffer->committing);
2456         local_inc(&cpu_buffer->commits);
2457 }
2458 
2459 static __always_inline void
2460 rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
2461 {
2462         unsigned long max_count;
2463 
2464         /*
2465          * We only race with interrupts and NMIs on this CPU.
2466          * If we own the commit event, then we can commit
2467          * all others that interrupted us, since the interruptions
2468          * are in stack format (they finish before they come
2469          * back to us). This allows us to do a simple loop to
2470          * assign the commit to the tail.
2471          */
2472  again:
2473         max_count = cpu_buffer->nr_pages * 100;
2474 
2475         while (cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)) {
2476                 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
2477                         return;
2478                 if (RB_WARN_ON(cpu_buffer,
2479                                rb_is_reader_page(cpu_buffer->tail_page)))
2480                         return;
2481                 local_set(&cpu_buffer->commit_page->page->commit,
2482                           rb_page_write(cpu_buffer->commit_page));
2483                 rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
2484                 /* Only update the write stamp if the page has an event */
2485                 if (rb_page_write(cpu_buffer->commit_page))
2486                         cpu_buffer->write_stamp =
2487                                 cpu_buffer->commit_page->page->time_stamp;
2488                 /* add barrier to keep gcc from optimizing too much */
2489                 barrier();
2490         }
2491         while (rb_commit_index(cpu_buffer) !=
2492                rb_page_write(cpu_buffer->commit_page)) {
2493 
2494                 local_set(&cpu_buffer->commit_page->page->commit,
2495                           rb_page_write(cpu_buffer->commit_page));
2496                 RB_WARN_ON(cpu_buffer,
2497                            local_read(&cpu_buffer->commit_page->page->commit) &
2498                            ~RB_WRITE_MASK);
2499                 barrier();
2500         }
2501 
2502         /* again, keep gcc from optimizing */
2503         barrier();
2504 
2505         /*
2506          * If an interrupt came in just after the first while loop
2507          * and pushed the tail page forward, we will be left with
2508          * a dangling commit that will never go forward.
2509          */
2510         if (unlikely(cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)))
2511                 goto again;
2512 }
2513 
2514 static __always_inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
2515 {
2516         unsigned long commits;
2517 
2518         if (RB_WARN_ON(cpu_buffer,
2519                        !local_read(&cpu_buffer->committing)))
2520                 return;
2521 
2522  again:
2523         commits = local_read(&cpu_buffer->commits);
2524         /* synchronize with interrupts */
2525         barrier();
2526         if (local_read(&cpu_buffer->committing) == 1)
2527                 rb_set_commit_to_write(cpu_buffer);
2528 
2529         local_dec(&cpu_buffer->committing);
2530 
2531         /* synchronize with interrupts */
2532         barrier();
2533 
2534         /*
2535          * Need to account for interrupts coming in between the
2536          * updating of the commit page and the clearing of the
2537          * committing counter.
2538          */
2539         if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
2540             !local_read(&cpu_buffer->committing)) {
2541                 local_inc(&cpu_buffer->committing);
2542                 goto again;
2543         }
2544 }
2545 
2546 static inline void rb_event_discard(struct ring_buffer_event *event)
2547 {
2548         if (extended_time(event))
2549                 event = skip_time_extend(event);
2550 
2551         /* array[0] holds the actual length for the discarded event */
2552         event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
2553         event->type_len = RINGBUF_TYPE_PADDING;
2554         /* time delta must be non zero */
2555         if (!event->time_delta)
2556                 event->time_delta = 1;
2557 }
2558 
2559 static __always_inline bool
2560 rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2561                    struct ring_buffer_event *event)
2562 {
2563         unsigned long addr = (unsigned long)event;
2564         unsigned long index;
2565 
2566         index = rb_event_index(event);
2567         addr &= PAGE_MASK;
2568 
2569         return cpu_buffer->commit_page->page == (void *)addr &&
2570                 rb_commit_index(cpu_buffer) == index;
2571 }
2572 
2573 static __always_inline void
2574 rb_update_write_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2575                       struct ring_buffer_event *event)
2576 {
2577         u64 delta;
2578 
2579         /*
2580          * The event first in the commit queue updates the
2581          * time stamp.
2582          */
2583         if (rb_event_is_commit(cpu_buffer, event)) {
2584                 /*
2585                  * A commit event that is first on a page
2586                  * updates the write timestamp with the page stamp
2587                  */
2588                 if (!rb_event_index(event))
2589                         cpu_buffer->write_stamp =
2590                                 cpu_buffer->commit_page->page->time_stamp;
2591                 else if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
2592                         delta = ring_buffer_event_time_stamp(event);
2593                         cpu_buffer->write_stamp += delta;
2594                 } else if (event->type_len == RINGBUF_TYPE_TIME_STAMP) {
2595                         delta = ring_buffer_event_time_stamp(event);
2596                         cpu_buffer->write_stamp = delta;
2597                 } else
2598                         cpu_buffer->write_stamp += event->time_delta;
2599         }
2600 }
2601 
2602 static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
2603                       struct ring_buffer_event *event)
2604 {
2605         local_inc(&cpu_buffer->entries);
2606         rb_update_write_stamp(cpu_buffer, event);
2607         rb_end_commit(cpu_buffer);
2608 }
2609 
2610 static __always_inline void
2611 rb_wakeups(struct trace_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer)
2612 {
2613         size_t nr_pages;
2614         size_t dirty;
2615         size_t full;
2616 
2617         if (buffer->irq_work.waiters_pending) {
2618                 buffer->irq_work.waiters_pending = false;
2619                 /* irq_work_queue() supplies it's own memory barriers */
2620                 irq_work_queue(&buffer->irq_work.work);
2621         }
2622 
2623         if (cpu_buffer->irq_work.waiters_pending) {
2624                 cpu_buffer->irq_work.waiters_pending = false;
2625                 /* irq_work_queue() supplies it's own memory barriers */
2626                 irq_work_queue(&cpu_buffer->irq_work.work);
2627         }
2628 
2629         if (cpu_buffer->last_pages_touch == local_read(&cpu_buffer->pages_touched))
2630                 return;
2631 
2632         if (cpu_buffer->reader_page == cpu_buffer->commit_page)
2633                 return;
2634 
2635         if (!cpu_buffer->irq_work.full_waiters_pending)
2636                 return;
2637 
2638         cpu_buffer->last_pages_touch = local_read(&cpu_buffer->pages_touched);
2639 
2640         full = cpu_buffer->shortest_full;
2641         nr_pages = cpu_buffer->nr_pages;
2642         dirty = ring_buffer_nr_dirty_pages(buffer, cpu_buffer->cpu);
2643         if (full && nr_pages && (dirty * 100) <= full * nr_pages)
2644                 return;
2645 
2646         cpu_buffer->irq_work.wakeup_full = true;
2647         cpu_buffer->irq_work.full_waiters_pending = false;
2648         /* irq_work_queue() supplies it's own memory barriers */
2649         irq_work_queue(&cpu_buffer->irq_work.work);
2650 }
2651 
2652 /*
2653  * The lock and unlock are done within a preempt disable section.
2654  * The current_context per_cpu variable can only be modified
2655  * by the current task between lock and unlock. But it can
2656  * be modified more than once via an interrupt. To pass this
2657  * information from the lock to the unlock without having to
2658  * access the 'in_interrupt()' functions again (which do show
2659  * a bit of overhead in something as critical as function tracing,
2660  * we use a bitmask trick.
2661  *
2662  *  bit 0 =  NMI context
2663  *  bit 1 =  IRQ context
2664  *  bit 2 =  SoftIRQ context
2665  *  bit 3 =  normal context.
2666  *
2667  * This works because this is the order of contexts that can
2668  * preempt other contexts. A SoftIRQ never preempts an IRQ
2669  * context.
2670  *
2671  * When the context is determined, the corresponding bit is
2672  * checked and set (if it was set, then a recursion of that context
2673  * happened).
2674  *
2675  * On unlock, we need to clear this bit. To do so, just subtract
2676  * 1 from the current_context and AND it to itself.
2677  *
2678  * (binary)
2679  *  101 - 1 = 100
2680  *  101 & 100 = 100 (clearing bit zero)
2681  *
2682  *  1010 - 1 = 1001
2683  *  1010 & 1001 = 1000 (clearing bit 1)
2684  *
2685  * The least significant bit can be cleared this way, and it
2686  * just so happens that it is the same bit corresponding to
2687  * the current context.
2688  */
2689 
2690 static __always_inline int
2691 trace_recursive_lock(struct ring_buffer_per_cpu *cpu_buffer)
2692 {
2693         unsigned int val = cpu_buffer->current_context;
2694         unsigned long pc = preempt_count();
2695         int bit;
2696 
2697         if (!(pc & (NMI_MASK | HARDIRQ_MASK | SOFTIRQ_OFFSET)))
2698                 bit = RB_CTX_NORMAL;
2699         else
2700                 bit = pc & NMI_MASK ? RB_CTX_NMI :
2701                         pc & HARDIRQ_MASK ? RB_CTX_IRQ : RB_CTX_SOFTIRQ;
2702 
2703         if (unlikely(val & (1 << (bit + cpu_buffer->nest))))
2704                 return 1;
2705 
2706         val |= (1 << (bit + cpu_buffer->nest));
2707         cpu_buffer->current_context = val;
2708 
2709         return 0;
2710 }
2711 
2712 static __always_inline void
2713 trace_recursive_unlock(struct ring_buffer_per_cpu *cpu_buffer)
2714 {
2715         cpu_buffer->current_context &=
2716                 cpu_buffer->current_context - (1 << cpu_buffer->nest);
2717 }
2718 
2719 /* The recursive locking above uses 4 bits */
2720 #define NESTED_BITS 4
2721 
2722 /**
2723  * ring_buffer_nest_start - Allow to trace while nested
2724  * @buffer: The ring buffer to modify
2725  *
2726  * The ring buffer has a safety mechanism to prevent recursion.
2727  * But there may be a case where a trace needs to be done while
2728  * tracing something else. In this case, calling this function
2729  * will allow this function to nest within a currently active
2730  * ring_buffer_lock_reserve().
2731  *
2732  * Call this function before calling another ring_buffer_lock_reserve() and
2733  * call ring_buffer_nest_end() after the nested ring_buffer_unlock_commit().
2734  */
2735 void ring_buffer_nest_start(struct trace_buffer *buffer)
2736 {
2737         struct ring_buffer_per_cpu *cpu_buffer;
2738         int cpu;
2739 
2740         /* Enabled by ring_buffer_nest_end() */
2741         preempt_disable_notrace();
2742         cpu = raw_smp_processor_id();
2743         cpu_buffer = buffer->buffers[cpu];
2744         /* This is the shift value for the above recursive locking */
2745         cpu_buffer->nest += NESTED_BITS;
2746 }
2747 
2748 /**
2749  * ring_buffer_nest_end - Allow to trace while nested
2750  * @buffer: The ring buffer to modify
2751  *
2752  * Must be called after ring_buffer_nest_start() and after the
2753  * ring_buffer_unlock_commit().
2754  */
2755 void ring_buffer_nest_end(struct trace_buffer *buffer)
2756 {
2757         struct ring_buffer_per_cpu *cpu_buffer;
2758         int cpu;
2759 
2760         /* disabled by ring_buffer_nest_start() */
2761         cpu = raw_smp_processor_id();
2762         cpu_buffer = buffer->buffers[cpu];
2763         /* This is the shift value for the above recursive locking */
2764         cpu_buffer->nest -= NESTED_BITS;
2765         preempt_enable_notrace();
2766 }
2767 
2768 /**
2769  * ring_buffer_unlock_commit - commit a reserved
2770  * @buffer: The buffer to commit to
2771  * @event: The event pointer to commit.
2772  *
2773  * This commits the data to the ring buffer, and releases any locks held.
2774  *
2775  * Must be paired with ring_buffer_lock_reserve.
2776  */
2777 int ring_buffer_unlock_commit(struct trace_buffer *buffer,
2778                               struct ring_buffer_event *event)
2779 {
2780         struct ring_buffer_per_cpu *cpu_buffer;
2781         int cpu = raw_smp_processor_id();
2782 
2783         cpu_buffer = buffer->buffers[cpu];
2784 
2785         rb_commit(cpu_buffer, event);
2786 
2787         rb_wakeups(buffer, cpu_buffer);
2788 
2789         trace_recursive_unlock(cpu_buffer);
2790 
2791         preempt_enable_notrace();
2792 
2793         return 0;
2794 }
2795 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
2796 
2797 static noinline void
2798 rb_handle_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2799                     struct rb_event_info *info)
2800 {
2801         WARN_ONCE(info->delta > (1ULL << 59),
2802                   KERN_WARNING "Delta way too big! %llu ts=%llu write stamp = %llu\n%s",
2803                   (unsigned long long)info->delta,
2804                   (unsigned long long)info->ts,
2805                   (unsigned long long)cpu_buffer->write_stamp,
2806                   sched_clock_stable() ? "" :
2807                   "If you just came from a suspend/resume,\n"
2808                   "please switch to the trace global clock:\n"
2809                   "  echo global > /sys/kernel/debug/tracing/trace_clock\n"
2810                   "or add trace_clock=global to the kernel command line\n");
2811         info->add_timestamp = 1;
2812 }
2813 
2814 static struct ring_buffer_event *
2815 __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
2816                   struct rb_event_info *info)
2817 {
2818         struct ring_buffer_event *event;
2819         struct buffer_page *tail_page;
2820         unsigned long tail, write;
2821 
2822         /*
2823          * If the time delta since the last event is too big to
2824          * hold in the time field of the event, then we append a
2825          * TIME EXTEND event ahead of the data event.
2826          */
2827         if (unlikely(info->add_timestamp))
2828                 info->length += RB_LEN_TIME_EXTEND;
2829 
2830         /* Don't let the compiler play games with cpu_buffer->tail_page */
2831         tail_page = info->tail_page = READ_ONCE(cpu_buffer->tail_page);
2832         write = local_add_return(info->length, &tail_page->write);
2833 
2834         /* set write to only the index of the write */
2835         write &= RB_WRITE_MASK;
2836         tail = write - info->length;
2837 
2838         /*
2839          * If this is the first commit on the page, then it has the same
2840          * timestamp as the page itself.
2841          */
2842         if (!tail && !ring_buffer_time_stamp_abs(cpu_buffer->buffer))
2843                 info->delta = 0;
2844 
2845         /* See if we shot pass the end of this buffer page */
2846         if (unlikely(write > BUF_PAGE_SIZE))
2847                 return rb_move_tail(cpu_buffer, tail, info);
2848 
2849         /* We reserved something on the buffer */
2850 
2851         event = __rb_page_index(tail_page, tail);
2852         rb_update_event(cpu_buffer, event, info);
2853 
2854         local_inc(&tail_page->entries);
2855 
2856         /*
2857          * If this is the first commit on the page, then update
2858          * its timestamp.
2859          */
2860         if (!tail)
2861                 tail_page->page->time_stamp = info->ts;
2862 
2863         /* account for these added bytes */
2864         local_add(info->length, &cpu_buffer->entries_bytes);
2865 
2866         return event;
2867 }
2868 
2869 static __always_inline struct ring_buffer_event *
2870 rb_reserve_next_event(struct trace_buffer *buffer,
2871                       struct ring_buffer_per_cpu *cpu_buffer,
2872                       unsigned long length)
2873 {
2874         struct ring_buffer_event *event;
2875         struct rb_event_info info;
2876         int nr_loops = 0;
2877         u64 diff;
2878 
2879         rb_start_commit(cpu_buffer);
2880 
2881 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2882         /*
2883          * Due to the ability to swap a cpu buffer from a buffer
2884          * it is possible it was swapped before we committed.
2885          * (committing stops a swap). We check for it here and
2886          * if it happened, we have to fail the write.
2887          */
2888         barrier();
2889         if (unlikely(READ_ONCE(cpu_buffer->buffer) != buffer)) {
2890                 local_dec(&cpu_buffer->committing);
2891                 local_dec(&cpu_buffer->commits);
2892                 return NULL;
2893         }
2894 #endif
2895 
2896         info.length = rb_calculate_event_length(length);
2897  again:
2898         info.add_timestamp = 0;
2899         info.delta = 0;
2900 
2901         /*
2902          * We allow for interrupts to reenter here and do a trace.
2903          * If one does, it will cause this original code to loop
2904          * back here. Even with heavy interrupts happening, this
2905          * should only happen a few times in a row. If this happens
2906          * 1000 times in a row, there must be either an interrupt
2907          * storm or we have something buggy.
2908          * Bail!
2909          */
2910         if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
2911                 goto out_fail;
2912 
2913         info.ts = rb_time_stamp(cpu_buffer->buffer);
2914         diff = info.ts - cpu_buffer->write_stamp;
2915 
2916         /* make sure this diff is calculated here */
2917         barrier();
2918 
2919         if (ring_buffer_time_stamp_abs(buffer)) {
2920                 info.delta = info.ts;
2921                 rb_handle_timestamp(cpu_buffer, &info);
2922         } else /* Did the write stamp get updated already? */
2923                 if (likely(info.ts >= cpu_buffer->write_stamp)) {
2924                 info.delta = diff;
2925                 if (unlikely(test_time_stamp(info.delta)))
2926                         rb_handle_timestamp(cpu_buffer, &info);
2927         }
2928 
2929         event = __rb_reserve_next(cpu_buffer, &info);
2930 
2931         if (unlikely(PTR_ERR(event) == -EAGAIN)) {
2932                 if (info.add_timestamp)
2933                         info.length -= RB_LEN_TIME_EXTEND;
2934                 goto again;
2935         }
2936 
2937         if (!event)
2938                 goto out_fail;
2939 
2940         return event;
2941 
2942  out_fail:
2943         rb_end_commit(cpu_buffer);
2944         return NULL;
2945 }
2946 
2947 /**
2948  * ring_buffer_lock_reserve - reserve a part of the buffer
2949  * @buffer: the ring buffer to reserve from
2950  * @length: the length of the data to reserve (excluding event header)
2951  *
2952  * Returns a reserved event on the ring buffer to copy directly to.
2953  * The user of this interface will need to get the body to write into
2954  * and can use the ring_buffer_event_data() interface.
2955  *
2956  * The length is the length of the data needed, not the event length
2957  * which also includes the event header.
2958  *
2959  * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2960  * If NULL is returned, then nothing has been allocated or locked.
2961  */
2962 struct ring_buffer_event *
2963 ring_buffer_lock_reserve(struct trace_buffer *buffer, unsigned long length)
2964 {
2965         struct ring_buffer_per_cpu *cpu_buffer;
2966         struct ring_buffer_event *event;
2967         int cpu;
2968 
2969         /* If we are tracing schedule, we don't want to recurse */
2970         preempt_disable_notrace();
2971 
2972         if (unlikely(atomic_read(&buffer->record_disabled)))
2973                 goto out;
2974 
2975         cpu = raw_smp_processor_id();
2976 
2977         if (unlikely(!cpumask_test_cpu(cpu, buffer->cpumask)))
2978                 goto out;
2979 
2980         cpu_buffer = buffer->buffers[cpu];
2981 
2982         if (unlikely(atomic_read(&cpu_buffer->record_disabled)))
2983                 goto out;
2984 
2985         if (unlikely(length > BUF_MAX_DATA_SIZE))
2986                 goto out;
2987 
2988         if (unlikely(trace_recursive_lock(cpu_buffer)))
2989                 goto out;
2990 
2991         event = rb_reserve_next_event(buffer, cpu_buffer, length);
2992         if (!event)
2993                 goto out_unlock;
2994 
2995         return event;
2996 
2997  out_unlock:
2998         trace_recursive_unlock(cpu_buffer);
2999  out:
3000         preempt_enable_notrace();
3001         return NULL;
3002 }
3003 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
3004 
3005 /*
3006  * Decrement the entries to the page that an event is on.
3007  * The event does not even need to exist, only the pointer
3008  * to the page it is on. This may only be called before the commit
3009  * takes place.
3010  */
3011 static inline void
3012 rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
3013                    struct ring_buffer_event *event)
3014 {
3015         unsigned long addr = (unsigned long)event;
3016         struct buffer_page *bpage = cpu_buffer->commit_page;
3017         struct buffer_page *start;
3018 
3019         addr &= PAGE_MASK;
3020 
3021         /* Do the likely case first */
3022         if (likely(bpage->page == (void *)addr)) {
3023                 local_dec(&bpage->entries);
3024                 return;
3025         }
3026 
3027         /*
3028          * Because the commit page may be on the reader page we
3029          * start with the next page and check the end loop there.
3030          */
3031         rb_inc_page(cpu_buffer, &bpage);
3032         start = bpage;
3033         do {
3034                 if (bpage->page == (void *)addr) {
3035                         local_dec(&bpage->entries);
3036                         return;
3037                 }
3038                 rb_inc_page(cpu_buffer, &bpage);
3039         } while (bpage != start);
3040 
3041         /* commit not part of this buffer?? */
3042         RB_WARN_ON(cpu_buffer, 1);
3043 }
3044 
3045 /**
3046  * ring_buffer_commit_discard - discard an event that has not been committed
3047  * @buffer: the ring buffer
3048  * @event: non committed event to discard
3049  *
3050  * Sometimes an event that is in the ring buffer needs to be ignored.
3051  * This function lets the user discard an event in the ring buffer
3052  * and then that event will not be read later.
3053  *
3054  * This function only works if it is called before the item has been
3055  * committed. It will try to free the event from the ring buffer
3056  * if another event has not been added behind it.
3057  *
3058  * If another event has been added behind it, it will set the event
3059  * up as discarded, and perform the commit.
3060  *
3061  * If this function is called, do not call ring_buffer_unlock_commit on
3062  * the event.
3063  */
3064 void ring_buffer_discard_commit(struct trace_buffer *buffer,
3065                                 struct ring_buffer_event *event)
3066 {
3067         struct ring_buffer_per_cpu *cpu_buffer;
3068         int cpu;
3069 
3070         /* The event is discarded regardless */
3071         rb_event_discard(event);
3072 
3073         cpu = smp_processor_id();
3074         cpu_buffer = buffer->buffers[cpu];
3075 
3076         /*
3077          * This must only be called if the event has not been
3078          * committed yet. Thus we can assume that preemption
3079          * is still disabled.
3080          */
3081         RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
3082 
3083         rb_decrement_entry(cpu_buffer, event);
3084         if (rb_try_to_discard(cpu_buffer, event))
3085                 goto out;
3086 
3087         /*
3088          * The commit is still visible by the reader, so we
3089          * must still update the timestamp.
3090          */
3091         rb_update_write_stamp(cpu_buffer, event);
3092  out:
3093         rb_end_commit(cpu_buffer);
3094 
3095         trace_recursive_unlock(cpu_buffer);
3096 
3097         preempt_enable_notrace();
3098 
3099 }
3100 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
3101 
3102 /**
3103  * ring_buffer_write - write data to the buffer without reserving
3104  * @buffer: The ring buffer to write to.
3105  * @length: The length of the data being written (excluding the event header)
3106  * @data: The data to write to the buffer.
3107  *
3108  * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
3109  * one function. If you already have the data to write to the buffer, it
3110  * may be easier to simply call this function.
3111  *
3112  * Note, like ring_buffer_lock_reserve, the length is the length of the data
3113  * and not the length of the event which would hold the header.
3114  */
3115 int ring_buffer_write(struct trace_buffer *buffer,
3116                       unsigned long length,
3117                       void *data)
3118 {
3119         struct ring_buffer_per_cpu *cpu_buffer;
3120         struct ring_buffer_event *event;
3121         void *body;
3122         int ret = -EBUSY;
3123         int cpu;
3124 
3125         preempt_disable_notrace();
3126 
3127         if (atomic_read(&buffer->record_disabled))
3128                 goto out;
3129 
3130         cpu = raw_smp_processor_id();
3131 
3132         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3133                 goto out;
3134 
3135         cpu_buffer = buffer->buffers[cpu];
3136 
3137         if (atomic_read(&cpu_buffer->record_disabled))
3138                 goto out;
3139 
3140         if (length > BUF_MAX_DATA_SIZE)
3141                 goto out;
3142 
3143         if (unlikely(trace_recursive_lock(cpu_buffer)))
3144                 goto out;
3145 
3146         event = rb_reserve_next_event(buffer, cpu_buffer, length);
3147         if (!event)
3148                 goto out_unlock;
3149 
3150         body = rb_event_data(event);
3151 
3152         memcpy(body, data, length);
3153 
3154         rb_commit(cpu_buffer, event);
3155 
3156         rb_wakeups(buffer, cpu_buffer);
3157 
3158         ret = 0;
3159 
3160  out_unlock:
3161         trace_recursive_unlock(cpu_buffer);
3162 
3163  out:
3164         preempt_enable_notrace();
3165 
3166         return ret;
3167 }
3168 EXPORT_SYMBOL_GPL(ring_buffer_write);
3169 
3170 static bool rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
3171 {
3172         struct buffer_page *reader = cpu_buffer->reader_page;
3173         struct buffer_page *head = rb_set_head_page(cpu_buffer);
3174         struct buffer_page *commit = cpu_buffer->commit_page;
3175 
3176         /* In case of error, head will be NULL */
3177         if (unlikely(!head))
3178                 return true;
3179 
3180         return reader->read == rb_page_commit(reader) &&
3181                 (commit == reader ||
3182                  (commit == head &&
3183                   head->read == rb_page_commit(commit)));
3184 }
3185 
3186 /**
3187  * ring_buffer_record_disable - stop all writes into the buffer
3188  * @buffer: The ring buffer to stop writes to.
3189  *
3190  * This prevents all writes to the buffer. Any attempt to write
3191  * to the buffer after this will fail and return NULL.
3192  *
3193  * The caller should call synchronize_rcu() after this.
3194  */
3195 void ring_buffer_record_disable(struct trace_buffer *buffer)
3196 {
3197         atomic_inc(&buffer->record_disabled);
3198 }
3199 EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
3200 
3201 /**
3202  * ring_buffer_record_enable - enable writes to the buffer
3203  * @buffer: The ring buffer to enable writes
3204  *
3205  * Note, multiple disables will need the same number of enables
3206  * to truly enable the writing (much like preempt_disable).
3207  */
3208 void ring_buffer_record_enable(struct trace_buffer *buffer)
3209 {
3210         atomic_dec(&buffer->record_disabled);
3211 }
3212 EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
3213 
3214 /**
3215  * ring_buffer_record_off - stop all writes into the buffer
3216  * @buffer: The ring buffer to stop writes to.
3217  *
3218  * This prevents all writes to the buffer. Any attempt to write
3219  * to the buffer after this will fail and return NULL.
3220  *
3221  * This is different than ring_buffer_record_disable() as
3222  * it works like an on/off switch, where as the disable() version
3223  * must be paired with a enable().
3224  */
3225 void ring_buffer_record_off(struct trace_buffer *buffer)
3226 {
3227         unsigned int rd;
3228         unsigned int new_rd;
3229 
3230         do {
3231                 rd = atomic_read(&buffer->record_disabled);
3232                 new_rd = rd | RB_BUFFER_OFF;
3233         } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3234 }
3235 EXPORT_SYMBOL_GPL(ring_buffer_record_off);
3236 
3237 /**
3238  * ring_buffer_record_on - restart writes into the buffer
3239  * @buffer: The ring buffer to start writes to.
3240  *
3241  * This enables all writes to the buffer that was disabled by
3242  * ring_buffer_record_off().
3243  *
3244  * This is different than ring_buffer_record_enable() as
3245  * it works like an on/off switch, where as the enable() version
3246  * must be paired with a disable().
3247  */
3248 void ring_buffer_record_on(struct trace_buffer *buffer)
3249 {
3250         unsigned int rd;
3251         unsigned int new_rd;
3252 
3253         do {
3254                 rd = atomic_read(&buffer->record_disabled);
3255                 new_rd = rd & ~RB_BUFFER_OFF;
3256         } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3257 }
3258 EXPORT_SYMBOL_GPL(ring_buffer_record_on);
3259 
3260 /**
3261  * ring_buffer_record_is_on - return true if the ring buffer can write
3262  * @buffer: The ring buffer to see if write is enabled
3263  *
3264  * Returns true if the ring buffer is in a state that it accepts writes.
3265  */
3266 bool ring_buffer_record_is_on(struct trace_buffer *buffer)
3267 {
3268         return !atomic_read(&buffer->record_disabled);
3269 }
3270 
3271 /**
3272  * ring_buffer_record_is_set_on - return true if the ring buffer is set writable
3273  * @buffer: The ring buffer to see if write is set enabled
3274  *
3275  * Returns true if the ring buffer is set writable by ring_buffer_record_on().
3276  * Note that this does NOT mean it is in a writable state.
3277  *
3278  * It may return true when the ring buffer has been disabled by
3279  * ring_buffer_record_disable(), as that is a temporary disabling of
3280  * the ring buffer.
3281  */
3282 bool ring_buffer_record_is_set_on(struct trace_buffer *buffer)
3283 {
3284         return !(atomic_read(&buffer->record_disabled) & RB_BUFFER_OFF);
3285 }
3286 
3287 /**
3288  * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
3289  * @buffer: The ring buffer to stop writes to.
3290  * @cpu: The CPU buffer to stop
3291  *
3292  * This prevents all writes to the buffer. Any attempt to write
3293  * to the buffer after this will fail and return NULL.
3294  *
3295  * The caller should call synchronize_rcu() after this.
3296  */
3297 void ring_buffer_record_disable_cpu(struct trace_buffer *buffer, int cpu)
3298 {
3299         struct ring_buffer_per_cpu *cpu_buffer;
3300 
3301         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3302                 return;
3303 
3304         cpu_buffer = buffer->buffers[cpu];
3305         atomic_inc(&cpu_buffer->record_disabled);
3306 }
3307 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
3308 
3309 /**
3310  * ring_buffer_record_enable_cpu - enable writes to the buffer
3311  * @buffer: The ring buffer to enable writes
3312  * @cpu: The CPU to enable.
3313  *
3314  * Note, multiple disables will need the same number of enables
3315  * to truly enable the writing (much like preempt_disable).
3316  */
3317 void ring_buffer_record_enable_cpu(struct trace_buffer *buffer, int cpu)
3318 {
3319         struct ring_buffer_per_cpu *cpu_buffer;
3320 
3321         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3322                 return;
3323 
3324         cpu_buffer = buffer->buffers[cpu];
3325         atomic_dec(&cpu_buffer->record_disabled);
3326 }
3327 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
3328 
3329 /*
3330  * The total entries in the ring buffer is the running counter
3331  * of entries entered into the ring buffer, minus the sum of
3332  * the entries read from the ring buffer and the number of
3333  * entries that were overwritten.
3334  */
3335 static inline unsigned long
3336 rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
3337 {
3338         return local_read(&cpu_buffer->entries) -
3339                 (local_read(&cpu_buffer->overrun) + cpu_buffer->read);
3340 }
3341 
3342 /**
3343  * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
3344  * @buffer: The ring buffer
3345  * @cpu: The per CPU buffer to read from.
3346  */
3347 u64 ring_buffer_oldest_event_ts(struct trace_buffer *buffer, int cpu)
3348 {
3349         unsigned long flags;
3350         struct ring_buffer_per_cpu *cpu_buffer;
3351         struct buffer_page *bpage;
3352         u64 ret = 0;
3353 
3354         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3355                 return 0;
3356 
3357         cpu_buffer = buffer->buffers[cpu];
3358         raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3359         /*
3360          * if the tail is on reader_page, oldest time stamp is on the reader
3361          * page
3362          */
3363         if (cpu_buffer->tail_page == cpu_buffer->reader_page)
3364                 bpage = cpu_buffer->reader_page;
3365         else
3366                 bpage = rb_set_head_page(cpu_buffer);
3367         if (bpage)
3368                 ret = bpage->page->time_stamp;
3369         raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3370 
3371         return ret;
3372 }
3373 EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
3374 
3375 /**
3376  * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
3377  * @buffer: The ring buffer
3378  * @cpu: The per CPU buffer to read from.
3379  */
3380 unsigned long ring_buffer_bytes_cpu(struct trace_buffer *buffer, int cpu)
3381 {
3382         struct ring_buffer_per_cpu *cpu_buffer;
3383         unsigned long ret;
3384 
3385         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3386                 return 0;
3387 
3388         cpu_buffer = buffer->buffers[cpu];
3389         ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
3390 
3391         return ret;
3392 }
3393 EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
3394 
3395 /**
3396  * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
3397  * @buffer: The ring buffer
3398  * @cpu: The per CPU buffer to get the entries from.
3399  */
3400 unsigned long ring_buffer_entries_cpu(struct trace_buffer *buffer, int cpu)
3401 {
3402         struct ring_buffer_per_cpu *cpu_buffer;
3403 
3404         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3405                 return 0;
3406 
3407         cpu_buffer = buffer->buffers[cpu];
3408 
3409         return rb_num_of_entries(cpu_buffer);
3410 }
3411 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
3412 
3413 /**
3414  * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
3415  * buffer wrapping around (only if RB_FL_OVERWRITE is on).
3416  * @buffer: The ring buffer
3417  * @cpu: The per CPU buffer to get the number of overruns from
3418  */
3419 unsigned long ring_buffer_overrun_cpu(struct trace_buffer *buffer, int cpu)
3420 {
3421         struct ring_buffer_per_cpu *cpu_buffer;
3422         unsigned long ret;
3423 
3424         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3425                 return 0;
3426 
3427         cpu_buffer = buffer->buffers[cpu];
3428         ret = local_read(&cpu_buffer->overrun);
3429 
3430         return ret;
3431 }
3432 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
3433 
3434 /**
3435  * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
3436  * commits failing due to the buffer wrapping around while there are uncommitted
3437  * events, such as during an interrupt storm.
3438  * @buffer: The ring buffer
3439  * @cpu: The per CPU buffer to get the number of overruns from
3440  */
3441 unsigned long
3442 ring_buffer_commit_overrun_cpu(struct trace_buffer *buffer, int cpu)
3443 {
3444         struct ring_buffer_per_cpu *cpu_buffer;
3445         unsigned long ret;
3446 
3447         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3448                 return 0;
3449 
3450         cpu_buffer = buffer->buffers[cpu];
3451         ret = local_read(&cpu_buffer->commit_overrun);
3452 
3453         return ret;
3454 }
3455 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
3456 
3457 /**
3458  * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
3459  * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
3460  * @buffer: The ring buffer
3461  * @cpu: The per CPU buffer to get the number of overruns from
3462  */
3463 unsigned long
3464 ring_buffer_dropped_events_cpu(struct trace_buffer *buffer, int cpu)
3465 {
3466         struct ring_buffer_per_cpu *cpu_buffer;
3467         unsigned long ret;
3468 
3469         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3470                 return 0;
3471 
3472         cpu_buffer = buffer->buffers[cpu];
3473         ret = local_read(&cpu_buffer->dropped_events);
3474 
3475         return ret;
3476 }
3477 EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
3478 
3479 /**
3480  * ring_buffer_read_events_cpu - get the number of events successfully read
3481  * @buffer: The ring buffer
3482  * @cpu: The per CPU buffer to get the number of events read
3483  */
3484 unsigned long
3485 ring_buffer_read_events_cpu(struct trace_buffer *buffer, int cpu)
3486 {
3487         struct ring_buffer_per_cpu *cpu_buffer;
3488 
3489         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3490                 return 0;
3491 
3492         cpu_buffer = buffer->buffers[cpu];
3493         return cpu_buffer->read;
3494 }
3495 EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu);
3496 
3497 /**
3498  * ring_buffer_entries - get the number of entries in a buffer
3499  * @buffer: The ring buffer
3500  *
3501  * Returns the total number of entries in the ring buffer
3502  * (all CPU entries)
3503  */
3504 unsigned long ring_buffer_entries(struct trace_buffer *buffer)
3505 {
3506         struct ring_buffer_per_cpu *cpu_buffer;
3507         unsigned long entries = 0;
3508         int cpu;
3509 
3510         /* if you care about this being correct, lock the buffer */
3511         for_each_buffer_cpu(buffer, cpu) {
3512                 cpu_buffer = buffer->buffers[cpu];
3513                 entries += rb_num_of_entries(cpu_buffer);
3514         }
3515 
3516         return entries;
3517 }
3518 EXPORT_SYMBOL_GPL(ring_buffer_entries);
3519 
3520 /**
3521  * ring_buffer_overruns - get the number of overruns in buffer
3522  * @buffer: The ring buffer
3523  *
3524  * Returns the total number of overruns in the ring buffer
3525  * (all CPU entries)
3526  */
3527 unsigned long ring_buffer_overruns(struct trace_buffer *buffer)
3528 {
3529         struct ring_buffer_per_cpu *cpu_buffer;
3530         unsigned long overruns = 0;
3531         int cpu;
3532 
3533         /* if you care about this being correct, lock the buffer */
3534         for_each_buffer_cpu(buffer, cpu) {
3535                 cpu_buffer = buffer->buffers[cpu];
3536                 overruns += local_read(&cpu_buffer->overrun);
3537         }
3538 
3539         return overruns;
3540 }
3541 EXPORT_SYMBOL_GPL(ring_buffer_overruns);
3542 
3543 static void rb_iter_reset(struct ring_buffer_iter *iter)
3544 {
3545         struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3546 
3547         /* Iterator usage is expected to have record disabled */
3548         iter->head_page = cpu_buffer->reader_page;
3549         iter->head = cpu_buffer->reader_page->read;
3550 
3551         iter->cache_reader_page = iter->head_page;
3552         iter->cache_read = cpu_buffer->read;
3553 
3554         if (iter->head)
3555                 iter->read_stamp = cpu_buffer->read_stamp;
3556         else
3557                 iter->read_stamp = iter->head_page->page->time_stamp;
3558 }
3559 
3560 /**
3561  * ring_buffer_iter_reset - reset an iterator
3562  * @iter: The iterator to reset
3563  *
3564  * Resets the iterator, so that it will start from the beginning
3565  * again.
3566  */
3567 void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
3568 {
3569         struct ring_buffer_per_cpu *cpu_buffer;
3570         unsigned long flags;
3571 
3572         if (!iter)
3573                 return;
3574 
3575         cpu_buffer = iter->cpu_buffer;
3576 
3577         raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3578         rb_iter_reset(iter);
3579         raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3580 }
3581 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
3582 
3583 /**
3584  * ring_buffer_iter_empty - check if an iterator has no more to read
3585  * @iter: The iterator to check
3586  */
3587 int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
3588 {
3589         struct ring_buffer_per_cpu *cpu_buffer;
3590         struct buffer_page *reader;
3591         struct buffer_page *head_page;
3592         struct buffer_page *commit_page;
3593         unsigned commit;
3594 
3595         cpu_buffer = iter->cpu_buffer;
3596 
3597         /* Remember, trace recording is off when iterator is in use */
3598         reader = cpu_buffer->reader_page;
3599         head_page = cpu_buffer->head_page;
3600         commit_page = cpu_buffer->commit_page;
3601         commit = rb_page_commit(commit_page);
3602 
3603         return ((iter->head_page == commit_page && iter->head == commit) ||
3604                 (iter->head_page == reader && commit_page == head_page &&
3605                  head_page->read == commit &&
3606                  iter->head == rb_page_commit(cpu_buffer->reader_page)));
3607 }
3608 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
3609 
3610 static void
3611 rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
3612                      struct ring_buffer_event *event)
3613 {
3614         u64 delta;
3615 
3616         switch (event->type_len) {
3617         case RINGBUF_TYPE_PADDING:
3618                 return;
3619 
3620         case RINGBUF_TYPE_TIME_EXTEND:
3621                 delta = ring_buffer_event_time_stamp(event);
3622                 cpu_buffer->read_stamp += delta;
3623                 return;
3624 
3625         case RINGBUF_TYPE_TIME_STAMP:
3626                 delta = ring_buffer_event_time_stamp(event);
3627                 cpu_buffer->read_stamp = delta;
3628                 return;
3629 
3630         case RINGBUF_TYPE_DATA:
3631                 cpu_buffer->read_stamp += event->time_delta;
3632                 return;
3633 
3634         default:
3635                 BUG();
3636         }
3637         return;
3638 }
3639 
3640 static void
3641 rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
3642                           struct ring_buffer_event *event)
3643 {
3644         u64 delta;
3645 
3646         switch (event->type_len) {
3647         case RINGBUF_TYPE_PADDING:
3648                 return;
3649 
3650         case RINGBUF_TYPE_TIME_EXTEND:
3651                 delta = ring_buffer_event_time_stamp(event);
3652                 iter->read_stamp += delta;
3653                 return;
3654 
3655         case RINGBUF_TYPE_TIME_STAMP:
3656                 delta = ring_buffer_event_time_stamp(event);
3657                 iter->read_stamp = delta;
3658                 return;
3659 
3660         case RINGBUF_TYPE_DATA:
3661                 iter->read_stamp += event->time_delta;
3662                 return;
3663 
3664         default:
3665                 BUG();
3666         }
3667         return;
3668 }
3669 
3670 static struct buffer_page *
3671 rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
3672 {
3673         struct buffer_page *reader = NULL;
3674         unsigned long overwrite;
3675         unsigned long flags;
3676         int nr_loops = 0;
3677         int ret;
3678 
3679         local_irq_save(flags);
3680         arch_spin_lock(&cpu_buffer->lock);
3681 
3682  again:
3683         /*
3684          * This should normally only loop twice. But because the
3685          * start of the reader inserts an empty page, it causes
3686          * a case where we will loop three times. There should be no
3687          * reason to loop four times (that I know of).
3688          */
3689         if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
3690                 reader = NULL;
3691                 goto out;
3692         }
3693 
3694         reader = cpu_buffer->reader_page;
3695 
3696         /* If there's more to read, return this page */
3697         if (cpu_buffer->reader_page->read < rb_page_size(reader))
3698                 goto out;
3699 
3700         /* Never should we have an index greater than the size */
3701         if (RB_WARN_ON(cpu_buffer,
3702                        cpu_buffer->reader_page->read > rb_page_size(reader)))
3703                 goto out;
3704 
3705         /* check if we caught up to the tail */
3706         reader = NULL;
3707         if (cpu_buffer->commit_page == cpu_buffer->reader_page)
3708                 goto out;
3709 
3710         /* Don't bother swapping if the ring buffer is empty */
3711         if (rb_num_of_entries(cpu_buffer) == 0)
3712                 goto out;
3713 
3714         /*
3715          * Reset the reader page to size zero.
3716          */
3717         local_set(&cpu_buffer->reader_page->write, 0);
3718         local_set(&cpu_buffer->reader_page->entries, 0);
3719         local_set(&cpu_buffer->reader_page->page->commit, 0);
3720         cpu_buffer->reader_page->real_end = 0;
3721 
3722  spin:
3723         /*
3724          * Splice the empty reader page into the list around the head.
3725          */
3726         reader = rb_set_head_page(cpu_buffer);
3727         if (!reader)
3728                 goto out;
3729         cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
3730         cpu_buffer->reader_page->list.prev = reader->list.prev;
3731 
3732         /*
3733          * cpu_buffer->pages just needs to point to the buffer, it
3734          *  has no specific buffer page to point to. Lets move it out
3735          *  of our way so we don't accidentally swap it.
3736          */
3737         cpu_buffer->pages = reader->list.prev;
3738 
3739         /* The reader page will be pointing to the new head */
3740         rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list);
3741 
3742         /*
3743          * We want to make sure we read the overruns after we set up our
3744          * pointers to the next object. The writer side does a
3745          * cmpxchg to cross pages which acts as the mb on the writer
3746          * side. Note, the reader will constantly fail the swap
3747          * while the writer is updating the pointers, so this
3748          * guarantees that the overwrite recorded here is the one we
3749          * want to compare with the last_overrun.
3750          */
3751         smp_mb();
3752         overwrite = local_read(&(cpu_buffer->overrun));
3753 
3754         /*
3755          * Here's the tricky part.
3756          *
3757          * We need to move the pointer past the header page.
3758          * But we can only do that if a writer is not currently
3759          * moving it. The page before the header page has the
3760          * flag bit '1' set if it is pointing to the page we want.
3761          * but if the writer is in the process of moving it
3762          * than it will be '2' or already moved ''.
3763          */
3764 
3765         ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
3766 
3767         /*
3768          * If we did not convert it, then we must try again.
3769          */
3770         if (!ret)
3771                 goto spin;
3772 
3773         /*
3774          * Yay! We succeeded in replacing the page.
3775          *
3776          * Now make the new head point back to the reader page.
3777          */
3778         rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
3779         rb_inc_page(cpu_buffer, &cpu_buffer->head_page);
3780 
3781         local_inc(&cpu_buffer->pages_read);
3782 
3783         /* Finally update the reader page to the new head */
3784         cpu_buffer->reader_page = reader;
3785         cpu_buffer->reader_page->read = 0;
3786 
3787         if (overwrite != cpu_buffer->last_overrun) {
3788                 cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
3789                 cpu_buffer->last_overrun = overwrite;
3790         }
3791 
3792         goto again;
3793 
3794  out:
3795         /* Update the read_stamp on the first event */
3796         if (reader && reader->read == 0)
3797                 cpu_buffer->read_stamp = reader->page->time_stamp;
3798 
3799         arch_spin_unlock(&cpu_buffer->lock);
3800         local_irq_restore(flags);
3801 
3802         return reader;
3803 }
3804 
3805 static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
3806 {
3807         struct ring_buffer_event *event;
3808         struct buffer_page *reader;
3809         unsigned length;
3810 
3811         reader = rb_get_reader_page(cpu_buffer);
3812 
3813         /* This function should not be called when buffer is empty */
3814         if (RB_WARN_ON(cpu_buffer, !reader))
3815                 return;
3816 
3817         event = rb_reader_event(cpu_buffer);
3818 
3819         if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
3820                 cpu_buffer->read++;
3821 
3822         rb_update_read_stamp(cpu_buffer, event);
3823 
3824         length = rb_event_length(event);
3825         cpu_buffer->reader_page->read += length;
3826 }
3827 
3828 static void rb_advance_iter(struct ring_buffer_iter *iter)
3829 {
3830         struct ring_buffer_per_cpu *cpu_buffer;
3831         struct ring_buffer_event *event;
3832         unsigned length;
3833 
3834         cpu_buffer = iter->cpu_buffer;
3835 
3836         /*
3837          * Check if we are at the end of the buffer.
3838          */
3839         if (iter->head >= rb_page_size(iter->head_page)) {
3840                 /* discarded commits can make the page empty */
3841                 if (iter->head_page == cpu_buffer->commit_page)
3842                         return;
3843                 rb_inc_iter(iter);
3844                 return;
3845         }
3846 
3847         event = rb_iter_head_event(iter);
3848 
3849         length = rb_event_length(event);
3850 
3851         /*
3852          * This should not be called to advance the header if we are
3853          * at the tail of the buffer.
3854          */
3855         if (RB_WARN_ON(cpu_buffer,
3856                        (iter->head_page == cpu_buffer->commit_page) &&
3857                        (iter->head + length > rb_commit_index(cpu_buffer))))
3858                 return;
3859 
3860         rb_update_iter_read_stamp(iter, event);
3861 
3862         iter->head += length;
3863 
3864         /* check for end of page padding */
3865         if ((iter->head >= rb_page_size(iter->head_page)) &&
3866             (iter->head_page != cpu_buffer->commit_page))
3867                 rb_inc_iter(iter);
3868 }
3869 
3870 static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
3871 {
3872         return cpu_buffer->lost_events;
3873 }
3874 
3875 static struct ring_buffer_event *
3876 rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
3877                unsigned long *lost_events)
3878 {
3879         struct ring_buffer_event *event;
3880         struct buffer_page *reader;
3881         int nr_loops = 0;
3882 
3883         if (ts)
3884                 *ts = 0;
3885  again:
3886         /*
3887          * We repeat when a time extend is encountered.
3888          * Since the time extend is always attached to a data event,
3889          * we should never loop more than once.
3890          * (We never hit the following condition more than twice).
3891          */
3892         if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3893                 return NULL;
3894 
3895         reader = rb_get_reader_page(cpu_buffer);
3896         if (!reader)
3897                 return NULL;
3898 
3899         event = rb_reader_event(cpu_buffer);
3900 
3901         switch (event->type_len) {
3902         case RINGBUF_TYPE_PADDING:
3903                 if (rb_null_event(event))
3904                         RB_WARN_ON(cpu_buffer, 1);
3905                 /*
3906                  * Because the writer could be discarding every
3907                  * event it creates (which would probably be bad)
3908                  * if we were to go back to "again" then we may never
3909                  * catch up, and will trigger the warn on, or lock
3910                  * the box. Return the padding, and we will release
3911                  * the current locks, and try again.
3912                  */
3913                 return event;
3914 
3915         case RINGBUF_TYPE_TIME_EXTEND:
3916                 /* Internal data, OK to advance */
3917                 rb_advance_reader(cpu_buffer);
3918                 goto again;
3919 
3920         case RINGBUF_TYPE_TIME_STAMP:
3921                 if (ts) {
3922                         *ts = ring_buffer_event_time_stamp(event);
3923                         ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3924                                                          cpu_buffer->cpu, ts);
3925                 }
3926                 /* Internal data, OK to advance */
3927                 rb_advance_reader(cpu_buffer);
3928                 goto again;
3929 
3930         case RINGBUF_TYPE_DATA:
3931                 if (ts && !(*ts)) {
3932                         *ts = cpu_buffer->read_stamp + event->time_delta;
3933                         ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3934                                                          cpu_buffer->cpu, ts);
3935                 }
3936                 if (lost_events)
3937                         *lost_events = rb_lost_events(cpu_buffer);
3938                 return event;
3939 
3940         default:
3941                 BUG();
3942         }
3943 
3944         return NULL;
3945 }
3946 EXPORT_SYMBOL_GPL(ring_buffer_peek);
3947 
3948 static struct ring_buffer_event *
3949 rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3950 {
3951         struct trace_buffer *buffer;
3952         struct ring_buffer_per_cpu *cpu_buffer;
3953         struct ring_buffer_event *event;
3954         int nr_loops = 0;
3955 
3956         if (ts)
3957                 *ts = 0;
3958 
3959         cpu_buffer = iter->cpu_buffer;
3960         buffer = cpu_buffer->buffer;
3961 
3962         /*
3963          * Check if someone performed a consuming read to
3964          * the buffer. A consuming read invalidates the iterator
3965          * and we need to reset the iterator in this case.
3966          */
3967         if (unlikely(iter->cache_read != cpu_buffer->read ||
3968                      iter->cache_reader_page != cpu_buffer->reader_page))
3969                 rb_iter_reset(iter);
3970 
3971  again:
3972         if (ring_buffer_iter_empty(iter))
3973                 return NULL;
3974 
3975         /*
3976          * We repeat when a time extend is encountered or we hit
3977          * the end of the page. Since the time extend is always attached
3978          * to a data event, we should never loop more than three times.
3979          * Once for going to next page, once on time extend, and
3980          * finally once to get the event.
3981          * (We never hit the following condition more than thrice).
3982          */
3983         if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3))
3984                 return NULL;
3985 
3986         if (rb_per_cpu_empty(cpu_buffer))
3987                 return NULL;
3988 
3989         if (iter->head >= rb_page_size(iter->head_page)) {
3990                 rb_inc_iter(iter);
3991                 goto again;
3992         }
3993 
3994         event = rb_iter_head_event(iter);
3995 
3996         switch (event->type_len) {
3997         case RINGBUF_TYPE_PADDING:
3998                 if (rb_null_event(event)) {
3999                         rb_inc_iter(iter);
4000                         goto again;
4001                 }
4002                 rb_advance_iter(iter);
4003                 return event;
4004 
4005         case RINGBUF_TYPE_TIME_EXTEND:
4006                 /* Internal data, OK to advance */
4007                 rb_advance_iter(iter);
4008                 goto again;
4009 
4010         case RINGBUF_TYPE_TIME_STAMP:
4011                 if (ts) {
4012                         *ts = ring_buffer_event_time_stamp(event);
4013                         ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
4014                                                          cpu_buffer->cpu, ts);
4015                 }
4016                 /* Internal data, OK to advance */
4017                 rb_advance_iter(iter);
4018                 goto again;
4019 
4020         case RINGBUF_TYPE_DATA:
4021                 if (ts && !(*ts)) {
4022                         *ts = iter->read_stamp + event->time_delta;
4023                         ring_buffer_normalize_time_stamp(buffer,
4024                                                          cpu_buffer->cpu, ts);
4025                 }
4026                 return event;
4027 
4028         default:
4029                 BUG();
4030         }
4031 
4032         return NULL;
4033 }
4034 EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
4035 
4036 static inline bool rb_reader_lock(struct ring_buffer_per_cpu *cpu_buffer)
4037 {
4038         if (likely(!in_nmi())) {
4039                 raw_spin_lock(&cpu_buffer->reader_lock);
4040                 return true;
4041         }
4042 
4043         /*
4044          * If an NMI die dumps out the content of the ring buffer
4045          * trylock must be used to prevent a deadlock if the NMI
4046          * preempted a task that holds the ring buffer locks. If
4047          * we get the lock then all is fine, if not, then continue
4048          * to do the read, but this can corrupt the ring buffer,
4049          * so it must be permanently disabled from future writes.
4050          * Reading from NMI is a oneshot deal.
4051          */
4052         if (raw_spin_trylock(&cpu_buffer->reader_lock))
4053                 return true;
4054 
4055         /* Continue without locking, but disable the ring buffer */
4056         atomic_inc(&cpu_buffer->record_disabled);
4057         return false;
4058 }
4059 
4060 static inline void
4061 rb_reader_unlock(struct ring_buffer_per_cpu *cpu_buffer, bool locked)
4062 {
4063         if (likely(locked))
4064                 raw_spin_unlock(&cpu_buffer->reader_lock);
4065         return;
4066 }
4067 
4068 /**
4069  * ring_buffer_peek - peek at the next event to be read
4070  * @buffer: The ring buffer to read
4071  * @cpu: The cpu to peak at
4072  * @ts: The timestamp counter of this event.
4073  * @lost_events: a variable to store if events were lost (may be NULL)
4074  *
4075  * This will return the event that will be read next, but does
4076  * not consume the data.
4077  */
4078 struct ring_buffer_event *
4079 ring_buffer_peek(struct trace_buffer *buffer, int cpu, u64 *ts,
4080                  unsigned long *lost_events)
4081 {
4082         struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4083         struct ring_buffer_event *event;
4084         unsigned long flags;
4085         bool dolock;
4086 
4087         if (!cpumask_test_cpu(cpu, buffer->cpumask))
4088                 return NULL;
4089 
4090  again:
4091         local_irq_save(flags);
4092         dolock = rb_reader_lock(cpu_buffer);
4093         event = rb_buffer_peek(cpu_buffer, ts, lost_events);
4094         if (event && event->type_len == RINGBUF_TYPE_PADDING)
4095                 rb_advance_reader(cpu_buffer);
4096         rb_reader_unlock(cpu_buffer, dolock);
4097         local_irq_restore(flags);
4098 
4099         if (event && event->type_len == RINGBUF_TYPE_PADDING)
4100                 goto again;
4101 
4102         return event;
4103 }
4104 
4105 /**
4106  * ring_buffer_iter_peek - peek at the next event to be read
4107  * @iter: The ring buffer iterator
4108  * @ts: The timestamp counter of this event.
4109  *
4110  * This will return the event that will be read next, but does
4111  * not increment the iterator.
4112  */
4113 struct ring_buffer_event *
4114 ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
4115 {
4116         struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4117         struct ring_buffer_event *event;
4118         unsigned long flags;
4119 
4120  again:
4121         raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4122         event = rb_iter_peek(iter, ts);
4123         raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4124 
4125         if (event && event->type_len == RINGBUF_TYPE_PADDING)
4126                 goto again;
4127 
4128         return event;
4129 }
4130 
4131 /**
4132  * ring_buffer_consume - return an event and consume it
4133  * @buffer: The ring buffer to get the next event from
4134  * @cpu: the cpu to read the buffer from
4135  * @ts: a variable to store the timestamp (may be NULL)
4136  * @lost_events: a variable to store if events were lost (may be NULL)
4137  *
4138  * Returns the next event in the ring buffer, and that event is consumed.
4139  * Meaning, that sequential reads will keep returning a different event,
4140  * and eventually empty the ring buffer if the producer is slower.
4141  */
4142 struct ring_buffer_event *
4143 ring_buffer_consume(struct trace_buffer *buffer, int cpu, u64 *ts,
4144                     unsigned long *lost_events)
4145 {
4146         struct ring_buffer_per_cpu *cpu_buffer;
4147         struct ring_buffer_event *event = NULL;
4148         unsigned long flags;
4149         bool dolock;
4150 
4151  again:
4152         /* might be called in atomic */
4153         preempt_disable();
4154 
4155         if (!cpumask_test_cpu(cpu, buffer->cpumask))
4156                 goto out;
4157 
4158         cpu_buffer = buffer->buffers[cpu];
4159         local_irq_save(flags);
4160         dolock = rb_reader_lock(cpu_buffer);
4161 
4162         event = rb_buffer_peek(cpu_buffer, ts, lost_events);
4163         if (event) {
4164                 cpu_buffer->lost_events = 0;
4165                 rb_advance_reader(cpu_buffer);
4166         }
4167 
4168         rb_reader_unlock(cpu_buffer, dolock);
4169         local_irq_restore(flags);
4170 
4171  out:
4172         preempt_enable();
4173 
4174         if (event && event->type_len == RINGBUF_TYPE_PADDING)
4175                 goto again;
4176 
4177         return event;
4178 }
4179 EXPORT_SYMBOL_GPL(ring_buffer_consume);
4180 
4181 /**
4182  * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
4183  * @buffer: The ring buffer to read from
4184  * @cpu: The cpu buffer to iterate over
4185  * @flags: gfp flags to use for memory allocation
4186  *
4187  * This performs the initial preparations necessary to iterate
4188  * through the buffer.  Memory is allocated, buffer recording
4189  * is disabled, and the iterator pointer is returned to the caller.
4190  *
4191  * Disabling buffer recording prevents the reading from being
4192  * corrupted. This is not a consuming read, so a producer is not
4193  * expected.
4194  *
4195  * After a sequence of ring_buffer_read_prepare calls, the user is
4196  * expected to make at least one call to ring_buffer_read_prepare_sync.
4197  * Afterwards, ring_buffer_read_start is invoked to get things going
4198  * for real.
4199  *
4200  * This overall must be paired with ring_buffer_read_finish.
4201  */
4202 struct ring_buffer_iter *
4203 ring_buffer_read_prepare(struct trace_buffer *buffer, int cpu, gfp_t flags)
4204 {
4205         struct ring_buffer_per_cpu *cpu_buffer;
4206         struct ring_buffer_iter *iter;
4207 
4208         if (!cpumask_test_cpu(cpu, buffer->cpumask))
4209                 return NULL;
4210 
4211         iter = kmalloc(sizeof(*iter), flags);
4212         if (!iter)
4213                 return NULL;
4214 
4215         cpu_buffer = buffer->buffers[cpu];
4216 
4217         iter->cpu_buffer = cpu_buffer;
4218 
4219         atomic_inc(&buffer->resize_disabled);
4220         atomic_inc(&cpu_buffer->record_disabled);
4221 
4222         return iter;
4223 }
4224 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
4225 
4226 /**
4227  * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
4228  *
4229  * All previously invoked ring_buffer_read_prepare calls to prepare
4230  * iterators will be synchronized.  Afterwards, read_buffer_read_start
4231  * calls on those iterators are allowed.
4232  */
4233 void
4234 ring_buffer_read_prepare_sync(void)
4235 {
4236         synchronize_rcu();
4237 }
4238 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
4239 
4240 /**
4241  * ring_buffer_read_start - start a non consuming read of the buffer
4242  * @iter: The iterator returned by ring_buffer_read_prepare
4243  *
4244  * This finalizes the startup of an iteration through the buffer.
4245  * The iterator comes from a call to ring_buffer_read_prepare and
4246  * an intervening ring_buffer_read_prepare_sync must have been
4247  * performed.
4248  *
4249  * Must be paired with ring_buffer_read_finish.
4250  */
4251 void
4252 ring_buffer_read_start(struct ring_buffer_iter *iter)
4253 {
4254         struct ring_buffer_per_cpu *cpu_buffer;
4255         unsigned long flags;
4256 
4257         if (!iter)
4258                 return;
4259 
4260         cpu_buffer = iter->cpu_buffer;
4261 
4262         raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4263         arch_spin_lock(&cpu_buffer->lock);
4264         rb_iter_reset(iter);
4265         arch_spin_unlock(&cpu_buffer->lock);
4266         raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4267 }
4268 EXPORT_SYMBOL_GPL(ring_buffer_read_start);
4269 
4270 /**
4271  * ring_buffer_read_finish - finish reading the iterator of the buffer
4272  * @iter: The iterator retrieved by ring_buffer_start
4273  *
4274  * This re-enables the recording to the buffer, and frees the
4275  * iterator.
4276  */
4277 void
4278 ring_buffer_read_finish(struct ring_buffer_iter *iter)
4279 {
4280         struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4281         unsigned long flags;
4282 
4283         /*
4284          * Ring buffer is disabled from recording, here's a good place
4285          * to check the integrity of the ring buffer.
4286          * Must prevent readers from trying to read, as the check
4287          * clears the HEAD page and readers require it.
4288          */
4289         raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4290         rb_check_pages(cpu_buffer);
4291         raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4292 
4293         atomic_dec(&cpu_buffer->record_disabled);
4294         atomic_dec(&cpu_buffer->buffer->resize_disabled);
4295         kfree(iter);
4296 }
4297 EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
4298 
4299 /**
4300  * ring_buffer_read - read the next item in the ring buffer by the iterator
4301  * @iter: The ring buffer iterator
4302  * @ts: The time stamp of the event read.
4303  *
4304  * This reads the next event in the ring buffer and increments the iterator.
4305  */
4306 struct ring_buffer_event *
4307 ring_buffer_read(struct ring_buffer_iter *iter, u64 *ts)
4308 {
4309         struct ring_buffer_event *event;
4310         struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4311         unsigned long flags;
4312 
4313         raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4314  again:
4315         event = rb_iter_peek(iter, ts);
4316         if (!event)
4317                 goto out;
4318 
4319         if (event->type_len == RINGBUF_TYPE_PADDING)
4320                 goto again;
4321 
4322         rb_advance_iter(iter);
4323  out:
4324         raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4325 
4326         return event;
4327 }
4328 EXPORT_SYMBOL_GPL(ring_buffer_read);
4329 
4330 /**
4331  * ring_buffer_size - return the size of the ring buffer (in bytes)
4332  * @buffer: The ring buffer.
4333  * @cpu: The CPU to get ring buffer size from.
4334  */
4335 unsigned long ring_buffer_size(struct trace_buffer *buffer, int cpu)
4336 {
4337         /*
4338          * Earlier, this method returned
4339          *      BUF_PAGE_SIZE * buffer->nr_pages
4340          * Since the nr_pages field is now removed, we have converted this to
4341          * return the per cpu buffer value.
4342          */
4343         if (!cpumask_test_cpu(cpu, buffer->cpumask))
4344                 return 0;
4345 
4346         return BUF_PAGE_SIZE * buffer->buffers[cpu]->nr_pages;
4347 }
4348 EXPORT_SYMBOL_GPL(ring_buffer_size);
4349 
4350 static void
4351 rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
4352 {
4353         rb_head_page_deactivate(cpu_buffer);
4354 
4355         cpu_buffer->head_page
4356                 = list_entry(cpu_buffer->pages, struct buffer_page, list);
4357         local_set(&cpu_buffer->head_page->write, 0);
4358         local_set(&cpu_buffer->head_page->entries, 0);
4359         local_set(&cpu_buffer->head_page->page->commit, 0);
4360 
4361         cpu_buffer->head_page->read = 0;
4362 
4363         cpu_buffer->tail_page = cpu_buffer->head_page;
4364         cpu_buffer->commit_page = cpu_buffer->head_page;
4365 
4366         INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
4367         INIT_LIST_HEAD(&cpu_buffer->new_pages);
4368         local_set(&cpu_buffer->reader_page->write, 0);
4369         local_set(&cpu_buffer->reader_page->entries, 0);
4370         local_set(&cpu_buffer->reader_page->page->commit, 0);
4371         cpu_buffer->reader_page->read = 0;
4372 
4373         local_set(&cpu_buffer->entries_bytes, 0);
4374         local_set(&cpu_buffer->overrun, 0);
4375         local_set(&cpu_buffer->commit_overrun, 0);
4376         local_set(&cpu_buffer->dropped_events, 0);
4377         local_set(&cpu_buffer->entries, 0);
4378         local_set(&cpu_buffer->committing, 0);
4379         local_set(&cpu_buffer->commits, 0);
4380         local_set(&cpu_buffer->pages_touched, 0);
4381         local_set(&cpu_buffer->pages_read, 0);
4382         cpu_buffer->last_pages_touch = 0;
4383         cpu_buffer->shortest_full = 0;
4384         cpu_buffer->read = 0;
4385         cpu_buffer->read_bytes = 0;
4386 
4387         cpu_buffer->write_stamp = 0;
4388         cpu_buffer->read_stamp = 0;
4389 
4390         cpu_buffer->lost_events = 0;
4391         cpu_buffer->last_overrun = 0;
4392 
4393         rb_head_page_activate(cpu_buffer);
4394 }
4395 
4396 /**
4397  * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
4398  * @buffer: The ring buffer to reset a per cpu buffer of
4399  * @cpu: The CPU buffer to be reset
4400  */
4401 void ring_buffer_reset_cpu(struct trace_buffer *buffer, int cpu)
4402 {
4403         struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4404         unsigned long flags;
4405 
4406         if (!cpumask_test_cpu(cpu, buffer->cpumask))
4407                 return;
4408 
4409         atomic_inc(&buffer->resize_disabled);
4410         atomic_inc(&cpu_buffer->record_disabled);
4411 
4412         /* Make sure all commits have finished */
4413         synchronize_rcu();
4414 
4415         raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4416 
4417         if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
4418                 goto out;
4419 
4420         arch_spin_lock(&cpu_buffer->lock);
4421 
4422         rb_reset_cpu(cpu_buffer);
4423 
4424         arch_spin_unlock(&cpu_buffer->lock);
4425 
4426  out:
4427         raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4428 
4429         atomic_dec(&cpu_buffer->record_disabled);
4430         atomic_dec(&buffer->resize_disabled);
4431 }
4432 EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
4433 
4434 /**
4435  * ring_buffer_reset - reset a ring buffer
4436  * @buffer: The ring buffer to reset all cpu buffers
4437  */
4438 void ring_buffer_reset(struct trace_buffer *buffer)
4439 {
4440         int cpu;
4441 
4442         for_each_buffer_cpu(buffer, cpu)
4443                 ring_buffer_reset_cpu(buffer, cpu);
4444 }
4445 EXPORT_SYMBOL_GPL(ring_buffer_reset);
4446 
4447 /**
4448  * rind_buffer_empty - is the ring buffer empty?
4449  * @buffer: The ring buffer to test
4450  */
4451 bool ring_buffer_empty(struct trace_buffer *buffer)
4452 {
4453         struct ring_buffer_per_cpu *cpu_buffer;
4454         unsigned long flags;
4455         bool dolock;
4456         int cpu;
4457         int ret;
4458 
4459         /* yes this is racy, but if you don't like the race, lock the buffer */
4460         for_each_buffer_cpu(buffer, cpu) {
4461                 cpu_buffer = buffer->buffers[cpu];
4462                 local_irq_save(flags);
4463                 dolock = rb_reader_lock(cpu_buffer);
4464                 ret = rb_per_cpu_empty(cpu_buffer);
4465                 rb_reader_unlock(cpu_buffer, dolock);
4466                 local_irq_restore(flags);
4467 
4468                 if (!ret)
4469                         return false;
4470         }
4471 
4472         return true;
4473 }
4474 EXPORT_SYMBOL_GPL(ring_buffer_empty);
4475 
4476 /**
4477  * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
4478  * @buffer: The ring buffer
4479  * @cpu: The CPU buffer to test
4480  */
4481 bool ring_buffer_empty_cpu(struct trace_buffer *buffer, int cpu)
4482 {
4483         struct ring_buffer_per_cpu *cpu_buffer;
4484         unsigned long flags;
4485         bool dolock;
4486         int ret;
4487 
4488         if (!cpumask_test_cpu(cpu, buffer->cpumask))
4489                 return true;
4490 
4491         cpu_buffer = buffer->buffers[cpu];
4492         local_irq_save(flags);
4493         dolock = rb_reader_lock(cpu_buffer);
4494         ret = rb_per_cpu_empty(cpu_buffer);
4495         rb_reader_unlock(cpu_buffer, dolock);
4496         local_irq_restore(flags);
4497 
4498         return ret;
4499 }
4500 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
4501 
4502 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
4503 /**
4504  * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
4505  * @buffer_a: One buffer to swap with
4506  * @buffer_b: The other buffer to swap with
4507  * @cpu: the CPU of the buffers to swap
4508  *
4509  * This function is useful for tracers that want to take a "snapshot"
4510  * of a CPU buffer and has another back up buffer lying around.
4511  * it is expected that the tracer handles the cpu buffer not being
4512  * used at the moment.
4513  */
4514 int ring_buffer_swap_cpu(struct trace_buffer *buffer_a,
4515                          struct trace_buffer *buffer_b, int cpu)
4516 {
4517         struct ring_buffer_per_cpu *cpu_buffer_a;
4518         struct ring_buffer_per_cpu *cpu_buffer_b;
4519         int ret = -EINVAL;
4520 
4521         if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
4522             !cpumask_test_cpu(cpu, buffer_b->cpumask))
4523                 goto out;
4524 
4525         cpu_buffer_a = buffer_a->buffers[cpu];
4526         cpu_buffer_b = buffer_b->buffers[cpu];
4527 
4528         /* At least make sure the two buffers are somewhat the same */
4529         if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
4530                 goto out;
4531 
4532         ret = -EAGAIN;
4533 
4534         if (atomic_read(&buffer_a->record_disabled))
4535                 goto out;
4536 
4537         if (atomic_read(&buffer_b->record_disabled))
4538                 goto out;
4539 
4540         if (atomic_read(&cpu_buffer_a->record_disabled))
4541                 goto out;
4542 
4543         if (atomic_read(&cpu_buffer_b->record_disabled))
4544                 goto out;
4545 
4546         /*
4547          * We can't do a synchronize_rcu here because this
4548          * function can be called in atomic context.
4549          * Normally this will be called from the same CPU as cpu.
4550          * If not it's up to the caller to protect this.
4551          */
4552         atomic_inc(&cpu_buffer_a->record_disabled);
4553         atomic_inc(&cpu_buffer_b->record_disabled);
4554 
4555         ret = -EBUSY;
4556         if (local_read(&cpu_buffer_a->committing))
4557                 goto out_dec;
4558         if (local_read(&cpu_buffer_b->committing))
4559                 goto out_dec;
4560 
4561         buffer_a->buffers[cpu] = cpu_buffer_b;
4562         buffer_b->buffers[cpu] = cpu_buffer_a;
4563 
4564         cpu_buffer_b->buffer = buffer_a;
4565         cpu_buffer_a->buffer = buffer_b;
4566 
4567         ret = 0;
4568 
4569 out_dec:
4570         atomic_dec(&cpu_buffer_a->record_disabled);
4571         atomic_dec(&cpu_buffer_b->record_disabled);
4572 out:
4573         return ret;
4574 }
4575 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
4576 #endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
4577 
4578 /**
4579  * ring_buffer_alloc_read_page - allocate a page to read from buffer
4580  * @buffer: the buffer to allocate for.
4581  * @cpu: the cpu buffer to allocate.
4582  *
4583  * This function is used in conjunction with ring_buffer_read_page.
4584  * When reading a full page from the ring buffer, these functions
4585  * can be used to speed up the process. The calling function should
4586  * allocate a few pages first with this function. Then when it
4587  * needs to get pages from the ring buffer, it passes the result
4588  * of this function into ring_buffer_read_page, which will swap
4589  * the page that was allocated, with the read page of the buffer.
4590  *
4591  * Returns:
4592  *  The page allocated, or ERR_PTR
4593  */
4594 void *ring_buffer_alloc_read_page(struct trace_buffer *buffer, int cpu)
4595 {
4596         struct ring_buffer_per_cpu *cpu_buffer;
4597         struct buffer_data_page *bpage = NULL;
4598         unsigned long flags;
4599         struct page *page;
4600 
4601         if (!cpumask_test_cpu(cpu, buffer->cpumask))
4602                 return ERR_PTR(-ENODEV);
4603 
4604         cpu_buffer = buffer->buffers[cpu];
4605         local_irq_save(flags);
4606         arch_spin_lock(&cpu_buffer->lock);
4607 
4608         if (cpu_buffer->free_page) {
4609                 bpage = cpu_buffer->free_page;
4610                 cpu_buffer->free_page = NULL;
4611         }
4612 
4613         arch_spin_unlock(&cpu_buffer->lock);
4614         local_irq_restore(flags);
4615 
4616         if (bpage)
4617                 goto out;
4618 
4619         page = alloc_pages_node(cpu_to_node(cpu),
4620                                 GFP_KERNEL | __GFP_NORETRY, 0);
4621         if (!page)
4622                 return ERR_PTR(-ENOMEM);
4623 
4624         bpage = page_address(page);
4625 
4626  out:
4627         rb_init_page(bpage);
4628 
4629         return bpage;
4630 }
4631 EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
4632 
4633 /**
4634  * ring_buffer_free_read_page - free an allocated read page
4635  * @buffer: the buffer the page was allocate for
4636  * @cpu: the cpu buffer the page came from
4637  * @data: the page to free
4638  *
4639  * Free a page allocated from ring_buffer_alloc_read_page.
4640  */
4641 void ring_buffer_free_read_page(struct trace_buffer *buffer, int cpu, void *data)
4642 {
4643         struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4644         struct buffer_data_page *bpage = data;
4645         struct page *page = virt_to_page(bpage);
4646         unsigned long flags;
4647 
4648         /* If the page is still in use someplace else, we can't reuse it */
4649         if (page_ref_count(page) > 1)
4650                 goto out;
4651 
4652         local_irq_save(flags);
4653         arch_spin_lock(&cpu_buffer->lock);
4654 
4655         if (!cpu_buffer->free_page) {
4656                 cpu_buffer->free_page = bpage;
4657                 bpage = NULL;
4658         }
4659 
4660         arch_spin_unlock(&cpu_buffer->lock);
4661         local_irq_restore(flags);
4662 
4663  out:
4664         free_page((unsigned long)bpage);
4665 }
4666 EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
4667 
4668 /**
4669  * ring_buffer_read_page - extract a page from the ring buffer
4670  * @buffer: buffer to extract from
4671  * @data_page: the page to use allocated from ring_buffer_alloc_read_page
4672  * @len: amount to extract
4673  * @cpu: the cpu of the buffer to extract
4674  * @full: should the extraction only happen when the page is full.
4675  *
4676  * This function will pull out a page from the ring buffer and consume it.
4677  * @data_page must be the address of the variable that was returned
4678  * from ring_buffer_alloc_read_page. This is because the page might be used
4679  * to swap with a page in the ring buffer.
4680  *
4681  * for example:
4682  *      rpage = ring_buffer_alloc_read_page(buffer, cpu);
4683  *      if (IS_ERR(rpage))
4684  *              return PTR_ERR(rpage);
4685  *      ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
4686  *      if (ret >= 0)
4687  *              process_page(rpage, ret);
4688  *
4689  * When @full is set, the function will not return true unless
4690  * the writer is off the reader page.
4691  *
4692  * Note: it is up to the calling functions to handle sleeps and wakeups.
4693  *  The ring buffer can be used anywhere in the kernel and can not
4694  *  blindly call wake_up. The layer that uses the ring buffer must be
4695  *  responsible for that.
4696  *
4697  * Returns:
4698  *  >=0 if data has been transferred, returns the offset of consumed data.
4699  *  <0 if no data has been transferred.
4700  */
4701 int ring_buffer_read_page(struct trace_buffer *buffer,
4702                           void **data_page, size_t len, int cpu, int full)
4703 {
4704         struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4705         struct ring_buffer_event *event;
4706         struct buffer_data_page *bpage;
4707         struct buffer_page *reader;
4708         unsigned long missed_events;
4709         unsigned long flags;
4710         unsigned int commit;
4711         unsigned int read;
4712         u64 save_timestamp;
4713         int ret = -1;
4714 
4715         if (!cpumask_test_cpu(cpu, buffer->cpumask))
4716                 goto out;
4717 
4718         /*
4719          * If len is not big enough to hold the page header, then
4720          * we can not copy anything.
4721          */
4722         if (len <= BUF_PAGE_HDR_SIZE)
4723                 goto out;
4724 
4725         len -= BUF_PAGE_HDR_SIZE;
4726 
4727         if (!data_page)
4728                 goto out;
4729 
4730         bpage = *data_page;
4731         if (!bpage)
4732                 goto out;
4733 
4734         raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4735 
4736         reader = rb_get_reader_page(cpu_buffer);
4737         if (!reader)
4738                 goto out_unlock;
4739 
4740         event = rb_reader_event(cpu_buffer);
4741 
4742         read = reader->read;
4743         commit = rb_page_commit(reader);
4744 
4745         /* Check if any events were dropped */
4746         missed_events = cpu_buffer->lost_events;
4747 
4748         /*
4749          * If this page has been partially read or
4750          * if len is not big enough to read the rest of the page or
4751          * a writer is still on the page, then
4752          * we must copy the data from the page to the buffer.
4753          * Otherwise, we can simply swap the page with the one passed in.
4754          */
4755         if (read || (len < (commit - read)) ||
4756             cpu_buffer->reader_page == cpu_buffer->commit_page) {
4757                 struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
4758                 unsigned int rpos = read;
4759                 unsigned int pos = 0;
4760                 unsigned int size;
4761 
4762                 if (full)
4763                         goto out_unlock;
4764 
4765                 if (len > (commit - read))
4766                         len = (commit - read);
4767 
4768                 /* Always keep the time extend and data together */
4769                 size = rb_event_ts_length(event);
4770 
4771                 if (len < size)
4772                         goto out_unlock;
4773 
4774                 /* save the current timestamp, since the user will need it */
4775                 save_timestamp = cpu_buffer->read_stamp;
4776 
4777                 /* Need to copy one event at a time */
4778                 do {
4779                         /* We need the size of one event, because
4780                          * rb_advance_reader only advances by one event,
4781                          * whereas rb_event_ts_length may include the size of
4782                          * one or two events.
4783                          * We have already ensured there's enough space if this
4784                          * is a time extend. */
4785                         size = rb_event_length(event);
4786                         memcpy(bpage->data + pos, rpage->data + rpos, size);
4787 
4788                         len -= size;
4789 
4790                         rb_advance_reader(cpu_buffer);
4791                         rpos = reader->read;
4792                         pos += size;
4793 
4794                         if (rpos >= commit)
4795                                 break;
4796 
4797                         event = rb_reader_event(cpu_buffer);
4798                         /* Always keep the time extend and data together */
4799                         size = rb_event_ts_length(event);
4800                 } while (len >= size);
4801 
4802                 /* update bpage */
4803                 local_set(&bpage->commit, pos);
4804                 bpage->time_stamp = save_timestamp;
4805 
4806                 /* we copied everything to the beginning */
4807                 read = 0;
4808         } else {
4809                 /* update the entry counter */
4810                 cpu_buffer->read += rb_page_entries(reader);
4811                 cpu_buffer->read_bytes += BUF_PAGE_SIZE;
4812 
4813                 /* swap the pages */
4814                 rb_init_page(bpage);
4815                 bpage = reader->page;
4816                 reader->page = *data_page;
4817                 local_set(&reader->write, 0);
4818                 local_set(&reader->entries, 0);
4819                 reader->read = 0;
4820                 *data_page = bpage;
4821 
4822                 /*
4823                  * Use the real_end for the data size,
4824                  * This gives us a chance to store the lost events
4825                  * on the page.
4826                  */
4827                 if (reader->real_end)
4828                         local_set(&bpage->commit, reader->real_end);
4829         }
4830         ret = read;
4831 
4832         cpu_buffer->lost_events = 0;
4833 
4834         commit = local_read(&bpage->commit);
4835         /*
4836          * Set a flag in the commit field if we lost events
4837          */
4838         if (missed_events) {
4839                 /* If there is room at the end of the page to save the
4840                  * missed events, then record it there.
4841                  */
4842                 if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
4843                         memcpy(&bpage->data[commit], &missed_events,
4844                                sizeof(missed_events));
4845                         local_add(RB_MISSED_STORED, &bpage->commit);
4846                         commit += sizeof(missed_events);
4847                 }
4848                 local_add(RB_MISSED_EVENTS, &bpage->commit);
4849         }
4850 
4851         /*
4852          * This page may be off to user land. Zero it out here.
4853          */
4854         if (commit < BUF_PAGE_SIZE)
4855                 memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
4856 
4857  out_unlock:
4858         raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4859 
4860  out:
4861         return ret;
4862 }
4863 EXPORT_SYMBOL_GPL(ring_buffer_read_page);
4864 
4865 /*
4866  * We only allocate new buffers, never free them if the CPU goes down.
4867  * If we were to free the buffer, then the user would lose any trace that was in
4868  * the buffer.
4869  */
4870 int trace_rb_cpu_prepare(unsigned int cpu, struct hlist_node *node)
4871 {
4872         struct trace_buffer *buffer;
4873         long nr_pages_same;
4874         int cpu_i;
4875         unsigned long nr_pages;
4876 
4877         buffer = container_of(node, struct trace_buffer, node);
4878         if (cpumask_test_cpu(cpu, buffer->cpumask))
4879                 return 0;
4880 
4881         nr_pages = 0;
4882         nr_pages_same = 1;
4883         /* check if all cpu sizes are same */
4884         for_each_buffer_cpu(buffer, cpu_i) {
4885                 /* fill in the size from first enabled cpu */
4886                 if (nr_pages == 0)
4887                         nr_pages = buffer->buffers[cpu_i]->nr_pages;
4888                 if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
4889                         nr_pages_same = 0;
4890                         break;
4891                 }
4892         }
4893         /* allocate minimum pages, user can later expand it */
4894         if (!nr_pages_same)
4895                 nr_pages = 2;
4896         buffer->buffers[cpu] =
4897                 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
4898         if (!buffer->buffers[cpu]) {
4899                 WARN(1, "failed to allocate ring buffer on CPU %u\n",
4900                      cpu);
4901                 return -ENOMEM;
4902         }
4903         smp_wmb();
4904         cpumask_set_cpu(cpu, buffer->cpumask);
4905         return 0;
4906 }
4907 
4908 #ifdef CONFIG_RING_BUFFER_STARTUP_TEST
4909 /*
4910  * This is a basic integrity check of the ring buffer.
4911  * Late in the boot cycle this test will run when configured in.
4912  * It will kick off a thread per CPU that will go into a loop
4913  * writing to the per cpu ring buffer various sizes of data.
4914  * Some of the data will be large items, some small.
4915  *
4916  * Another thread is created that goes into a spin, sending out
4917  * IPIs to the other CPUs to also write into the ring buffer.
4918  * this is to test the nesting ability of the buffer.
4919  *
4920  * Basic stats are recorded and reported. If something in the
4921  * ring buffer should happen that's not expected, a big warning
4922  * is displayed and all ring buffers are disabled.
4923  */
4924 static struct task_struct *rb_threads[NR_CPUS] __initdata;
4925 
4926 struct rb_test_data {
4927         struct trace_buffer *buffer;
4928         unsigned long           events;
4929         unsigned long           bytes_written;
4930         unsigned long           bytes_alloc;
4931         unsigned long           bytes_dropped;
4932         unsigned long           events_nested;
4933         unsigned long           bytes_written_nested;
4934         unsigned long           bytes_alloc_nested;
4935         unsigned long           bytes_dropped_nested;
4936         int                     min_size_nested;
4937         int                     max_size_nested;
4938         int                     max_size;
4939         int                     min_size;
4940         int                     cpu;
4941         int                     cnt;
4942 };
4943 
4944 static struct rb_test_data rb_data[NR_CPUS] __initdata;
4945 
4946 /* 1 meg per cpu */
4947 #define RB_TEST_BUFFER_SIZE     1048576
4948 
4949 static char rb_string[] __initdata =
4950         "abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
4951         "?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
4952         "!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
4953 
4954 static bool rb_test_started __initdata;
4955 
4956 struct rb_item {
4957         int size;
4958         char str[];
4959 };
4960 
4961 static __init int rb_write_something(struct rb_test_data *data, bool nested)
4962 {
4963         struct ring_buffer_event *event;
4964         struct rb_item *item;
4965         bool started;
4966         int event_len;
4967         int size;
4968         int len;
4969         int cnt;
4970 
4971         /* Have nested writes different that what is written */
4972         cnt = data->cnt + (nested ? 27 : 0);
4973 
4974         /* Multiply cnt by ~e, to make some unique increment */
4975         size = (cnt * 68 / 25) % (sizeof(rb_string) - 1);
4976 
4977         len = size + sizeof(struct rb_item);
4978 
4979         started = rb_test_started;
4980         /* read rb_test_started before checking buffer enabled */
4981         smp_rmb();
4982 
4983         event = ring_buffer_lock_reserve(data->buffer, len);
4984         if (!event) {
4985                 /* Ignore dropped events before test starts. */
4986                 if (started) {
4987                         if (nested)
4988                                 data->bytes_dropped += len;
4989                         else
4990                                 data->bytes_dropped_nested += len;
4991                 }
4992                 return len;
4993         }
4994 
4995         event_len = ring_buffer_event_length(event);
4996 
4997         if (RB_WARN_ON(data->buffer, event_len < len))
4998                 goto out;
4999 
5000         item = ring_buffer_event_data(event);
5001         item->size = size;
5002         memcpy(item->str, rb_string, size);
5003 
5004         if (nested) {
5005                 data->bytes_alloc_nested += event_len;
5006                 data->bytes_written_nested += len;
5007                 data->events_nested++;
5008                 if (!data->min_size_nested || len < data->min_size_nested)
5009                         data->min_size_nested = len;
5010                 if (len > data->max_size_nested)
5011                         data->max_size_nested = len;
5012         } else {
5013                 data->bytes_alloc += event_len;
5014                 data->bytes_written += len;
5015                 data->events++;
5016                 if (!data->min_size || len < data->min_size)
5017                         data->max_size = len;
5018                 if (len > data->max_size)
5019                         data->max_size = len;
5020         }
5021 
5022  out:
5023         ring_buffer_unlock_commit(data->buffer, event);
5024 
5025         return 0;
5026 }
5027 
5028 static __init int rb_test(void *arg)
5029 {
5030         struct rb_test_data *data = arg;
5031 
5032         while (!kthread_should_stop()) {
5033                 rb_write_something(data, false);
5034                 data->cnt++;
5035 
5036                 set_current_state(TASK_INTERRUPTIBLE);
5037                 /* Now sleep between a min of 100-300us and a max of 1ms */
5038                 usleep_range(((data->cnt % 3) + 1) * 100, 1000);
5039         }
5040 
5041         return 0;
5042 }
5043 
5044 static __init void rb_ipi(void *ignore)
5045 {
5046         struct rb_test_data *data;
5047         int cpu = smp_processor_id();
5048 
5049         data = &rb_data[cpu];
5050         rb_write_something(data, true);
5051 }
5052 
5053 static __init int rb_hammer_test(void *arg)
5054 {
5055         while (!kthread_should_stop()) {
5056 
5057                 /* Send an IPI to all cpus to write data! */
5058                 smp_call_function(rb_ipi, NULL, 1);
5059                 /* No sleep, but for non preempt, let others run */
5060                 schedule();
5061         }
5062 
5063         return 0;
5064 }
5065 
5066 static __init int test_ringbuffer(void)
5067 {
5068         struct task_struct *rb_hammer;
5069         struct trace_buffer *buffer;
5070         int cpu;
5071         int ret = 0;
5072 
5073         if (security_locked_down(LOCKDOWN_TRACEFS)) {
5074                 pr_warn("Lockdown is enabled, skipping ring buffer tests\n");
5075                 return 0;
5076         }
5077 
5078         pr_info("Running ring buffer tests...\n");
5079 
5080         buffer = ring_buffer_alloc(RB_TEST_BUFFER_SIZE, RB_FL_OVERWRITE);
5081         if (WARN_ON(!buffer))
5082                 return 0;
5083 
5084         /* Disable buffer so that threads can't write to it yet */
5085         ring_buffer_record_off(buffer);
5086 
5087         for_each_online_cpu(cpu) {
5088                 rb_data[cpu].buffer = buffer;
5089                 rb_data[cpu].cpu = cpu;
5090                 rb_data[cpu].cnt = cpu;
5091                 rb_threads[cpu] = kthread_create(rb_test, &rb_data[cpu],
5092                                                  "rbtester/%d", cpu);
5093                 if (WARN_ON(IS_ERR(rb_threads[cpu]))) {
5094                         pr_cont("FAILED\n");
5095                         ret = PTR_ERR(rb_threads[cpu]);
5096                         goto out_free;
5097                 }
5098 
5099                 kthread_bind(rb_threads[cpu], cpu);
5100                 wake_up_process(rb_threads[cpu]);
5101         }
5102 
5103         /* Now create the rb hammer! */
5104         rb_hammer = kthread_run(rb_hammer_test, NULL, "rbhammer");
5105         if (WARN_ON(IS_ERR(rb_hammer))) {
5106                 pr_cont("FAILED\n");
5107                 ret = PTR_ERR(rb_hammer);
5108                 goto out_free;
5109         }
5110 
5111         ring_buffer_record_on(buffer);
5112         /*
5113          * Show buffer is enabled before setting rb_test_started.
5114          * Yes there's a small race window where events could be
5115          * dropped and the thread wont catch it. But when a ring
5116          * buffer gets enabled, there will always be some kind of
5117          * delay before other CPUs see it. Thus, we don't care about
5118          * those dropped events. We care about events dropped after
5119          * the threads see that the buffer is active.
5120          */
5121         smp_wmb();
5122         rb_test_started = true;
5123 
5124         set_current_state(TASK_INTERRUPTIBLE);
5125         /* Just run for 10 seconds */;
5126         schedule_timeout(10 * HZ);
5127 
5128         kthread_stop(rb_hammer);
5129 
5130  out_free:
5131         for_each_online_cpu(cpu) {
5132                 if (!rb_threads[cpu])
5133                         break;
5134                 kthread_stop(rb_threads[cpu]);
5135         }
5136         if (ret) {
5137                 ring_buffer_free(buffer);
5138                 return ret;
5139         }
5140 
5141         /* Report! */
5142         pr_info("finished\n");
5143         for_each_online_cpu(cpu) {
5144                 struct ring_buffer_event *event;
5145                 struct rb_test_data *data = &rb_data[cpu];
5146                 struct rb_item *item;
5147                 unsigned long total_events;
5148                 unsigned long total_dropped;
5149                 unsigned long total_written;
5150                 unsigned long total_alloc;
5151                 unsigned long total_read = 0;
5152                 unsigned long total_size = 0;
5153                 unsigned long total_len = 0;
5154                 unsigned long total_lost = 0;
5155                 unsigned long lost;
5156                 int big_event_size;
5157                 int small_event_size;
5158 
5159                 ret = -1;
5160 
5161                 total_events = data->events + data->events_nested;
5162                 total_written = data->bytes_written + data->bytes_written_nested;
5163                 total_alloc = data->bytes_alloc + data->bytes_alloc_nested;
5164                 total_dropped = data->bytes_dropped + data->bytes_dropped_nested;
5165 
5166                 big_event_size = data->max_size + data->max_size_nested;
5167                 small_event_size = data->min_size + data->min_size_nested;
5168 
5169                 pr_info("CPU %d:\n", cpu);
5170                 pr_info("              events:    %ld\n", total_events);
5171                 pr_info("       dropped bytes:    %ld\n", total_dropped);
5172                 pr_info("       alloced bytes:    %ld\n", total_alloc);
5173                 pr_info("       written bytes:    %ld\n", total_written);
5174                 pr_info("       biggest event:    %d\n", big_event_size);
5175                 pr_info("      smallest event:    %d\n", small_event_size);
5176 
5177                 if (RB_WARN_ON(buffer, total_dropped))
5178                         break;
5179 
5180                 ret = 0;
5181 
5182                 while ((event = ring_buffer_consume(buffer, cpu, NULL, &lost))) {
5183                         total_lost += lost;
5184                         item = ring_buffer_event_data(event);
5185                         total_len += ring_buffer_event_length(event);
5186                         total_size += item->size + sizeof(struct rb_item);
5187                         if (memcmp(&item->str[0], rb_string, item->size) != 0) {
5188                                 pr_info("FAILED!\n");
5189                                 pr_info("buffer had: %.*s\n", item->size, item->str);
5190                                 pr_info("expected:   %.*s\n", item->size, rb_string);
5191                                 RB_WARN_ON(buffer, 1);
5192                                 ret = -1;
5193                                 break;
5194                         }
5195                         total_read++;
5196                 }
5197                 if (ret)
5198                         break;
5199 
5200                 ret = -1;
5201 
5202                 pr_info("         read events:   %ld\n", total_read);
5203                 pr_info("         lost events:   %ld\n", total_lost);
5204                 pr_info("        total events:   %ld\n", total_lost + total_read);
5205                 pr_info("  recorded len bytes:   %ld\n", total_len);
5206                 pr_info(" recorded size bytes:   %ld\n", total_size);
5207                 if (total_lost)
5208                         pr_info(" With dropped events, record len and size may not match\n"
5209                                 " alloced and written from above\n");
5210                 if (!total_lost) {
5211                         if (RB_WARN_ON(buffer, total_len != total_alloc ||
5212                                        total_size != total_written))
5213                                 break;
5214                 }
5215                 if (RB_WARN_ON(buffer, total_lost + total_read != total_events))
5216                         break;
5217 
5218                 ret = 0;
5219         }
5220         if (!ret)
5221                 pr_info("Ring buffer PASSED!\n");
5222 
5223         ring_buffer_free(buffer);
5224         return 0;
5225 }
5226 
5227 late_initcall(test_ringbuffer);
5228 #endif /* CONFIG_RING_BUFFER_STARTUP_TEST */
5229 

~ [ 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