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

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

Version: ~ [ linux-5.10-rc5 ] ~ [ linux-5.9.10 ] ~ [ linux-5.8.18 ] ~ [ linux-5.7.19 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.79 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.159 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.208 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.245 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.245 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.19.8 ] ~ [ linux-3.18.140 ] ~ [ linux-3.17.8 ] ~ [ linux-3.16.85 ] ~ [ linux-3.15.10 ] ~ [ linux-3.14.79 ] ~ [ linux-3.13.11 ] ~ [ linux-3.12.74 ] ~ [ linux-3.11.10 ] ~ [ linux-3.10.108 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

  1 /*
  2  * Generic ring buffer
  3  *
  4  * Copyright (C) 2008 Steven Rostedt <srostedt@redhat.com>
  5  */
  6 #include <linux/trace_events.h>
  7 #include <linux/ring_buffer.h>
  8 #include <linux/trace_clock.h>
  9 #include <linux/sched/clock.h>
 10 #include <linux/trace_seq.h>
 11 #include <linux/spinlock.h>
 12 #include <linux/irq_work.h>
 13 #include <linux/uaccess.h>
 14 #include <linux/hardirq.h>
 15 #include <linux/kthread.h>      /* for self test */
 16 #include <linux/module.h>
 17 #include <linux/percpu.h>
 18 #include <linux/mutex.h>
 19 #include <linux/delay.h>
 20 #include <linux/slab.h>
 21 #include <linux/init.h>
 22 #include <linux/hash.h>
 23 #include <linux/list.h>
 24 #include <linux/cpu.h>
 25 #include <linux/oom.h>
 26 
 27 #include <asm/local.h>
 28 
 29 static void update_pages_handler(struct work_struct *work);
 30 
 31 /*
 32  * The ring buffer header is special. We must manually up keep it.
 33  */
 34 int ring_buffer_print_entry_header(struct trace_seq *s)
 35 {
 36         trace_seq_puts(s, "# compressed entry header\n");
 37         trace_seq_puts(s, "\ttype_len    :    5 bits\n");
 38         trace_seq_puts(s, "\ttime_delta  :   27 bits\n");
 39         trace_seq_puts(s, "\tarray       :   32 bits\n");
 40         trace_seq_putc(s, '\n');
 41         trace_seq_printf(s, "\tpadding     : type == %d\n",
 42                          RINGBUF_TYPE_PADDING);
 43         trace_seq_printf(s, "\ttime_extend : type == %d\n",
 44                          RINGBUF_TYPE_TIME_EXTEND);
 45         trace_seq_printf(s, "\ttime_stamp : type == %d\n",
 46                          RINGBUF_TYPE_TIME_STAMP);
 47         trace_seq_printf(s, "\tdata max type_len  == %d\n",
 48                          RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
 49 
 50         return !trace_seq_has_overflowed(s);
 51 }
 52 
 53 /*
 54  * The ring buffer is made up of a list of pages. A separate list of pages is
 55  * allocated for each CPU. A writer may only write to a buffer that is
 56  * associated with the CPU it is currently executing on.  A reader may read
 57  * from any per cpu buffer.
 58  *
 59  * The reader is special. For each per cpu buffer, the reader has its own
 60  * reader page. When a reader has read the entire reader page, this reader
 61  * page is swapped with another page in the ring buffer.
 62  *
 63  * Now, as long as the writer is off the reader page, the reader can do what
 64  * ever it wants with that page. The writer will never write to that page
 65  * again (as long as it is out of the ring buffer).
 66  *
 67  * Here's some silly ASCII art.
 68  *
 69  *   +------+
 70  *   |reader|          RING BUFFER
 71  *   |page  |
 72  *   +------+        +---+   +---+   +---+
 73  *                   |   |-->|   |-->|   |
 74  *                   +---+   +---+   +---+
 75  *                     ^               |
 76  *                     |               |
 77  *                     +---------------+
 78  *
 79  *
 80  *   +------+
 81  *   |reader|          RING BUFFER
 82  *   |page  |------------------v
 83  *   +------+        +---+   +---+   +---+
 84  *                   |   |-->|   |-->|   |
 85  *                   +---+   +---+   +---+
 86  *                     ^               |
 87  *                     |               |
 88  *                     +---------------+
 89  *
 90  *
 91  *   +------+
 92  *   |reader|          RING BUFFER
 93  *   |page  |------------------v
 94  *   +------+        +---+   +---+   +---+
 95  *      ^            |   |-->|   |-->|   |
 96  *      |            +---+   +---+   +---+
 97  *      |                              |
 98  *      |                              |
 99  *      +------------------------------+
100  *
101  *
102  *   +------+
103  *   |buffer|          RING BUFFER
104  *   |page  |------------------v
105  *   +------+        +---+   +---+   +---+
106  *      ^            |   |   |   |-->|   |
107  *      |   New      +---+   +---+   +---+
108  *      |  Reader------^               |
109  *      |   page                       |
110  *      +------------------------------+
111  *
112  *
113  * After we make this swap, the reader can hand this page off to the splice
114  * code and be done with it. It can even allocate a new page if it needs to
115  * and swap that into the ring buffer.
116  *
117  * We will be using cmpxchg soon to make all this lockless.
118  *
119  */
120 
121 /* Used for individual buffers (after the counter) */
122 #define RB_BUFFER_OFF           (1 << 20)
123 
124 #define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
125 
126 #define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
127 #define RB_ALIGNMENT            4U
128 #define RB_MAX_SMALL_DATA       (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
129 #define RB_EVNT_MIN_SIZE        8U      /* two 32bit words */
130 
131 #ifndef CONFIG_HAVE_64BIT_ALIGNED_ACCESS
132 # define RB_FORCE_8BYTE_ALIGNMENT       0
133 # define RB_ARCH_ALIGNMENT              RB_ALIGNMENT
134 #else
135 # define RB_FORCE_8BYTE_ALIGNMENT       1
136 # define RB_ARCH_ALIGNMENT              8U
137 #endif
138 
139 #define RB_ALIGN_DATA           __aligned(RB_ARCH_ALIGNMENT)
140 
141 /* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
142 #define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
143 
144 enum {
145         RB_LEN_TIME_EXTEND = 8,
146         RB_LEN_TIME_STAMP =  8,
147 };
148 
149 #define skip_time_extend(event) \
150         ((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
151 
152 #define extended_time(event) \
153         (event->type_len >= RINGBUF_TYPE_TIME_EXTEND)
154 
155 static inline int rb_null_event(struct ring_buffer_event *event)
156 {
157         return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
158 }
159 
160 static void rb_event_set_padding(struct ring_buffer_event *event)
161 {
162         /* padding has a NULL time_delta */
163         event->type_len = RINGBUF_TYPE_PADDING;
164         event->time_delta = 0;
165 }
166 
167 static unsigned
168 rb_event_data_length(struct ring_buffer_event *event)
169 {
170         unsigned length;
171 
172         if (event->type_len)
173                 length = event->type_len * RB_ALIGNMENT;
174         else
175                 length = event->array[0];
176         return length + RB_EVNT_HDR_SIZE;
177 }
178 
179 /*
180  * Return the length of the given event. Will return
181  * the length of the time extend if the event is a
182  * time extend.
183  */
184 static inline unsigned
185 rb_event_length(struct ring_buffer_event *event)
186 {
187         switch (event->type_len) {
188         case RINGBUF_TYPE_PADDING:
189                 if (rb_null_event(event))
190                         /* undefined */
191                         return -1;
192                 return  event->array[0] + RB_EVNT_HDR_SIZE;
193 
194         case RINGBUF_TYPE_TIME_EXTEND:
195                 return RB_LEN_TIME_EXTEND;
196 
197         case RINGBUF_TYPE_TIME_STAMP:
198                 return RB_LEN_TIME_STAMP;
199 
200         case RINGBUF_TYPE_DATA:
201                 return rb_event_data_length(event);
202         default:
203                 BUG();
204         }
205         /* not hit */
206         return 0;
207 }
208 
209 /*
210  * Return total length of time extend and data,
211  *   or just the event length for all other events.
212  */
213 static inline unsigned
214 rb_event_ts_length(struct ring_buffer_event *event)
215 {
216         unsigned len = 0;
217 
218         if (extended_time(event)) {
219                 /* time extends include the data event after it */
220                 len = RB_LEN_TIME_EXTEND;
221                 event = skip_time_extend(event);
222         }
223         return len + rb_event_length(event);
224 }
225 
226 /**
227  * ring_buffer_event_length - return the length of the event
228  * @event: the event to get the length of
229  *
230  * Returns the size of the data load of a data event.
231  * If the event is something other than a data event, it
232  * returns the size of the event itself. With the exception
233  * of a TIME EXTEND, where it still returns the size of the
234  * data load of the data event after it.
235  */
236 unsigned ring_buffer_event_length(struct ring_buffer_event *event)
237 {
238         unsigned length;
239 
240         if (extended_time(event))
241                 event = skip_time_extend(event);
242 
243         length = rb_event_length(event);
244         if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
245                 return length;
246         length -= RB_EVNT_HDR_SIZE;
247         if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
248                 length -= sizeof(event->array[0]);
249         return length;
250 }
251 EXPORT_SYMBOL_GPL(ring_buffer_event_length);
252 
253 /* inline for ring buffer fast paths */
254 static __always_inline void *
255 rb_event_data(struct ring_buffer_event *event)
256 {
257         if (extended_time(event))
258                 event = skip_time_extend(event);
259         BUG_ON(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
260         /* If length is in len field, then array[0] has the data */
261         if (event->type_len)
262                 return (void *)&event->array[0];
263         /* Otherwise length is in array[0] and array[1] has the data */
264         return (void *)&event->array[1];
265 }
266 
267 /**
268  * ring_buffer_event_data - return the data of the event
269  * @event: the event to get the data from
270  */
271 void *ring_buffer_event_data(struct ring_buffer_event *event)
272 {
273         return rb_event_data(event);
274 }
275 EXPORT_SYMBOL_GPL(ring_buffer_event_data);
276 
277 #define for_each_buffer_cpu(buffer, cpu)                \
278         for_each_cpu(cpu, buffer->cpumask)
279 
280 #define TS_SHIFT        27
281 #define TS_MASK         ((1ULL << TS_SHIFT) - 1)
282 #define TS_DELTA_TEST   (~TS_MASK)
283 
284 /**
285  * ring_buffer_event_time_stamp - return the event's extended timestamp
286  * @event: the event to get the timestamp of
287  *
288  * Returns the extended timestamp associated with a data event.
289  * An extended time_stamp is a 64-bit timestamp represented
290  * internally in a special way that makes the best use of space
291  * contained within a ring buffer event.  This function decodes
292  * it and maps it to a straight u64 value.
293  */
294 u64 ring_buffer_event_time_stamp(struct ring_buffer_event *event)
295 {
296         u64 ts;
297 
298         ts = event->array[0];
299         ts <<= TS_SHIFT;
300         ts += event->time_delta;
301 
302         return ts;
303 }
304 
305 /* Flag when events were overwritten */
306 #define RB_MISSED_EVENTS        (1 << 31)
307 /* Missed count stored at end */
308 #define RB_MISSED_STORED        (1 << 30)
309 
310 #define RB_MISSED_FLAGS         (RB_MISSED_EVENTS|RB_MISSED_STORED)
311 
312 struct buffer_data_page {
313         u64              time_stamp;    /* page time stamp */
314         local_t          commit;        /* write committed index */
315         unsigned char    data[] RB_ALIGN_DATA;  /* data of buffer page */
316 };
317 
318 /*
319  * Note, the buffer_page list must be first. The buffer pages
320  * are allocated in cache lines, which means that each buffer
321  * page will be at the beginning of a cache line, and thus
322  * the least significant bits will be zero. We use this to
323  * add flags in the list struct pointers, to make the ring buffer
324  * lockless.
325  */
326 struct buffer_page {
327         struct list_head list;          /* list of buffer pages */
328         local_t          write;         /* index for next write */
329         unsigned         read;          /* index for next read */
330         local_t          entries;       /* entries on this page */
331         unsigned long    real_end;      /* real end of data */
332         struct buffer_data_page *page;  /* Actual data page */
333 };
334 
335 /*
336  * The buffer page counters, write and entries, must be reset
337  * atomically when crossing page boundaries. To synchronize this
338  * update, two counters are inserted into the number. One is
339  * the actual counter for the write position or count on the page.
340  *
341  * The other is a counter of updaters. Before an update happens
342  * the update partition of the counter is incremented. This will
343  * allow the updater to update the counter atomically.
344  *
345  * The counter is 20 bits, and the state data is 12.
346  */
347 #define RB_WRITE_MASK           0xfffff
348 #define RB_WRITE_INTCNT         (1 << 20)
349 
350 static void rb_init_page(struct buffer_data_page *bpage)
351 {
352         local_set(&bpage->commit, 0);
353 }
354 
355 /**
356  * ring_buffer_page_len - the size of data on the page.
357  * @page: The page to read
358  *
359  * Returns the amount of data on the page, including buffer page header.
360  */
361 size_t ring_buffer_page_len(void *page)
362 {
363         struct buffer_data_page *bpage = page;
364 
365         return (local_read(&bpage->commit) & ~RB_MISSED_FLAGS)
366                 + BUF_PAGE_HDR_SIZE;
367 }
368 
369 /*
370  * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
371  * this issue out.
372  */
373 static void free_buffer_page(struct buffer_page *bpage)
374 {
375         free_page((unsigned long)bpage->page);
376         kfree(bpage);
377 }
378 
379 /*
380  * We need to fit the time_stamp delta into 27 bits.
381  */
382 static inline int test_time_stamp(u64 delta)
383 {
384         if (delta & TS_DELTA_TEST)
385                 return 1;
386         return 0;
387 }
388 
389 #define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
390 
391 /* Max payload is BUF_PAGE_SIZE - header (8bytes) */
392 #define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
393 
394 int ring_buffer_print_page_header(struct trace_seq *s)
395 {
396         struct buffer_data_page field;
397 
398         trace_seq_printf(s, "\tfield: u64 timestamp;\t"
399                          "offset:0;\tsize:%u;\tsigned:%u;\n",
400                          (unsigned int)sizeof(field.time_stamp),
401                          (unsigned int)is_signed_type(u64));
402 
403         trace_seq_printf(s, "\tfield: local_t commit;\t"
404                          "offset:%u;\tsize:%u;\tsigned:%u;\n",
405                          (unsigned int)offsetof(typeof(field), commit),
406                          (unsigned int)sizeof(field.commit),
407                          (unsigned int)is_signed_type(long));
408 
409         trace_seq_printf(s, "\tfield: int overwrite;\t"
410                          "offset:%u;\tsize:%u;\tsigned:%u;\n",
411                          (unsigned int)offsetof(typeof(field), commit),
412                          1,
413                          (unsigned int)is_signed_type(long));
414 
415         trace_seq_printf(s, "\tfield: char data;\t"
416                          "offset:%u;\tsize:%u;\tsigned:%u;\n",
417                          (unsigned int)offsetof(typeof(field), data),
418                          (unsigned int)BUF_PAGE_SIZE,
419                          (unsigned int)is_signed_type(char));
420 
421         return !trace_seq_has_overflowed(s);
422 }
423 
424 struct rb_irq_work {
425         struct irq_work                 work;
426         wait_queue_head_t               waiters;
427         wait_queue_head_t               full_waiters;
428         bool                            waiters_pending;
429         bool                            full_waiters_pending;
430         bool                            wakeup_full;
431 };
432 
433 /*
434  * Structure to hold event state and handle nested events.
435  */
436 struct rb_event_info {
437         u64                     ts;
438         u64                     delta;
439         unsigned long           length;
440         struct buffer_page      *tail_page;
441         int                     add_timestamp;
442 };
443 
444 /*
445  * Used for which event context the event is in.
446  *  NMI     = 0
447  *  IRQ     = 1
448  *  SOFTIRQ = 2
449  *  NORMAL  = 3
450  *
451  * See trace_recursive_lock() comment below for more details.
452  */
453 enum {
454         RB_CTX_NMI,
455         RB_CTX_IRQ,
456         RB_CTX_SOFTIRQ,
457         RB_CTX_NORMAL,
458         RB_CTX_MAX
459 };
460 
461 /*
462  * head_page == tail_page && head == tail then buffer is empty.
463  */
464 struct ring_buffer_per_cpu {
465         int                             cpu;
466         atomic_t                        record_disabled;
467         struct ring_buffer              *buffer;
468         raw_spinlock_t                  reader_lock;    /* serialize readers */
469         arch_spinlock_t                 lock;
470         struct lock_class_key           lock_key;
471         struct buffer_data_page         *free_page;
472         unsigned long                   nr_pages;
473         unsigned int                    current_context;
474         struct list_head                *pages;
475         struct buffer_page              *head_page;     /* read from head */
476         struct buffer_page              *tail_page;     /* write to tail */
477         struct buffer_page              *commit_page;   /* committed pages */
478         struct buffer_page              *reader_page;
479         unsigned long                   lost_events;
480         unsigned long                   last_overrun;
481         unsigned long                   nest;
482         local_t                         entries_bytes;
483         local_t                         entries;
484         local_t                         overrun;
485         local_t                         commit_overrun;
486         local_t                         dropped_events;
487         local_t                         committing;
488         local_t                         commits;
489         unsigned long                   read;
490         unsigned long                   read_bytes;
491         u64                             write_stamp;
492         u64                             read_stamp;
493         /* ring buffer pages to update, > 0 to add, < 0 to remove */
494         long                            nr_pages_to_update;
495         struct list_head                new_pages; /* new pages to add */
496         struct work_struct              update_pages_work;
497         struct completion               update_done;
498 
499         struct rb_irq_work              irq_work;
500 };
501 
502 struct ring_buffer {
503         unsigned                        flags;
504         int                             cpus;
505         atomic_t                        record_disabled;
506         atomic_t                        resize_disabled;
507         cpumask_var_t                   cpumask;
508 
509         struct lock_class_key           *reader_lock_key;
510 
511         struct mutex                    mutex;
512 
513         struct ring_buffer_per_cpu      **buffers;
514 
515         struct hlist_node               node;
516         u64                             (*clock)(void);
517 
518         struct rb_irq_work              irq_work;
519         bool                            time_stamp_abs;
520 };
521 
522 struct ring_buffer_iter {
523         struct ring_buffer_per_cpu      *cpu_buffer;
524         unsigned long                   head;
525         struct buffer_page              *head_page;
526         struct buffer_page              *cache_reader_page;
527         unsigned long                   cache_read;
528         u64                             read_stamp;
529 };
530 
531 /*
532  * rb_wake_up_waiters - wake up tasks waiting for ring buffer input
533  *
534  * Schedules a delayed work to wake up any task that is blocked on the
535  * ring buffer waiters queue.
536  */
537 static void rb_wake_up_waiters(struct irq_work *work)
538 {
539         struct rb_irq_work *rbwork = container_of(work, struct rb_irq_work, work);
540 
541         wake_up_all(&rbwork->waiters);
542         if (rbwork->wakeup_full) {
543                 rbwork->wakeup_full = false;
544                 wake_up_all(&rbwork->full_waiters);
545         }
546 }
547 
548 /**
549  * ring_buffer_wait - wait for input to the ring buffer
550  * @buffer: buffer to wait on
551  * @cpu: the cpu buffer to wait on
552  * @full: wait until a full page is available, if @cpu != RING_BUFFER_ALL_CPUS
553  *
554  * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
555  * as data is added to any of the @buffer's cpu buffers. Otherwise
556  * it will wait for data to be added to a specific cpu buffer.
557  */
558 int ring_buffer_wait(struct ring_buffer *buffer, int cpu, bool full)
559 {
560         struct ring_buffer_per_cpu *uninitialized_var(cpu_buffer);
561         DEFINE_WAIT(wait);
562         struct rb_irq_work *work;
563         int ret = 0;
564 
565         /*
566          * Depending on what the caller is waiting for, either any
567          * data in any cpu buffer, or a specific buffer, put the
568          * caller on the appropriate wait queue.
569          */
570         if (cpu == RING_BUFFER_ALL_CPUS) {
571                 work = &buffer->irq_work;
572                 /* Full only makes sense on per cpu reads */
573                 full = false;
574         } else {
575                 if (!cpumask_test_cpu(cpu, buffer->cpumask))
576                         return -ENODEV;
577                 cpu_buffer = buffer->buffers[cpu];
578                 work = &cpu_buffer->irq_work;
579         }
580 
581 
582         while (true) {
583                 if (full)
584                         prepare_to_wait(&work->full_waiters, &wait, TASK_INTERRUPTIBLE);
585                 else
586                         prepare_to_wait(&work->waiters, &wait, TASK_INTERRUPTIBLE);
587 
588                 /*
589                  * The events can happen in critical sections where
590                  * checking a work queue can cause deadlocks.
591                  * After adding a task to the queue, this flag is set
592                  * only to notify events to try to wake up the queue
593                  * using irq_work.
594                  *
595                  * We don't clear it even if the buffer is no longer
596                  * empty. The flag only causes the next event to run
597                  * irq_work to do the work queue wake up. The worse
598                  * that can happen if we race with !trace_empty() is that
599                  * an event will cause an irq_work to try to wake up
600                  * an empty queue.
601                  *
602                  * There's no reason to protect this flag either, as
603                  * the work queue and irq_work logic will do the necessary
604                  * synchronization for the wake ups. The only thing
605                  * that is necessary is that the wake up happens after
606                  * a task has been queued. It's OK for spurious wake ups.
607                  */
608                 if (full)
609                         work->full_waiters_pending = true;
610                 else
611                         work->waiters_pending = true;
612 
613                 if (signal_pending(current)) {
614                         ret = -EINTR;
615                         break;
616                 }
617 
618                 if (cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer))
619                         break;
620 
621                 if (cpu != RING_BUFFER_ALL_CPUS &&
622                     !ring_buffer_empty_cpu(buffer, cpu)) {
623                         unsigned long flags;
624                         bool pagebusy;
625 
626                         if (!full)
627                                 break;
628 
629                         raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
630                         pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
631                         raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
632 
633                         if (!pagebusy)
634                                 break;
635                 }
636 
637                 schedule();
638         }
639 
640         if (full)
641                 finish_wait(&work->full_waiters, &wait);
642         else
643                 finish_wait(&work->waiters, &wait);
644 
645         return ret;
646 }
647 
648 /**
649  * ring_buffer_poll_wait - poll on buffer input
650  * @buffer: buffer to wait on
651  * @cpu: the cpu buffer to wait on
652  * @filp: the file descriptor
653  * @poll_table: The poll descriptor
654  *
655  * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
656  * as data is added to any of the @buffer's cpu buffers. Otherwise
657  * it will wait for data to be added to a specific cpu buffer.
658  *
659  * Returns EPOLLIN | EPOLLRDNORM if data exists in the buffers,
660  * zero otherwise.
661  */
662 __poll_t ring_buffer_poll_wait(struct ring_buffer *buffer, int cpu,
663                           struct file *filp, poll_table *poll_table)
664 {
665         struct ring_buffer_per_cpu *cpu_buffer;
666         struct rb_irq_work *work;
667 
668         if (cpu == RING_BUFFER_ALL_CPUS)
669                 work = &buffer->irq_work;
670         else {
671                 if (!cpumask_test_cpu(cpu, buffer->cpumask))
672                         return -EINVAL;
673 
674                 cpu_buffer = buffer->buffers[cpu];
675                 work = &cpu_buffer->irq_work;
676         }
677 
678         poll_wait(filp, &work->waiters, poll_table);
679         work->waiters_pending = true;
680         /*
681          * There's a tight race between setting the waiters_pending and
682          * checking if the ring buffer is empty.  Once the waiters_pending bit
683          * is set, the next event will wake the task up, but we can get stuck
684          * if there's only a single event in.
685          *
686          * FIXME: Ideally, we need a memory barrier on the writer side as well,
687          * but adding a memory barrier to all events will cause too much of a
688          * performance hit in the fast path.  We only need a memory barrier when
689          * the buffer goes from empty to having content.  But as this race is
690          * extremely small, and it's not a problem if another event comes in, we
691          * will fix it later.
692          */
693         smp_mb();
694 
695         if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
696             (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
697                 return EPOLLIN | EPOLLRDNORM;
698         return 0;
699 }
700 
701 /* buffer may be either ring_buffer or ring_buffer_per_cpu */
702 #define RB_WARN_ON(b, cond)                                             \
703         ({                                                              \
704                 int _____ret = unlikely(cond);                          \
705                 if (_____ret) {                                         \
706                         if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
707                                 struct ring_buffer_per_cpu *__b =       \
708                                         (void *)b;                      \
709                                 atomic_inc(&__b->buffer->record_disabled); \
710                         } else                                          \
711                                 atomic_inc(&b->record_disabled);        \
712                         WARN_ON(1);                                     \
713                 }                                                       \
714                 _____ret;                                               \
715         })
716 
717 /* Up this if you want to test the TIME_EXTENTS and normalization */
718 #define DEBUG_SHIFT 0
719 
720 static inline u64 rb_time_stamp(struct ring_buffer *buffer)
721 {
722         /* shift to debug/test normalization and TIME_EXTENTS */
723         return buffer->clock() << DEBUG_SHIFT;
724 }
725 
726 u64 ring_buffer_time_stamp(struct ring_buffer *buffer, int cpu)
727 {
728         u64 time;
729 
730         preempt_disable_notrace();
731         time = rb_time_stamp(buffer);
732         preempt_enable_no_resched_notrace();
733 
734         return time;
735 }
736 EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
737 
738 void ring_buffer_normalize_time_stamp(struct ring_buffer *buffer,
739                                       int cpu, u64 *ts)
740 {
741         /* Just stupid testing the normalize function and deltas */
742         *ts >>= DEBUG_SHIFT;
743 }
744 EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
745 
746 /*
747  * Making the ring buffer lockless makes things tricky.
748  * Although writes only happen on the CPU that they are on,
749  * and they only need to worry about interrupts. Reads can
750  * happen on any CPU.
751  *
752  * The reader page is always off the ring buffer, but when the
753  * reader finishes with a page, it needs to swap its page with
754  * a new one from the buffer. The reader needs to take from
755  * the head (writes go to the tail). But if a writer is in overwrite
756  * mode and wraps, it must push the head page forward.
757  *
758  * Here lies the problem.
759  *
760  * The reader must be careful to replace only the head page, and
761  * not another one. As described at the top of the file in the
762  * ASCII art, the reader sets its old page to point to the next
763  * page after head. It then sets the page after head to point to
764  * the old reader page. But if the writer moves the head page
765  * during this operation, the reader could end up with the tail.
766  *
767  * We use cmpxchg to help prevent this race. We also do something
768  * special with the page before head. We set the LSB to 1.
769  *
770  * When the writer must push the page forward, it will clear the
771  * bit that points to the head page, move the head, and then set
772  * the bit that points to the new head page.
773  *
774  * We also don't want an interrupt coming in and moving the head
775  * page on another writer. Thus we use the second LSB to catch
776  * that too. Thus:
777  *
778  * head->list->prev->next        bit 1          bit 0
779  *                              -------        -------
780  * Normal page                     0              0
781  * Points to head page             0              1
782  * New head page                   1              0
783  *
784  * Note we can not trust the prev pointer of the head page, because:
785  *
786  * +----+       +-----+        +-----+
787  * |    |------>|  T  |---X--->|  N  |
788  * |    |<------|     |        |     |
789  * +----+       +-----+        +-----+
790  *   ^                           ^ |
791  *   |          +-----+          | |
792  *   +----------|  R  |----------+ |
793  *              |     |<-----------+
794  *              +-----+
795  *
796  * Key:  ---X-->  HEAD flag set in pointer
797  *         T      Tail page
798  *         R      Reader page
799  *         N      Next page
800  *
801  * (see __rb_reserve_next() to see where this happens)
802  *
803  *  What the above shows is that the reader just swapped out
804  *  the reader page with a page in the buffer, but before it
805  *  could make the new header point back to the new page added
806  *  it was preempted by a writer. The writer moved forward onto
807  *  the new page added by the reader and is about to move forward
808  *  again.
809  *
810  *  You can see, it is legitimate for the previous pointer of
811  *  the head (or any page) not to point back to itself. But only
812  *  temporarially.
813  */
814 
815 #define RB_PAGE_NORMAL          0UL
816 #define RB_PAGE_HEAD            1UL
817 #define RB_PAGE_UPDATE          2UL
818 
819 
820 #define RB_FLAG_MASK            3UL
821 
822 /* PAGE_MOVED is not part of the mask */
823 #define RB_PAGE_MOVED           4UL
824 
825 /*
826  * rb_list_head - remove any bit
827  */
828 static struct list_head *rb_list_head(struct list_head *list)
829 {
830         unsigned long val = (unsigned long)list;
831 
832         return (struct list_head *)(val & ~RB_FLAG_MASK);
833 }
834 
835 /*
836  * rb_is_head_page - test if the given page is the head page
837  *
838  * Because the reader may move the head_page pointer, we can
839  * not trust what the head page is (it may be pointing to
840  * the reader page). But if the next page is a header page,
841  * its flags will be non zero.
842  */
843 static inline int
844 rb_is_head_page(struct ring_buffer_per_cpu *cpu_buffer,
845                 struct buffer_page *page, struct list_head *list)
846 {
847         unsigned long val;
848 
849         val = (unsigned long)list->next;
850 
851         if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
852                 return RB_PAGE_MOVED;
853 
854         return val & RB_FLAG_MASK;
855 }
856 
857 /*
858  * rb_is_reader_page
859  *
860  * The unique thing about the reader page, is that, if the
861  * writer is ever on it, the previous pointer never points
862  * back to the reader page.
863  */
864 static bool rb_is_reader_page(struct buffer_page *page)
865 {
866         struct list_head *list = page->list.prev;
867 
868         return rb_list_head(list->next) != &page->list;
869 }
870 
871 /*
872  * rb_set_list_to_head - set a list_head to be pointing to head.
873  */
874 static void rb_set_list_to_head(struct ring_buffer_per_cpu *cpu_buffer,
875                                 struct list_head *list)
876 {
877         unsigned long *ptr;
878 
879         ptr = (unsigned long *)&list->next;
880         *ptr |= RB_PAGE_HEAD;
881         *ptr &= ~RB_PAGE_UPDATE;
882 }
883 
884 /*
885  * rb_head_page_activate - sets up head page
886  */
887 static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
888 {
889         struct buffer_page *head;
890 
891         head = cpu_buffer->head_page;
892         if (!head)
893                 return;
894 
895         /*
896          * Set the previous list pointer to have the HEAD flag.
897          */
898         rb_set_list_to_head(cpu_buffer, head->list.prev);
899 }
900 
901 static void rb_list_head_clear(struct list_head *list)
902 {
903         unsigned long *ptr = (unsigned long *)&list->next;
904 
905         *ptr &= ~RB_FLAG_MASK;
906 }
907 
908 /*
909  * rb_head_page_dactivate - clears head page ptr (for free list)
910  */
911 static void
912 rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
913 {
914         struct list_head *hd;
915 
916         /* Go through the whole list and clear any pointers found. */
917         rb_list_head_clear(cpu_buffer->pages);
918 
919         list_for_each(hd, cpu_buffer->pages)
920                 rb_list_head_clear(hd);
921 }
922 
923 static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
924                             struct buffer_page *head,
925                             struct buffer_page *prev,
926                             int old_flag, int new_flag)
927 {
928         struct list_head *list;
929         unsigned long val = (unsigned long)&head->list;
930         unsigned long ret;
931 
932         list = &prev->list;
933 
934         val &= ~RB_FLAG_MASK;
935 
936         ret = cmpxchg((unsigned long *)&list->next,
937                       val | old_flag, val | new_flag);
938 
939         /* check if the reader took the page */
940         if ((ret & ~RB_FLAG_MASK) != val)
941                 return RB_PAGE_MOVED;
942 
943         return ret & RB_FLAG_MASK;
944 }
945 
946 static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
947                                    struct buffer_page *head,
948                                    struct buffer_page *prev,
949                                    int old_flag)
950 {
951         return rb_head_page_set(cpu_buffer, head, prev,
952                                 old_flag, RB_PAGE_UPDATE);
953 }
954 
955 static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
956                                  struct buffer_page *head,
957                                  struct buffer_page *prev,
958                                  int old_flag)
959 {
960         return rb_head_page_set(cpu_buffer, head, prev,
961                                 old_flag, RB_PAGE_HEAD);
962 }
963 
964 static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
965                                    struct buffer_page *head,
966                                    struct buffer_page *prev,
967                                    int old_flag)
968 {
969         return rb_head_page_set(cpu_buffer, head, prev,
970                                 old_flag, RB_PAGE_NORMAL);
971 }
972 
973 static inline void rb_inc_page(struct ring_buffer_per_cpu *cpu_buffer,
974                                struct buffer_page **bpage)
975 {
976         struct list_head *p = rb_list_head((*bpage)->list.next);
977 
978         *bpage = list_entry(p, struct buffer_page, list);
979 }
980 
981 static struct buffer_page *
982 rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
983 {
984         struct buffer_page *head;
985         struct buffer_page *page;
986         struct list_head *list;
987         int i;
988 
989         if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
990                 return NULL;
991 
992         /* sanity check */
993         list = cpu_buffer->pages;
994         if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
995                 return NULL;
996 
997         page = head = cpu_buffer->head_page;
998         /*
999          * It is possible that the writer moves the header behind
1000          * where we started, and we miss in one loop.
1001          * A second loop should grab the header, but we'll do
1002          * three loops just because I'm paranoid.
1003          */
1004         for (i = 0; i < 3; i++) {
1005                 do {
1006                         if (rb_is_head_page(cpu_buffer, page, page->list.prev)) {
1007                                 cpu_buffer->head_page = page;
1008                                 return page;
1009                         }
1010                         rb_inc_page(cpu_buffer, &page);
1011                 } while (page != head);
1012         }
1013 
1014         RB_WARN_ON(cpu_buffer, 1);
1015 
1016         return NULL;
1017 }
1018 
1019 static int rb_head_page_replace(struct buffer_page *old,
1020                                 struct buffer_page *new)
1021 {
1022         unsigned long *ptr = (unsigned long *)&old->list.prev->next;
1023         unsigned long val;
1024         unsigned long ret;
1025 
1026         val = *ptr & ~RB_FLAG_MASK;
1027         val |= RB_PAGE_HEAD;
1028 
1029         ret = cmpxchg(ptr, val, (unsigned long)&new->list);
1030 
1031         return ret == val;
1032 }
1033 
1034 /*
1035  * rb_tail_page_update - move the tail page forward
1036  */
1037 static void rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
1038                                struct buffer_page *tail_page,
1039                                struct buffer_page *next_page)
1040 {
1041         unsigned long old_entries;
1042         unsigned long old_write;
1043 
1044         /*
1045          * The tail page now needs to be moved forward.
1046          *
1047          * We need to reset the tail page, but without messing
1048          * with possible erasing of data brought in by interrupts
1049          * that have moved the tail page and are currently on it.
1050          *
1051          * We add a counter to the write field to denote this.
1052          */
1053         old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
1054         old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
1055 
1056         /*
1057          * Just make sure we have seen our old_write and synchronize
1058          * with any interrupts that come in.
1059          */
1060         barrier();
1061 
1062         /*
1063          * If the tail page is still the same as what we think
1064          * it is, then it is up to us to update the tail
1065          * pointer.
1066          */
1067         if (tail_page == READ_ONCE(cpu_buffer->tail_page)) {
1068                 /* Zero the write counter */
1069                 unsigned long val = old_write & ~RB_WRITE_MASK;
1070                 unsigned long eval = old_entries & ~RB_WRITE_MASK;
1071 
1072                 /*
1073                  * This will only succeed if an interrupt did
1074                  * not come in and change it. In which case, we
1075                  * do not want to modify it.
1076                  *
1077                  * We add (void) to let the compiler know that we do not care
1078                  * about the return value of these functions. We use the
1079                  * cmpxchg to only update if an interrupt did not already
1080                  * do it for us. If the cmpxchg fails, we don't care.
1081                  */
1082                 (void)local_cmpxchg(&next_page->write, old_write, val);
1083                 (void)local_cmpxchg(&next_page->entries, old_entries, eval);
1084 
1085                 /*
1086                  * No need to worry about races with clearing out the commit.
1087                  * it only can increment when a commit takes place. But that
1088                  * only happens in the outer most nested commit.
1089                  */
1090                 local_set(&next_page->page->commit, 0);
1091 
1092                 /* Again, either we update tail_page or an interrupt does */
1093                 (void)cmpxchg(&cpu_buffer->tail_page, tail_page, next_page);
1094         }
1095 }
1096 
1097 static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
1098                           struct buffer_page *bpage)
1099 {
1100         unsigned long val = (unsigned long)bpage;
1101 
1102         if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK))
1103                 return 1;
1104 
1105         return 0;
1106 }
1107 
1108 /**
1109  * rb_check_list - make sure a pointer to a list has the last bits zero
1110  */
1111 static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer,
1112                          struct list_head *list)
1113 {
1114         if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev))
1115                 return 1;
1116         if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next))
1117                 return 1;
1118         return 0;
1119 }
1120 
1121 /**
1122  * rb_check_pages - integrity check of buffer pages
1123  * @cpu_buffer: CPU buffer with pages to test
1124  *
1125  * As a safety measure we check to make sure the data pages have not
1126  * been corrupted.
1127  */
1128 static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
1129 {
1130         struct list_head *head = cpu_buffer->pages;
1131         struct buffer_page *bpage, *tmp;
1132 
1133         /* Reset the head page if it exists */
1134         if (cpu_buffer->head_page)
1135                 rb_set_head_page(cpu_buffer);
1136 
1137         rb_head_page_deactivate(cpu_buffer);
1138 
1139         if (RB_WARN_ON(cpu_buffer, head->next->prev != head))
1140                 return -1;
1141         if (RB_WARN_ON(cpu_buffer, head->prev->next != head))
1142                 return -1;
1143 
1144         if (rb_check_list(cpu_buffer, head))
1145                 return -1;
1146 
1147         list_for_each_entry_safe(bpage, tmp, head, list) {
1148                 if (RB_WARN_ON(cpu_buffer,
1149                                bpage->list.next->prev != &bpage->list))
1150                         return -1;
1151                 if (RB_WARN_ON(cpu_buffer,
1152                                bpage->list.prev->next != &bpage->list))
1153                         return -1;
1154                 if (rb_check_list(cpu_buffer, &bpage->list))
1155                         return -1;
1156         }
1157 
1158         rb_head_page_activate(cpu_buffer);
1159 
1160         return 0;
1161 }
1162 
1163 static int __rb_allocate_pages(long nr_pages, struct list_head *pages, int cpu)
1164 {
1165         struct buffer_page *bpage, *tmp;
1166         bool user_thread = current->mm != NULL;
1167         gfp_t mflags;
1168         long i;
1169 
1170         /*
1171          * Check if the available memory is there first.
1172          * Note, si_mem_available() only gives us a rough estimate of available
1173          * memory. It may not be accurate. But we don't care, we just want
1174          * to prevent doing any allocation when it is obvious that it is
1175          * not going to succeed.
1176          */
1177         i = si_mem_available();
1178         if (i < nr_pages)
1179                 return -ENOMEM;
1180 
1181         /*
1182          * __GFP_RETRY_MAYFAIL flag makes sure that the allocation fails
1183          * gracefully without invoking oom-killer and the system is not
1184          * destabilized.
1185          */
1186         mflags = GFP_KERNEL | __GFP_RETRY_MAYFAIL;
1187 
1188         /*
1189          * If a user thread allocates too much, and si_mem_available()
1190          * reports there's enough memory, even though there is not.
1191          * Make sure the OOM killer kills this thread. This can happen
1192          * even with RETRY_MAYFAIL because another task may be doing
1193          * an allocation after this task has taken all memory.
1194          * This is the task the OOM killer needs to take out during this
1195          * loop, even if it was triggered by an allocation somewhere else.
1196          */
1197         if (user_thread)
1198                 set_current_oom_origin();
1199         for (i = 0; i < nr_pages; i++) {
1200                 struct page *page;
1201 
1202                 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1203                                     mflags, cpu_to_node(cpu));
1204                 if (!bpage)
1205                         goto free_pages;
1206 
1207                 list_add(&bpage->list, pages);
1208 
1209                 page = alloc_pages_node(cpu_to_node(cpu), mflags, 0);
1210                 if (!page)
1211                         goto free_pages;
1212                 bpage->page = page_address(page);
1213                 rb_init_page(bpage->page);
1214 
1215                 if (user_thread && fatal_signal_pending(current))
1216                         goto free_pages;
1217         }
1218         if (user_thread)
1219                 clear_current_oom_origin();
1220 
1221         return 0;
1222 
1223 free_pages:
1224         list_for_each_entry_safe(bpage, tmp, pages, list) {
1225                 list_del_init(&bpage->list);
1226                 free_buffer_page(bpage);
1227         }
1228         if (user_thread)
1229                 clear_current_oom_origin();
1230 
1231         return -ENOMEM;
1232 }
1233 
1234 static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1235                              unsigned long nr_pages)
1236 {
1237         LIST_HEAD(pages);
1238 
1239         WARN_ON(!nr_pages);
1240 
1241         if (__rb_allocate_pages(nr_pages, &pages, cpu_buffer->cpu))
1242                 return -ENOMEM;
1243 
1244         /*
1245          * The ring buffer page list is a circular list that does not
1246          * start and end with a list head. All page list items point to
1247          * other pages.
1248          */
1249         cpu_buffer->pages = pages.next;
1250         list_del(&pages);
1251 
1252         cpu_buffer->nr_pages = nr_pages;
1253 
1254         rb_check_pages(cpu_buffer);
1255 
1256         return 0;
1257 }
1258 
1259 static struct ring_buffer_per_cpu *
1260 rb_allocate_cpu_buffer(struct ring_buffer *buffer, long nr_pages, int cpu)
1261 {
1262         struct ring_buffer_per_cpu *cpu_buffer;
1263         struct buffer_page *bpage;
1264         struct page *page;
1265         int ret;
1266 
1267         cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1268                                   GFP_KERNEL, cpu_to_node(cpu));
1269         if (!cpu_buffer)
1270                 return NULL;
1271 
1272         cpu_buffer->cpu = cpu;
1273         cpu_buffer->buffer = buffer;
1274         raw_spin_lock_init(&cpu_buffer->reader_lock);
1275         lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1276         cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1277         INIT_WORK(&cpu_buffer->update_pages_work, update_pages_handler);
1278         init_completion(&cpu_buffer->update_done);
1279         init_irq_work(&cpu_buffer->irq_work.work, rb_wake_up_waiters);
1280         init_waitqueue_head(&cpu_buffer->irq_work.waiters);
1281         init_waitqueue_head(&cpu_buffer->irq_work.full_waiters);
1282 
1283         bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1284                             GFP_KERNEL, cpu_to_node(cpu));
1285         if (!bpage)
1286                 goto fail_free_buffer;
1287 
1288         rb_check_bpage(cpu_buffer, bpage);
1289 
1290         cpu_buffer->reader_page = bpage;
1291         page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0);
1292         if (!page)
1293                 goto fail_free_reader;
1294         bpage->page = page_address(page);
1295         rb_init_page(bpage->page);
1296 
1297         INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1298         INIT_LIST_HEAD(&cpu_buffer->new_pages);
1299 
1300         ret = rb_allocate_pages(cpu_buffer, nr_pages);
1301         if (ret < 0)
1302                 goto fail_free_reader;
1303 
1304         cpu_buffer->head_page
1305                 = list_entry(cpu_buffer->pages, struct buffer_page, list);
1306         cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1307 
1308         rb_head_page_activate(cpu_buffer);
1309 
1310         return cpu_buffer;
1311 
1312  fail_free_reader:
1313         free_buffer_page(cpu_buffer->reader_page);
1314 
1315  fail_free_buffer:
1316         kfree(cpu_buffer);
1317         return NULL;
1318 }
1319 
1320 static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1321 {
1322         struct list_head *head = cpu_buffer->pages;
1323         struct buffer_page *bpage, *tmp;
1324 
1325         free_buffer_page(cpu_buffer->reader_page);
1326 
1327         rb_head_page_deactivate(cpu_buffer);
1328 
1329         if (head) {
1330                 list_for_each_entry_safe(bpage, tmp, head, list) {
1331                         list_del_init(&bpage->list);
1332                         free_buffer_page(bpage);
1333                 }
1334                 bpage = list_entry(head, struct buffer_page, list);
1335                 free_buffer_page(bpage);
1336         }
1337 
1338         kfree(cpu_buffer);
1339 }
1340 
1341 /**
1342  * __ring_buffer_alloc - allocate a new ring_buffer
1343  * @size: the size in bytes per cpu that is needed.
1344  * @flags: attributes to set for the ring buffer.
1345  *
1346  * Currently the only flag that is available is the RB_FL_OVERWRITE
1347  * flag. This flag means that the buffer will overwrite old data
1348  * when the buffer wraps. If this flag is not set, the buffer will
1349  * drop data when the tail hits the head.
1350  */
1351 struct ring_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1352                                         struct lock_class_key *key)
1353 {
1354         struct ring_buffer *buffer;
1355         long nr_pages;
1356         int bsize;
1357         int cpu;
1358         int ret;
1359 
1360         /* keep it in its own cache line */
1361         buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1362                          GFP_KERNEL);
1363         if (!buffer)
1364                 return NULL;
1365 
1366         if (!zalloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1367                 goto fail_free_buffer;
1368 
1369         nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1370         buffer->flags = flags;
1371         buffer->clock = trace_clock_local;
1372         buffer->reader_lock_key = key;
1373 
1374         init_irq_work(&buffer->irq_work.work, rb_wake_up_waiters);
1375         init_waitqueue_head(&buffer->irq_work.waiters);
1376 
1377         /* need at least two pages */
1378         if (nr_pages < 2)
1379                 nr_pages = 2;
1380 
1381         buffer->cpus = nr_cpu_ids;
1382 
1383         bsize = sizeof(void *) * nr_cpu_ids;
1384         buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1385                                   GFP_KERNEL);
1386         if (!buffer->buffers)
1387                 goto fail_free_cpumask;
1388 
1389         cpu = raw_smp_processor_id();
1390         cpumask_set_cpu(cpu, buffer->cpumask);
1391         buffer->buffers[cpu] = rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
1392         if (!buffer->buffers[cpu])
1393                 goto fail_free_buffers;
1394 
1395         ret = cpuhp_state_add_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1396         if (ret < 0)
1397                 goto fail_free_buffers;
1398 
1399         mutex_init(&buffer->mutex);
1400 
1401         return buffer;
1402 
1403  fail_free_buffers:
1404         for_each_buffer_cpu(buffer, cpu) {
1405                 if (buffer->buffers[cpu])
1406                         rb_free_cpu_buffer(buffer->buffers[cpu]);
1407         }
1408         kfree(buffer->buffers);
1409 
1410  fail_free_cpumask:
1411         free_cpumask_var(buffer->cpumask);
1412 
1413  fail_free_buffer:
1414         kfree(buffer);
1415         return NULL;
1416 }
1417 EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1418 
1419 /**
1420  * ring_buffer_free - free a ring buffer.
1421  * @buffer: the buffer to free.
1422  */
1423 void
1424 ring_buffer_free(struct ring_buffer *buffer)
1425 {
1426         int cpu;
1427 
1428         cpuhp_state_remove_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1429 
1430         for_each_buffer_cpu(buffer, cpu)
1431                 rb_free_cpu_buffer(buffer->buffers[cpu]);
1432 
1433         kfree(buffer->buffers);
1434         free_cpumask_var(buffer->cpumask);
1435 
1436         kfree(buffer);
1437 }
1438 EXPORT_SYMBOL_GPL(ring_buffer_free);
1439 
1440 void ring_buffer_set_clock(struct ring_buffer *buffer,
1441                            u64 (*clock)(void))
1442 {
1443         buffer->clock = clock;
1444 }
1445 
1446 void ring_buffer_set_time_stamp_abs(struct ring_buffer *buffer, bool abs)
1447 {
1448         buffer->time_stamp_abs = abs;
1449 }
1450 
1451 bool ring_buffer_time_stamp_abs(struct ring_buffer *buffer)
1452 {
1453         return buffer->time_stamp_abs;
1454 }
1455 
1456 static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1457 
1458 static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1459 {
1460         return local_read(&bpage->entries) & RB_WRITE_MASK;
1461 }
1462 
1463 static inline unsigned long rb_page_write(struct buffer_page *bpage)
1464 {
1465         return local_read(&bpage->write) & RB_WRITE_MASK;
1466 }
1467 
1468 static int
1469 rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned long nr_pages)
1470 {
1471         struct list_head *tail_page, *to_remove, *next_page;
1472         struct buffer_page *to_remove_page, *tmp_iter_page;
1473         struct buffer_page *last_page, *first_page;
1474         unsigned long nr_removed;
1475         unsigned long head_bit;
1476         int page_entries;
1477 
1478         head_bit = 0;
1479 
1480         raw_spin_lock_irq(&cpu_buffer->reader_lock);
1481         atomic_inc(&cpu_buffer->record_disabled);
1482         /*
1483          * We don't race with the readers since we have acquired the reader
1484          * lock. We also don't race with writers after disabling recording.
1485          * This makes it easy to figure out the first and the last page to be
1486          * removed from the list. We unlink all the pages in between including
1487          * the first and last pages. This is done in a busy loop so that we
1488          * lose the least number of traces.
1489          * The pages are freed after we restart recording and unlock readers.
1490          */
1491         tail_page = &cpu_buffer->tail_page->list;
1492 
1493         /*
1494          * tail page might be on reader page, we remove the next page
1495          * from the ring buffer
1496          */
1497         if (cpu_buffer->tail_page == cpu_buffer->reader_page)
1498                 tail_page = rb_list_head(tail_page->next);
1499         to_remove = tail_page;
1500 
1501         /* start of pages to remove */
1502         first_page = list_entry(rb_list_head(to_remove->next),
1503                                 struct buffer_page, list);
1504 
1505         for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
1506                 to_remove = rb_list_head(to_remove)->next;
1507                 head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
1508         }
1509 
1510         next_page = rb_list_head(to_remove)->next;
1511 
1512         /*
1513          * Now we remove all pages between tail_page and next_page.
1514          * Make sure that we have head_bit value preserved for the
1515          * next page
1516          */
1517         tail_page->next = (struct list_head *)((unsigned long)next_page |
1518                                                 head_bit);
1519         next_page = rb_list_head(next_page);
1520         next_page->prev = tail_page;
1521 
1522         /* make sure pages points to a valid page in the ring buffer */
1523         cpu_buffer->pages = next_page;
1524 
1525         /* update head page */
1526         if (head_bit)
1527                 cpu_buffer->head_page = list_entry(next_page,
1528                                                 struct buffer_page, list);
1529 
1530         /*
1531          * change read pointer to make sure any read iterators reset
1532          * themselves
1533          */
1534         cpu_buffer->read = 0;
1535 
1536         /* pages are removed, resume tracing and then free the pages */
1537         atomic_dec(&cpu_buffer->record_disabled);
1538         raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1539 
1540         RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
1541 
1542         /* last buffer page to remove */
1543         last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
1544                                 list);
1545         tmp_iter_page = first_page;
1546 
1547         do {
1548                 to_remove_page = tmp_iter_page;
1549                 rb_inc_page(cpu_buffer, &tmp_iter_page);
1550 
1551                 /* update the counters */
1552                 page_entries = rb_page_entries(to_remove_page);
1553                 if (page_entries) {
1554                         /*
1555                          * If something was added to this page, it was full
1556                          * since it is not the tail page. So we deduct the
1557                          * bytes consumed in ring buffer from here.
1558                          * Increment overrun to account for the lost events.
1559                          */
1560                         local_add(page_entries, &cpu_buffer->overrun);
1561                         local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1562                 }
1563 
1564                 /*
1565                  * We have already removed references to this list item, just
1566                  * free up the buffer_page and its page
1567                  */
1568                 free_buffer_page(to_remove_page);
1569                 nr_removed--;
1570 
1571         } while (to_remove_page != last_page);
1572 
1573         RB_WARN_ON(cpu_buffer, nr_removed);
1574 
1575         return nr_removed == 0;
1576 }
1577 
1578 static int
1579 rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
1580 {
1581         struct list_head *pages = &cpu_buffer->new_pages;
1582         int retries, success;
1583 
1584         raw_spin_lock_irq(&cpu_buffer->reader_lock);
1585         /*
1586          * We are holding the reader lock, so the reader page won't be swapped
1587          * in the ring buffer. Now we are racing with the writer trying to
1588          * move head page and the tail page.
1589          * We are going to adapt the reader page update process where:
1590          * 1. We first splice the start and end of list of new pages between
1591          *    the head page and its previous page.
1592          * 2. We cmpxchg the prev_page->next to point from head page to the
1593          *    start of new pages list.
1594          * 3. Finally, we update the head->prev to the end of new list.
1595          *
1596          * We will try this process 10 times, to make sure that we don't keep
1597          * spinning.
1598          */
1599         retries = 10;
1600         success = 0;
1601         while (retries--) {
1602                 struct list_head *head_page, *prev_page, *r;
1603                 struct list_head *last_page, *first_page;
1604                 struct list_head *head_page_with_bit;
1605 
1606                 head_page = &rb_set_head_page(cpu_buffer)->list;
1607                 if (!head_page)
1608                         break;
1609                 prev_page = head_page->prev;
1610 
1611                 first_page = pages->next;
1612                 last_page  = pages->prev;
1613 
1614                 head_page_with_bit = (struct list_head *)
1615                                      ((unsigned long)head_page | RB_PAGE_HEAD);
1616 
1617                 last_page->next = head_page_with_bit;
1618                 first_page->prev = prev_page;
1619 
1620                 r = cmpxchg(&prev_page->next, head_page_with_bit, first_page);
1621 
1622                 if (r == head_page_with_bit) {
1623                         /*
1624                          * yay, we replaced the page pointer to our new list,
1625                          * now, we just have to update to head page's prev
1626                          * pointer to point to end of list
1627                          */
1628                         head_page->prev = last_page;
1629                         success = 1;
1630                         break;
1631                 }
1632         }
1633 
1634         if (success)
1635                 INIT_LIST_HEAD(pages);
1636         /*
1637          * If we weren't successful in adding in new pages, warn and stop
1638          * tracing
1639          */
1640         RB_WARN_ON(cpu_buffer, !success);
1641         raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1642 
1643         /* free pages if they weren't inserted */
1644         if (!success) {
1645                 struct buffer_page *bpage, *tmp;
1646                 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1647                                          list) {
1648                         list_del_init(&bpage->list);
1649                         free_buffer_page(bpage);
1650                 }
1651         }
1652         return success;
1653 }
1654 
1655 static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
1656 {
1657         int success;
1658 
1659         if (cpu_buffer->nr_pages_to_update > 0)
1660                 success = rb_insert_pages(cpu_buffer);
1661         else
1662                 success = rb_remove_pages(cpu_buffer,
1663                                         -cpu_buffer->nr_pages_to_update);
1664 
1665         if (success)
1666                 cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
1667 }
1668 
1669 static void update_pages_handler(struct work_struct *work)
1670 {
1671         struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
1672                         struct ring_buffer_per_cpu, update_pages_work);
1673         rb_update_pages(cpu_buffer);
1674         complete(&cpu_buffer->update_done);
1675 }
1676 
1677 /**
1678  * ring_buffer_resize - resize the ring buffer
1679  * @buffer: the buffer to resize.
1680  * @size: the new size.
1681  * @cpu_id: the cpu buffer to resize
1682  *
1683  * Minimum size is 2 * BUF_PAGE_SIZE.
1684  *
1685  * Returns 0 on success and < 0 on failure.
1686  */
1687 int ring_buffer_resize(struct ring_buffer *buffer, unsigned long size,
1688                         int cpu_id)
1689 {
1690         struct ring_buffer_per_cpu *cpu_buffer;
1691         unsigned long nr_pages;
1692         int cpu, err = 0;
1693 
1694         /*
1695          * Always succeed at resizing a non-existent buffer:
1696          */
1697         if (!buffer)
1698                 return size;
1699 
1700         /* Make sure the requested buffer exists */
1701         if (cpu_id != RING_BUFFER_ALL_CPUS &&
1702             !cpumask_test_cpu(cpu_id, buffer->cpumask))
1703                 return size;
1704 
1705         nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1706 
1707         /* we need a minimum of two pages */
1708         if (nr_pages < 2)
1709                 nr_pages = 2;
1710 
1711         size = nr_pages * BUF_PAGE_SIZE;
1712 
1713         /*
1714          * Don't succeed if resizing is disabled, as a reader might be
1715          * manipulating the ring buffer and is expecting a sane state while
1716          * this is true.
1717          */
1718         if (atomic_read(&buffer->resize_disabled))
1719                 return -EBUSY;
1720 
1721         /* prevent another thread from changing buffer sizes */
1722         mutex_lock(&buffer->mutex);
1723 
1724         if (cpu_id == RING_BUFFER_ALL_CPUS) {
1725                 /* calculate the pages to update */
1726                 for_each_buffer_cpu(buffer, cpu) {
1727                         cpu_buffer = buffer->buffers[cpu];
1728 
1729                         cpu_buffer->nr_pages_to_update = nr_pages -
1730                                                         cpu_buffer->nr_pages;
1731                         /*
1732                          * nothing more to do for removing pages or no update
1733                          */
1734                         if (cpu_buffer->nr_pages_to_update <= 0)
1735                                 continue;
1736                         /*
1737                          * to add pages, make sure all new pages can be
1738                          * allocated without receiving ENOMEM
1739                          */
1740                         INIT_LIST_HEAD(&cpu_buffer->new_pages);
1741                         if (__rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1742                                                 &cpu_buffer->new_pages, cpu)) {
1743                                 /* not enough memory for new pages */
1744                                 err = -ENOMEM;
1745                                 goto out_err;
1746                         }
1747                 }
1748 
1749                 get_online_cpus();
1750                 /*
1751                  * Fire off all the required work handlers
1752                  * We can't schedule on offline CPUs, but it's not necessary
1753                  * since we can change their buffer sizes without any race.
1754                  */
1755                 for_each_buffer_cpu(buffer, cpu) {
1756                         cpu_buffer = buffer->buffers[cpu];
1757                         if (!cpu_buffer->nr_pages_to_update)
1758                                 continue;
1759 
1760                         /* Can't run something on an offline CPU. */
1761                         if (!cpu_online(cpu)) {
1762                                 rb_update_pages(cpu_buffer);
1763                                 cpu_buffer->nr_pages_to_update = 0;
1764                         } else {
1765                                 schedule_work_on(cpu,
1766                                                 &cpu_buffer->update_pages_work);
1767                         }
1768                 }
1769 
1770                 /* wait for all the updates to complete */
1771                 for_each_buffer_cpu(buffer, cpu) {
1772                         cpu_buffer = buffer->buffers[cpu];
1773                         if (!cpu_buffer->nr_pages_to_update)
1774                                 continue;
1775 
1776                         if (cpu_online(cpu))
1777                                 wait_for_completion(&cpu_buffer->update_done);
1778                         cpu_buffer->nr_pages_to_update = 0;
1779                 }
1780 
1781                 put_online_cpus();
1782         } else {
1783                 /* Make sure this CPU has been intitialized */
1784                 if (!cpumask_test_cpu(cpu_id, buffer->cpumask))
1785                         goto out;
1786 
1787                 cpu_buffer = buffer->buffers[cpu_id];
1788 
1789                 if (nr_pages == cpu_buffer->nr_pages)
1790                         goto out;
1791 
1792                 cpu_buffer->nr_pages_to_update = nr_pages -
1793                                                 cpu_buffer->nr_pages;
1794 
1795                 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1796                 if (cpu_buffer->nr_pages_to_update > 0 &&
1797                         __rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1798                                             &cpu_buffer->new_pages, cpu_id)) {
1799                         err = -ENOMEM;
1800                         goto out_err;
1801                 }
1802 
1803                 get_online_cpus();
1804 
1805                 /* Can't run something on an offline CPU. */
1806                 if (!cpu_online(cpu_id))
1807                         rb_update_pages(cpu_buffer);
1808                 else {
1809                         schedule_work_on(cpu_id,
1810                                          &cpu_buffer->update_pages_work);
1811                         wait_for_completion(&cpu_buffer->update_done);
1812                 }
1813 
1814                 cpu_buffer->nr_pages_to_update = 0;
1815                 put_online_cpus();
1816         }
1817 
1818  out:
1819         /*
1820          * The ring buffer resize can happen with the ring buffer
1821          * enabled, so that the update disturbs the tracing as little
1822          * as possible. But if the buffer is disabled, we do not need
1823          * to worry about that, and we can take the time to verify
1824          * that the buffer is not corrupt.
1825          */
1826         if (atomic_read(&buffer->record_disabled)) {
1827                 atomic_inc(&buffer->record_disabled);
1828                 /*
1829                  * Even though the buffer was disabled, we must make sure
1830                  * that it is truly disabled before calling rb_check_pages.
1831                  * There could have been a race between checking
1832                  * record_disable and incrementing it.
1833                  */
1834                 synchronize_sched();
1835                 for_each_buffer_cpu(buffer, cpu) {
1836                         cpu_buffer = buffer->buffers[cpu];
1837                         rb_check_pages(cpu_buffer);
1838                 }
1839                 atomic_dec(&buffer->record_disabled);
1840         }
1841 
1842         mutex_unlock(&buffer->mutex);
1843         return size;
1844 
1845  out_err:
1846         for_each_buffer_cpu(buffer, cpu) {
1847                 struct buffer_page *bpage, *tmp;
1848 
1849                 cpu_buffer = buffer->buffers[cpu];
1850                 cpu_buffer->nr_pages_to_update = 0;
1851 
1852                 if (list_empty(&cpu_buffer->new_pages))
1853                         continue;
1854 
1855                 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1856                                         list) {
1857                         list_del_init(&bpage->list);
1858                         free_buffer_page(bpage);
1859                 }
1860         }
1861         mutex_unlock(&buffer->mutex);
1862         return err;
1863 }
1864 EXPORT_SYMBOL_GPL(ring_buffer_resize);
1865 
1866 void ring_buffer_change_overwrite(struct ring_buffer *buffer, int val)
1867 {
1868         mutex_lock(&buffer->mutex);
1869         if (val)
1870                 buffer->flags |= RB_FL_OVERWRITE;
1871         else
1872                 buffer->flags &= ~RB_FL_OVERWRITE;
1873         mutex_unlock(&buffer->mutex);
1874 }
1875 EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
1876 
1877 static __always_inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
1878 {
1879         return bpage->page->data + index;
1880 }
1881 
1882 static __always_inline struct ring_buffer_event *
1883 rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
1884 {
1885         return __rb_page_index(cpu_buffer->reader_page,
1886                                cpu_buffer->reader_page->read);
1887 }
1888 
1889 static __always_inline struct ring_buffer_event *
1890 rb_iter_head_event(struct ring_buffer_iter *iter)
1891 {
1892         return __rb_page_index(iter->head_page, iter->head);
1893 }
1894 
1895 static __always_inline unsigned rb_page_commit(struct buffer_page *bpage)
1896 {
1897         return local_read(&bpage->page->commit);
1898 }
1899 
1900 /* Size is determined by what has been committed */
1901 static __always_inline unsigned rb_page_size(struct buffer_page *bpage)
1902 {
1903         return rb_page_commit(bpage);
1904 }
1905 
1906 static __always_inline unsigned
1907 rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
1908 {
1909         return rb_page_commit(cpu_buffer->commit_page);
1910 }
1911 
1912 static __always_inline unsigned
1913 rb_event_index(struct ring_buffer_event *event)
1914 {
1915         unsigned long addr = (unsigned long)event;
1916 
1917         return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
1918 }
1919 
1920 static void rb_inc_iter(struct ring_buffer_iter *iter)
1921 {
1922         struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
1923 
1924         /*
1925          * The iterator could be on the reader page (it starts there).
1926          * But the head could have moved, since the reader was
1927          * found. Check for this case and assign the iterator
1928          * to the head page instead of next.
1929          */
1930         if (iter->head_page == cpu_buffer->reader_page)
1931                 iter->head_page = rb_set_head_page(cpu_buffer);
1932         else
1933                 rb_inc_page(cpu_buffer, &iter->head_page);
1934 
1935         iter->read_stamp = iter->head_page->page->time_stamp;
1936         iter->head = 0;
1937 }
1938 
1939 /*
1940  * rb_handle_head_page - writer hit the head page
1941  *
1942  * Returns: +1 to retry page
1943  *           0 to continue
1944  *          -1 on error
1945  */
1946 static int
1947 rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
1948                     struct buffer_page *tail_page,
1949                     struct buffer_page *next_page)
1950 {
1951         struct buffer_page *new_head;
1952         int entries;
1953         int type;
1954         int ret;
1955 
1956         entries = rb_page_entries(next_page);
1957 
1958         /*
1959          * The hard part is here. We need to move the head
1960          * forward, and protect against both readers on
1961          * other CPUs and writers coming in via interrupts.
1962          */
1963         type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
1964                                        RB_PAGE_HEAD);
1965 
1966         /*
1967          * type can be one of four:
1968          *  NORMAL - an interrupt already moved it for us
1969          *  HEAD   - we are the first to get here.
1970          *  UPDATE - we are the interrupt interrupting
1971          *           a current move.
1972          *  MOVED  - a reader on another CPU moved the next
1973          *           pointer to its reader page. Give up
1974          *           and try again.
1975          */
1976 
1977         switch (type) {
1978         case RB_PAGE_HEAD:
1979                 /*
1980                  * We changed the head to UPDATE, thus
1981                  * it is our responsibility to update
1982                  * the counters.
1983                  */
1984                 local_add(entries, &cpu_buffer->overrun);
1985                 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1986 
1987                 /*
1988                  * The entries will be zeroed out when we move the
1989                  * tail page.
1990                  */
1991 
1992                 /* still more to do */
1993                 break;
1994 
1995         case RB_PAGE_UPDATE:
1996                 /*
1997                  * This is an interrupt that interrupt the
1998                  * previous update. Still more to do.
1999                  */
2000                 break;
2001         case RB_PAGE_NORMAL:
2002                 /*
2003                  * An interrupt came in before the update
2004                  * and processed this for us.
2005                  * Nothing left to do.
2006                  */
2007                 return 1;
2008         case RB_PAGE_MOVED:
2009                 /*
2010                  * The reader is on another CPU and just did
2011                  * a swap with our next_page.
2012                  * Try again.
2013                  */
2014                 return 1;
2015         default:
2016                 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
2017                 return -1;
2018         }
2019 
2020         /*
2021          * Now that we are here, the old head pointer is
2022          * set to UPDATE. This will keep the reader from
2023          * swapping the head page with the reader page.
2024          * The reader (on another CPU) will spin till
2025          * we are finished.
2026          *
2027          * We just need to protect against interrupts
2028          * doing the job. We will set the next pointer
2029          * to HEAD. After that, we set the old pointer
2030          * to NORMAL, but only if it was HEAD before.
2031          * otherwise we are an interrupt, and only
2032          * want the outer most commit to reset it.
2033          */
2034         new_head = next_page;
2035         rb_inc_page(cpu_buffer, &new_head);
2036 
2037         ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
2038                                     RB_PAGE_NORMAL);
2039 
2040         /*
2041          * Valid returns are:
2042          *  HEAD   - an interrupt came in and already set it.
2043          *  NORMAL - One of two things:
2044          *            1) We really set it.
2045          *            2) A bunch of interrupts came in and moved
2046          *               the page forward again.
2047          */
2048         switch (ret) {
2049         case RB_PAGE_HEAD:
2050         case RB_PAGE_NORMAL:
2051                 /* OK */
2052                 break;
2053         default:
2054                 RB_WARN_ON(cpu_buffer, 1);
2055                 return -1;
2056         }
2057 
2058         /*
2059          * It is possible that an interrupt came in,
2060          * set the head up, then more interrupts came in
2061          * and moved it again. When we get back here,
2062          * the page would have been set to NORMAL but we
2063          * just set it back to HEAD.
2064          *
2065          * How do you detect this? Well, if that happened
2066          * the tail page would have moved.
2067          */
2068         if (ret == RB_PAGE_NORMAL) {
2069                 struct buffer_page *buffer_tail_page;
2070 
2071                 buffer_tail_page = READ_ONCE(cpu_buffer->tail_page);
2072                 /*
2073                  * If the tail had moved passed next, then we need
2074                  * to reset the pointer.
2075                  */
2076                 if (buffer_tail_page != tail_page &&
2077                     buffer_tail_page != next_page)
2078                         rb_head_page_set_normal(cpu_buffer, new_head,
2079                                                 next_page,
2080                                                 RB_PAGE_HEAD);
2081         }
2082 
2083         /*
2084          * If this was the outer most commit (the one that
2085          * changed the original pointer from HEAD to UPDATE),
2086          * then it is up to us to reset it to NORMAL.
2087          */
2088         if (type == RB_PAGE_HEAD) {
2089                 ret = rb_head_page_set_normal(cpu_buffer, next_page,
2090                                               tail_page,
2091                                               RB_PAGE_UPDATE);
2092                 if (RB_WARN_ON(cpu_buffer,
2093                                ret != RB_PAGE_UPDATE))
2094                         return -1;
2095         }
2096 
2097         return 0;
2098 }
2099 
2100 static inline void
2101 rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2102               unsigned long tail, struct rb_event_info *info)
2103 {
2104         struct buffer_page *tail_page = info->tail_page;
2105         struct ring_buffer_event *event;
2106         unsigned long length = info->length;
2107 
2108         /*
2109          * Only the event that crossed the page boundary
2110          * must fill the old tail_page with padding.
2111          */
2112         if (tail >= BUF_PAGE_SIZE) {
2113                 /*
2114                  * If the page was filled, then we still need
2115                  * to update the real_end. Reset it to zero
2116                  * and the reader will ignore it.
2117                  */
2118                 if (tail == BUF_PAGE_SIZE)
2119                         tail_page->real_end = 0;
2120 
2121                 local_sub(length, &tail_page->write);
2122                 return;
2123         }
2124 
2125         event = __rb_page_index(tail_page, tail);
2126 
2127         /* account for padding bytes */
2128         local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes);
2129 
2130         /*
2131          * Save the original length to the meta data.
2132          * This will be used by the reader to add lost event
2133          * counter.
2134          */
2135         tail_page->real_end = tail;
2136 
2137         /*
2138          * If this event is bigger than the minimum size, then
2139          * we need to be careful that we don't subtract the
2140          * write counter enough to allow another writer to slip
2141          * in on this page.
2142          * We put in a discarded commit instead, to make sure
2143          * that this space is not used again.
2144          *
2145          * If we are less than the minimum size, we don't need to
2146          * worry about it.
2147          */
2148         if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
2149                 /* No room for any events */
2150 
2151                 /* Mark the rest of the page with padding */
2152                 rb_event_set_padding(event);
2153 
2154                 /* Set the write back to the previous setting */
2155                 local_sub(length, &tail_page->write);
2156                 return;
2157         }
2158 
2159         /* Put in a discarded event */
2160         event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
2161         event->type_len = RINGBUF_TYPE_PADDING;
2162         /* time delta must be non zero */
2163         event->time_delta = 1;
2164 
2165         /* Set write to end of buffer */
2166         length = (tail + length) - BUF_PAGE_SIZE;
2167         local_sub(length, &tail_page->write);
2168 }
2169 
2170 static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer);
2171 
2172 /*
2173  * This is the slow path, force gcc not to inline it.
2174  */
2175 static noinline struct ring_buffer_event *
2176 rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2177              unsigned long tail, struct rb_event_info *info)
2178 {
2179         struct buffer_page *tail_page = info->tail_page;
2180         struct buffer_page *commit_page = cpu_buffer->commit_page;
2181         struct ring_buffer *buffer = cpu_buffer->buffer;
2182         struct buffer_page *next_page;
2183         int ret;
2184 
2185         next_page = tail_page;
2186 
2187         rb_inc_page(cpu_buffer, &next_page);
2188 
2189         /*
2190          * If for some reason, we had an interrupt storm that made
2191          * it all the way around the buffer, bail, and warn
2192          * about it.
2193          */
2194         if (unlikely(next_page == commit_page)) {
2195                 local_inc(&cpu_buffer->commit_overrun);
2196                 goto out_reset;
2197         }
2198 
2199         /*
2200          * This is where the fun begins!
2201          *
2202          * We are fighting against races between a reader that
2203          * could be on another CPU trying to swap its reader
2204          * page with the buffer head.
2205          *
2206          * We are also fighting against interrupts coming in and
2207          * moving the head or tail on us as well.
2208          *
2209          * If the next page is the head page then we have filled
2210          * the buffer, unless the commit page is still on the
2211          * reader page.
2212          */
2213         if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) {
2214 
2215                 /*
2216                  * If the commit is not on the reader page, then
2217                  * move the header page.
2218                  */
2219                 if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2220                         /*
2221                          * If we are not in overwrite mode,
2222                          * this is easy, just stop here.
2223                          */
2224                         if (!(buffer->flags & RB_FL_OVERWRITE)) {
2225                                 local_inc(&cpu_buffer->dropped_events);
2226                                 goto out_reset;
2227                         }
2228 
2229                         ret = rb_handle_head_page(cpu_buffer,
2230                                                   tail_page,
2231                                                   next_page);
2232                         if (ret < 0)
2233                                 goto out_reset;
2234                         if (ret)
2235                                 goto out_again;
2236                 } else {
2237                         /*
2238                          * We need to be careful here too. The
2239                          * commit page could still be on the reader
2240                          * page. We could have a small buffer, and
2241                          * have filled up the buffer with events
2242                          * from interrupts and such, and wrapped.
2243                          *
2244                          * Note, if the tail page is also the on the
2245                          * reader_page, we let it move out.
2246                          */
2247                         if (unlikely((cpu_buffer->commit_page !=
2248                                       cpu_buffer->tail_page) &&
2249                                      (cpu_buffer->commit_page ==
2250                                       cpu_buffer->reader_page))) {
2251                                 local_inc(&cpu_buffer->commit_overrun);
2252                                 goto out_reset;
2253                         }
2254                 }
2255         }
2256 
2257         rb_tail_page_update(cpu_buffer, tail_page, next_page);
2258 
2259  out_again:
2260 
2261         rb_reset_tail(cpu_buffer, tail, info);
2262 
2263         /* Commit what we have for now. */
2264         rb_end_commit(cpu_buffer);
2265         /* rb_end_commit() decs committing */
2266         local_inc(&cpu_buffer->committing);
2267 
2268         /* fail and let the caller try again */
2269         return ERR_PTR(-EAGAIN);
2270 
2271  out_reset:
2272         /* reset write */
2273         rb_reset_tail(cpu_buffer, tail, info);
2274 
2275         return NULL;
2276 }
2277 
2278 /* Slow path, do not inline */
2279 static noinline struct ring_buffer_event *
2280 rb_add_time_stamp(struct ring_buffer_event *event, u64 delta, bool abs)
2281 {
2282         if (abs)
2283                 event->type_len = RINGBUF_TYPE_TIME_STAMP;
2284         else
2285                 event->type_len = RINGBUF_TYPE_TIME_EXTEND;
2286 
2287         /* Not the first event on the page, or not delta? */
2288         if (abs || rb_event_index(event)) {
2289                 event->time_delta = delta & TS_MASK;
2290                 event->array[0] = delta >> TS_SHIFT;
2291         } else {
2292                 /* nope, just zero it */
2293                 event->time_delta = 0;
2294                 event->array[0] = 0;
2295         }
2296 
2297         return skip_time_extend(event);
2298 }
2299 
2300 static inline bool rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2301                                      struct ring_buffer_event *event);
2302 
2303 /**
2304  * rb_update_event - update event type and data
2305  * @event: the event to update
2306  * @type: the type of event
2307  * @length: the size of the event field in the ring buffer
2308  *
2309  * Update the type and data fields of the event. The length
2310  * is the actual size that is written to the ring buffer,
2311  * and with this, we can determine what to place into the
2312  * data field.
2313  */
2314 static void
2315 rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
2316                 struct ring_buffer_event *event,
2317                 struct rb_event_info *info)
2318 {
2319         unsigned length = info->length;
2320         u64 delta = info->delta;
2321 
2322         /* Only a commit updates the timestamp */
2323         if (unlikely(!rb_event_is_commit(cpu_buffer, event)))
2324                 delta = 0;
2325 
2326         /*
2327          * If we need to add a timestamp, then we
2328          * add it to the start of the resevered space.
2329          */
2330         if (unlikely(info->add_timestamp)) {
2331                 bool abs = ring_buffer_time_stamp_abs(cpu_buffer->buffer);
2332 
2333                 event = rb_add_time_stamp(event, info->delta, abs);
2334                 length -= RB_LEN_TIME_EXTEND;
2335                 delta = 0;
2336         }
2337 
2338         event->time_delta = delta;
2339         length -= RB_EVNT_HDR_SIZE;
2340         if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
2341                 event->type_len = 0;
2342                 event->array[0] = length;
2343         } else
2344                 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
2345 }
2346 
2347 static unsigned rb_calculate_event_length(unsigned length)
2348 {
2349         struct ring_buffer_event event; /* Used only for sizeof array */
2350 
2351         /* zero length can cause confusions */
2352         if (!length)
2353                 length++;
2354 
2355         if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
2356                 length += sizeof(event.array[0]);
2357 
2358         length += RB_EVNT_HDR_SIZE;
2359         length = ALIGN(length, RB_ARCH_ALIGNMENT);
2360 
2361         /*
2362          * In case the time delta is larger than the 27 bits for it
2363          * in the header, we need to add a timestamp. If another
2364          * event comes in when trying to discard this one to increase
2365          * the length, then the timestamp will be added in the allocated
2366          * space of this event. If length is bigger than the size needed
2367          * for the TIME_EXTEND, then padding has to be used. The events
2368          * length must be either RB_LEN_TIME_EXTEND, or greater than or equal
2369          * to RB_LEN_TIME_EXTEND + 8, as 8 is the minimum size for padding.
2370          * As length is a multiple of 4, we only need to worry if it
2371          * is 12 (RB_LEN_TIME_EXTEND + 4).
2372          */
2373         if (length == RB_LEN_TIME_EXTEND + RB_ALIGNMENT)
2374                 length += RB_ALIGNMENT;
2375 
2376         return length;
2377 }
2378 
2379 #ifndef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2380 static inline bool sched_clock_stable(void)
2381 {
2382         return true;
2383 }
2384 #endif
2385 
2386 static inline int
2387 rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2388                   struct ring_buffer_event *event)
2389 {
2390         unsigned long new_index, old_index;
2391         struct buffer_page *bpage;
2392         unsigned long index;
2393         unsigned long addr;
2394 
2395         new_index = rb_event_index(event);
2396         old_index = new_index + rb_event_ts_length(event);
2397         addr = (unsigned long)event;
2398         addr &= PAGE_MASK;
2399 
2400         bpage = READ_ONCE(cpu_buffer->tail_page);
2401 
2402         if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2403                 unsigned long write_mask =
2404                         local_read(&bpage->write) & ~RB_WRITE_MASK;
2405                 unsigned long event_length = rb_event_length(event);
2406                 /*
2407                  * This is on the tail page. It is possible that
2408                  * a write could come in and move the tail page
2409                  * and write to the next page. That is fine
2410                  * because we just shorten what is on this page.
2411                  */
2412                 old_index += write_mask;
2413                 new_index += write_mask;
2414                 index = local_cmpxchg(&bpage->write, old_index, new_index);
2415                 if (index == old_index) {
2416                         /* update counters */
2417                         local_sub(event_length, &cpu_buffer->entries_bytes);
2418                         return 1;
2419                 }
2420         }
2421 
2422         /* could not discard */
2423         return 0;
2424 }
2425 
2426 static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2427 {
2428         local_inc(&cpu_buffer->committing);
2429         local_inc(&cpu_buffer->commits);
2430 }
2431 
2432 static __always_inline void
2433 rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
2434 {
2435         unsigned long max_count;
2436 
2437         /*
2438          * We only race with interrupts and NMIs on this CPU.
2439          * If we own the commit event, then we can commit
2440          * all others that interrupted us, since the interruptions
2441          * are in stack format (they finish before they come
2442          * back to us). This allows us to do a simple loop to
2443          * assign the commit to the tail.
2444          */
2445  again:
2446         max_count = cpu_buffer->nr_pages * 100;
2447 
2448         while (cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)) {
2449                 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
2450                         return;
2451                 if (RB_WARN_ON(cpu_buffer,
2452                                rb_is_reader_page(cpu_buffer->tail_page)))
2453                         return;
2454                 local_set(&cpu_buffer->commit_page->page->commit,
2455                           rb_page_write(cpu_buffer->commit_page));
2456                 rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
2457                 /* Only update the write stamp if the page has an event */
2458                 if (rb_page_write(cpu_buffer->commit_page))
2459                         cpu_buffer->write_stamp =
2460                                 cpu_buffer->commit_page->page->time_stamp;
2461                 /* add barrier to keep gcc from optimizing too much */
2462                 barrier();
2463         }
2464         while (rb_commit_index(cpu_buffer) !=
2465                rb_page_write(cpu_buffer->commit_page)) {
2466 
2467                 local_set(&cpu_buffer->commit_page->page->commit,
2468                           rb_page_write(cpu_buffer->commit_page));
2469                 RB_WARN_ON(cpu_buffer,
2470                            local_read(&cpu_buffer->commit_page->page->commit) &
2471                            ~RB_WRITE_MASK);
2472                 barrier();
2473         }
2474 
2475         /* again, keep gcc from optimizing */
2476         barrier();
2477 
2478         /*
2479          * If an interrupt came in just after the first while loop
2480          * and pushed the tail page forward, we will be left with
2481          * a dangling commit that will never go forward.
2482          */
2483         if (unlikely(cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)))
2484                 goto again;
2485 }
2486 
2487 static __always_inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
2488 {
2489         unsigned long commits;
2490 
2491         if (RB_WARN_ON(cpu_buffer,
2492                        !local_read(&cpu_buffer->committing)))
2493                 return;
2494 
2495  again:
2496         commits = local_read(&cpu_buffer->commits);
2497         /* synchronize with interrupts */
2498         barrier();
2499         if (local_read(&cpu_buffer->committing) == 1)
2500                 rb_set_commit_to_write(cpu_buffer);
2501 
2502         local_dec(&cpu_buffer->committing);
2503 
2504         /* synchronize with interrupts */
2505         barrier();
2506 
2507         /*
2508          * Need to account for interrupts coming in between the
2509          * updating of the commit page and the clearing of the
2510          * committing counter.
2511          */
2512         if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
2513             !local_read(&cpu_buffer->committing)) {
2514                 local_inc(&cpu_buffer->committing);
2515                 goto again;
2516         }
2517 }
2518 
2519 static inline void rb_event_discard(struct ring_buffer_event *event)
2520 {
2521         if (extended_time(event))
2522                 event = skip_time_extend(event);
2523 
2524         /* array[0] holds the actual length for the discarded event */
2525         event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
2526         event->type_len = RINGBUF_TYPE_PADDING;
2527         /* time delta must be non zero */
2528         if (!event->time_delta)
2529                 event->time_delta = 1;
2530 }
2531 
2532 static __always_inline bool
2533 rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2534                    struct ring_buffer_event *event)
2535 {
2536         unsigned long addr = (unsigned long)event;
2537         unsigned long index;
2538 
2539         index = rb_event_index(event);
2540         addr &= PAGE_MASK;
2541 
2542         return cpu_buffer->commit_page->page == (void *)addr &&
2543                 rb_commit_index(cpu_buffer) == index;
2544 }
2545 
2546 static __always_inline void
2547 rb_update_write_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2548                       struct ring_buffer_event *event)
2549 {
2550         u64 delta;
2551 
2552         /*
2553          * The event first in the commit queue updates the
2554          * time stamp.
2555          */
2556         if (rb_event_is_commit(cpu_buffer, event)) {
2557                 /*
2558                  * A commit event that is first on a page
2559                  * updates the write timestamp with the page stamp
2560                  */
2561                 if (!rb_event_index(event))
2562                         cpu_buffer->write_stamp =
2563                                 cpu_buffer->commit_page->page->time_stamp;
2564                 else if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
2565                         delta = ring_buffer_event_time_stamp(event);
2566                         cpu_buffer->write_stamp += delta;
2567                 } else if (event->type_len == RINGBUF_TYPE_TIME_STAMP) {
2568                         delta = ring_buffer_event_time_stamp(event);
2569                         cpu_buffer->write_stamp = delta;
2570                 } else
2571                         cpu_buffer->write_stamp += event->time_delta;
2572         }
2573 }
2574 
2575 static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
2576                       struct ring_buffer_event *event)
2577 {
2578         local_inc(&cpu_buffer->entries);
2579         rb_update_write_stamp(cpu_buffer, event);
2580         rb_end_commit(cpu_buffer);
2581 }
2582 
2583 static __always_inline void
2584 rb_wakeups(struct ring_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer)
2585 {
2586         bool pagebusy;
2587 
2588         if (buffer->irq_work.waiters_pending) {
2589                 buffer->irq_work.waiters_pending = false;
2590                 /* irq_work_queue() supplies it's own memory barriers */
2591                 irq_work_queue(&buffer->irq_work.work);
2592         }
2593 
2594         if (cpu_buffer->irq_work.waiters_pending) {
2595                 cpu_buffer->irq_work.waiters_pending = false;
2596                 /* irq_work_queue() supplies it's own memory barriers */
2597                 irq_work_queue(&cpu_buffer->irq_work.work);
2598         }
2599 
2600         pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
2601 
2602         if (!pagebusy && cpu_buffer->irq_work.full_waiters_pending) {
2603                 cpu_buffer->irq_work.wakeup_full = true;
2604                 cpu_buffer->irq_work.full_waiters_pending = false;
2605                 /* irq_work_queue() supplies it's own memory barriers */
2606                 irq_work_queue(&cpu_buffer->irq_work.work);
2607         }
2608 }
2609 
2610 /*
2611  * The lock and unlock are done within a preempt disable section.
2612  * The current_context per_cpu variable can only be modified
2613  * by the current task between lock and unlock. But it can
2614  * be modified more than once via an interrupt. To pass this
2615  * information from the lock to the unlock without having to
2616  * access the 'in_interrupt()' functions again (which do show
2617  * a bit of overhead in something as critical as function tracing,
2618  * we use a bitmask trick.
2619  *
2620  *  bit 0 =  NMI context
2621  *  bit 1 =  IRQ context
2622  *  bit 2 =  SoftIRQ context
2623  *  bit 3 =  normal context.
2624  *
2625  * This works because this is the order of contexts that can
2626  * preempt other contexts. A SoftIRQ never preempts an IRQ
2627  * context.
2628  *
2629  * When the context is determined, the corresponding bit is
2630  * checked and set (if it was set, then a recursion of that context
2631  * happened).
2632  *
2633  * On unlock, we need to clear this bit. To do so, just subtract
2634  * 1 from the current_context and AND it to itself.
2635  *
2636  * (binary)
2637  *  101 - 1 = 100
2638  *  101 & 100 = 100 (clearing bit zero)
2639  *
2640  *  1010 - 1 = 1001
2641  *  1010 & 1001 = 1000 (clearing bit 1)
2642  *
2643  * The least significant bit can be cleared this way, and it
2644  * just so happens that it is the same bit corresponding to
2645  * the current context.
2646  */
2647 
2648 static __always_inline int
2649 trace_recursive_lock(struct ring_buffer_per_cpu *cpu_buffer)
2650 {
2651         unsigned int val = cpu_buffer->current_context;
2652         unsigned long pc = preempt_count();
2653         int bit;
2654 
2655         if (!(pc & (NMI_MASK | HARDIRQ_MASK | SOFTIRQ_OFFSET)))
2656                 bit = RB_CTX_NORMAL;
2657         else
2658                 bit = pc & NMI_MASK ? RB_CTX_NMI :
2659                         pc & HARDIRQ_MASK ? RB_CTX_IRQ : RB_CTX_SOFTIRQ;
2660 
2661         if (unlikely(val & (1 << (bit + cpu_buffer->nest))))
2662                 return 1;
2663 
2664         val |= (1 << (bit + cpu_buffer->nest));
2665         cpu_buffer->current_context = val;
2666 
2667         return 0;
2668 }
2669 
2670 static __always_inline void
2671 trace_recursive_unlock(struct ring_buffer_per_cpu *cpu_buffer)
2672 {
2673         cpu_buffer->current_context &=
2674                 cpu_buffer->current_context - (1 << cpu_buffer->nest);
2675 }
2676 
2677 /* The recursive locking above uses 4 bits */
2678 #define NESTED_BITS 4
2679 
2680 /**
2681  * ring_buffer_nest_start - Allow to trace while nested
2682  * @buffer: The ring buffer to modify
2683  *
2684  * The ring buffer has a safty mechanism to prevent recursion.
2685  * But there may be a case where a trace needs to be done while
2686  * tracing something else. In this case, calling this function
2687  * will allow this function to nest within a currently active
2688  * ring_buffer_lock_reserve().
2689  *
2690  * Call this function before calling another ring_buffer_lock_reserve() and
2691  * call ring_buffer_nest_end() after the nested ring_buffer_unlock_commit().
2692  */
2693 void ring_buffer_nest_start(struct ring_buffer *buffer)
2694 {
2695         struct ring_buffer_per_cpu *cpu_buffer;
2696         int cpu;
2697 
2698         /* Enabled by ring_buffer_nest_end() */
2699         preempt_disable_notrace();
2700         cpu = raw_smp_processor_id();
2701         cpu_buffer = buffer->buffers[cpu];
2702         /* This is the shift value for the above recusive locking */
2703         cpu_buffer->nest += NESTED_BITS;
2704 }
2705 
2706 /**
2707  * ring_buffer_nest_end - Allow to trace while nested
2708  * @buffer: The ring buffer to modify
2709  *
2710  * Must be called after ring_buffer_nest_start() and after the
2711  * ring_buffer_unlock_commit().
2712  */
2713 void ring_buffer_nest_end(struct ring_buffer *buffer)
2714 {
2715         struct ring_buffer_per_cpu *cpu_buffer;
2716         int cpu;
2717 
2718         /* disabled by ring_buffer_nest_start() */
2719         cpu = raw_smp_processor_id();
2720         cpu_buffer = buffer->buffers[cpu];
2721         /* This is the shift value for the above recusive locking */
2722         cpu_buffer->nest -= NESTED_BITS;
2723         preempt_enable_notrace();
2724 }
2725 
2726 /**
2727  * ring_buffer_unlock_commit - commit a reserved
2728  * @buffer: The buffer to commit to
2729  * @event: The event pointer to commit.
2730  *
2731  * This commits the data to the ring buffer, and releases any locks held.
2732  *
2733  * Must be paired with ring_buffer_lock_reserve.
2734  */
2735 int ring_buffer_unlock_commit(struct ring_buffer *buffer,
2736                               struct ring_buffer_event *event)
2737 {
2738         struct ring_buffer_per_cpu *cpu_buffer;
2739         int cpu = raw_smp_processor_id();
2740 
2741         cpu_buffer = buffer->buffers[cpu];
2742 
2743         rb_commit(cpu_buffer, event);
2744 
2745         rb_wakeups(buffer, cpu_buffer);
2746 
2747         trace_recursive_unlock(cpu_buffer);
2748 
2749         preempt_enable_notrace();
2750 
2751         return 0;
2752 }
2753 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
2754 
2755 static noinline void
2756 rb_handle_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2757                     struct rb_event_info *info)
2758 {
2759         WARN_ONCE(info->delta > (1ULL << 59),
2760                   KERN_WARNING "Delta way too big! %llu ts=%llu write stamp = %llu\n%s",
2761                   (unsigned long long)info->delta,
2762                   (unsigned long long)info->ts,
2763                   (unsigned long long)cpu_buffer->write_stamp,
2764                   sched_clock_stable() ? "" :
2765                   "If you just came from a suspend/resume,\n"
2766                   "please switch to the trace global clock:\n"
2767                   "  echo global > /sys/kernel/debug/tracing/trace_clock\n"
2768                   "or add trace_clock=global to the kernel command line\n");
2769         info->add_timestamp = 1;
2770 }
2771 
2772 static struct ring_buffer_event *
2773 __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
2774                   struct rb_event_info *info)
2775 {
2776         struct ring_buffer_event *event;
2777         struct buffer_page *tail_page;
2778         unsigned long tail, write;
2779 
2780         /*
2781          * If the time delta since the last event is too big to
2782          * hold in the time field of the event, then we append a
2783          * TIME EXTEND event ahead of the data event.
2784          */
2785         if (unlikely(info->add_timestamp))
2786                 info->length += RB_LEN_TIME_EXTEND;
2787 
2788         /* Don't let the compiler play games with cpu_buffer->tail_page */
2789         tail_page = info->tail_page = READ_ONCE(cpu_buffer->tail_page);
2790         write = local_add_return(info->length, &tail_page->write);
2791 
2792         /* set write to only the index of the write */
2793         write &= RB_WRITE_MASK;
2794         tail = write - info->length;
2795 
2796         /*
2797          * If this is the first commit on the page, then it has the same
2798          * timestamp as the page itself.
2799          */
2800         if (!tail && !ring_buffer_time_stamp_abs(cpu_buffer->buffer))
2801                 info->delta = 0;
2802 
2803         /* See if we shot pass the end of this buffer page */
2804         if (unlikely(write > BUF_PAGE_SIZE))
2805                 return rb_move_tail(cpu_buffer, tail, info);
2806 
2807         /* We reserved something on the buffer */
2808 
2809         event = __rb_page_index(tail_page, tail);
2810         rb_update_event(cpu_buffer, event, info);
2811 
2812         local_inc(&tail_page->entries);
2813 
2814         /*
2815          * If this is the first commit on the page, then update
2816          * its timestamp.
2817          */
2818         if (!tail)
2819                 tail_page->page->time_stamp = info->ts;
2820 
2821         /* account for these added bytes */
2822         local_add(info->length, &cpu_buffer->entries_bytes);
2823 
2824         return event;
2825 }
2826 
2827 static __always_inline struct ring_buffer_event *
2828 rb_reserve_next_event(struct ring_buffer *buffer,
2829                       struct ring_buffer_per_cpu *cpu_buffer,
2830                       unsigned long length)
2831 {
2832         struct ring_buffer_event *event;
2833         struct rb_event_info info;
2834         int nr_loops = 0;
2835         u64 diff;
2836 
2837         rb_start_commit(cpu_buffer);
2838 
2839 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2840         /*
2841          * Due to the ability to swap a cpu buffer from a buffer
2842          * it is possible it was swapped before we committed.
2843          * (committing stops a swap). We check for it here and
2844          * if it happened, we have to fail the write.
2845          */
2846         barrier();
2847         if (unlikely(READ_ONCE(cpu_buffer->buffer) != buffer)) {
2848                 local_dec(&cpu_buffer->committing);
2849                 local_dec(&cpu_buffer->commits);
2850                 return NULL;
2851         }
2852 #endif
2853 
2854         info.length = rb_calculate_event_length(length);
2855  again:
2856         info.add_timestamp = 0;
2857         info.delta = 0;
2858 
2859         /*
2860          * We allow for interrupts to reenter here and do a trace.
2861          * If one does, it will cause this original code to loop
2862          * back here. Even with heavy interrupts happening, this
2863          * should only happen a few times in a row. If this happens
2864          * 1000 times in a row, there must be either an interrupt
2865          * storm or we have something buggy.
2866          * Bail!
2867          */
2868         if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
2869                 goto out_fail;
2870 
2871         info.ts = rb_time_stamp(cpu_buffer->buffer);
2872         diff = info.ts - cpu_buffer->write_stamp;
2873 
2874         /* make sure this diff is calculated here */
2875         barrier();
2876 
2877         if (ring_buffer_time_stamp_abs(buffer)) {
2878                 info.delta = info.ts;
2879                 rb_handle_timestamp(cpu_buffer, &info);
2880         } else /* Did the write stamp get updated already? */
2881                 if (likely(info.ts >= cpu_buffer->write_stamp)) {
2882                 info.delta = diff;
2883                 if (unlikely(test_time_stamp(info.delta)))
2884                         rb_handle_timestamp(cpu_buffer, &info);
2885         }
2886 
2887         event = __rb_reserve_next(cpu_buffer, &info);
2888 
2889         if (unlikely(PTR_ERR(event) == -EAGAIN)) {
2890                 if (info.add_timestamp)
2891                         info.length -= RB_LEN_TIME_EXTEND;
2892                 goto again;
2893         }
2894 
2895         if (!event)
2896                 goto out_fail;
2897 
2898         return event;
2899 
2900  out_fail:
2901         rb_end_commit(cpu_buffer);
2902         return NULL;
2903 }
2904 
2905 /**
2906  * ring_buffer_lock_reserve - reserve a part of the buffer
2907  * @buffer: the ring buffer to reserve from
2908  * @length: the length of the data to reserve (excluding event header)
2909  *
2910  * Returns a reseverd event on the ring buffer to copy directly to.
2911  * The user of this interface will need to get the body to write into
2912  * and can use the ring_buffer_event_data() interface.
2913  *
2914  * The length is the length of the data needed, not the event length
2915  * which also includes the event header.
2916  *
2917  * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2918  * If NULL is returned, then nothing has been allocated or locked.
2919  */
2920 struct ring_buffer_event *
2921 ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length)
2922 {
2923         struct ring_buffer_per_cpu *cpu_buffer;
2924         struct ring_buffer_event *event;
2925         int cpu;
2926 
2927         /* If we are tracing schedule, we don't want to recurse */
2928         preempt_disable_notrace();
2929 
2930         if (unlikely(atomic_read(&buffer->record_disabled)))
2931                 goto out;
2932 
2933         cpu = raw_smp_processor_id();
2934 
2935         if (unlikely(!cpumask_test_cpu(cpu, buffer->cpumask)))
2936                 goto out;
2937 
2938         cpu_buffer = buffer->buffers[cpu];
2939 
2940         if (unlikely(atomic_read(&cpu_buffer->record_disabled)))
2941                 goto out;
2942 
2943         if (unlikely(length > BUF_MAX_DATA_SIZE))
2944                 goto out;
2945 
2946         if (unlikely(trace_recursive_lock(cpu_buffer)))
2947                 goto out;
2948 
2949         event = rb_reserve_next_event(buffer, cpu_buffer, length);
2950         if (!event)
2951                 goto out_unlock;
2952 
2953         return event;
2954 
2955  out_unlock:
2956         trace_recursive_unlock(cpu_buffer);
2957  out:
2958         preempt_enable_notrace();
2959         return NULL;
2960 }
2961 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
2962 
2963 /*
2964  * Decrement the entries to the page that an event is on.
2965  * The event does not even need to exist, only the pointer
2966  * to the page it is on. This may only be called before the commit
2967  * takes place.
2968  */
2969 static inline void
2970 rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
2971                    struct ring_buffer_event *event)
2972 {
2973         unsigned long addr = (unsigned long)event;
2974         struct buffer_page *bpage = cpu_buffer->commit_page;
2975         struct buffer_page *start;
2976 
2977         addr &= PAGE_MASK;
2978 
2979         /* Do the likely case first */
2980         if (likely(bpage->page == (void *)addr)) {
2981                 local_dec(&bpage->entries);
2982                 return;
2983         }
2984 
2985         /*
2986          * Because the commit page may be on the reader page we
2987          * start with the next page and check the end loop there.
2988          */
2989         rb_inc_page(cpu_buffer, &bpage);
2990         start = bpage;
2991         do {
2992                 if (bpage->page == (void *)addr) {
2993                         local_dec(&bpage->entries);
2994                         return;
2995                 }
2996                 rb_inc_page(cpu_buffer, &bpage);
2997         } while (bpage != start);
2998 
2999         /* commit not part of this buffer?? */
3000         RB_WARN_ON(cpu_buffer, 1);
3001 }
3002 
3003 /**
3004  * ring_buffer_commit_discard - discard an event that has not been committed
3005  * @buffer: the ring buffer
3006  * @event: non committed event to discard
3007  *
3008  * Sometimes an event that is in the ring buffer needs to be ignored.
3009  * This function lets the user discard an event in the ring buffer
3010  * and then that event will not be read later.
3011  *
3012  * This function only works if it is called before the the item has been
3013  * committed. It will try to free the event from the ring buffer
3014  * if another event has not been added behind it.
3015  *
3016  * If another event has been added behind it, it will set the event
3017  * up as discarded, and perform the commit.
3018  *
3019  * If this function is called, do not call ring_buffer_unlock_commit on
3020  * the event.
3021  */
3022 void ring_buffer_discard_commit(struct ring_buffer *buffer,
3023                                 struct ring_buffer_event *event)
3024 {
3025         struct ring_buffer_per_cpu *cpu_buffer;
3026         int cpu;
3027 
3028         /* The event is discarded regardless */
3029         rb_event_discard(event);
3030 
3031         cpu = smp_processor_id();
3032         cpu_buffer = buffer->buffers[cpu];
3033 
3034         /*
3035          * This must only be called if the event has not been
3036          * committed yet. Thus we can assume that preemption
3037          * is still disabled.
3038          */
3039         RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
3040 
3041         rb_decrement_entry(cpu_buffer, event);
3042         if (rb_try_to_discard(cpu_buffer, event))
3043                 goto out;
3044 
3045         /*
3046          * The commit is still visible by the reader, so we
3047          * must still update the timestamp.
3048          */
3049         rb_update_write_stamp(cpu_buffer, event);
3050  out:
3051         rb_end_commit(cpu_buffer);
3052 
3053         trace_recursive_unlock(cpu_buffer);
3054 
3055         preempt_enable_notrace();
3056 
3057 }
3058 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
3059 
3060 /**
3061  * ring_buffer_write - write data to the buffer without reserving
3062  * @buffer: The ring buffer to write to.
3063  * @length: The length of the data being written (excluding the event header)
3064  * @data: The data to write to the buffer.
3065  *
3066  * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
3067  * one function. If you already have the data to write to the buffer, it
3068  * may be easier to simply call this function.
3069  *
3070  * Note, like ring_buffer_lock_reserve, the length is the length of the data
3071  * and not the length of the event which would hold the header.
3072  */
3073 int ring_buffer_write(struct ring_buffer *buffer,
3074                       unsigned long length,
3075                       void *data)
3076 {
3077         struct ring_buffer_per_cpu *cpu_buffer;
3078         struct ring_buffer_event *event;
3079         void *body;
3080         int ret = -EBUSY;
3081         int cpu;
3082 
3083         preempt_disable_notrace();
3084 
3085         if (atomic_read(&buffer->record_disabled))
3086                 goto out;
3087 
3088         cpu = raw_smp_processor_id();
3089 
3090         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3091                 goto out;
3092 
3093         cpu_buffer = buffer->buffers[cpu];
3094 
3095         if (atomic_read(&cpu_buffer->record_disabled))
3096                 goto out;
3097 
3098         if (length > BUF_MAX_DATA_SIZE)
3099                 goto out;
3100 
3101         if (unlikely(trace_recursive_lock(cpu_buffer)))
3102                 goto out;
3103 
3104         event = rb_reserve_next_event(buffer, cpu_buffer, length);
3105         if (!event)
3106                 goto out_unlock;
3107 
3108         body = rb_event_data(event);
3109 
3110         memcpy(body, data, length);
3111 
3112         rb_commit(cpu_buffer, event);
3113 
3114         rb_wakeups(buffer, cpu_buffer);
3115 
3116         ret = 0;
3117 
3118  out_unlock:
3119         trace_recursive_unlock(cpu_buffer);
3120 
3121  out:
3122         preempt_enable_notrace();
3123 
3124         return ret;
3125 }
3126 EXPORT_SYMBOL_GPL(ring_buffer_write);
3127 
3128 static bool rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
3129 {
3130         struct buffer_page *reader = cpu_buffer->reader_page;
3131         struct buffer_page *head = rb_set_head_page(cpu_buffer);
3132         struct buffer_page *commit = cpu_buffer->commit_page;
3133 
3134         /* In case of error, head will be NULL */
3135         if (unlikely(!head))
3136                 return true;
3137 
3138         return reader->read == rb_page_commit(reader) &&
3139                 (commit == reader ||
3140                  (commit == head &&
3141                   head->read == rb_page_commit(commit)));
3142 }
3143 
3144 /**
3145  * ring_buffer_record_disable - stop all writes into the buffer
3146  * @buffer: The ring buffer to stop writes to.
3147  *
3148  * This prevents all writes to the buffer. Any attempt to write
3149  * to the buffer after this will fail and return NULL.
3150  *
3151  * The caller should call synchronize_sched() after this.
3152  */
3153 void ring_buffer_record_disable(struct ring_buffer *buffer)
3154 {
3155         atomic_inc(&buffer->record_disabled);
3156 }
3157 EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
3158 
3159 /**
3160  * ring_buffer_record_enable - enable writes to the buffer
3161  * @buffer: The ring buffer to enable writes
3162  *
3163  * Note, multiple disables will need the same number of enables
3164  * to truly enable the writing (much like preempt_disable).
3165  */
3166 void ring_buffer_record_enable(struct ring_buffer *buffer)
3167 {
3168         atomic_dec(&buffer->record_disabled);
3169 }
3170 EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
3171 
3172 /**
3173  * ring_buffer_record_off - stop all writes into the buffer
3174  * @buffer: The ring buffer to stop writes to.
3175  *
3176  * This prevents all writes to the buffer. Any attempt to write
3177  * to the buffer after this will fail and return NULL.
3178  *
3179  * This is different than ring_buffer_record_disable() as
3180  * it works like an on/off switch, where as the disable() version
3181  * must be paired with a enable().
3182  */
3183 void ring_buffer_record_off(struct ring_buffer *buffer)
3184 {
3185         unsigned int rd;
3186         unsigned int new_rd;
3187 
3188         do {
3189                 rd = atomic_read(&buffer->record_disabled);
3190                 new_rd = rd | RB_BUFFER_OFF;
3191         } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3192 }
3193 EXPORT_SYMBOL_GPL(ring_buffer_record_off);
3194 
3195 /**
3196  * ring_buffer_record_on - restart writes into the buffer
3197  * @buffer: The ring buffer to start writes to.
3198  *
3199  * This enables all writes to the buffer that was disabled by
3200  * ring_buffer_record_off().
3201  *
3202  * This is different than ring_buffer_record_enable() as
3203  * it works like an on/off switch, where as the enable() version
3204  * must be paired with a disable().
3205  */
3206 void ring_buffer_record_on(struct ring_buffer *buffer)
3207 {
3208         unsigned int rd;
3209         unsigned int new_rd;
3210 
3211         do {
3212                 rd = atomic_read(&buffer->record_disabled);
3213                 new_rd = rd & ~RB_BUFFER_OFF;
3214         } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3215 }
3216 EXPORT_SYMBOL_GPL(ring_buffer_record_on);
3217 
3218 /**
3219  * ring_buffer_record_is_on - return true if the ring buffer can write
3220  * @buffer: The ring buffer to see if write is enabled
3221  *
3222  * Returns true if the ring buffer is in a state that it accepts writes.
3223  */
3224 int ring_buffer_record_is_on(struct ring_buffer *buffer)
3225 {
3226         return !atomic_read(&buffer->record_disabled);
3227 }
3228 
3229 /**
3230  * ring_buffer_record_is_set_on - return true if the ring buffer is set writable
3231  * @buffer: The ring buffer to see if write is set enabled
3232  *
3233  * Returns true if the ring buffer is set writable by ring_buffer_record_on().
3234  * Note that this does NOT mean it is in a writable state.
3235  *
3236  * It may return true when the ring buffer has been disabled by
3237  * ring_buffer_record_disable(), as that is a temporary disabling of
3238  * the ring buffer.
3239  */
3240 int ring_buffer_record_is_set_on(struct ring_buffer *buffer)
3241 {
3242         return !(atomic_read(&buffer->record_disabled) & RB_BUFFER_OFF);
3243 }
3244 
3245 /**
3246  * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
3247  * @buffer: The ring buffer to stop writes to.
3248  * @cpu: The CPU buffer to stop
3249  *
3250  * This prevents all writes to the buffer. Any attempt to write
3251  * to the buffer after this will fail and return NULL.
3252  *
3253  * The caller should call synchronize_sched() after this.
3254  */
3255 void ring_buffer_record_disable_cpu(struct ring_buffer *buffer, int cpu)
3256 {
3257         struct ring_buffer_per_cpu *cpu_buffer;
3258 
3259         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3260                 return;
3261 
3262         cpu_buffer = buffer->buffers[cpu];
3263         atomic_inc(&cpu_buffer->record_disabled);
3264 }
3265 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
3266 
3267 /**
3268  * ring_buffer_record_enable_cpu - enable writes to the buffer
3269  * @buffer: The ring buffer to enable writes
3270  * @cpu: The CPU to enable.
3271  *
3272  * Note, multiple disables will need the same number of enables
3273  * to truly enable the writing (much like preempt_disable).
3274  */
3275 void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu)
3276 {
3277         struct ring_buffer_per_cpu *cpu_buffer;
3278 
3279         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3280                 return;
3281 
3282         cpu_buffer = buffer->buffers[cpu];
3283         atomic_dec(&cpu_buffer->record_disabled);
3284 }
3285 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
3286 
3287 /*
3288  * The total entries in the ring buffer is the running counter
3289  * of entries entered into the ring buffer, minus the sum of
3290  * the entries read from the ring buffer and the number of
3291  * entries that were overwritten.
3292  */
3293 static inline unsigned long
3294 rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
3295 {
3296         return local_read(&cpu_buffer->entries) -
3297                 (local_read(&cpu_buffer->overrun) + cpu_buffer->read);
3298 }
3299 
3300 /**
3301  * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
3302  * @buffer: The ring buffer
3303  * @cpu: The per CPU buffer to read from.
3304  */
3305 u64 ring_buffer_oldest_event_ts(struct ring_buffer *buffer, int cpu)
3306 {
3307         unsigned long flags;
3308         struct ring_buffer_per_cpu *cpu_buffer;
3309         struct buffer_page *bpage;
3310         u64 ret = 0;
3311 
3312         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3313                 return 0;
3314 
3315         cpu_buffer = buffer->buffers[cpu];
3316         raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3317         /*
3318          * if the tail is on reader_page, oldest time stamp is on the reader
3319          * page
3320          */
3321         if (cpu_buffer->tail_page == cpu_buffer->reader_page)
3322                 bpage = cpu_buffer->reader_page;
3323         else
3324                 bpage = rb_set_head_page(cpu_buffer);
3325         if (bpage)
3326                 ret = bpage->page->time_stamp;
3327         raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3328 
3329         return ret;
3330 }
3331 EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
3332 
3333 /**
3334  * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
3335  * @buffer: The ring buffer
3336  * @cpu: The per CPU buffer to read from.
3337  */
3338 unsigned long ring_buffer_bytes_cpu(struct ring_buffer *buffer, int cpu)
3339 {
3340         struct ring_buffer_per_cpu *cpu_buffer;
3341         unsigned long ret;
3342 
3343         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3344                 return 0;
3345 
3346         cpu_buffer = buffer->buffers[cpu];
3347         ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
3348 
3349         return ret;
3350 }
3351 EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
3352 
3353 /**
3354  * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
3355  * @buffer: The ring buffer
3356  * @cpu: The per CPU buffer to get the entries from.
3357  */
3358 unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu)
3359 {
3360         struct ring_buffer_per_cpu *cpu_buffer;
3361 
3362         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3363                 return 0;
3364 
3365         cpu_buffer = buffer->buffers[cpu];
3366 
3367         return rb_num_of_entries(cpu_buffer);
3368 }
3369 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
3370 
3371 /**
3372  * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
3373  * buffer wrapping around (only if RB_FL_OVERWRITE is on).
3374  * @buffer: The ring buffer
3375  * @cpu: The per CPU buffer to get the number of overruns from
3376  */
3377 unsigned long ring_buffer_overrun_cpu(struct ring_buffer *buffer, int cpu)
3378 {
3379         struct ring_buffer_per_cpu *cpu_buffer;
3380         unsigned long ret;
3381 
3382         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3383                 return 0;
3384 
3385         cpu_buffer = buffer->buffers[cpu];
3386         ret = local_read(&cpu_buffer->overrun);
3387 
3388         return ret;
3389 }
3390 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
3391 
3392 /**
3393  * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
3394  * commits failing due to the buffer wrapping around while there are uncommitted
3395  * events, such as during an interrupt storm.
3396  * @buffer: The ring buffer
3397  * @cpu: The per CPU buffer to get the number of overruns from
3398  */
3399 unsigned long
3400 ring_buffer_commit_overrun_cpu(struct ring_buffer *buffer, int cpu)
3401 {
3402         struct ring_buffer_per_cpu *cpu_buffer;
3403         unsigned long ret;
3404 
3405         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3406                 return 0;
3407 
3408         cpu_buffer = buffer->buffers[cpu];
3409         ret = local_read(&cpu_buffer->commit_overrun);
3410 
3411         return ret;
3412 }
3413 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
3414 
3415 /**
3416  * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
3417  * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
3418  * @buffer: The ring buffer
3419  * @cpu: The per CPU buffer to get the number of overruns from
3420  */
3421 unsigned long
3422 ring_buffer_dropped_events_cpu(struct ring_buffer *buffer, int cpu)
3423 {
3424         struct ring_buffer_per_cpu *cpu_buffer;
3425         unsigned long ret;
3426 
3427         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3428                 return 0;
3429 
3430         cpu_buffer = buffer->buffers[cpu];
3431         ret = local_read(&cpu_buffer->dropped_events);
3432 
3433         return ret;
3434 }
3435 EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
3436 
3437 /**
3438  * ring_buffer_read_events_cpu - get the number of events successfully read
3439  * @buffer: The ring buffer
3440  * @cpu: The per CPU buffer to get the number of events read
3441  */
3442 unsigned long
3443 ring_buffer_read_events_cpu(struct ring_buffer *buffer, int cpu)
3444 {
3445         struct ring_buffer_per_cpu *cpu_buffer;
3446 
3447         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3448                 return 0;
3449 
3450         cpu_buffer = buffer->buffers[cpu];
3451         return cpu_buffer->read;
3452 }
3453 EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu);
3454 
3455 /**
3456  * ring_buffer_entries - get the number of entries in a buffer
3457  * @buffer: The ring buffer
3458  *
3459  * Returns the total number of entries in the ring buffer
3460  * (all CPU entries)
3461  */
3462 unsigned long ring_buffer_entries(struct ring_buffer *buffer)
3463 {
3464         struct ring_buffer_per_cpu *cpu_buffer;
3465         unsigned long entries = 0;
3466         int cpu;
3467 
3468         /* if you care about this being correct, lock the buffer */
3469         for_each_buffer_cpu(buffer, cpu) {
3470                 cpu_buffer = buffer->buffers[cpu];
3471                 entries += rb_num_of_entries(cpu_buffer);
3472         }
3473 
3474         return entries;
3475 }
3476 EXPORT_SYMBOL_GPL(ring_buffer_entries);
3477 
3478 /**
3479  * ring_buffer_overruns - get the number of overruns in buffer
3480  * @buffer: The ring buffer
3481  *
3482  * Returns the total number of overruns in the ring buffer
3483  * (all CPU entries)
3484  */
3485 unsigned long ring_buffer_overruns(struct ring_buffer *buffer)
3486 {
3487         struct ring_buffer_per_cpu *cpu_buffer;
3488         unsigned long overruns = 0;
3489         int cpu;
3490 
3491         /* if you care about this being correct, lock the buffer */
3492         for_each_buffer_cpu(buffer, cpu) {
3493                 cpu_buffer = buffer->buffers[cpu];
3494                 overruns += local_read(&cpu_buffer->overrun);
3495         }
3496 
3497         return overruns;
3498 }
3499 EXPORT_SYMBOL_GPL(ring_buffer_overruns);
3500 
3501 static void rb_iter_reset(struct ring_buffer_iter *iter)
3502 {
3503         struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3504 
3505         /* Iterator usage is expected to have record disabled */
3506         iter->head_page = cpu_buffer->reader_page;
3507         iter->head = cpu_buffer->reader_page->read;
3508 
3509         iter->cache_reader_page = iter->head_page;
3510         iter->cache_read = cpu_buffer->read;
3511 
3512         if (iter->head)
3513                 iter->read_stamp = cpu_buffer->read_stamp;
3514         else
3515                 iter->read_stamp = iter->head_page->page->time_stamp;
3516 }
3517 
3518 /**
3519  * ring_buffer_iter_reset - reset an iterator
3520  * @iter: The iterator to reset
3521  *
3522  * Resets the iterator, so that it will start from the beginning
3523  * again.
3524  */
3525 void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
3526 {
3527         struct ring_buffer_per_cpu *cpu_buffer;
3528         unsigned long flags;
3529 
3530         if (!iter)
3531                 return;
3532 
3533         cpu_buffer = iter->cpu_buffer;
3534 
3535         raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3536         rb_iter_reset(iter);
3537         raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3538 }
3539 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
3540 
3541 /**
3542  * ring_buffer_iter_empty - check if an iterator has no more to read
3543  * @iter: The iterator to check
3544  */
3545 int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
3546 {
3547         struct ring_buffer_per_cpu *cpu_buffer;
3548         struct buffer_page *reader;
3549         struct buffer_page *head_page;
3550         struct buffer_page *commit_page;
3551         unsigned commit;
3552 
3553         cpu_buffer = iter->cpu_buffer;
3554 
3555         /* Remember, trace recording is off when iterator is in use */
3556         reader = cpu_buffer->reader_page;
3557         head_page = cpu_buffer->head_page;
3558         commit_page = cpu_buffer->commit_page;
3559         commit = rb_page_commit(commit_page);
3560 
3561         return ((iter->head_page == commit_page && iter->head == commit) ||
3562                 (iter->head_page == reader && commit_page == head_page &&
3563                  head_page->read == commit &&
3564                  iter->head == rb_page_commit(cpu_buffer->reader_page)));
3565 }
3566 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
3567 
3568 static void
3569 rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
3570                      struct ring_buffer_event *event)
3571 {
3572         u64 delta;
3573 
3574         switch (event->type_len) {
3575         case RINGBUF_TYPE_PADDING:
3576                 return;
3577 
3578         case RINGBUF_TYPE_TIME_EXTEND:
3579                 delta = ring_buffer_event_time_stamp(event);
3580                 cpu_buffer->read_stamp += delta;
3581                 return;
3582 
3583         case RINGBUF_TYPE_TIME_STAMP:
3584                 delta = ring_buffer_event_time_stamp(event);
3585                 cpu_buffer->read_stamp = delta;
3586                 return;
3587 
3588         case RINGBUF_TYPE_DATA:
3589                 cpu_buffer->read_stamp += event->time_delta;
3590                 return;
3591 
3592         default:
3593                 BUG();
3594         }
3595         return;
3596 }
3597 
3598 static void
3599 rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
3600                           struct ring_buffer_event *event)
3601 {
3602         u64 delta;
3603 
3604         switch (event->type_len) {
3605         case RINGBUF_TYPE_PADDING:
3606                 return;
3607 
3608         case RINGBUF_TYPE_TIME_EXTEND:
3609                 delta = ring_buffer_event_time_stamp(event);
3610                 iter->read_stamp += delta;
3611                 return;
3612 
3613         case RINGBUF_TYPE_TIME_STAMP:
3614                 delta = ring_buffer_event_time_stamp(event);
3615                 iter->read_stamp = delta;
3616                 return;
3617 
3618         case RINGBUF_TYPE_DATA:
3619                 iter->read_stamp += event->time_delta;
3620                 return;
3621 
3622         default:
3623                 BUG();
3624         }
3625         return;
3626 }
3627 
3628 static struct buffer_page *
3629 rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
3630 {
3631         struct buffer_page *reader = NULL;
3632         unsigned long overwrite;
3633         unsigned long flags;
3634         int nr_loops = 0;
3635         int ret;
3636 
3637         local_irq_save(flags);
3638         arch_spin_lock(&cpu_buffer->lock);
3639 
3640  again:
3641         /*
3642          * This should normally only loop twice. But because the
3643          * start of the reader inserts an empty page, it causes
3644          * a case where we will loop three times. There should be no
3645          * reason to loop four times (that I know of).
3646          */
3647         if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
3648                 reader = NULL;
3649                 goto out;
3650         }
3651 
3652         reader = cpu_buffer->reader_page;
3653 
3654         /* If there's more to read, return this page */
3655         if (cpu_buffer->reader_page->read < rb_page_size(reader))
3656                 goto out;
3657 
3658         /* Never should we have an index greater than the size */
3659         if (RB_WARN_ON(cpu_buffer,
3660                        cpu_buffer->reader_page->read > rb_page_size(reader)))
3661                 goto out;
3662 
3663         /* check if we caught up to the tail */
3664         reader = NULL;
3665         if (cpu_buffer->commit_page == cpu_buffer->reader_page)
3666                 goto out;
3667 
3668         /* Don't bother swapping if the ring buffer is empty */
3669         if (rb_num_of_entries(cpu_buffer) == 0)
3670                 goto out;
3671 
3672         /*
3673          * Reset the reader page to size zero.
3674          */
3675         local_set(&cpu_buffer->reader_page->write, 0);
3676         local_set(&cpu_buffer->reader_page->entries, 0);
3677         local_set(&cpu_buffer->reader_page->page->commit, 0);
3678         cpu_buffer->reader_page->real_end = 0;
3679 
3680  spin:
3681         /*
3682          * Splice the empty reader page into the list around the head.
3683          */
3684         reader = rb_set_head_page(cpu_buffer);
3685         if (!reader)
3686                 goto out;
3687         cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
3688         cpu_buffer->reader_page->list.prev = reader->list.prev;
3689 
3690         /*
3691          * cpu_buffer->pages just needs to point to the buffer, it
3692          *  has no specific buffer page to point to. Lets move it out
3693          *  of our way so we don't accidentally swap it.
3694          */
3695         cpu_buffer->pages = reader->list.prev;
3696 
3697         /* The reader page will be pointing to the new head */
3698         rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list);
3699 
3700         /*
3701          * We want to make sure we read the overruns after we set up our
3702          * pointers to the next object. The writer side does a
3703          * cmpxchg to cross pages which acts as the mb on the writer
3704          * side. Note, the reader will constantly fail the swap
3705          * while the writer is updating the pointers, so this
3706          * guarantees that the overwrite recorded here is the one we
3707          * want to compare with the last_overrun.
3708          */
3709         smp_mb();
3710         overwrite = local_read(&(cpu_buffer->overrun));
3711 
3712         /*
3713          * Here's the tricky part.
3714          *
3715          * We need to move the pointer past the header page.
3716          * But we can only do that if a writer is not currently
3717          * moving it. The page before the header page has the
3718          * flag bit '1' set if it is pointing to the page we want.
3719          * but if the writer is in the process of moving it
3720          * than it will be '2' or already moved ''.
3721          */
3722 
3723         ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
3724 
3725         /*
3726          * If we did not convert it, then we must try again.
3727          */
3728         if (!ret)
3729                 goto spin;
3730 
3731         /*
3732          * Yeah! We succeeded in replacing the page.
3733          *
3734          * Now make the new head point back to the reader page.
3735          */
3736         rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
3737         rb_inc_page(cpu_buffer, &cpu_buffer->head_page);
3738 
3739         /* Finally update the reader page to the new head */
3740         cpu_buffer->reader_page = reader;
3741         cpu_buffer->reader_page->read = 0;
3742 
3743         if (overwrite != cpu_buffer->last_overrun) {
3744                 cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
3745                 cpu_buffer->last_overrun = overwrite;
3746         }
3747 
3748         goto again;
3749 
3750  out:
3751         /* Update the read_stamp on the first event */
3752         if (reader && reader->read == 0)
3753                 cpu_buffer->read_stamp = reader->page->time_stamp;
3754 
3755         arch_spin_unlock(&cpu_buffer->lock);
3756         local_irq_restore(flags);
3757 
3758         return reader;
3759 }
3760 
3761 static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
3762 {
3763         struct ring_buffer_event *event;
3764         struct buffer_page *reader;
3765         unsigned length;
3766 
3767         reader = rb_get_reader_page(cpu_buffer);
3768 
3769         /* This function should not be called when buffer is empty */
3770         if (RB_WARN_ON(cpu_buffer, !reader))
3771                 return;
3772 
3773         event = rb_reader_event(cpu_buffer);
3774 
3775         if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
3776                 cpu_buffer->read++;
3777 
3778         rb_update_read_stamp(cpu_buffer, event);
3779 
3780         length = rb_event_length(event);
3781         cpu_buffer->reader_page->read += length;
3782 }
3783 
3784 static void rb_advance_iter(struct ring_buffer_iter *iter)
3785 {
3786         struct ring_buffer_per_cpu *cpu_buffer;
3787         struct ring_buffer_event *event;
3788         unsigned length;
3789 
3790         cpu_buffer = iter->cpu_buffer;
3791 
3792         /*
3793          * Check if we are at the end of the buffer.
3794          */
3795         if (iter->head >= rb_page_size(iter->head_page)) {
3796                 /* discarded commits can make the page empty */
3797                 if (iter->head_page == cpu_buffer->commit_page)
3798                         return;
3799                 rb_inc_iter(iter);
3800                 return;
3801         }
3802 
3803         event = rb_iter_head_event(iter);
3804 
3805         length = rb_event_length(event);
3806 
3807         /*
3808          * This should not be called to advance the header if we are
3809          * at the tail of the buffer.
3810          */
3811         if (RB_WARN_ON(cpu_buffer,
3812                        (iter->head_page == cpu_buffer->commit_page) &&
3813                        (iter->head + length > rb_commit_index(cpu_buffer))))
3814                 return;
3815 
3816         rb_update_iter_read_stamp(iter, event);
3817 
3818         iter->head += length;
3819 
3820         /* check for end of page padding */
3821         if ((iter->head >= rb_page_size(iter->head_page)) &&
3822             (iter->head_page != cpu_buffer->commit_page))
3823                 rb_inc_iter(iter);
3824 }
3825 
3826 static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
3827 {
3828         return cpu_buffer->lost_events;
3829 }
3830 
3831 static struct ring_buffer_event *
3832 rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
3833                unsigned long *lost_events)
3834 {
3835         struct ring_buffer_event *event;
3836         struct buffer_page *reader;
3837         int nr_loops = 0;
3838 
3839         if (ts)
3840                 *ts = 0;
3841  again:
3842         /*
3843          * We repeat when a time extend is encountered.
3844          * Since the time extend is always attached to a data event,
3845          * we should never loop more than once.
3846          * (We never hit the following condition more than twice).
3847          */
3848         if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3849                 return NULL;
3850 
3851         reader = rb_get_reader_page(cpu_buffer);
3852         if (!reader)
3853                 return NULL;
3854 
3855         event = rb_reader_event(cpu_buffer);
3856 
3857         switch (event->type_len) {
3858         case RINGBUF_TYPE_PADDING:
3859                 if (rb_null_event(event))
3860                         RB_WARN_ON(cpu_buffer, 1);
3861                 /*
3862                  * Because the writer could be discarding every
3863                  * event it creates (which would probably be bad)
3864                  * if we were to go back to "again" then we may never
3865                  * catch up, and will trigger the warn on, or lock
3866                  * the box. Return the padding, and we will release
3867                  * the current locks, and try again.
3868                  */
3869                 return event;
3870 
3871         case RINGBUF_TYPE_TIME_EXTEND:
3872                 /* Internal data, OK to advance */
3873                 rb_advance_reader(cpu_buffer);
3874                 goto again;
3875 
3876         case RINGBUF_TYPE_TIME_STAMP:
3877                 if (ts) {
3878                         *ts = ring_buffer_event_time_stamp(event);
3879                         ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3880                                                          cpu_buffer->cpu, ts);
3881                 }
3882                 /* Internal data, OK to advance */
3883                 rb_advance_reader(cpu_buffer);
3884                 goto again;
3885 
3886         case RINGBUF_TYPE_DATA:
3887                 if (ts && !(*ts)) {
3888                         *ts = cpu_buffer->read_stamp + event->time_delta;
3889                         ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3890                                                          cpu_buffer->cpu, ts);
3891                 }
3892                 if (lost_events)
3893                         *lost_events = rb_lost_events(cpu_buffer);
3894                 return event;
3895 
3896         default:
3897                 BUG();
3898         }
3899 
3900         return NULL;
3901 }
3902 EXPORT_SYMBOL_GPL(ring_buffer_peek);
3903 
3904 static struct ring_buffer_event *
3905 rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3906 {
3907         struct ring_buffer *buffer;
3908         struct ring_buffer_per_cpu *cpu_buffer;
3909         struct ring_buffer_event *event;
3910         int nr_loops = 0;
3911 
3912         if (ts)
3913                 *ts = 0;
3914 
3915         cpu_buffer = iter->cpu_buffer;
3916         buffer = cpu_buffer->buffer;
3917 
3918         /*
3919          * Check if someone performed a consuming read to
3920          * the buffer. A consuming read invalidates the iterator
3921          * and we need to reset the iterator in this case.
3922          */
3923         if (unlikely(iter->cache_read != cpu_buffer->read ||
3924                      iter->cache_reader_page != cpu_buffer->reader_page))
3925                 rb_iter_reset(iter);
3926 
3927  again:
3928         if (ring_buffer_iter_empty(iter))
3929                 return NULL;
3930 
3931         /*
3932          * We repeat when a time extend is encountered or we hit
3933          * the end of the page. Since the time extend is always attached
3934          * to a data event, we should never loop more than three times.
3935          * Once for going to next page, once on time extend, and
3936          * finally once to get the event.
3937          * (We never hit the following condition more than thrice).
3938          */
3939         if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3))
3940                 return NULL;
3941 
3942         if (rb_per_cpu_empty(cpu_buffer))
3943                 return NULL;
3944 
3945         if (iter->head >= rb_page_size(iter->head_page)) {
3946                 rb_inc_iter(iter);
3947                 goto again;
3948         }
3949 
3950         event = rb_iter_head_event(iter);
3951 
3952         switch (event->type_len) {
3953         case RINGBUF_TYPE_PADDING:
3954                 if (rb_null_event(event)) {
3955                         rb_inc_iter(iter);
3956                         goto again;
3957                 }
3958                 rb_advance_iter(iter);
3959                 return event;
3960 
3961         case RINGBUF_TYPE_TIME_EXTEND:
3962                 /* Internal data, OK to advance */
3963                 rb_advance_iter(iter);
3964                 goto again;
3965 
3966         case RINGBUF_TYPE_TIME_STAMP:
3967                 if (ts) {
3968                         *ts = ring_buffer_event_time_stamp(event);
3969                         ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3970                                                          cpu_buffer->cpu, ts);
3971                 }
3972                 /* Internal data, OK to advance */
3973                 rb_advance_iter(iter);
3974                 goto again;
3975 
3976         case RINGBUF_TYPE_DATA:
3977                 if (ts && !(*ts)) {
3978                         *ts = iter->read_stamp + event->time_delta;
3979                         ring_buffer_normalize_time_stamp(buffer,
3980                                                          cpu_buffer->cpu, ts);
3981                 }
3982                 return event;
3983 
3984         default:
3985                 BUG();
3986         }
3987 
3988         return NULL;
3989 }
3990 EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
3991 
3992 static inline bool rb_reader_lock(struct ring_buffer_per_cpu *cpu_buffer)
3993 {
3994         if (likely(!in_nmi())) {
3995                 raw_spin_lock(&cpu_buffer->reader_lock);
3996                 return true;
3997         }
3998 
3999         /*
4000          * If an NMI die dumps out the content of the ring buffer
4001          * trylock must be used to prevent a deadlock if the NMI
4002          * preempted a task that holds the ring buffer locks. If
4003          * we get the lock then all is fine, if not, then continue
4004          * to do the read, but this can corrupt the ring buffer,
4005          * so it must be permanently disabled from future writes.
4006          * Reading from NMI is a oneshot deal.
4007          */
4008         if (raw_spin_trylock(&cpu_buffer->reader_lock))
4009                 return true;
4010 
4011         /* Continue without locking, but disable the ring buffer */
4012         atomic_inc(&cpu_buffer->record_disabled);
4013         return false;
4014 }
4015 
4016 static inline void
4017 rb_reader_unlock(struct ring_buffer_per_cpu *cpu_buffer, bool locked)
4018 {
4019         if (likely(locked))
4020                 raw_spin_unlock(&cpu_buffer->reader_lock);
4021         return;
4022 }
4023 
4024 /**
4025  * ring_buffer_peek - peek at the next event to be read
4026  * @buffer: The ring buffer to read
4027  * @cpu: The cpu to peak at
4028  * @ts: The timestamp counter of this event.
4029  * @lost_events: a variable to store if events were lost (may be NULL)
4030  *
4031  * This will return the event that will be read next, but does
4032  * not consume the data.
4033  */
4034 struct ring_buffer_event *
4035 ring_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts,
4036                  unsigned long *lost_events)
4037 {
4038         struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4039         struct ring_buffer_event *event;
4040         unsigned long flags;
4041         bool dolock;
4042 
4043         if (!cpumask_test_cpu(cpu, buffer->cpumask))
4044                 return NULL;
4045 
4046  again:
4047         local_irq_save(flags);
4048         dolock = rb_reader_lock(cpu_buffer);
4049         event = rb_buffer_peek(cpu_buffer, ts, lost_events);
4050         if (event && event->type_len == RINGBUF_TYPE_PADDING)
4051                 rb_advance_reader(cpu_buffer);
4052         rb_reader_unlock(cpu_buffer, dolock);
4053         local_irq_restore(flags);
4054 
4055         if (event && event->type_len == RINGBUF_TYPE_PADDING)
4056                 goto again;
4057 
4058         return event;
4059 }
4060 
4061 /**
4062  * ring_buffer_iter_peek - peek at the next event to be read
4063  * @iter: The ring buffer iterator
4064  * @ts: The timestamp counter of this event.
4065  *
4066  * This will return the event that will be read next, but does
4067  * not increment the iterator.
4068  */
4069 struct ring_buffer_event *
4070 ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
4071 {
4072         struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4073         struct ring_buffer_event *event;
4074         unsigned long flags;
4075 
4076  again:
4077         raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4078         event = rb_iter_peek(iter, ts);
4079         raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4080 
4081         if (event && event->type_len == RINGBUF_TYPE_PADDING)
4082                 goto again;
4083 
4084         return event;
4085 }
4086 
4087 /**
4088  * ring_buffer_consume - return an event and consume it
4089  * @buffer: The ring buffer to get the next event from
4090  * @cpu: the cpu to read the buffer from
4091  * @ts: a variable to store the timestamp (may be NULL)
4092  * @lost_events: a variable to store if events were lost (may be NULL)
4093  *
4094  * Returns the next event in the ring buffer, and that event is consumed.
4095  * Meaning, that sequential reads will keep returning a different event,
4096  * and eventually empty the ring buffer if the producer is slower.
4097  */
4098 struct ring_buffer_event *
4099 ring_buffer_consume(struct ring_buffer *buffer, int cpu, u64 *ts,
4100                     unsigned long *lost_events)
4101 {
4102         struct ring_buffer_per_cpu *cpu_buffer;
4103         struct ring_buffer_event *event = NULL;
4104         unsigned long flags;
4105         bool dolock;
4106 
4107  again:
4108         /* might be called in atomic */
4109         preempt_disable();
4110 
4111         if (!cpumask_test_cpu(cpu, buffer->cpumask))
4112                 goto out;
4113 
4114         cpu_buffer = buffer->buffers[cpu];
4115         local_irq_save(flags);
4116         dolock = rb_reader_lock(cpu_buffer);
4117 
4118         event = rb_buffer_peek(cpu_buffer, ts, lost_events);
4119         if (event) {
4120                 cpu_buffer->lost_events = 0;
4121                 rb_advance_reader(cpu_buffer);
4122         }
4123 
4124         rb_reader_unlock(cpu_buffer, dolock);
4125         local_irq_restore(flags);
4126 
4127  out:
4128         preempt_enable();
4129 
4130         if (event && event->type_len == RINGBUF_TYPE_PADDING)
4131                 goto again;
4132 
4133         return event;
4134 }
4135 EXPORT_SYMBOL_GPL(ring_buffer_consume);
4136 
4137 /**
4138  * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
4139  * @buffer: The ring buffer to read from
4140  * @cpu: The cpu buffer to iterate over
4141  *
4142  * This performs the initial preparations necessary to iterate
4143  * through the buffer.  Memory is allocated, buffer recording
4144  * is disabled, and the iterator pointer is returned to the caller.
4145  *
4146  * Disabling buffer recordng prevents the reading from being
4147  * corrupted. This is not a consuming read, so a producer is not
4148  * expected.
4149  *
4150  * After a sequence of ring_buffer_read_prepare calls, the user is
4151  * expected to make at least one call to ring_buffer_read_prepare_sync.
4152  * Afterwards, ring_buffer_read_start is invoked to get things going
4153  * for real.
4154  *
4155  * This overall must be paired with ring_buffer_read_finish.
4156  */
4157 struct ring_buffer_iter *
4158 ring_buffer_read_prepare(struct ring_buffer *buffer, int cpu)
4159 {
4160         struct ring_buffer_per_cpu *cpu_buffer;
4161         struct ring_buffer_iter *iter;
4162 
4163         if (!cpumask_test_cpu(cpu, buffer->cpumask))
4164                 return NULL;
4165 
4166         iter = kmalloc(sizeof(*iter), GFP_KERNEL);
4167         if (!iter)
4168                 return NULL;
4169 
4170         cpu_buffer = buffer->buffers[cpu];
4171 
4172         iter->cpu_buffer = cpu_buffer;
4173 
4174         atomic_inc(&buffer->resize_disabled);
4175         atomic_inc(&cpu_buffer->record_disabled);
4176 
4177         return iter;
4178 }
4179 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
4180 
4181 /**
4182  * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
4183  *
4184  * All previously invoked ring_buffer_read_prepare calls to prepare
4185  * iterators will be synchronized.  Afterwards, read_buffer_read_start
4186  * calls on those iterators are allowed.
4187  */
4188 void
4189 ring_buffer_read_prepare_sync(void)
4190 {
4191         synchronize_sched();
4192 }
4193 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
4194 
4195 /**
4196  * ring_buffer_read_start - start a non consuming read of the buffer
4197  * @iter: The iterator returned by ring_buffer_read_prepare
4198  *
4199  * This finalizes the startup of an iteration through the buffer.
4200  * The iterator comes from a call to ring_buffer_read_prepare and
4201  * an intervening ring_buffer_read_prepare_sync must have been
4202  * performed.
4203  *
4204  * Must be paired with ring_buffer_read_finish.
4205  */
4206 void
4207 ring_buffer_read_start(struct ring_buffer_iter *iter)
4208 {
4209         struct ring_buffer_per_cpu *cpu_buffer;
4210         unsigned long flags;
4211 
4212         if (!iter)
4213                 return;
4214 
4215         cpu_buffer = iter->cpu_buffer;
4216 
4217         raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4218         arch_spin_lock(&cpu_buffer->lock);
4219         rb_iter_reset(iter);
4220         arch_spin_unlock(&cpu_buffer->lock);
4221         raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4222 }
4223 EXPORT_SYMBOL_GPL(ring_buffer_read_start);
4224 
4225 /**
4226  * ring_buffer_read_finish - finish reading the iterator of the buffer
4227  * @iter: The iterator retrieved by ring_buffer_start
4228  *
4229  * This re-enables the recording to the buffer, and frees the
4230  * iterator.
4231  */
4232 void
4233 ring_buffer_read_finish(struct ring_buffer_iter *iter)
4234 {
4235         struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4236         unsigned long flags;
4237 
4238         /*
4239          * Ring buffer is disabled from recording, here's a good place
4240          * to check the integrity of the ring buffer.
4241          * Must prevent readers from trying to read, as the check
4242          * clears the HEAD page and readers require it.
4243          */
4244         raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4245         rb_check_pages(cpu_buffer);
4246         raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4247 
4248         atomic_dec(&cpu_buffer->record_disabled);
4249         atomic_dec(&cpu_buffer->buffer->resize_disabled);
4250         kfree(iter);
4251 }
4252 EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
4253 
4254 /**
4255  * ring_buffer_read - read the next item in the ring buffer by the iterator
4256  * @iter: The ring buffer iterator
4257  * @ts: The time stamp of the event read.
4258  *
4259  * This reads the next event in the ring buffer and increments the iterator.
4260  */
4261 struct ring_buffer_event *
4262 ring_buffer_read(struct ring_buffer_iter *iter, u64 *ts)
4263 {
4264         struct ring_buffer_event *event;
4265         struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4266         unsigned long flags;
4267 
4268         raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4269  again:
4270         event = rb_iter_peek(iter, ts);
4271         if (!event)
4272                 goto out;
4273 
4274         if (event->type_len == RINGBUF_TYPE_PADDING)
4275                 goto again;
4276 
4277         rb_advance_iter(iter);
4278  out:
4279         raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4280 
4281         return event;
4282 }
4283 EXPORT_SYMBOL_GPL(ring_buffer_read);
4284 
4285 /**
4286  * ring_buffer_size - return the size of the ring buffer (in bytes)
4287  * @buffer: The ring buffer.
4288  */
4289 unsigned long ring_buffer_size(struct ring_buffer *buffer, int cpu)
4290 {
4291         /*
4292          * Earlier, this method returned
4293          *      BUF_PAGE_SIZE * buffer->nr_pages
4294          * Since the nr_pages field is now removed, we have converted this to
4295          * return the per cpu buffer value.
4296          */
4297         if (!cpumask_test_cpu(cpu, buffer->cpumask))
4298                 return 0;
4299 
4300         return BUF_PAGE_SIZE * buffer->buffers[cpu]->nr_pages;
4301 }
4302 EXPORT_SYMBOL_GPL(ring_buffer_size);
4303 
4304 static void
4305 rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
4306 {
4307         rb_head_page_deactivate(cpu_buffer);
4308 
4309         cpu_buffer->head_page
4310                 = list_entry(cpu_buffer->pages, struct buffer_page, list);
4311         local_set(&cpu_buffer->head_page->write, 0);
4312         local_set(&cpu_buffer->head_page->entries, 0);
4313         local_set(&cpu_buffer->head_page->page->commit, 0);
4314 
4315         cpu_buffer->head_page->read = 0;
4316 
4317         cpu_buffer->tail_page = cpu_buffer->head_page;
4318         cpu_buffer->commit_page = cpu_buffer->head_page;
4319 
4320         INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
4321         INIT_LIST_HEAD(&cpu_buffer->new_pages);
4322         local_set(&cpu_buffer->reader_page->write, 0);
4323         local_set(&cpu_buffer->reader_page->entries, 0);
4324         local_set(&cpu_buffer->reader_page->page->commit, 0);
4325         cpu_buffer->reader_page->read = 0;
4326 
4327         local_set(&cpu_buffer->entries_bytes, 0);
4328         local_set(&cpu_buffer->overrun, 0);
4329         local_set(&cpu_buffer->commit_overrun, 0);
4330         local_set(&cpu_buffer->dropped_events, 0);
4331         local_set(&cpu_buffer->entries, 0);
4332         local_set(&cpu_buffer->committing, 0);
4333         local_set(&cpu_buffer->commits, 0);
4334         cpu_buffer->read = 0;
4335         cpu_buffer->read_bytes = 0;
4336 
4337         cpu_buffer->write_stamp = 0;
4338         cpu_buffer->read_stamp = 0;
4339 
4340         cpu_buffer->lost_events = 0;
4341         cpu_buffer->last_overrun = 0;
4342 
4343         rb_head_page_activate(cpu_buffer);
4344 }
4345 
4346 /**
4347  * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
4348  * @buffer: The ring buffer to reset a per cpu buffer of
4349  * @cpu: The CPU buffer to be reset
4350  */
4351 void ring_buffer_reset_cpu(struct ring_buffer *buffer, int cpu)
4352 {
4353         struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4354         unsigned long flags;
4355 
4356         if (!cpumask_test_cpu(cpu, buffer->cpumask))
4357                 return;
4358 
4359         atomic_inc(&buffer->resize_disabled);
4360         atomic_inc(&cpu_buffer->record_disabled);
4361 
4362         /* Make sure all commits have finished */
4363         synchronize_sched();
4364 
4365         raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4366 
4367         if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
4368                 goto out;
4369 
4370         arch_spin_lock(&cpu_buffer->lock);
4371 
4372         rb_reset_cpu(cpu_buffer);
4373 
4374         arch_spin_unlock(&cpu_buffer->lock);
4375 
4376  out:
4377         raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4378 
4379         atomic_dec(&cpu_buffer->record_disabled);
4380         atomic_dec(&buffer->resize_disabled);
4381 }
4382 EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
4383 
4384 /**
4385  * ring_buffer_reset - reset a ring buffer
4386  * @buffer: The ring buffer to reset all cpu buffers
4387  */
4388 void ring_buffer_reset(struct ring_buffer *buffer)
4389 {
4390         int cpu;
4391 
4392         for_each_buffer_cpu(buffer, cpu)
4393                 ring_buffer_reset_cpu(buffer, cpu);
4394 }
4395 EXPORT_SYMBOL_GPL(ring_buffer_reset);
4396 
4397 /**
4398  * rind_buffer_empty - is the ring buffer empty?
4399  * @buffer: The ring buffer to test
4400  */
4401 bool ring_buffer_empty(struct ring_buffer *buffer)
4402 {
4403         struct ring_buffer_per_cpu *cpu_buffer;
4404         unsigned long flags;
4405         bool dolock;
4406         int cpu;
4407         int ret;
4408 
4409         /* yes this is racy, but if you don't like the race, lock the buffer */
4410         for_each_buffer_cpu(buffer, cpu) {
4411                 cpu_buffer = buffer->buffers[cpu];
4412                 local_irq_save(flags);
4413                 dolock = rb_reader_lock(cpu_buffer);
4414                 ret = rb_per_cpu_empty(cpu_buffer);
4415                 rb_reader_unlock(cpu_buffer, dolock);
4416                 local_irq_restore(flags);
4417 
4418                 if (!ret)
4419                         return false;
4420         }
4421 
4422         return true;
4423 }
4424 EXPORT_SYMBOL_GPL(ring_buffer_empty);
4425 
4426 /**
4427  * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
4428  * @buffer: The ring buffer
4429  * @cpu: The CPU buffer to test
4430  */
4431 bool ring_buffer_empty_cpu(struct ring_buffer *buffer, int cpu)
4432 {
4433         struct ring_buffer_per_cpu *cpu_buffer;
4434         unsigned long flags;
4435         bool dolock;
4436         int ret;
4437 
4438         if (!cpumask_test_cpu(cpu, buffer->cpumask))
4439                 return true;
4440 
4441         cpu_buffer = buffer->buffers[cpu];
4442         local_irq_save(flags);
4443         dolock = rb_reader_lock(cpu_buffer);
4444         ret = rb_per_cpu_empty(cpu_buffer);
4445         rb_reader_unlock(cpu_buffer, dolock);
4446         local_irq_restore(flags);
4447 
4448         return ret;
4449 }
4450 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
4451 
4452 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
4453 /**
4454  * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
4455  * @buffer_a: One buffer to swap with
4456  * @buffer_b: The other buffer to swap with
4457  *
4458  * This function is useful for tracers that want to take a "snapshot"
4459  * of a CPU buffer and has another back up buffer lying around.
4460  * it is expected that the tracer handles the cpu buffer not being
4461  * used at the moment.
4462  */
4463 int ring_buffer_swap_cpu(struct ring_buffer *buffer_a,
4464                          struct ring_buffer *buffer_b, int cpu)
4465 {
4466         struct ring_buffer_per_cpu *cpu_buffer_a;
4467         struct ring_buffer_per_cpu *cpu_buffer_b;
4468         int ret = -EINVAL;
4469 
4470         if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
4471             !cpumask_test_cpu(cpu, buffer_b->cpumask))
4472                 goto out;
4473 
4474         cpu_buffer_a = buffer_a->buffers[cpu];
4475         cpu_buffer_b = buffer_b->buffers[cpu];
4476 
4477         /* At least make sure the two buffers are somewhat the same */
4478         if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
4479                 goto out;
4480 
4481         ret = -EAGAIN;
4482 
4483         if (atomic_read(&buffer_a->record_disabled))
4484                 goto out;
4485 
4486         if (atomic_read(&buffer_b->record_disabled))
4487                 goto out;
4488 
4489         if (atomic_read(&cpu_buffer_a->record_disabled))
4490                 goto out;
4491 
4492         if (atomic_read(&cpu_buffer_b->record_disabled))
4493                 goto out;
4494 
4495         /*
4496          * We can't do a synchronize_sched here because this
4497          * function can be called in atomic context.
4498          * Normally this will be called from the same CPU as cpu.
4499          * If not it's up to the caller to protect this.
4500          */
4501         atomic_inc(&cpu_buffer_a->record_disabled);
4502         atomic_inc(&cpu_buffer_b->record_disabled);
4503 
4504         ret = -EBUSY;
4505         if (local_read(&cpu_buffer_a->committing))
4506                 goto out_dec;
4507         if (local_read(&cpu_buffer_b->committing))
4508                 goto out_dec;
4509 
4510         buffer_a->buffers[cpu] = cpu_buffer_b;
4511         buffer_b->buffers[cpu] = cpu_buffer_a;
4512 
4513         cpu_buffer_b->buffer = buffer_a;
4514         cpu_buffer_a->buffer = buffer_b;
4515 
4516         ret = 0;
4517 
4518 out_dec:
4519         atomic_dec(&cpu_buffer_a->record_disabled);
4520         atomic_dec(&cpu_buffer_b->record_disabled);
4521 out:
4522         return ret;
4523 }
4524 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
4525 #endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
4526 
4527 /**
4528  * ring_buffer_alloc_read_page - allocate a page to read from buffer
4529  * @buffer: the buffer to allocate for.
4530  * @cpu: the cpu buffer to allocate.
4531  *
4532  * This function is used in conjunction with ring_buffer_read_page.
4533  * When reading a full page from the ring buffer, these functions
4534  * can be used to speed up the process. The calling function should
4535  * allocate a few pages first with this function. Then when it
4536  * needs to get pages from the ring buffer, it passes the result
4537  * of this function into ring_buffer_read_page, which will swap
4538  * the page that was allocated, with the read page of the buffer.
4539  *
4540  * Returns:
4541  *  The page allocated, or ERR_PTR
4542  */
4543 void *ring_buffer_alloc_read_page(struct ring_buffer *buffer, int cpu)
4544 {
4545         struct ring_buffer_per_cpu *cpu_buffer;
4546         struct buffer_data_page *bpage = NULL;
4547         unsigned long flags;
4548         struct page *page;
4549 
4550         if (!cpumask_test_cpu(cpu, buffer->cpumask))
4551                 return ERR_PTR(-ENODEV);
4552 
4553         cpu_buffer = buffer->buffers[cpu];
4554         local_irq_save(flags);
4555         arch_spin_lock(&cpu_buffer->lock);
4556 
4557         if (cpu_buffer->free_page) {
4558                 bpage = cpu_buffer->free_page;
4559                 cpu_buffer->free_page = NULL;
4560         }
4561 
4562         arch_spin_unlock(&cpu_buffer->lock);
4563         local_irq_restore(flags);
4564 
4565         if (bpage)
4566                 goto out;
4567 
4568         page = alloc_pages_node(cpu_to_node(cpu),
4569                                 GFP_KERNEL | __GFP_NORETRY, 0);
4570         if (!page)
4571                 return ERR_PTR(-ENOMEM);
4572 
4573         bpage = page_address(page);
4574 
4575  out:
4576         rb_init_page(bpage);
4577 
4578         return bpage;
4579 }
4580 EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
4581 
4582 /**
4583  * ring_buffer_free_read_page - free an allocated read page
4584  * @buffer: the buffer the page was allocate for
4585  * @cpu: the cpu buffer the page came from
4586  * @data: the page to free
4587  *
4588  * Free a page allocated from ring_buffer_alloc_read_page.
4589  */
4590 void ring_buffer_free_read_page(struct ring_buffer *buffer, int cpu, void *data)
4591 {
4592         struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4593         struct buffer_data_page *bpage = data;
4594         struct page *page = virt_to_page(bpage);
4595         unsigned long flags;
4596 
4597         /* If the page is still in use someplace else, we can't reuse it */
4598         if (page_ref_count(page) > 1)
4599                 goto out;
4600 
4601         local_irq_save(flags);
4602         arch_spin_lock(&cpu_buffer->lock);
4603 
4604         if (!cpu_buffer->free_page) {
4605                 cpu_buffer->free_page = bpage;
4606                 bpage = NULL;
4607         }
4608 
4609         arch_spin_unlock(&cpu_buffer->lock);
4610         local_irq_restore(flags);
4611 
4612  out:
4613         free_page((unsigned long)bpage);
4614 }
4615 EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
4616 
4617 /**
4618  * ring_buffer_read_page - extract a page from the ring buffer
4619  * @buffer: buffer to extract from
4620  * @data_page: the page to use allocated from ring_buffer_alloc_read_page
4621  * @len: amount to extract
4622  * @cpu: the cpu of the buffer to extract
4623  * @full: should the extraction only happen when the page is full.
4624  *
4625  * This function will pull out a page from the ring buffer and consume it.
4626  * @data_page must be the address of the variable that was returned
4627  * from ring_buffer_alloc_read_page. This is because the page might be used
4628  * to swap with a page in the ring buffer.
4629  *
4630  * for example:
4631  *      rpage = ring_buffer_alloc_read_page(buffer, cpu);
4632  *      if (IS_ERR(rpage))
4633  *              return PTR_ERR(rpage);
4634  *      ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
4635  *      if (ret >= 0)
4636  *              process_page(rpage, ret);
4637  *
4638  * When @full is set, the function will not return true unless
4639  * the writer is off the reader page.
4640  *
4641  * Note: it is up to the calling functions to handle sleeps and wakeups.
4642  *  The ring buffer can be used anywhere in the kernel and can not
4643  *  blindly call wake_up. The layer that uses the ring buffer must be
4644  *  responsible for that.
4645  *
4646  * Returns:
4647  *  >=0 if data has been transferred, returns the offset of consumed data.
4648  *  <0 if no data has been transferred.
4649  */
4650 int ring_buffer_read_page(struct ring_buffer *buffer,
4651                           void **data_page, size_t len, int cpu, int full)
4652 {
4653         struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4654         struct ring_buffer_event *event;
4655         struct buffer_data_page *bpage;
4656         struct buffer_page *reader;
4657         unsigned long missed_events;
4658         unsigned long flags;
4659         unsigned int commit;
4660         unsigned int read;
4661         u64 save_timestamp;
4662         int ret = -1;
4663 
4664         if (!cpumask_test_cpu(cpu, buffer->cpumask))
4665                 goto out;
4666 
4667         /*
4668          * If len is not big enough to hold the page header, then
4669          * we can not copy anything.
4670          */
4671         if (len <= BUF_PAGE_HDR_SIZE)
4672                 goto out;
4673 
4674         len -= BUF_PAGE_HDR_SIZE;
4675 
4676         if (!data_page)
4677                 goto out;
4678 
4679         bpage = *data_page;
4680         if (!bpage)
4681                 goto out;
4682 
4683         raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4684 
4685         reader = rb_get_reader_page(cpu_buffer);
4686         if (!reader)
4687                 goto out_unlock;
4688 
4689         event = rb_reader_event(cpu_buffer);
4690 
4691         read = reader->read;
4692         commit = rb_page_commit(reader);
4693 
4694         /* Check if any events were dropped */
4695         missed_events = cpu_buffer->lost_events;
4696 
4697         /*
4698          * If this page has been partially read or
4699          * if len is not big enough to read the rest of the page or
4700          * a writer is still on the page, then
4701          * we must copy the data from the page to the buffer.
4702          * Otherwise, we can simply swap the page with the one passed in.
4703          */
4704         if (read || (len < (commit - read)) ||
4705             cpu_buffer->reader_page == cpu_buffer->commit_page) {
4706                 struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
4707                 unsigned int rpos = read;
4708                 unsigned int pos = 0;
4709                 unsigned int size;
4710 
4711                 if (full)
4712                         goto out_unlock;
4713 
4714                 if (len > (commit - read))
4715                         len = (commit - read);
4716 
4717                 /* Always keep the time extend and data together */
4718                 size = rb_event_ts_length(event);
4719 
4720                 if (len < size)
4721                         goto out_unlock;
4722 
4723                 /* save the current timestamp, since the user will need it */
4724                 save_timestamp = cpu_buffer->read_stamp;
4725 
4726                 /* Need to copy one event at a time */
4727                 do {
4728                         /* We need the size of one event, because
4729                          * rb_advance_reader only advances by one event,
4730                          * whereas rb_event_ts_length may include the size of
4731                          * one or two events.
4732                          * We have already ensured there's enough space if this
4733                          * is a time extend. */
4734                         size = rb_event_length(event);
4735                         memcpy(bpage->data + pos, rpage->data + rpos, size);
4736 
4737                         len -= size;
4738 
4739                         rb_advance_reader(cpu_buffer);
4740                         rpos = reader->read;
4741                         pos += size;
4742 
4743                         if (rpos >= commit)
4744                                 break;
4745 
4746                         event = rb_reader_event(cpu_buffer);
4747                         /* Always keep the time extend and data together */
4748                         size = rb_event_ts_length(event);
4749                 } while (len >= size);
4750 
4751                 /* update bpage */
4752                 local_set(&bpage->commit, pos);
4753                 bpage->time_stamp = save_timestamp;
4754 
4755                 /* we copied everything to the beginning */
4756                 read = 0;
4757         } else {
4758                 /* update the entry counter */
4759                 cpu_buffer->read += rb_page_entries(reader);
4760                 cpu_buffer->read_bytes += BUF_PAGE_SIZE;
4761 
4762                 /* swap the pages */
4763                 rb_init_page(bpage);
4764                 bpage = reader->page;
4765                 reader->page = *data_page;
4766                 local_set(&reader->write, 0);
4767                 local_set(&reader->entries, 0);
4768                 reader->read = 0;
4769                 *data_page = bpage;
4770 
4771                 /*
4772                  * Use the real_end for the data size,
4773                  * This gives us a chance to store the lost events
4774                  * on the page.
4775                  */
4776                 if (reader->real_end)
4777                         local_set(&bpage->commit, reader->real_end);
4778         }
4779         ret = read;
4780 
4781         cpu_buffer->lost_events = 0;
4782 
4783         commit = local_read(&bpage->commit);
4784         /*
4785          * Set a flag in the commit field if we lost events
4786          */
4787         if (missed_events) {
4788                 /* If there is room at the end of the page to save the
4789                  * missed events, then record it there.
4790                  */
4791                 if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
4792                         memcpy(&bpage->data[commit], &missed_events,
4793                                sizeof(missed_events));
4794                         local_add(RB_MISSED_STORED, &bpage->commit);
4795                         commit += sizeof(missed_events);
4796                 }
4797                 local_add(RB_MISSED_EVENTS, &bpage->commit);
4798         }
4799 
4800         /*
4801          * This page may be off to user land. Zero it out here.
4802          */
4803         if (commit < BUF_PAGE_SIZE)
4804                 memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
4805 
4806  out_unlock:
4807         raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4808 
4809  out:
4810         return ret;
4811 }
4812 EXPORT_SYMBOL_GPL(ring_buffer_read_page);
4813 
4814 /*
4815  * We only allocate new buffers, never free them if the CPU goes down.
4816  * If we were to free the buffer, then the user would lose any trace that was in
4817  * the buffer.
4818  */
4819 int trace_rb_cpu_prepare(unsigned int cpu, struct hlist_node *node)
4820 {
4821         struct ring_buffer *buffer;
4822         long nr_pages_same;
4823         int cpu_i;
4824         unsigned long nr_pages;
4825 
4826         buffer = container_of(node, struct ring_buffer, node);
4827         if (cpumask_test_cpu(cpu, buffer->cpumask))
4828                 return 0;
4829 
4830         nr_pages = 0;
4831         nr_pages_same = 1;
4832         /* check if all cpu sizes are same */
4833         for_each_buffer_cpu(buffer, cpu_i) {
4834                 /* fill in the size from first enabled cpu */
4835                 if (nr_pages == 0)
4836                         nr_pages = buffer->buffers[cpu_i]->nr_pages;
4837                 if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
4838                         nr_pages_same = 0;
4839                         break;
4840                 }
4841         }
4842         /* allocate minimum pages, user can later expand it */
4843         if (!nr_pages_same)
4844                 nr_pages = 2;
4845         buffer->buffers[cpu] =
4846                 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
4847         if (!buffer->buffers[cpu]) {
4848                 WARN(1, "failed to allocate ring buffer on CPU %u\n",
4849                      cpu);
4850                 return -ENOMEM;
4851         }
4852         smp_wmb();
4853         cpumask_set_cpu(cpu, buffer->cpumask);
4854         return 0;
4855 }
4856 
4857 #ifdef CONFIG_RING_BUFFER_STARTUP_TEST
4858 /*
4859  * This is a basic integrity check of the ring buffer.
4860  * Late in the boot cycle this test will run when configured in.
4861  * It will kick off a thread per CPU that will go into a loop
4862  * writing to the per cpu ring buffer various sizes of data.
4863  * Some of the data will be large items, some small.
4864  *
4865  * Another thread is created that goes into a spin, sending out
4866  * IPIs to the other CPUs to also write into the ring buffer.
4867  * this is to test the nesting ability of the buffer.
4868  *
4869  * Basic stats are recorded and reported. If something in the
4870  * ring buffer should happen that's not expected, a big warning
4871  * is displayed and all ring buffers are disabled.
4872  */
4873 static struct task_struct *rb_threads[NR_CPUS] __initdata;
4874 
4875 struct rb_test_data {
4876         struct ring_buffer      *buffer;
4877         unsigned long           events;
4878         unsigned long           bytes_written;
4879         unsigned long           bytes_alloc;
4880         unsigned long           bytes_dropped;
4881         unsigned long           events_nested;
4882         unsigned long           bytes_written_nested;
4883         unsigned long           bytes_alloc_nested;
4884         unsigned long           bytes_dropped_nested;
4885         int                     min_size_nested;
4886         int                     max_size_nested;
4887         int                     max_size;
4888         int                     min_size;
4889         int                     cpu;
4890         int                     cnt;
4891 };
4892 
4893 static struct rb_test_data rb_data[NR_CPUS] __initdata;
4894 
4895 /* 1 meg per cpu */
4896 #define RB_TEST_BUFFER_SIZE     1048576
4897 
4898 static char rb_string[] __initdata =
4899         "abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
4900         "?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
4901         "!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
4902 
4903 static bool rb_test_started __initdata;
4904 
4905 struct rb_item {
4906         int size;
4907         char str[];
4908 };
4909 
4910 static __init int rb_write_something(struct rb_test_data *data, bool nested)
4911 {
4912         struct ring_buffer_event *event;
4913         struct rb_item *item;
4914         bool started;
4915         int event_len;
4916         int size;
4917         int len;
4918         int cnt;
4919 
4920         /* Have nested writes different that what is written */
4921         cnt = data->cnt + (nested ? 27 : 0);
4922 
4923         /* Multiply cnt by ~e, to make some unique increment */
4924         size = (data->cnt * 68 / 25) % (sizeof(rb_string) - 1);
4925 
4926         len = size + sizeof(struct rb_item);
4927 
4928         started = rb_test_started;
4929         /* read rb_test_started before checking buffer enabled */
4930         smp_rmb();
4931 
4932         event = ring_buffer_lock_reserve(data->buffer, len);
4933         if (!event) {
4934                 /* Ignore dropped events before test starts. */
4935                 if (started) {
4936                         if (nested)
4937                                 data->bytes_dropped += len;
4938                         else
4939                                 data->bytes_dropped_nested += len;
4940                 }
4941                 return len;
4942         }
4943 
4944         event_len = ring_buffer_event_length(event);
4945 
4946         if (RB_WARN_ON(data->buffer, event_len < len))
4947                 goto out;
4948 
4949         item = ring_buffer_event_data(event);
4950         item->size = size;
4951         memcpy(item->str, rb_string, size);
4952 
4953         if (nested) {
4954                 data->bytes_alloc_nested += event_len;
4955                 data->bytes_written_nested += len;
4956                 data->events_nested++;
4957                 if (!data->min_size_nested || len < data->min_size_nested)
4958                         data->min_size_nested = len;
4959                 if (len > data->max_size_nested)
4960                         data->max_size_nested = len;
4961         } else {
4962                 data->bytes_alloc += event_len;
4963                 data->bytes_written += len;
4964                 data->events++;
4965                 if (!data->min_size || len < data->min_size)
4966                         data->max_size = len;
4967                 if (len > data->max_size)
4968                         data->max_size = len;
4969         }
4970 
4971  out:
4972         ring_buffer_unlock_commit(data->buffer, event);
4973 
4974         return 0;
4975 }
4976 
4977 static __init int rb_test(void *arg)
4978 {
4979         struct rb_test_data *data = arg;
4980 
4981         while (!kthread_should_stop()) {
4982                 rb_write_something(data, false);
4983                 data->cnt++;
4984 
4985                 set_current_state(TASK_INTERRUPTIBLE);
4986                 /* Now sleep between a min of 100-300us and a max of 1ms */
4987                 usleep_range(((data->cnt % 3) + 1) * 100, 1000);
4988         }
4989 
4990         return 0;
4991 }
4992 
4993 static __init void rb_ipi(void *ignore)
4994 {
4995         struct rb_test_data *data;
4996         int cpu = smp_processor_id();
4997 
4998         data = &rb_data[cpu];
4999         rb_write_something(data, true);
5000 }
5001 
5002 static __init int rb_hammer_test(void *arg)
5003 {
5004         while (!kthread_should_stop()) {
5005 
5006                 /* Send an IPI to all cpus to write data! */
5007                 smp_call_function(rb_ipi, NULL, 1);
5008                 /* No sleep, but for non preempt, let others run */
5009                 schedule();
5010         }
5011 
5012         return 0;
5013 }
5014 
5015 static __init int test_ringbuffer(void)
5016 {
5017         struct task_struct *rb_hammer;
5018         struct ring_buffer *buffer;
5019         int cpu;
5020         int ret = 0;
5021 
5022         pr_info("Running ring buffer tests...\n");
5023 
5024         buffer = ring_buffer_alloc(RB_TEST_BUFFER_SIZE, RB_FL_OVERWRITE);
5025         if (WARN_ON(!buffer))
5026                 return 0;
5027 
5028         /* Disable buffer so that threads can't write to it yet */
5029         ring_buffer_record_off(buffer);
5030 
5031         for_each_online_cpu(cpu) {
5032                 rb_data[cpu].buffer = buffer;
5033                 rb_data[cpu].cpu = cpu;
5034                 rb_data[cpu].cnt = cpu;
5035                 rb_threads[cpu] = kthread_create(rb_test, &rb_data[cpu],
5036                                                  "rbtester/%d", cpu);
5037                 if (WARN_ON(IS_ERR(rb_threads[cpu]))) {
5038                         pr_cont("FAILED\n");
5039                         ret = PTR_ERR(rb_threads[cpu]);
5040                         goto out_free;
5041                 }
5042 
5043                 kthread_bind(rb_threads[cpu], cpu);
5044                 wake_up_process(rb_threads[cpu]);
5045         }
5046 
5047         /* Now create the rb hammer! */
5048         rb_hammer = kthread_run(rb_hammer_test, NULL, "rbhammer");
5049         if (WARN_ON(IS_ERR(rb_hammer))) {
5050                 pr_cont("FAILED\n");
5051                 ret = PTR_ERR(rb_hammer);
5052                 goto out_free;
5053         }
5054 
5055         ring_buffer_record_on(buffer);
5056         /*
5057          * Show buffer is enabled before setting rb_test_started.
5058          * Yes there's a small race window where events could be
5059          * dropped and the thread wont catch it. But when a ring
5060          * buffer gets enabled, there will always be some kind of
5061          * delay before other CPUs see it. Thus, we don't care about
5062          * those dropped events. We care about events dropped after
5063          * the threads see that the buffer is active.
5064          */
5065         smp_wmb();
5066         rb_test_started = true;
5067 
5068         set_current_state(TASK_INTERRUPTIBLE);
5069         /* Just run for 10 seconds */;
5070         schedule_timeout(10 * HZ);
5071 
5072         kthread_stop(rb_hammer);
5073 
5074  out_free:
5075         for_each_online_cpu(cpu) {
5076                 if (!rb_threads[cpu])
5077                         break;
5078                 kthread_stop(rb_threads[cpu]);
5079         }
5080         if (ret) {
5081                 ring_buffer_free(buffer);
5082                 return ret;
5083         }
5084 
5085         /* Report! */
5086         pr_info("finished\n");
5087         for_each_online_cpu(cpu) {
5088                 struct ring_buffer_event *event;
5089                 struct rb_test_data *data = &rb_data[cpu];
5090                 struct rb_item *item;
5091                 unsigned long total_events;
5092                 unsigned long total_dropped;
5093                 unsigned long total_written;
5094                 unsigned long total_alloc;
5095                 unsigned long total_read = 0;
5096                 unsigned long total_size = 0;
5097                 unsigned long total_len = 0;
5098                 unsigned long total_lost = 0;
5099                 unsigned long lost;
5100                 int big_event_size;
5101                 int small_event_size;
5102 
5103                 ret = -1;
5104 
5105                 total_events = data->events + data->events_nested;
5106                 total_written = data->bytes_written + data->bytes_written_nested;
5107                 total_alloc = data->bytes_alloc + data->bytes_alloc_nested;
5108                 total_dropped = data->bytes_dropped + data->bytes_dropped_nested;
5109 
5110                 big_event_size = data->max_size + data->max_size_nested;
5111                 small_event_size = data->min_size + data->min_size_nested;
5112 
5113                 pr_info("CPU %d:\n", cpu);
5114                 pr_info("              events:    %ld\n", total_events);
5115                 pr_info("       dropped bytes:    %ld\n", total_dropped);
5116                 pr_info("       alloced bytes:    %ld\n", total_alloc);
5117                 pr_info("       written bytes:    %ld\n", total_written);
5118                 pr_info("       biggest event:    %d\n", big_event_size);
5119                 pr_info("      smallest event:    %d\n", small_event_size);
5120 
5121                 if (RB_WARN_ON(buffer, total_dropped))
5122                         break;
5123 
5124                 ret = 0;
5125 
5126                 while ((event = ring_buffer_consume(buffer, cpu, NULL, &lost))) {
5127                         total_lost += lost;
5128                         item = ring_buffer_event_data(event);
5129                         total_len += ring_buffer_event_length(event);
5130                         total_size += item->size + sizeof(struct rb_item);
5131                         if (memcmp(&item->str[0], rb_string, item->size) != 0) {
5132                                 pr_info("FAILED!\n");
5133                                 pr_info("buffer had: %.*s\n", item->size, item->str);
5134                                 pr_info("expected:   %.*s\n", item->size, rb_string);
5135                                 RB_WARN_ON(buffer, 1);
5136                                 ret = -1;
5137                                 break;
5138                         }
5139                         total_read++;
5140                 }
5141                 if (ret)
5142                         break;
5143 
5144                 ret = -1;
5145 
5146                 pr_info("         read events:   %ld\n", total_read);
5147                 pr_info("         lost events:   %ld\n", total_lost);
5148                 pr_info("        total events:   %ld\n", total_lost + total_read);
5149                 pr_info("  recorded len bytes:   %ld\n", total_len);
5150                 pr_info(" recorded size bytes:   %ld\n", total_size);
5151                 if (total_lost)
5152                         pr_info(" With dropped events, record len and size may not match\n"
5153                                 " alloced and written from above\n");
5154                 if (!total_lost) {
5155                         if (RB_WARN_ON(buffer, total_len != total_alloc ||
5156                                        total_size != total_written))
5157                                 break;
5158                 }
5159                 if (RB_WARN_ON(buffer, total_lost + total_read != total_events))
5160                         break;
5161 
5162                 ret = 0;
5163         }
5164         if (!ret)
5165                 pr_info("Ring buffer PASSED!\n");
5166 
5167         ring_buffer_free(buffer);
5168         return 0;
5169 }
5170 
5171 late_initcall(test_ringbuffer);
5172 #endif /* CONFIG_RING_BUFFER_STARTUP_TEST */
5173 

~ [ 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