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

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

Version: ~ [ linux-5.5-rc7 ] ~ [ linux-5.4.13 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.97 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.166 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.210 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.210 ] ~ [ 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-3.16.81 ] ~ [ 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-3.9.11 ] ~ [ linux-3.8.13 ] ~ [ linux-3.7.10 ] ~ [ linux-3.6.11 ] ~ [ linux-3.5.7 ] ~ [ linux-3.4.113 ] ~ [ linux-3.3.8 ] ~ [ linux-3.2.102 ] ~ [ linux-3.1.10 ] ~ [ linux-3.0.101 ] ~ [ 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  *  temporarily.
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_deactivate - 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                 cond_resched();
1549 
1550                 to_remove_page = tmp_iter_page;
1551                 rb_inc_page(cpu_buffer, &tmp_iter_page);
1552 
1553                 /* update the counters */
1554                 page_entries = rb_page_entries(to_remove_page);
1555                 if (page_entries) {
1556                         /*
1557                          * If something was added to this page, it was full
1558                          * since it is not the tail page. So we deduct the
1559                          * bytes consumed in ring buffer from here.
1560                          * Increment overrun to account for the lost events.
1561                          */
1562                         local_add(page_entries, &cpu_buffer->overrun);
1563                         local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1564                 }
1565 
1566                 /*
1567                  * We have already removed references to this list item, just
1568                  * free up the buffer_page and its page
1569                  */
1570                 free_buffer_page(to_remove_page);
1571                 nr_removed--;
1572 
1573         } while (to_remove_page != last_page);
1574 
1575         RB_WARN_ON(cpu_buffer, nr_removed);
1576 
1577         return nr_removed == 0;
1578 }
1579 
1580 static int
1581 rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
1582 {
1583         struct list_head *pages = &cpu_buffer->new_pages;
1584         int retries, success;
1585 
1586         raw_spin_lock_irq(&cpu_buffer->reader_lock);
1587         /*
1588          * We are holding the reader lock, so the reader page won't be swapped
1589          * in the ring buffer. Now we are racing with the writer trying to
1590          * move head page and the tail page.
1591          * We are going to adapt the reader page update process where:
1592          * 1. We first splice the start and end of list of new pages between
1593          *    the head page and its previous page.
1594          * 2. We cmpxchg the prev_page->next to point from head page to the
1595          *    start of new pages list.
1596          * 3. Finally, we update the head->prev to the end of new list.
1597          *
1598          * We will try this process 10 times, to make sure that we don't keep
1599          * spinning.
1600          */
1601         retries = 10;
1602         success = 0;
1603         while (retries--) {
1604                 struct list_head *head_page, *prev_page, *r;
1605                 struct list_head *last_page, *first_page;
1606                 struct list_head *head_page_with_bit;
1607 
1608                 head_page = &rb_set_head_page(cpu_buffer)->list;
1609                 if (!head_page)
1610                         break;
1611                 prev_page = head_page->prev;
1612 
1613                 first_page = pages->next;
1614                 last_page  = pages->prev;
1615 
1616                 head_page_with_bit = (struct list_head *)
1617                                      ((unsigned long)head_page | RB_PAGE_HEAD);
1618 
1619                 last_page->next = head_page_with_bit;
1620                 first_page->prev = prev_page;
1621 
1622                 r = cmpxchg(&prev_page->next, head_page_with_bit, first_page);
1623 
1624                 if (r == head_page_with_bit) {
1625                         /*
1626                          * yay, we replaced the page pointer to our new list,
1627                          * now, we just have to update to head page's prev
1628                          * pointer to point to end of list
1629                          */
1630                         head_page->prev = last_page;
1631                         success = 1;
1632                         break;
1633                 }
1634         }
1635 
1636         if (success)
1637                 INIT_LIST_HEAD(pages);
1638         /*
1639          * If we weren't successful in adding in new pages, warn and stop
1640          * tracing
1641          */
1642         RB_WARN_ON(cpu_buffer, !success);
1643         raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1644 
1645         /* free pages if they weren't inserted */
1646         if (!success) {
1647                 struct buffer_page *bpage, *tmp;
1648                 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1649                                          list) {
1650                         list_del_init(&bpage->list);
1651                         free_buffer_page(bpage);
1652                 }
1653         }
1654         return success;
1655 }
1656 
1657 static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
1658 {
1659         int success;
1660 
1661         if (cpu_buffer->nr_pages_to_update > 0)
1662                 success = rb_insert_pages(cpu_buffer);
1663         else
1664                 success = rb_remove_pages(cpu_buffer,
1665                                         -cpu_buffer->nr_pages_to_update);
1666 
1667         if (success)
1668                 cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
1669 }
1670 
1671 static void update_pages_handler(struct work_struct *work)
1672 {
1673         struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
1674                         struct ring_buffer_per_cpu, update_pages_work);
1675         rb_update_pages(cpu_buffer);
1676         complete(&cpu_buffer->update_done);
1677 }
1678 
1679 /**
1680  * ring_buffer_resize - resize the ring buffer
1681  * @buffer: the buffer to resize.
1682  * @size: the new size.
1683  * @cpu_id: the cpu buffer to resize
1684  *
1685  * Minimum size is 2 * BUF_PAGE_SIZE.
1686  *
1687  * Returns 0 on success and < 0 on failure.
1688  */
1689 int ring_buffer_resize(struct ring_buffer *buffer, unsigned long size,
1690                         int cpu_id)
1691 {
1692         struct ring_buffer_per_cpu *cpu_buffer;
1693         unsigned long nr_pages;
1694         int cpu, err = 0;
1695 
1696         /*
1697          * Always succeed at resizing a non-existent buffer:
1698          */
1699         if (!buffer)
1700                 return size;
1701 
1702         /* Make sure the requested buffer exists */
1703         if (cpu_id != RING_BUFFER_ALL_CPUS &&
1704             !cpumask_test_cpu(cpu_id, buffer->cpumask))
1705                 return size;
1706 
1707         nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1708 
1709         /* we need a minimum of two pages */
1710         if (nr_pages < 2)
1711                 nr_pages = 2;
1712 
1713         size = nr_pages * BUF_PAGE_SIZE;
1714 
1715         /*
1716          * Don't succeed if resizing is disabled, as a reader might be
1717          * manipulating the ring buffer and is expecting a sane state while
1718          * this is true.
1719          */
1720         if (atomic_read(&buffer->resize_disabled))
1721                 return -EBUSY;
1722 
1723         /* prevent another thread from changing buffer sizes */
1724         mutex_lock(&buffer->mutex);
1725 
1726         if (cpu_id == RING_BUFFER_ALL_CPUS) {
1727                 /* calculate the pages to update */
1728                 for_each_buffer_cpu(buffer, cpu) {
1729                         cpu_buffer = buffer->buffers[cpu];
1730 
1731                         cpu_buffer->nr_pages_to_update = nr_pages -
1732                                                         cpu_buffer->nr_pages;
1733                         /*
1734                          * nothing more to do for removing pages or no update
1735                          */
1736                         if (cpu_buffer->nr_pages_to_update <= 0)
1737                                 continue;
1738                         /*
1739                          * to add pages, make sure all new pages can be
1740                          * allocated without receiving ENOMEM
1741                          */
1742                         INIT_LIST_HEAD(&cpu_buffer->new_pages);
1743                         if (__rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1744                                                 &cpu_buffer->new_pages, cpu)) {
1745                                 /* not enough memory for new pages */
1746                                 err = -ENOMEM;
1747                                 goto out_err;
1748                         }
1749                 }
1750 
1751                 get_online_cpus();
1752                 /*
1753                  * Fire off all the required work handlers
1754                  * We can't schedule on offline CPUs, but it's not necessary
1755                  * since we can change their buffer sizes without any race.
1756                  */
1757                 for_each_buffer_cpu(buffer, cpu) {
1758                         cpu_buffer = buffer->buffers[cpu];
1759                         if (!cpu_buffer->nr_pages_to_update)
1760                                 continue;
1761 
1762                         /* Can't run something on an offline CPU. */
1763                         if (!cpu_online(cpu)) {
1764                                 rb_update_pages(cpu_buffer);
1765                                 cpu_buffer->nr_pages_to_update = 0;
1766                         } else {
1767                                 schedule_work_on(cpu,
1768                                                 &cpu_buffer->update_pages_work);
1769                         }
1770                 }
1771 
1772                 /* wait for all the updates to complete */
1773                 for_each_buffer_cpu(buffer, cpu) {
1774                         cpu_buffer = buffer->buffers[cpu];
1775                         if (!cpu_buffer->nr_pages_to_update)
1776                                 continue;
1777 
1778                         if (cpu_online(cpu))
1779                                 wait_for_completion(&cpu_buffer->update_done);
1780                         cpu_buffer->nr_pages_to_update = 0;
1781                 }
1782 
1783                 put_online_cpus();
1784         } else {
1785                 /* Make sure this CPU has been initialized */
1786                 if (!cpumask_test_cpu(cpu_id, buffer->cpumask))
1787                         goto out;
1788 
1789                 cpu_buffer = buffer->buffers[cpu_id];
1790 
1791                 if (nr_pages == cpu_buffer->nr_pages)
1792                         goto out;
1793 
1794                 cpu_buffer->nr_pages_to_update = nr_pages -
1795                                                 cpu_buffer->nr_pages;
1796 
1797                 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1798                 if (cpu_buffer->nr_pages_to_update > 0 &&
1799                         __rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1800                                             &cpu_buffer->new_pages, cpu_id)) {
1801                         err = -ENOMEM;
1802                         goto out_err;
1803                 }
1804 
1805                 get_online_cpus();
1806 
1807                 /* Can't run something on an offline CPU. */
1808                 if (!cpu_online(cpu_id))
1809                         rb_update_pages(cpu_buffer);
1810                 else {
1811                         schedule_work_on(cpu_id,
1812                                          &cpu_buffer->update_pages_work);
1813                         wait_for_completion(&cpu_buffer->update_done);
1814                 }
1815 
1816                 cpu_buffer->nr_pages_to_update = 0;
1817                 put_online_cpus();
1818         }
1819 
1820  out:
1821         /*
1822          * The ring buffer resize can happen with the ring buffer
1823          * enabled, so that the update disturbs the tracing as little
1824          * as possible. But if the buffer is disabled, we do not need
1825          * to worry about that, and we can take the time to verify
1826          * that the buffer is not corrupt.
1827          */
1828         if (atomic_read(&buffer->record_disabled)) {
1829                 atomic_inc(&buffer->record_disabled);
1830                 /*
1831                  * Even though the buffer was disabled, we must make sure
1832                  * that it is truly disabled before calling rb_check_pages.
1833                  * There could have been a race between checking
1834                  * record_disable and incrementing it.
1835                  */
1836                 synchronize_sched();
1837                 for_each_buffer_cpu(buffer, cpu) {
1838                         cpu_buffer = buffer->buffers[cpu];
1839                         rb_check_pages(cpu_buffer);
1840                 }
1841                 atomic_dec(&buffer->record_disabled);
1842         }
1843 
1844         mutex_unlock(&buffer->mutex);
1845         return size;
1846 
1847  out_err:
1848         for_each_buffer_cpu(buffer, cpu) {
1849                 struct buffer_page *bpage, *tmp;
1850 
1851                 cpu_buffer = buffer->buffers[cpu];
1852                 cpu_buffer->nr_pages_to_update = 0;
1853 
1854                 if (list_empty(&cpu_buffer->new_pages))
1855                         continue;
1856 
1857                 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1858                                         list) {
1859                         list_del_init(&bpage->list);
1860                         free_buffer_page(bpage);
1861                 }
1862         }
1863         mutex_unlock(&buffer->mutex);
1864         return err;
1865 }
1866 EXPORT_SYMBOL_GPL(ring_buffer_resize);
1867 
1868 void ring_buffer_change_overwrite(struct ring_buffer *buffer, int val)
1869 {
1870         mutex_lock(&buffer->mutex);
1871         if (val)
1872                 buffer->flags |= RB_FL_OVERWRITE;
1873         else
1874                 buffer->flags &= ~RB_FL_OVERWRITE;
1875         mutex_unlock(&buffer->mutex);
1876 }
1877 EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
1878 
1879 static __always_inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
1880 {
1881         return bpage->page->data + index;
1882 }
1883 
1884 static __always_inline struct ring_buffer_event *
1885 rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
1886 {
1887         return __rb_page_index(cpu_buffer->reader_page,
1888                                cpu_buffer->reader_page->read);
1889 }
1890 
1891 static __always_inline struct ring_buffer_event *
1892 rb_iter_head_event(struct ring_buffer_iter *iter)
1893 {
1894         return __rb_page_index(iter->head_page, iter->head);
1895 }
1896 
1897 static __always_inline unsigned rb_page_commit(struct buffer_page *bpage)
1898 {
1899         return local_read(&bpage->page->commit);
1900 }
1901 
1902 /* Size is determined by what has been committed */
1903 static __always_inline unsigned rb_page_size(struct buffer_page *bpage)
1904 {
1905         return rb_page_commit(bpage);
1906 }
1907 
1908 static __always_inline unsigned
1909 rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
1910 {
1911         return rb_page_commit(cpu_buffer->commit_page);
1912 }
1913 
1914 static __always_inline unsigned
1915 rb_event_index(struct ring_buffer_event *event)
1916 {
1917         unsigned long addr = (unsigned long)event;
1918 
1919         return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
1920 }
1921 
1922 static void rb_inc_iter(struct ring_buffer_iter *iter)
1923 {
1924         struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
1925 
1926         /*
1927          * The iterator could be on the reader page (it starts there).
1928          * But the head could have moved, since the reader was
1929          * found. Check for this case and assign the iterator
1930          * to the head page instead of next.
1931          */
1932         if (iter->head_page == cpu_buffer->reader_page)
1933                 iter->head_page = rb_set_head_page(cpu_buffer);
1934         else
1935                 rb_inc_page(cpu_buffer, &iter->head_page);
1936 
1937         iter->read_stamp = iter->head_page->page->time_stamp;
1938         iter->head = 0;
1939 }
1940 
1941 /*
1942  * rb_handle_head_page - writer hit the head page
1943  *
1944  * Returns: +1 to retry page
1945  *           0 to continue
1946  *          -1 on error
1947  */
1948 static int
1949 rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
1950                     struct buffer_page *tail_page,
1951                     struct buffer_page *next_page)
1952 {
1953         struct buffer_page *new_head;
1954         int entries;
1955         int type;
1956         int ret;
1957 
1958         entries = rb_page_entries(next_page);
1959 
1960         /*
1961          * The hard part is here. We need to move the head
1962          * forward, and protect against both readers on
1963          * other CPUs and writers coming in via interrupts.
1964          */
1965         type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
1966                                        RB_PAGE_HEAD);
1967 
1968         /*
1969          * type can be one of four:
1970          *  NORMAL - an interrupt already moved it for us
1971          *  HEAD   - we are the first to get here.
1972          *  UPDATE - we are the interrupt interrupting
1973          *           a current move.
1974          *  MOVED  - a reader on another CPU moved the next
1975          *           pointer to its reader page. Give up
1976          *           and try again.
1977          */
1978 
1979         switch (type) {
1980         case RB_PAGE_HEAD:
1981                 /*
1982                  * We changed the head to UPDATE, thus
1983                  * it is our responsibility to update
1984                  * the counters.
1985                  */
1986                 local_add(entries, &cpu_buffer->overrun);
1987                 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1988 
1989                 /*
1990                  * The entries will be zeroed out when we move the
1991                  * tail page.
1992                  */
1993 
1994                 /* still more to do */
1995                 break;
1996 
1997         case RB_PAGE_UPDATE:
1998                 /*
1999                  * This is an interrupt that interrupt the
2000                  * previous update. Still more to do.
2001                  */
2002                 break;
2003         case RB_PAGE_NORMAL:
2004                 /*
2005                  * An interrupt came in before the update
2006                  * and processed this for us.
2007                  * Nothing left to do.
2008                  */
2009                 return 1;
2010         case RB_PAGE_MOVED:
2011                 /*
2012                  * The reader is on another CPU and just did
2013                  * a swap with our next_page.
2014                  * Try again.
2015                  */
2016                 return 1;
2017         default:
2018                 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
2019                 return -1;
2020         }
2021 
2022         /*
2023          * Now that we are here, the old head pointer is
2024          * set to UPDATE. This will keep the reader from
2025          * swapping the head page with the reader page.
2026          * The reader (on another CPU) will spin till
2027          * we are finished.
2028          *
2029          * We just need to protect against interrupts
2030          * doing the job. We will set the next pointer
2031          * to HEAD. After that, we set the old pointer
2032          * to NORMAL, but only if it was HEAD before.
2033          * otherwise we are an interrupt, and only
2034          * want the outer most commit to reset it.
2035          */
2036         new_head = next_page;
2037         rb_inc_page(cpu_buffer, &new_head);
2038 
2039         ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
2040                                     RB_PAGE_NORMAL);
2041 
2042         /*
2043          * Valid returns are:
2044          *  HEAD   - an interrupt came in and already set it.
2045          *  NORMAL - One of two things:
2046          *            1) We really set it.
2047          *            2) A bunch of interrupts came in and moved
2048          *               the page forward again.
2049          */
2050         switch (ret) {
2051         case RB_PAGE_HEAD:
2052         case RB_PAGE_NORMAL:
2053                 /* OK */
2054                 break;
2055         default:
2056                 RB_WARN_ON(cpu_buffer, 1);
2057                 return -1;
2058         }
2059 
2060         /*
2061          * It is possible that an interrupt came in,
2062          * set the head up, then more interrupts came in
2063          * and moved it again. When we get back here,
2064          * the page would have been set to NORMAL but we
2065          * just set it back to HEAD.
2066          *
2067          * How do you detect this? Well, if that happened
2068          * the tail page would have moved.
2069          */
2070         if (ret == RB_PAGE_NORMAL) {
2071                 struct buffer_page *buffer_tail_page;
2072 
2073                 buffer_tail_page = READ_ONCE(cpu_buffer->tail_page);
2074                 /*
2075                  * If the tail had moved passed next, then we need
2076                  * to reset the pointer.
2077                  */
2078                 if (buffer_tail_page != tail_page &&
2079                     buffer_tail_page != next_page)
2080                         rb_head_page_set_normal(cpu_buffer, new_head,
2081                                                 next_page,
2082                                                 RB_PAGE_HEAD);
2083         }
2084 
2085         /*
2086          * If this was the outer most commit (the one that
2087          * changed the original pointer from HEAD to UPDATE),
2088          * then it is up to us to reset it to NORMAL.
2089          */
2090         if (type == RB_PAGE_HEAD) {
2091                 ret = rb_head_page_set_normal(cpu_buffer, next_page,
2092                                               tail_page,
2093                                               RB_PAGE_UPDATE);
2094                 if (RB_WARN_ON(cpu_buffer,
2095                                ret != RB_PAGE_UPDATE))
2096                         return -1;
2097         }
2098 
2099         return 0;
2100 }
2101 
2102 static inline void
2103 rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2104               unsigned long tail, struct rb_event_info *info)
2105 {
2106         struct buffer_page *tail_page = info->tail_page;
2107         struct ring_buffer_event *event;
2108         unsigned long length = info->length;
2109 
2110         /*
2111          * Only the event that crossed the page boundary
2112          * must fill the old tail_page with padding.
2113          */
2114         if (tail >= BUF_PAGE_SIZE) {
2115                 /*
2116                  * If the page was filled, then we still need
2117                  * to update the real_end. Reset it to zero
2118                  * and the reader will ignore it.
2119                  */
2120                 if (tail == BUF_PAGE_SIZE)
2121                         tail_page->real_end = 0;
2122 
2123                 local_sub(length, &tail_page->write);
2124                 return;
2125         }
2126 
2127         event = __rb_page_index(tail_page, tail);
2128 
2129         /* account for padding bytes */
2130         local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes);
2131 
2132         /*
2133          * Save the original length to the meta data.
2134          * This will be used by the reader to add lost event
2135          * counter.
2136          */
2137         tail_page->real_end = tail;
2138 
2139         /*
2140          * If this event is bigger than the minimum size, then
2141          * we need to be careful that we don't subtract the
2142          * write counter enough to allow another writer to slip
2143          * in on this page.
2144          * We put in a discarded commit instead, to make sure
2145          * that this space is not used again.
2146          *
2147          * If we are less than the minimum size, we don't need to
2148          * worry about it.
2149          */
2150         if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
2151                 /* No room for any events */
2152 
2153                 /* Mark the rest of the page with padding */
2154                 rb_event_set_padding(event);
2155 
2156                 /* Set the write back to the previous setting */
2157                 local_sub(length, &tail_page->write);
2158                 return;
2159         }
2160 
2161         /* Put in a discarded event */
2162         event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
2163         event->type_len = RINGBUF_TYPE_PADDING;
2164         /* time delta must be non zero */
2165         event->time_delta = 1;
2166 
2167         /* Set write to end of buffer */
2168         length = (tail + length) - BUF_PAGE_SIZE;
2169         local_sub(length, &tail_page->write);
2170 }
2171 
2172 static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer);
2173 
2174 /*
2175  * This is the slow path, force gcc not to inline it.
2176  */
2177 static noinline struct ring_buffer_event *
2178 rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2179              unsigned long tail, struct rb_event_info *info)
2180 {
2181         struct buffer_page *tail_page = info->tail_page;
2182         struct buffer_page *commit_page = cpu_buffer->commit_page;
2183         struct ring_buffer *buffer = cpu_buffer->buffer;
2184         struct buffer_page *next_page;
2185         int ret;
2186 
2187         next_page = tail_page;
2188 
2189         rb_inc_page(cpu_buffer, &next_page);
2190 
2191         /*
2192          * If for some reason, we had an interrupt storm that made
2193          * it all the way around the buffer, bail, and warn
2194          * about it.
2195          */
2196         if (unlikely(next_page == commit_page)) {
2197                 local_inc(&cpu_buffer->commit_overrun);
2198                 goto out_reset;
2199         }
2200 
2201         /*
2202          * This is where the fun begins!
2203          *
2204          * We are fighting against races between a reader that
2205          * could be on another CPU trying to swap its reader
2206          * page with the buffer head.
2207          *
2208          * We are also fighting against interrupts coming in and
2209          * moving the head or tail on us as well.
2210          *
2211          * If the next page is the head page then we have filled
2212          * the buffer, unless the commit page is still on the
2213          * reader page.
2214          */
2215         if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) {
2216 
2217                 /*
2218                  * If the commit is not on the reader page, then
2219                  * move the header page.
2220                  */
2221                 if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2222                         /*
2223                          * If we are not in overwrite mode,
2224                          * this is easy, just stop here.
2225                          */
2226                         if (!(buffer->flags & RB_FL_OVERWRITE)) {
2227                                 local_inc(&cpu_buffer->dropped_events);
2228                                 goto out_reset;
2229                         }
2230 
2231                         ret = rb_handle_head_page(cpu_buffer,
2232                                                   tail_page,
2233                                                   next_page);
2234                         if (ret < 0)
2235                                 goto out_reset;
2236                         if (ret)
2237                                 goto out_again;
2238                 } else {
2239                         /*
2240                          * We need to be careful here too. The
2241                          * commit page could still be on the reader
2242                          * page. We could have a small buffer, and
2243                          * have filled up the buffer with events
2244                          * from interrupts and such, and wrapped.
2245                          *
2246                          * Note, if the tail page is also the on the
2247                          * reader_page, we let it move out.
2248                          */
2249                         if (unlikely((cpu_buffer->commit_page !=
2250                                       cpu_buffer->tail_page) &&
2251                                      (cpu_buffer->commit_page ==
2252                                       cpu_buffer->reader_page))) {
2253                                 local_inc(&cpu_buffer->commit_overrun);
2254                                 goto out_reset;
2255                         }
2256                 }
2257         }
2258 
2259         rb_tail_page_update(cpu_buffer, tail_page, next_page);
2260 
2261  out_again:
2262 
2263         rb_reset_tail(cpu_buffer, tail, info);
2264 
2265         /* Commit what we have for now. */
2266         rb_end_commit(cpu_buffer);
2267         /* rb_end_commit() decs committing */
2268         local_inc(&cpu_buffer->committing);
2269 
2270         /* fail and let the caller try again */
2271         return ERR_PTR(-EAGAIN);
2272 
2273  out_reset:
2274         /* reset write */
2275         rb_reset_tail(cpu_buffer, tail, info);
2276 
2277         return NULL;
2278 }
2279 
2280 /* Slow path, do not inline */
2281 static noinline struct ring_buffer_event *
2282 rb_add_time_stamp(struct ring_buffer_event *event, u64 delta, bool abs)
2283 {
2284         if (abs)
2285                 event->type_len = RINGBUF_TYPE_TIME_STAMP;
2286         else
2287                 event->type_len = RINGBUF_TYPE_TIME_EXTEND;
2288 
2289         /* Not the first event on the page, or not delta? */
2290         if (abs || rb_event_index(event)) {
2291                 event->time_delta = delta & TS_MASK;
2292                 event->array[0] = delta >> TS_SHIFT;
2293         } else {
2294                 /* nope, just zero it */
2295                 event->time_delta = 0;
2296                 event->array[0] = 0;
2297         }
2298 
2299         return skip_time_extend(event);
2300 }
2301 
2302 static inline bool rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2303                                      struct ring_buffer_event *event);
2304 
2305 /**
2306  * rb_update_event - update event type and data
2307  * @event: the event to update
2308  * @type: the type of event
2309  * @length: the size of the event field in the ring buffer
2310  *
2311  * Update the type and data fields of the event. The length
2312  * is the actual size that is written to the ring buffer,
2313  * and with this, we can determine what to place into the
2314  * data field.
2315  */
2316 static void
2317 rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
2318                 struct ring_buffer_event *event,
2319                 struct rb_event_info *info)
2320 {
2321         unsigned length = info->length;
2322         u64 delta = info->delta;
2323 
2324         /* Only a commit updates the timestamp */
2325         if (unlikely(!rb_event_is_commit(cpu_buffer, event)))
2326                 delta = 0;
2327 
2328         /*
2329          * If we need to add a timestamp, then we
2330          * add it to the start of the reserved space.
2331          */
2332         if (unlikely(info->add_timestamp)) {
2333                 bool abs = ring_buffer_time_stamp_abs(cpu_buffer->buffer);
2334 
2335                 event = rb_add_time_stamp(event, info->delta, abs);
2336                 length -= RB_LEN_TIME_EXTEND;
2337                 delta = 0;
2338         }
2339 
2340         event->time_delta = delta;
2341         length -= RB_EVNT_HDR_SIZE;
2342         if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
2343                 event->type_len = 0;
2344                 event->array[0] = length;
2345         } else
2346                 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
2347 }
2348 
2349 static unsigned rb_calculate_event_length(unsigned length)
2350 {
2351         struct ring_buffer_event event; /* Used only for sizeof array */
2352 
2353         /* zero length can cause confusions */
2354         if (!length)
2355                 length++;
2356 
2357         if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
2358                 length += sizeof(event.array[0]);
2359 
2360         length += RB_EVNT_HDR_SIZE;
2361         length = ALIGN(length, RB_ARCH_ALIGNMENT);
2362 
2363         /*
2364          * In case the time delta is larger than the 27 bits for it
2365          * in the header, we need to add a timestamp. If another
2366          * event comes in when trying to discard this one to increase
2367          * the length, then the timestamp will be added in the allocated
2368          * space of this event. If length is bigger than the size needed
2369          * for the TIME_EXTEND, then padding has to be used. The events
2370          * length must be either RB_LEN_TIME_EXTEND, or greater than or equal
2371          * to RB_LEN_TIME_EXTEND + 8, as 8 is the minimum size for padding.
2372          * As length is a multiple of 4, we only need to worry if it
2373          * is 12 (RB_LEN_TIME_EXTEND + 4).
2374          */
2375         if (length == RB_LEN_TIME_EXTEND + RB_ALIGNMENT)
2376                 length += RB_ALIGNMENT;
2377 
2378         return length;
2379 }
2380 
2381 #ifndef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2382 static inline bool sched_clock_stable(void)
2383 {
2384         return true;
2385 }
2386 #endif
2387 
2388 static inline int
2389 rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2390                   struct ring_buffer_event *event)
2391 {
2392         unsigned long new_index, old_index;
2393         struct buffer_page *bpage;
2394         unsigned long index;
2395         unsigned long addr;
2396 
2397         new_index = rb_event_index(event);
2398         old_index = new_index + rb_event_ts_length(event);
2399         addr = (unsigned long)event;
2400         addr &= PAGE_MASK;
2401 
2402         bpage = READ_ONCE(cpu_buffer->tail_page);
2403 
2404         if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2405                 unsigned long write_mask =
2406                         local_read(&bpage->write) & ~RB_WRITE_MASK;
2407                 unsigned long event_length = rb_event_length(event);
2408                 /*
2409                  * This is on the tail page. It is possible that
2410                  * a write could come in and move the tail page
2411                  * and write to the next page. That is fine
2412                  * because we just shorten what is on this page.
2413                  */
2414                 old_index += write_mask;
2415                 new_index += write_mask;
2416                 index = local_cmpxchg(&bpage->write, old_index, new_index);
2417                 if (index == old_index) {
2418                         /* update counters */
2419                         local_sub(event_length, &cpu_buffer->entries_bytes);
2420                         return 1;
2421                 }
2422         }
2423 
2424         /* could not discard */
2425         return 0;
2426 }
2427 
2428 static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2429 {
2430         local_inc(&cpu_buffer->committing);
2431         local_inc(&cpu_buffer->commits);
2432 }
2433 
2434 static __always_inline void
2435 rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
2436 {
2437         unsigned long max_count;
2438 
2439         /*
2440          * We only race with interrupts and NMIs on this CPU.
2441          * If we own the commit event, then we can commit
2442          * all others that interrupted us, since the interruptions
2443          * are in stack format (they finish before they come
2444          * back to us). This allows us to do a simple loop to
2445          * assign the commit to the tail.
2446          */
2447  again:
2448         max_count = cpu_buffer->nr_pages * 100;
2449 
2450         while (cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)) {
2451                 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
2452                         return;
2453                 if (RB_WARN_ON(cpu_buffer,
2454                                rb_is_reader_page(cpu_buffer->tail_page)))
2455                         return;
2456                 local_set(&cpu_buffer->commit_page->page->commit,
2457                           rb_page_write(cpu_buffer->commit_page));
2458                 rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
2459                 /* Only update the write stamp if the page has an event */
2460                 if (rb_page_write(cpu_buffer->commit_page))
2461                         cpu_buffer->write_stamp =
2462                                 cpu_buffer->commit_page->page->time_stamp;
2463                 /* add barrier to keep gcc from optimizing too much */
2464                 barrier();
2465         }
2466         while (rb_commit_index(cpu_buffer) !=
2467                rb_page_write(cpu_buffer->commit_page)) {
2468 
2469                 local_set(&cpu_buffer->commit_page->page->commit,
2470                           rb_page_write(cpu_buffer->commit_page));
2471                 RB_WARN_ON(cpu_buffer,
2472                            local_read(&cpu_buffer->commit_page->page->commit) &
2473                            ~RB_WRITE_MASK);
2474                 barrier();
2475         }
2476 
2477         /* again, keep gcc from optimizing */
2478         barrier();
2479 
2480         /*
2481          * If an interrupt came in just after the first while loop
2482          * and pushed the tail page forward, we will be left with
2483          * a dangling commit that will never go forward.
2484          */
2485         if (unlikely(cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)))
2486                 goto again;
2487 }
2488 
2489 static __always_inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
2490 {
2491         unsigned long commits;
2492 
2493         if (RB_WARN_ON(cpu_buffer,
2494                        !local_read(&cpu_buffer->committing)))
2495                 return;
2496 
2497  again:
2498         commits = local_read(&cpu_buffer->commits);
2499         /* synchronize with interrupts */
2500         barrier();
2501         if (local_read(&cpu_buffer->committing) == 1)
2502                 rb_set_commit_to_write(cpu_buffer);
2503 
2504         local_dec(&cpu_buffer->committing);
2505 
2506         /* synchronize with interrupts */
2507         barrier();
2508 
2509         /*
2510          * Need to account for interrupts coming in between the
2511          * updating of the commit page and the clearing of the
2512          * committing counter.
2513          */
2514         if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
2515             !local_read(&cpu_buffer->committing)) {
2516                 local_inc(&cpu_buffer->committing);
2517                 goto again;
2518         }
2519 }
2520 
2521 static inline void rb_event_discard(struct ring_buffer_event *event)
2522 {
2523         if (extended_time(event))
2524                 event = skip_time_extend(event);
2525 
2526         /* array[0] holds the actual length for the discarded event */
2527         event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
2528         event->type_len = RINGBUF_TYPE_PADDING;
2529         /* time delta must be non zero */
2530         if (!event->time_delta)
2531                 event->time_delta = 1;
2532 }
2533 
2534 static __always_inline bool
2535 rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2536                    struct ring_buffer_event *event)
2537 {
2538         unsigned long addr = (unsigned long)event;
2539         unsigned long index;
2540 
2541         index = rb_event_index(event);
2542         addr &= PAGE_MASK;
2543 
2544         return cpu_buffer->commit_page->page == (void *)addr &&
2545                 rb_commit_index(cpu_buffer) == index;
2546 }
2547 
2548 static __always_inline void
2549 rb_update_write_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2550                       struct ring_buffer_event *event)
2551 {
2552         u64 delta;
2553 
2554         /*
2555          * The event first in the commit queue updates the
2556          * time stamp.
2557          */
2558         if (rb_event_is_commit(cpu_buffer, event)) {
2559                 /*
2560                  * A commit event that is first on a page
2561                  * updates the write timestamp with the page stamp
2562                  */
2563                 if (!rb_event_index(event))
2564                         cpu_buffer->write_stamp =
2565                                 cpu_buffer->commit_page->page->time_stamp;
2566                 else if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
2567                         delta = ring_buffer_event_time_stamp(event);
2568                         cpu_buffer->write_stamp += delta;
2569                 } else if (event->type_len == RINGBUF_TYPE_TIME_STAMP) {
2570                         delta = ring_buffer_event_time_stamp(event);
2571                         cpu_buffer->write_stamp = delta;
2572                 } else
2573                         cpu_buffer->write_stamp += event->time_delta;
2574         }
2575 }
2576 
2577 static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
2578                       struct ring_buffer_event *event)
2579 {
2580         local_inc(&cpu_buffer->entries);
2581         rb_update_write_stamp(cpu_buffer, event);
2582         rb_end_commit(cpu_buffer);
2583 }
2584 
2585 static __always_inline void
2586 rb_wakeups(struct ring_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer)
2587 {
2588         bool pagebusy;
2589 
2590         if (buffer->irq_work.waiters_pending) {
2591                 buffer->irq_work.waiters_pending = false;
2592                 /* irq_work_queue() supplies it's own memory barriers */
2593                 irq_work_queue(&buffer->irq_work.work);
2594         }
2595 
2596         if (cpu_buffer->irq_work.waiters_pending) {
2597                 cpu_buffer->irq_work.waiters_pending = false;
2598                 /* irq_work_queue() supplies it's own memory barriers */
2599                 irq_work_queue(&cpu_buffer->irq_work.work);
2600         }
2601 
2602         pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
2603 
2604         if (!pagebusy && cpu_buffer->irq_work.full_waiters_pending) {
2605                 cpu_buffer->irq_work.wakeup_full = true;
2606                 cpu_buffer->irq_work.full_waiters_pending = false;
2607                 /* irq_work_queue() supplies it's own memory barriers */
2608                 irq_work_queue(&cpu_buffer->irq_work.work);
2609         }
2610 }
2611 
2612 /*
2613  * The lock and unlock are done within a preempt disable section.
2614  * The current_context per_cpu variable can only be modified
2615  * by the current task between lock and unlock. But it can
2616  * be modified more than once via an interrupt. To pass this
2617  * information from the lock to the unlock without having to
2618  * access the 'in_interrupt()' functions again (which do show
2619  * a bit of overhead in something as critical as function tracing,
2620  * we use a bitmask trick.
2621  *
2622  *  bit 0 =  NMI context
2623  *  bit 1 =  IRQ context
2624  *  bit 2 =  SoftIRQ context
2625  *  bit 3 =  normal context.
2626  *
2627  * This works because this is the order of contexts that can
2628  * preempt other contexts. A SoftIRQ never preempts an IRQ
2629  * context.
2630  *
2631  * When the context is determined, the corresponding bit is
2632  * checked and set (if it was set, then a recursion of that context
2633  * happened).
2634  *
2635  * On unlock, we need to clear this bit. To do so, just subtract
2636  * 1 from the current_context and AND it to itself.
2637  *
2638  * (binary)
2639  *  101 - 1 = 100
2640  *  101 & 100 = 100 (clearing bit zero)
2641  *
2642  *  1010 - 1 = 1001
2643  *  1010 & 1001 = 1000 (clearing bit 1)
2644  *
2645  * The least significant bit can be cleared this way, and it
2646  * just so happens that it is the same bit corresponding to
2647  * the current context.
2648  */
2649 
2650 static __always_inline int
2651 trace_recursive_lock(struct ring_buffer_per_cpu *cpu_buffer)
2652 {
2653         unsigned int val = cpu_buffer->current_context;
2654         unsigned long pc = preempt_count();
2655         int bit;
2656 
2657         if (!(pc & (NMI_MASK | HARDIRQ_MASK | SOFTIRQ_OFFSET)))
2658                 bit = RB_CTX_NORMAL;
2659         else
2660                 bit = pc & NMI_MASK ? RB_CTX_NMI :
2661                         pc & HARDIRQ_MASK ? RB_CTX_IRQ : RB_CTX_SOFTIRQ;
2662 
2663         if (unlikely(val & (1 << (bit + cpu_buffer->nest))))
2664                 return 1;
2665 
2666         val |= (1 << (bit + cpu_buffer->nest));
2667         cpu_buffer->current_context = val;
2668 
2669         return 0;
2670 }
2671 
2672 static __always_inline void
2673 trace_recursive_unlock(struct ring_buffer_per_cpu *cpu_buffer)
2674 {
2675         cpu_buffer->current_context &=
2676                 cpu_buffer->current_context - (1 << cpu_buffer->nest);
2677 }
2678 
2679 /* The recursive locking above uses 4 bits */
2680 #define NESTED_BITS 4
2681 
2682 /**
2683  * ring_buffer_nest_start - Allow to trace while nested
2684  * @buffer: The ring buffer to modify
2685  *
2686  * The ring buffer has a safety mechanism to prevent recursion.
2687  * But there may be a case where a trace needs to be done while
2688  * tracing something else. In this case, calling this function
2689  * will allow this function to nest within a currently active
2690  * ring_buffer_lock_reserve().
2691  *
2692  * Call this function before calling another ring_buffer_lock_reserve() and
2693  * call ring_buffer_nest_end() after the nested ring_buffer_unlock_commit().
2694  */
2695 void ring_buffer_nest_start(struct ring_buffer *buffer)
2696 {
2697         struct ring_buffer_per_cpu *cpu_buffer;
2698         int cpu;
2699 
2700         /* Enabled by ring_buffer_nest_end() */
2701         preempt_disable_notrace();
2702         cpu = raw_smp_processor_id();
2703         cpu_buffer = buffer->buffers[cpu];
2704         /* This is the shift value for the above recursive locking */
2705         cpu_buffer->nest += NESTED_BITS;
2706 }
2707 
2708 /**
2709  * ring_buffer_nest_end - Allow to trace while nested
2710  * @buffer: The ring buffer to modify
2711  *
2712  * Must be called after ring_buffer_nest_start() and after the
2713  * ring_buffer_unlock_commit().
2714  */
2715 void ring_buffer_nest_end(struct ring_buffer *buffer)
2716 {
2717         struct ring_buffer_per_cpu *cpu_buffer;
2718         int cpu;
2719 
2720         /* disabled by ring_buffer_nest_start() */
2721         cpu = raw_smp_processor_id();
2722         cpu_buffer = buffer->buffers[cpu];
2723         /* This is the shift value for the above recursive locking */
2724         cpu_buffer->nest -= NESTED_BITS;
2725         preempt_enable_notrace();
2726 }
2727 
2728 /**
2729  * ring_buffer_unlock_commit - commit a reserved
2730  * @buffer: The buffer to commit to
2731  * @event: The event pointer to commit.
2732  *
2733  * This commits the data to the ring buffer, and releases any locks held.
2734  *
2735  * Must be paired with ring_buffer_lock_reserve.
2736  */
2737 int ring_buffer_unlock_commit(struct ring_buffer *buffer,
2738                               struct ring_buffer_event *event)
2739 {
2740         struct ring_buffer_per_cpu *cpu_buffer;
2741         int cpu = raw_smp_processor_id();
2742 
2743         cpu_buffer = buffer->buffers[cpu];
2744 
2745         rb_commit(cpu_buffer, event);
2746 
2747         rb_wakeups(buffer, cpu_buffer);
2748 
2749         trace_recursive_unlock(cpu_buffer);
2750 
2751         preempt_enable_notrace();
2752 
2753         return 0;
2754 }
2755 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
2756 
2757 static noinline void
2758 rb_handle_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2759                     struct rb_event_info *info)
2760 {
2761         WARN_ONCE(info->delta > (1ULL << 59),
2762                   KERN_WARNING "Delta way too big! %llu ts=%llu write stamp = %llu\n%s",
2763                   (unsigned long long)info->delta,
2764                   (unsigned long long)info->ts,
2765                   (unsigned long long)cpu_buffer->write_stamp,
2766                   sched_clock_stable() ? "" :
2767                   "If you just came from a suspend/resume,\n"
2768                   "please switch to the trace global clock:\n"
2769                   "  echo global > /sys/kernel/debug/tracing/trace_clock\n"
2770                   "or add trace_clock=global to the kernel command line\n");
2771         info->add_timestamp = 1;
2772 }
2773 
2774 static struct ring_buffer_event *
2775 __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
2776                   struct rb_event_info *info)
2777 {
2778         struct ring_buffer_event *event;
2779         struct buffer_page *tail_page;
2780         unsigned long tail, write;
2781 
2782         /*
2783          * If the time delta since the last event is too big to
2784          * hold in the time field of the event, then we append a
2785          * TIME EXTEND event ahead of the data event.
2786          */
2787         if (unlikely(info->add_timestamp))
2788                 info->length += RB_LEN_TIME_EXTEND;
2789 
2790         /* Don't let the compiler play games with cpu_buffer->tail_page */
2791         tail_page = info->tail_page = READ_ONCE(cpu_buffer->tail_page);
2792         write = local_add_return(info->length, &tail_page->write);
2793 
2794         /* set write to only the index of the write */
2795         write &= RB_WRITE_MASK;
2796         tail = write - info->length;
2797 
2798         /*
2799          * If this is the first commit on the page, then it has the same
2800          * timestamp as the page itself.
2801          */
2802         if (!tail && !ring_buffer_time_stamp_abs(cpu_buffer->buffer))
2803                 info->delta = 0;
2804 
2805         /* See if we shot pass the end of this buffer page */
2806         if (unlikely(write > BUF_PAGE_SIZE))
2807                 return rb_move_tail(cpu_buffer, tail, info);
2808 
2809         /* We reserved something on the buffer */
2810 
2811         event = __rb_page_index(tail_page, tail);
2812         rb_update_event(cpu_buffer, event, info);
2813 
2814         local_inc(&tail_page->entries);
2815 
2816         /*
2817          * If this is the first commit on the page, then update
2818          * its timestamp.
2819          */
2820         if (!tail)
2821                 tail_page->page->time_stamp = info->ts;
2822 
2823         /* account for these added bytes */
2824         local_add(info->length, &cpu_buffer->entries_bytes);
2825 
2826         return event;
2827 }
2828 
2829 static __always_inline struct ring_buffer_event *
2830 rb_reserve_next_event(struct ring_buffer *buffer,
2831                       struct ring_buffer_per_cpu *cpu_buffer,
2832                       unsigned long length)
2833 {
2834         struct ring_buffer_event *event;
2835         struct rb_event_info info;
2836         int nr_loops = 0;
2837         u64 diff;
2838 
2839         rb_start_commit(cpu_buffer);
2840 
2841 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2842         /*
2843          * Due to the ability to swap a cpu buffer from a buffer
2844          * it is possible it was swapped before we committed.
2845          * (committing stops a swap). We check for it here and
2846          * if it happened, we have to fail the write.
2847          */
2848         barrier();
2849         if (unlikely(READ_ONCE(cpu_buffer->buffer) != buffer)) {
2850                 local_dec(&cpu_buffer->committing);
2851                 local_dec(&cpu_buffer->commits);
2852                 return NULL;
2853         }
2854 #endif
2855 
2856         info.length = rb_calculate_event_length(length);
2857  again:
2858         info.add_timestamp = 0;
2859         info.delta = 0;
2860 
2861         /*
2862          * We allow for interrupts to reenter here and do a trace.
2863          * If one does, it will cause this original code to loop
2864          * back here. Even with heavy interrupts happening, this
2865          * should only happen a few times in a row. If this happens
2866          * 1000 times in a row, there must be either an interrupt
2867          * storm or we have something buggy.
2868          * Bail!
2869          */
2870         if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
2871                 goto out_fail;
2872 
2873         info.ts = rb_time_stamp(cpu_buffer->buffer);
2874         diff = info.ts - cpu_buffer->write_stamp;
2875 
2876         /* make sure this diff is calculated here */
2877         barrier();
2878 
2879         if (ring_buffer_time_stamp_abs(buffer)) {
2880                 info.delta = info.ts;
2881                 rb_handle_timestamp(cpu_buffer, &info);
2882         } else /* Did the write stamp get updated already? */
2883                 if (likely(info.ts >= cpu_buffer->write_stamp)) {
2884                 info.delta = diff;
2885                 if (unlikely(test_time_stamp(info.delta)))
2886                         rb_handle_timestamp(cpu_buffer, &info);
2887         }
2888 
2889         event = __rb_reserve_next(cpu_buffer, &info);
2890 
2891         if (unlikely(PTR_ERR(event) == -EAGAIN)) {
2892                 if (info.add_timestamp)
2893                         info.length -= RB_LEN_TIME_EXTEND;
2894                 goto again;
2895         }
2896 
2897         if (!event)
2898                 goto out_fail;
2899 
2900         return event;
2901 
2902  out_fail:
2903         rb_end_commit(cpu_buffer);
2904         return NULL;
2905 }
2906 
2907 /**
2908  * ring_buffer_lock_reserve - reserve a part of the buffer
2909  * @buffer: the ring buffer to reserve from
2910  * @length: the length of the data to reserve (excluding event header)
2911  *
2912  * Returns a reserved event on the ring buffer to copy directly to.
2913  * The user of this interface will need to get the body to write into
2914  * and can use the ring_buffer_event_data() interface.
2915  *
2916  * The length is the length of the data needed, not the event length
2917  * which also includes the event header.
2918  *
2919  * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2920  * If NULL is returned, then nothing has been allocated or locked.
2921  */
2922 struct ring_buffer_event *
2923 ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length)
2924 {
2925         struct ring_buffer_per_cpu *cpu_buffer;
2926         struct ring_buffer_event *event;
2927         int cpu;
2928 
2929         /* If we are tracing schedule, we don't want to recurse */
2930         preempt_disable_notrace();
2931 
2932         if (unlikely(atomic_read(&buffer->record_disabled)))
2933                 goto out;
2934 
2935         cpu = raw_smp_processor_id();
2936 
2937         if (unlikely(!cpumask_test_cpu(cpu, buffer->cpumask)))
2938                 goto out;
2939 
2940         cpu_buffer = buffer->buffers[cpu];
2941 
2942         if (unlikely(atomic_read(&cpu_buffer->record_disabled)))
2943                 goto out;
2944 
2945         if (unlikely(length > BUF_MAX_DATA_SIZE))
2946                 goto out;
2947 
2948         if (unlikely(trace_recursive_lock(cpu_buffer)))
2949                 goto out;
2950 
2951         event = rb_reserve_next_event(buffer, cpu_buffer, length);
2952         if (!event)
2953                 goto out_unlock;
2954 
2955         return event;
2956 
2957  out_unlock:
2958         trace_recursive_unlock(cpu_buffer);
2959  out:
2960         preempt_enable_notrace();
2961         return NULL;
2962 }
2963 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
2964 
2965 /*
2966  * Decrement the entries to the page that an event is on.
2967  * The event does not even need to exist, only the pointer
2968  * to the page it is on. This may only be called before the commit
2969  * takes place.
2970  */
2971 static inline void
2972 rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
2973                    struct ring_buffer_event *event)
2974 {
2975         unsigned long addr = (unsigned long)event;
2976         struct buffer_page *bpage = cpu_buffer->commit_page;
2977         struct buffer_page *start;
2978 
2979         addr &= PAGE_MASK;
2980 
2981         /* Do the likely case first */
2982         if (likely(bpage->page == (void *)addr)) {
2983                 local_dec(&bpage->entries);
2984                 return;
2985         }
2986 
2987         /*
2988          * Because the commit page may be on the reader page we
2989          * start with the next page and check the end loop there.
2990          */
2991         rb_inc_page(cpu_buffer, &bpage);
2992         start = bpage;
2993         do {
2994                 if (bpage->page == (void *)addr) {
2995                         local_dec(&bpage->entries);
2996                         return;
2997                 }
2998                 rb_inc_page(cpu_buffer, &bpage);
2999         } while (bpage != start);
3000 
3001         /* commit not part of this buffer?? */
3002         RB_WARN_ON(cpu_buffer, 1);
3003 }
3004 
3005 /**
3006  * ring_buffer_commit_discard - discard an event that has not been committed
3007  * @buffer: the ring buffer
3008  * @event: non committed event to discard
3009  *
3010  * Sometimes an event that is in the ring buffer needs to be ignored.
3011  * This function lets the user discard an event in the ring buffer
3012  * and then that event will not be read later.
3013  *
3014  * This function only works if it is called before the item has been
3015  * committed. It will try to free the event from the ring buffer
3016  * if another event has not been added behind it.
3017  *
3018  * If another event has been added behind it, it will set the event
3019  * up as discarded, and perform the commit.
3020  *
3021  * If this function is called, do not call ring_buffer_unlock_commit on
3022  * the event.
3023  */
3024 void ring_buffer_discard_commit(struct ring_buffer *buffer,
3025                                 struct ring_buffer_event *event)
3026 {
3027         struct ring_buffer_per_cpu *cpu_buffer;
3028         int cpu;
3029 
3030         /* The event is discarded regardless */
3031         rb_event_discard(event);
3032 
3033         cpu = smp_processor_id();
3034         cpu_buffer = buffer->buffers[cpu];
3035 
3036         /*
3037          * This must only be called if the event has not been
3038          * committed yet. Thus we can assume that preemption
3039          * is still disabled.
3040          */
3041         RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
3042 
3043         rb_decrement_entry(cpu_buffer, event);
3044         if (rb_try_to_discard(cpu_buffer, event))
3045                 goto out;
3046 
3047         /*
3048          * The commit is still visible by the reader, so we
3049          * must still update the timestamp.
3050          */
3051         rb_update_write_stamp(cpu_buffer, event);
3052  out:
3053         rb_end_commit(cpu_buffer);
3054 
3055         trace_recursive_unlock(cpu_buffer);
3056 
3057         preempt_enable_notrace();
3058 
3059 }
3060 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
3061 
3062 /**
3063  * ring_buffer_write - write data to the buffer without reserving
3064  * @buffer: The ring buffer to write to.
3065  * @length: The length of the data being written (excluding the event header)
3066  * @data: The data to write to the buffer.
3067  *
3068  * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
3069  * one function. If you already have the data to write to the buffer, it
3070  * may be easier to simply call this function.
3071  *
3072  * Note, like ring_buffer_lock_reserve, the length is the length of the data
3073  * and not the length of the event which would hold the header.
3074  */
3075 int ring_buffer_write(struct ring_buffer *buffer,
3076                       unsigned long length,
3077                       void *data)
3078 {
3079         struct ring_buffer_per_cpu *cpu_buffer;
3080         struct ring_buffer_event *event;
3081         void *body;
3082         int ret = -EBUSY;
3083         int cpu;
3084 
3085         preempt_disable_notrace();
3086 
3087         if (atomic_read(&buffer->record_disabled))
3088                 goto out;
3089 
3090         cpu = raw_smp_processor_id();
3091 
3092         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3093                 goto out;
3094 
3095         cpu_buffer = buffer->buffers[cpu];
3096 
3097         if (atomic_read(&cpu_buffer->record_disabled))
3098                 goto out;
3099 
3100         if (length > BUF_MAX_DATA_SIZE)
3101                 goto out;
3102 
3103         if (unlikely(trace_recursive_lock(cpu_buffer)))
3104                 goto out;
3105 
3106         event = rb_reserve_next_event(buffer, cpu_buffer, length);
3107         if (!event)
3108                 goto out_unlock;
3109 
3110         body = rb_event_data(event);
3111 
3112         memcpy(body, data, length);
3113 
3114         rb_commit(cpu_buffer, event);
3115 
3116         rb_wakeups(buffer, cpu_buffer);
3117 
3118         ret = 0;
3119 
3120  out_unlock:
3121         trace_recursive_unlock(cpu_buffer);
3122 
3123  out:
3124         preempt_enable_notrace();
3125 
3126         return ret;
3127 }
3128 EXPORT_SYMBOL_GPL(ring_buffer_write);
3129 
3130 static bool rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
3131 {
3132         struct buffer_page *reader = cpu_buffer->reader_page;
3133         struct buffer_page *head = rb_set_head_page(cpu_buffer);
3134         struct buffer_page *commit = cpu_buffer->commit_page;
3135 
3136         /* In case of error, head will be NULL */
3137         if (unlikely(!head))
3138                 return true;
3139 
3140         return reader->read == rb_page_commit(reader) &&
3141                 (commit == reader ||
3142                  (commit == head &&
3143                   head->read == rb_page_commit(commit)));
3144 }
3145 
3146 /**
3147  * ring_buffer_record_disable - stop all writes into the buffer
3148  * @buffer: The ring buffer to stop writes to.
3149  *
3150  * This prevents all writes to the buffer. Any attempt to write
3151  * to the buffer after this will fail and return NULL.
3152  *
3153  * The caller should call synchronize_sched() after this.
3154  */
3155 void ring_buffer_record_disable(struct ring_buffer *buffer)
3156 {
3157         atomic_inc(&buffer->record_disabled);
3158 }
3159 EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
3160 
3161 /**
3162  * ring_buffer_record_enable - enable writes to the buffer
3163  * @buffer: The ring buffer to enable writes
3164  *
3165  * Note, multiple disables will need the same number of enables
3166  * to truly enable the writing (much like preempt_disable).
3167  */
3168 void ring_buffer_record_enable(struct ring_buffer *buffer)
3169 {
3170         atomic_dec(&buffer->record_disabled);
3171 }
3172 EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
3173 
3174 /**
3175  * ring_buffer_record_off - stop all writes into the buffer
3176  * @buffer: The ring buffer to stop writes to.
3177  *
3178  * This prevents all writes to the buffer. Any attempt to write
3179  * to the buffer after this will fail and return NULL.
3180  *
3181  * This is different than ring_buffer_record_disable() as
3182  * it works like an on/off switch, where as the disable() version
3183  * must be paired with a enable().
3184  */
3185 void ring_buffer_record_off(struct ring_buffer *buffer)
3186 {
3187         unsigned int rd;
3188         unsigned int new_rd;
3189 
3190         do {
3191                 rd = atomic_read(&buffer->record_disabled);
3192                 new_rd = rd | RB_BUFFER_OFF;
3193         } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3194 }
3195 EXPORT_SYMBOL_GPL(ring_buffer_record_off);
3196 
3197 /**
3198  * ring_buffer_record_on - restart writes into the buffer
3199  * @buffer: The ring buffer to start writes to.
3200  *
3201  * This enables all writes to the buffer that was disabled by
3202  * ring_buffer_record_off().
3203  *
3204  * This is different than ring_buffer_record_enable() as
3205  * it works like an on/off switch, where as the enable() version
3206  * must be paired with a disable().
3207  */
3208 void ring_buffer_record_on(struct ring_buffer *buffer)
3209 {
3210         unsigned int rd;
3211         unsigned int new_rd;
3212 
3213         do {
3214                 rd = atomic_read(&buffer->record_disabled);
3215                 new_rd = rd & ~RB_BUFFER_OFF;
3216         } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3217 }
3218 EXPORT_SYMBOL_GPL(ring_buffer_record_on);
3219 
3220 /**
3221  * ring_buffer_record_is_on - return true if the ring buffer can write
3222  * @buffer: The ring buffer to see if write is enabled
3223  *
3224  * Returns true if the ring buffer is in a state that it accepts writes.
3225  */
3226 int ring_buffer_record_is_on(struct ring_buffer *buffer)
3227 {
3228         return !atomic_read(&buffer->record_disabled);
3229 }
3230 
3231 /**
3232  * ring_buffer_record_is_set_on - return true if the ring buffer is set writable
3233  * @buffer: The ring buffer to see if write is set enabled
3234  *
3235  * Returns true if the ring buffer is set writable by ring_buffer_record_on().
3236  * Note that this does NOT mean it is in a writable state.
3237  *
3238  * It may return true when the ring buffer has been disabled by
3239  * ring_buffer_record_disable(), as that is a temporary disabling of
3240  * the ring buffer.
3241  */
3242 int ring_buffer_record_is_set_on(struct ring_buffer *buffer)
3243 {
3244         return !(atomic_read(&buffer->record_disabled) & RB_BUFFER_OFF);
3245 }
3246 
3247 /**
3248  * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
3249  * @buffer: The ring buffer to stop writes to.
3250  * @cpu: The CPU buffer to stop
3251  *
3252  * This prevents all writes to the buffer. Any attempt to write
3253  * to the buffer after this will fail and return NULL.
3254  *
3255  * The caller should call synchronize_sched() after this.
3256  */
3257 void ring_buffer_record_disable_cpu(struct ring_buffer *buffer, int cpu)
3258 {
3259         struct ring_buffer_per_cpu *cpu_buffer;
3260 
3261         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3262                 return;
3263 
3264         cpu_buffer = buffer->buffers[cpu];
3265         atomic_inc(&cpu_buffer->record_disabled);
3266 }
3267 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
3268 
3269 /**
3270  * ring_buffer_record_enable_cpu - enable writes to the buffer
3271  * @buffer: The ring buffer to enable writes
3272  * @cpu: The CPU to enable.
3273  *
3274  * Note, multiple disables will need the same number of enables
3275  * to truly enable the writing (much like preempt_disable).
3276  */
3277 void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu)
3278 {
3279         struct ring_buffer_per_cpu *cpu_buffer;
3280 
3281         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3282                 return;
3283 
3284         cpu_buffer = buffer->buffers[cpu];
3285         atomic_dec(&cpu_buffer->record_disabled);
3286 }
3287 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
3288 
3289 /*
3290  * The total entries in the ring buffer is the running counter
3291  * of entries entered into the ring buffer, minus the sum of
3292  * the entries read from the ring buffer and the number of
3293  * entries that were overwritten.
3294  */
3295 static inline unsigned long
3296 rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
3297 {
3298         return local_read(&cpu_buffer->entries) -
3299                 (local_read(&cpu_buffer->overrun) + cpu_buffer->read);
3300 }
3301 
3302 /**
3303  * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
3304  * @buffer: The ring buffer
3305  * @cpu: The per CPU buffer to read from.
3306  */
3307 u64 ring_buffer_oldest_event_ts(struct ring_buffer *buffer, int cpu)
3308 {
3309         unsigned long flags;
3310         struct ring_buffer_per_cpu *cpu_buffer;
3311         struct buffer_page *bpage;
3312         u64 ret = 0;
3313 
3314         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3315                 return 0;
3316 
3317         cpu_buffer = buffer->buffers[cpu];
3318         raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3319         /*
3320          * if the tail is on reader_page, oldest time stamp is on the reader
3321          * page
3322          */
3323         if (cpu_buffer->tail_page == cpu_buffer->reader_page)
3324                 bpage = cpu_buffer->reader_page;
3325         else
3326                 bpage = rb_set_head_page(cpu_buffer);
3327         if (bpage)
3328                 ret = bpage->page->time_stamp;
3329         raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3330 
3331         return ret;
3332 }
3333 EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
3334 
3335 /**
3336  * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
3337  * @buffer: The ring buffer
3338  * @cpu: The per CPU buffer to read from.
3339  */
3340 unsigned long ring_buffer_bytes_cpu(struct ring_buffer *buffer, int cpu)
3341 {
3342         struct ring_buffer_per_cpu *cpu_buffer;
3343         unsigned long ret;
3344 
3345         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3346                 return 0;
3347 
3348         cpu_buffer = buffer->buffers[cpu];
3349         ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
3350 
3351         return ret;
3352 }
3353 EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
3354 
3355 /**
3356  * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
3357  * @buffer: The ring buffer
3358  * @cpu: The per CPU buffer to get the entries from.
3359  */
3360 unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu)
3361 {
3362         struct ring_buffer_per_cpu *cpu_buffer;
3363 
3364         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3365                 return 0;
3366 
3367         cpu_buffer = buffer->buffers[cpu];
3368 
3369         return rb_num_of_entries(cpu_buffer);
3370 }
3371 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
3372 
3373 /**
3374  * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
3375  * buffer wrapping around (only if RB_FL_OVERWRITE is on).
3376  * @buffer: The ring buffer
3377  * @cpu: The per CPU buffer to get the number of overruns from
3378  */
3379 unsigned long ring_buffer_overrun_cpu(struct ring_buffer *buffer, int cpu)
3380 {
3381         struct ring_buffer_per_cpu *cpu_buffer;
3382         unsigned long ret;
3383 
3384         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3385                 return 0;
3386 
3387         cpu_buffer = buffer->buffers[cpu];
3388         ret = local_read(&cpu_buffer->overrun);
3389 
3390         return ret;
3391 }
3392 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
3393 
3394 /**
3395  * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
3396  * commits failing due to the buffer wrapping around while there are uncommitted
3397  * events, such as during an interrupt storm.
3398  * @buffer: The ring buffer
3399  * @cpu: The per CPU buffer to get the number of overruns from
3400  */
3401 unsigned long
3402 ring_buffer_commit_overrun_cpu(struct ring_buffer *buffer, int cpu)
3403 {
3404         struct ring_buffer_per_cpu *cpu_buffer;
3405         unsigned long ret;
3406 
3407         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3408                 return 0;
3409 
3410         cpu_buffer = buffer->buffers[cpu];
3411         ret = local_read(&cpu_buffer->commit_overrun);
3412 
3413         return ret;
3414 }
3415 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
3416 
3417 /**
3418  * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
3419  * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
3420  * @buffer: The ring buffer
3421  * @cpu: The per CPU buffer to get the number of overruns from
3422  */
3423 unsigned long
3424 ring_buffer_dropped_events_cpu(struct ring_buffer *buffer, int cpu)
3425 {
3426         struct ring_buffer_per_cpu *cpu_buffer;
3427         unsigned long ret;
3428 
3429         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3430                 return 0;
3431 
3432         cpu_buffer = buffer->buffers[cpu];
3433         ret = local_read(&cpu_buffer->dropped_events);
3434 
3435         return ret;
3436 }
3437 EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
3438 
3439 /**
3440  * ring_buffer_read_events_cpu - get the number of events successfully read
3441  * @buffer: The ring buffer
3442  * @cpu: The per CPU buffer to get the number of events read
3443  */
3444 unsigned long
3445 ring_buffer_read_events_cpu(struct ring_buffer *buffer, int cpu)
3446 {
3447         struct ring_buffer_per_cpu *cpu_buffer;
3448 
3449         if (!cpumask_test_cpu(cpu, buffer->cpumask))
3450                 return 0;
3451 
3452         cpu_buffer = buffer->buffers[cpu];
3453         return cpu_buffer->read;
3454 }
3455 EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu);
3456 
3457 /**
3458  * ring_buffer_entries - get the number of entries in a buffer
3459  * @buffer: The ring buffer
3460  *
3461  * Returns the total number of entries in the ring buffer
3462  * (all CPU entries)
3463  */
3464 unsigned long ring_buffer_entries(struct ring_buffer *buffer)
3465 {
3466         struct ring_buffer_per_cpu *cpu_buffer;
3467         unsigned long entries = 0;
3468         int cpu;
3469 
3470         /* if you care about this being correct, lock the buffer */
3471         for_each_buffer_cpu(buffer, cpu) {
3472                 cpu_buffer = buffer->buffers[cpu];
3473                 entries += rb_num_of_entries(cpu_buffer);
3474         }
3475 
3476         return entries;
3477 }
3478 EXPORT_SYMBOL_GPL(ring_buffer_entries);
3479 
3480 /**
3481  * ring_buffer_overruns - get the number of overruns in buffer
3482  * @buffer: The ring buffer
3483  *
3484  * Returns the total number of overruns in the ring buffer
3485  * (all CPU entries)
3486  */
3487 unsigned long ring_buffer_overruns(struct ring_buffer *buffer)
3488 {
3489         struct ring_buffer_per_cpu *cpu_buffer;
3490         unsigned long overruns = 0;
3491         int cpu;
3492 
3493         /* if you care about this being correct, lock the buffer */
3494         for_each_buffer_cpu(buffer, cpu) {
3495                 cpu_buffer = buffer->buffers[cpu];
3496                 overruns += local_read(&cpu_buffer->overrun);
3497         }
3498 
3499         return overruns;
3500 }
3501 EXPORT_SYMBOL_GPL(ring_buffer_overruns);
3502 
3503 static void rb_iter_reset(struct ring_buffer_iter *iter)
3504 {
3505         struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3506 
3507         /* Iterator usage is expected to have record disabled */
3508         iter->head_page = cpu_buffer->reader_page;
3509         iter->head = cpu_buffer->reader_page->read;
3510 
3511         iter->cache_reader_page = iter->head_page;
3512         iter->cache_read = cpu_buffer->read;
3513 
3514         if (iter->head)
3515                 iter->read_stamp = cpu_buffer->read_stamp;
3516         else
3517                 iter->read_stamp = iter->head_page->page->time_stamp;
3518 }
3519 
3520 /**
3521  * ring_buffer_iter_reset - reset an iterator
3522  * @iter: The iterator to reset
3523  *
3524  * Resets the iterator, so that it will start from the beginning
3525  * again.
3526  */
3527 void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
3528 {
3529         struct ring_buffer_per_cpu *cpu_buffer;
3530         unsigned long flags;
3531 
3532         if (!iter)
3533                 return;
3534 
3535         cpu_buffer = iter->cpu_buffer;
3536 
3537         raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3538         rb_iter_reset(iter);
3539         raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3540 }
3541 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
3542 
3543 /**
3544  * ring_buffer_iter_empty - check if an iterator has no more to read
3545  * @iter: The iterator to check
3546  */
3547 int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
3548 {
3549         struct ring_buffer_per_cpu *cpu_buffer;
3550         struct buffer_page *reader;
3551         struct buffer_page *head_page;
3552         struct buffer_page *commit_page;
3553         unsigned commit;
3554 
3555         cpu_buffer = iter->cpu_buffer;
3556 
3557         /* Remember, trace recording is off when iterator is in use */
3558         reader = cpu_buffer->reader_page;
3559         head_page = cpu_buffer->head_page;
3560         commit_page = cpu_buffer->commit_page;
3561         commit = rb_page_commit(commit_page);
3562 
3563         return ((iter->head_page == commit_page && iter->head == commit) ||
3564                 (iter->head_page == reader && commit_page == head_page &&
3565                  head_page->read == commit &&
3566                  iter->head == rb_page_commit(cpu_buffer->reader_page)));
3567 }
3568 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
3569 
3570 static void
3571 rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
3572                      struct ring_buffer_event *event)
3573 {
3574         u64 delta;
3575 
3576         switch (event->type_len) {
3577         case RINGBUF_TYPE_PADDING:
3578                 return;
3579 
3580         case RINGBUF_TYPE_TIME_EXTEND:
3581                 delta = ring_buffer_event_time_stamp(event);
3582                 cpu_buffer->read_stamp += delta;
3583                 return;
3584 
3585         case RINGBUF_TYPE_TIME_STAMP:
3586                 delta = ring_buffer_event_time_stamp(event);
3587                 cpu_buffer->read_stamp = delta;
3588                 return;
3589 
3590         case RINGBUF_TYPE_DATA:
3591                 cpu_buffer->read_stamp += event->time_delta;
3592                 return;
3593 
3594         default:
3595                 BUG();
3596         }
3597         return;
3598 }
3599 
3600 static void
3601 rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
3602                           struct ring_buffer_event *event)
3603 {
3604         u64 delta;
3605 
3606         switch (event->type_len) {
3607         case RINGBUF_TYPE_PADDING:
3608                 return;
3609 
3610         case RINGBUF_TYPE_TIME_EXTEND:
3611                 delta = ring_buffer_event_time_stamp(event);
3612                 iter->read_stamp += delta;
3613                 return;
3614 
3615         case RINGBUF_TYPE_TIME_STAMP:
3616                 delta = ring_buffer_event_time_stamp(event);
3617                 iter->read_stamp = delta;
3618                 return;
3619 
3620         case RINGBUF_TYPE_DATA:
3621                 iter->read_stamp += event->time_delta;
3622                 return;
3623 
3624         default:
3625                 BUG();
3626         }
3627         return;
3628 }
3629 
3630 static struct buffer_page *
3631 rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
3632 {
3633         struct buffer_page *reader = NULL;
3634         unsigned long overwrite;
3635         unsigned long flags;
3636         int nr_loops = 0;
3637         int ret;
3638 
3639         local_irq_save(flags);
3640         arch_spin_lock(&cpu_buffer->lock);
3641 
3642  again:
3643         /*
3644          * This should normally only loop twice. But because the
3645          * start of the reader inserts an empty page, it causes
3646          * a case where we will loop three times. There should be no
3647          * reason to loop four times (that I know of).
3648          */
3649         if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
3650                 reader = NULL;
3651                 goto out;
3652         }
3653 
3654         reader = cpu_buffer->reader_page;
3655 
3656         /* If there's more to read, return this page */
3657         if (cpu_buffer->reader_page->read < rb_page_size(reader))
3658                 goto out;
3659 
3660         /* Never should we have an index greater than the size */
3661         if (RB_WARN_ON(cpu_buffer,
3662                        cpu_buffer->reader_page->read > rb_page_size(reader)))
3663                 goto out;
3664 
3665         /* check if we caught up to the tail */
3666         reader = NULL;
3667         if (cpu_buffer->commit_page == cpu_buffer->reader_page)
3668                 goto out;
3669 
3670         /* Don't bother swapping if the ring buffer is empty */
3671         if (rb_num_of_entries(cpu_buffer) == 0)
3672                 goto out;
3673 
3674         /*
3675          * Reset the reader page to size zero.
3676          */
3677         local_set(&cpu_buffer->reader_page->write, 0);
3678         local_set(&cpu_buffer->reader_page->entries, 0);
3679         local_set(&cpu_buffer->reader_page->page->commit, 0);
3680         cpu_buffer->reader_page->real_end = 0;
3681 
3682  spin:
3683         /*
3684          * Splice the empty reader page into the list around the head.
3685          */
3686         reader = rb_set_head_page(cpu_buffer);
3687         if (!reader)
3688                 goto out;
3689         cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
3690         cpu_buffer->reader_page->list.prev = reader->list.prev;
3691 
3692         /*
3693          * cpu_buffer->pages just needs to point to the buffer, it
3694          *  has no specific buffer page to point to. Lets move it out
3695          *  of our way so we don't accidentally swap it.
3696          */
3697         cpu_buffer->pages = reader->list.prev;
3698 
3699         /* The reader page will be pointing to the new head */
3700         rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list);
3701 
3702         /*
3703          * We want to make sure we read the overruns after we set up our
3704          * pointers to the next object. The writer side does a
3705          * cmpxchg to cross pages which acts as the mb on the writer
3706          * side. Note, the reader will constantly fail the swap
3707          * while the writer is updating the pointers, so this
3708          * guarantees that the overwrite recorded here is the one we
3709          * want to compare with the last_overrun.
3710          */
3711         smp_mb();
3712         overwrite = local_read(&(cpu_buffer->overrun));
3713 
3714         /*
3715          * Here's the tricky part.
3716          *
3717          * We need to move the pointer past the header page.
3718          * But we can only do that if a writer is not currently
3719          * moving it. The page before the header page has the
3720          * flag bit '1' set if it is pointing to the page we want.
3721          * but if the writer is in the process of moving it
3722          * than it will be '2' or already moved ''.
3723          */
3724 
3725         ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
3726 
3727         /*
3728          * If we did not convert it, then we must try again.
3729          */
3730         if (!ret)
3731                 goto spin;
3732 
3733         /*
3734          * Yeah! We succeeded in replacing the page.
3735          *
3736          * Now make the new head point back to the reader page.
3737          */
3738         rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
3739         rb_inc_page(cpu_buffer, &cpu_buffer->head_page);
3740 
3741         /* Finally update the reader page to the new head */
3742         cpu_buffer->reader_page = reader;
3743         cpu_buffer->reader_page->read = 0;
3744 
3745         if (overwrite != cpu_buffer->last_overrun) {
3746                 cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
3747                 cpu_buffer->last_overrun = overwrite;
3748         }
3749 
3750         goto again;
3751 
3752  out:
3753         /* Update the read_stamp on the first event */
3754         if (reader && reader->read == 0)
3755                 cpu_buffer->read_stamp = reader->page->time_stamp;
3756 
3757         arch_spin_unlock(&cpu_buffer->lock);
3758         local_irq_restore(flags);
3759 
3760         return reader;
3761 }
3762 
3763 static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
3764 {
3765         struct ring_buffer_event *event;
3766         struct buffer_page *reader;
3767         unsigned length;
3768 
3769         reader = rb_get_reader_page(cpu_buffer);
3770 
3771         /* This function should not be called when buffer is empty */
3772         if (RB_WARN_ON(cpu_buffer, !reader))
3773                 return;
3774 
3775         event = rb_reader_event(cpu_buffer);
3776 
3777         if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
3778                 cpu_buffer->read++;
3779 
3780         rb_update_read_stamp(cpu_buffer, event);
3781 
3782         length = rb_event_length(event);
3783         cpu_buffer->reader_page->read += length;
3784 }
3785 
3786 static void rb_advance_iter(struct ring_buffer_iter *iter)
3787 {
3788         struct ring_buffer_per_cpu *cpu_buffer;
3789         struct ring_buffer_event *event;
3790         unsigned length;
3791 
3792         cpu_buffer = iter->cpu_buffer;
3793 
3794         /*
3795          * Check if we are at the end of the buffer.
3796          */
3797         if (iter->head >= rb_page_size(iter->head_page)) {
3798                 /* discarded commits can make the page empty */
3799                 if (iter->head_page == cpu_buffer->commit_page)
3800                         return;
3801                 rb_inc_iter(iter);
3802                 return;
3803         }
3804 
3805         event = rb_iter_head_event(iter);
3806 
3807         length = rb_event_length(event);
3808 
3809         /*
3810          * This should not be called to advance the header if we are
3811          * at the tail of the buffer.
3812          */
3813         if (RB_WARN_ON(cpu_buffer,
3814                        (iter->head_page == cpu_buffer->commit_page) &&
3815                        (iter->head + length > rb_commit_index(cpu_buffer))))
3816                 return;
3817 
3818         rb_update_iter_read_stamp(iter, event);
3819 
3820         iter->head += length;
3821 
3822         /* check for end of page padding */
3823         if ((iter->head >= rb_page_size(iter->head_page)) &&
3824             (iter->head_page != cpu_buffer->commit_page))
3825                 rb_inc_iter(iter);
3826 }
3827 
3828 static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
3829 {
3830         return cpu_buffer->lost_events;
3831 }
3832 
3833 static struct ring_buffer_event *
3834 rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
3835                unsigned long *lost_events)
3836 {
3837         struct ring_buffer_event *event;
3838         struct buffer_page *reader;
3839         int nr_loops = 0;
3840 
3841         if (ts)
3842                 *ts = 0;
3843  again:
3844         /*
3845          * We repeat when a time extend is encountered.
3846          * Since the time extend is always attached to a data event,
3847          * we should never loop more than once.
3848          * (We never hit the following condition more than twice).
3849          */
3850         if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3851                 return NULL;
3852 
3853         reader = rb_get_reader_page(cpu_buffer);
3854         if (!reader)
3855                 return NULL;
3856 
3857         event = rb_reader_event(cpu_buffer);
3858 
3859         switch (event->type_len) {
3860         case RINGBUF_TYPE_PADDING:
3861                 if (rb_null_event(event))
3862                         RB_WARN_ON(cpu_buffer, 1);
3863                 /*
3864                  * Because the writer could be discarding every
3865                  * event it creates (which would probably be bad)
3866                  * if we were to go back to "again" then we may never
3867                  * catch up, and will trigger the warn on, or lock
3868                  * the box. Return the padding, and we will release
3869                  * the current locks, and try again.
3870                  */
3871                 return event;
3872 
3873         case RINGBUF_TYPE_TIME_EXTEND:
3874                 /* Internal data, OK to advance */
3875                 rb_advance_reader(cpu_buffer);
3876                 goto again;
3877 
3878         case RINGBUF_TYPE_TIME_STAMP:
3879                 if (ts) {
3880                         *ts = ring_buffer_event_time_stamp(event);
3881                         ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3882                                                          cpu_buffer->cpu, ts);
3883                 }
3884                 /* Internal data, OK to advance */
3885                 rb_advance_reader(cpu_buffer);
3886                 goto again;
3887 
3888         case RINGBUF_TYPE_DATA:
3889                 if (ts && !(*ts)) {
3890                         *ts = cpu_buffer->read_stamp + event->time_delta;
3891                         ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3892                                                          cpu_buffer->cpu, ts);
3893                 }
3894                 if (lost_events)
3895                         *lost_events = rb_lost_events(cpu_buffer);
3896                 return event;
3897 
3898         default:
3899                 BUG();
3900         }
3901 
3902         return NULL;
3903 }
3904 EXPORT_SYMBOL_GPL(ring_buffer_peek);
3905 
3906 static struct ring_buffer_event *
3907 rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3908 {
3909         struct ring_buffer *buffer;
3910         struct ring_buffer_per_cpu *cpu_buffer;
3911         struct ring_buffer_event *event;
3912         int nr_loops = 0;
3913 
3914         if (ts)
3915                 *ts = 0;
3916 
3917         cpu_buffer = iter->cpu_buffer;
3918         buffer = cpu_buffer->buffer;
3919 
3920         /*
3921          * Check if someone performed a consuming read to
3922          * the buffer. A consuming read invalidates the iterator
3923          * and we need to reset the iterator in this case.
3924          */
3925         if (unlikely(iter->cache_read != cpu_buffer->read ||
3926                      iter->cache_reader_page != cpu_buffer->reader_page))
3927                 rb_iter_reset(iter);
3928 
3929  again:
3930         if (ring_buffer_iter_empty(iter))
3931                 return NULL;
3932 
3933         /*
3934          * We repeat when a time extend is encountered or we hit
3935          * the end of the page. Since the time extend is always attached
3936          * to a data event, we should never loop more than three times.
3937          * Once for going to next page, once on time extend, and
3938          * finally once to get the event.
3939          * (We never hit the following condition more than thrice).
3940          */
3941         if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3))
3942                 return NULL;
3943 
3944         if (rb_per_cpu_empty(cpu_buffer))
3945                 return NULL;
3946 
3947         if (iter->head >= rb_page_size(iter->head_page)) {
3948                 rb_inc_iter(iter);
3949                 goto again;
3950         }
3951 
3952         event = rb_iter_head_event(iter);
3953 
3954         switch (event->type_len) {
3955         case RINGBUF_TYPE_PADDING:
3956                 if (rb_null_event(event)) {
3957                         rb_inc_iter(iter);
3958                         goto again;
3959                 }
3960                 rb_advance_iter(iter);
3961                 return event;
3962 
3963         case RINGBUF_TYPE_TIME_EXTEND:
3964                 /* Internal data, OK to advance */
3965                 rb_advance_iter(iter);
3966                 goto again;
3967 
3968         case RINGBUF_TYPE_TIME_STAMP:
3969                 if (ts) {
3970                         *ts = ring_buffer_event_time_stamp(event);
3971                         ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3972                                                          cpu_buffer->cpu, ts);
3973                 }
3974                 /* Internal data, OK to advance */
3975                 rb_advance_iter(iter);
3976                 goto again;
3977 
3978         case RINGBUF_TYPE_DATA:
3979                 if (ts && !(*ts)) {
3980                         *ts = iter->read_stamp + event->time_delta;
3981                         ring_buffer_normalize_time_stamp(buffer,
3982                                                          cpu_buffer->cpu, ts);
3983                 }
3984                 return event;
3985 
3986         default:
3987                 BUG();
3988         }
3989 
3990         return NULL;
3991 }
3992 EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
3993 
3994 static inline bool rb_reader_lock(struct ring_buffer_per_cpu *cpu_buffer)
3995 {
3996         if (likely(!in_nmi())) {
3997                 raw_spin_lock(&cpu_buffer->reader_lock);
3998                 return true;
3999         }
4000 
4001         /*
4002          * If an NMI die dumps out the content of the ring buffer
4003          * trylock must be used to prevent a deadlock if the NMI
4004          * preempted a task that holds the ring buffer locks. If
4005          * we get the lock then all is fine, if not, then continue
4006          * to do the read, but this can corrupt the ring buffer,
4007          * so it must be permanently disabled from future writes.
4008          * Reading from NMI is a oneshot deal.
4009          */
4010         if (raw_spin_trylock(&cpu_buffer->reader_lock))
4011                 return true;
4012 
4013         /* Continue without locking, but disable the ring buffer */
4014         atomic_inc(&cpu_buffer->record_disabled);
4015         return false;
4016 }
4017 
4018 static inline void
4019 rb_reader_unlock(struct ring_buffer_per_cpu *cpu_buffer, bool locked)
4020 {
4021         if (likely(locked))
4022                 raw_spin_unlock(&cpu_buffer->reader_lock);
4023         return;
4024 }
4025 
4026 /**
4027  * ring_buffer_peek - peek at the next event to be read
4028  * @buffer: The ring buffer to read
4029  * @cpu: The cpu to peak at
4030  * @ts: The timestamp counter of this event.
4031  * @lost_events: a variable to store if events were lost (may be NULL)
4032  *
4033  * This will return the event that will be read next, but does
4034  * not consume the data.
4035  */
4036 struct ring_buffer_event *
4037 ring_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts,
4038                  unsigned long *lost_events)
4039 {
4040         struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4041         struct ring_buffer_event *event;
4042         unsigned long flags;
4043         bool dolock;
4044 
4045         if (!cpumask_test_cpu(cpu, buffer->cpumask))
4046                 return NULL;
4047 
4048  again:
4049         local_irq_save(flags);
4050         dolock = rb_reader_lock(cpu_buffer);
4051         event = rb_buffer_peek(cpu_buffer, ts, lost_events);
4052         if (event && event->type_len == RINGBUF_TYPE_PADDING)
4053                 rb_advance_reader(cpu_buffer);
4054         rb_reader_unlock(cpu_buffer, dolock);
4055         local_irq_restore(flags);
4056 
4057         if (event && event->type_len == RINGBUF_TYPE_PADDING)
4058                 goto again;
4059 
4060         return event;
4061 }
4062 
4063 /**
4064  * ring_buffer_iter_peek - peek at the next event to be read
4065  * @iter: The ring buffer iterator
4066  * @ts: The timestamp counter of this event.
4067  *
4068  * This will return the event that will be read next, but does
4069  * not increment the iterator.
4070  */
4071 struct ring_buffer_event *
4072 ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
4073 {
4074         struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4075         struct ring_buffer_event *event;
4076         unsigned long flags;
4077 
4078  again:
4079         raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4080         event = rb_iter_peek(iter, ts);
4081         raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4082 
4083         if (event && event->type_len == RINGBUF_TYPE_PADDING)
4084                 goto again;
4085 
4086         return event;
4087 }
4088 
4089 /**
4090  * ring_buffer_consume - return an event and consume it
4091  * @buffer: The ring buffer to get the next event from
4092  * @cpu: the cpu to read the buffer from
4093  * @ts: a variable to store the timestamp (may be NULL)
4094  * @lost_events: a variable to store if events were lost (may be NULL)
4095  *
4096  * Returns the next event in the ring buffer, and that event is consumed.
4097  * Meaning, that sequential reads will keep returning a different event,
4098  * and eventually empty the ring buffer if the producer is slower.
4099  */
4100 struct ring_buffer_event *
4101 ring_buffer_consume(struct ring_buffer *buffer, int cpu, u64 *ts,
4102                     unsigned long *lost_events)
4103 {
4104         struct ring_buffer_per_cpu *cpu_buffer;
4105         struct ring_buffer_event *event = NULL;
4106         unsigned long flags;
4107         bool dolock;
4108 
4109  again:
4110         /* might be called in atomic */
4111         preempt_disable();
4112 
4113         if (!cpumask_test_cpu(cpu, buffer->cpumask))
4114                 goto out;
4115 
4116         cpu_buffer = buffer->buffers[cpu];
4117         local_irq_save(flags);
4118         dolock = rb_reader_lock(cpu_buffer);
4119 
4120         event = rb_buffer_peek(cpu_buffer, ts, lost_events);
4121         if (event) {
4122                 cpu_buffer->lost_events = 0;
4123                 rb_advance_reader(cpu_buffer);
4124         }
4125 
4126         rb_reader_unlock(cpu_buffer, dolock);
4127         local_irq_restore(flags);
4128 
4129  out:
4130         preempt_enable();
4131 
4132         if (event && event->type_len == RINGBUF_TYPE_PADDING)
4133                 goto again;
4134 
4135         return event;
4136 }
4137 EXPORT_SYMBOL_GPL(ring_buffer_consume);
4138 
4139 /**
4140  * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
4141  * @buffer: The ring buffer to read from
4142  * @cpu: The cpu buffer to iterate over
4143  *
4144  * This performs the initial preparations necessary to iterate
4145  * through the buffer.  Memory is allocated, buffer recording
4146  * is disabled, and the iterator pointer is returned to the caller.
4147  *
4148  * Disabling buffer recording prevents the reading from being
4149  * corrupted. This is not a consuming read, so a producer is not
4150  * expected.
4151  *
4152  * After a sequence of ring_buffer_read_prepare calls, the user is
4153  * expected to make at least one call to ring_buffer_read_prepare_sync.
4154  * Afterwards, ring_buffer_read_start is invoked to get things going
4155  * for real.
4156  *
4157  * This overall must be paired with ring_buffer_read_finish.
4158  */
4159 struct ring_buffer_iter *
4160 ring_buffer_read_prepare(struct ring_buffer *buffer, int cpu)
4161 {
4162         struct ring_buffer_per_cpu *cpu_buffer;
4163         struct ring_buffer_iter *iter;
4164 
4165         if (!cpumask_test_cpu(cpu, buffer->cpumask))
4166                 return NULL;
4167 
4168         iter = kmalloc(sizeof(*iter), GFP_KERNEL);
4169         if (!iter)
4170                 return NULL;
4171 
4172         cpu_buffer = buffer->buffers[cpu];
4173 
4174         iter->cpu_buffer = cpu_buffer;
4175 
4176         atomic_inc(&buffer->resize_disabled);
4177         atomic_inc(&cpu_buffer->record_disabled);
4178 
4179         return iter;
4180 }
4181 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
4182 
4183 /**
4184  * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
4185  *
4186  * All previously invoked ring_buffer_read_prepare calls to prepare
4187  * iterators will be synchronized.  Afterwards, read_buffer_read_start
4188  * calls on those iterators are allowed.
4189  */
4190 void
4191 ring_buffer_read_prepare_sync(void)
4192 {
4193         synchronize_sched();
4194 }
4195 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
4196 
4197 /**
4198  * ring_buffer_read_start - start a non consuming read of the buffer
4199  * @iter: The iterator returned by ring_buffer_read_prepare
4200  *
4201  * This finalizes the startup of an iteration through the buffer.
4202  * The iterator comes from a call to ring_buffer_read_prepare and
4203  * an intervening ring_buffer_read_prepare_sync must have been
4204  * performed.
4205  *
4206  * Must be paired with ring_buffer_read_finish.
4207  */
4208 void
4209 ring_buffer_read_start(struct ring_buffer_iter *iter)
4210 {
4211         struct ring_buffer_per_cpu *cpu_buffer;
4212         unsigned long flags;
4213 
4214         if (!iter)
4215                 return;
4216 
4217         cpu_buffer = iter->cpu_buffer;
4218 
4219         raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4220         arch_spin_lock(&cpu_buffer->lock);
4221         rb_iter_reset(iter);
4222         arch_spin_unlock(&cpu_buffer->lock);
4223         raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4224 }
4225 EXPORT_SYMBOL_GPL(ring_buffer_read_start);
4226 
4227 /**
4228  * ring_buffer_read_finish - finish reading the iterator of the buffer
4229  * @iter: The iterator retrieved by ring_buffer_start
4230  *
4231  * This re-enables the recording to the buffer, and frees the
4232  * iterator.
4233  */
4234 void
4235 ring_buffer_read_finish(struct ring_buffer_iter *iter)
4236 {
4237         struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4238         unsigned long flags;
4239 
4240         /*
4241          * Ring buffer is disabled from recording, here's a good place
4242          * to check the integrity of the ring buffer.
4243          * Must prevent readers from trying to read, as the check
4244          * clears the HEAD page and readers require it.
4245          */
4246         raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4247         rb_check_pages(cpu_buffer);
4248         raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4249 
4250         atomic_dec(&cpu_buffer->record_disabled);
4251         atomic_dec(&cpu_buffer->buffer->resize_disabled);
4252         kfree(iter);
4253 }
4254 EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
4255 
4256 /**
4257  * ring_buffer_read - read the next item in the ring buffer by the iterator
4258  * @iter: The ring buffer iterator
4259  * @ts: The time stamp of the event read.
4260  *
4261  * This reads the next event in the ring buffer and increments the iterator.
4262  */
4263 struct ring_buffer_event *
4264 ring_buffer_read(struct ring_buffer_iter *iter, u64 *ts)
4265 {
4266         struct ring_buffer_event *event;
4267         struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4268         unsigned long flags;
4269 
4270         raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4271  again:
4272         event = rb_iter_peek(iter, ts);
4273         if (!event)
4274                 goto out;
4275 
4276         if (event->type_len == RINGBUF_TYPE_PADDING)
4277                 goto again;
4278 
4279         rb_advance_iter(iter);
4280  out:
4281         raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4282 
4283         return event;
4284 }
4285 EXPORT_SYMBOL_GPL(ring_buffer_read);
4286 
4287 /**
4288  * ring_buffer_size - return the size of the ring buffer (in bytes)
4289  * @buffer: The ring buffer.
4290  */
4291 unsigned long ring_buffer_size(struct ring_buffer *buffer, int cpu)
4292 {
4293         /*
4294          * Earlier, this method returned
4295          *      BUF_PAGE_SIZE * buffer->nr_pages
4296          * Since the nr_pages field is now removed, we have converted this to
4297          * return the per cpu buffer value.
4298          */
4299         if (!cpumask_test_cpu(cpu, buffer->cpumask))
4300                 return 0;
4301 
4302         return BUF_PAGE_SIZE * buffer->buffers[cpu]->nr_pages;
4303 }
4304 EXPORT_SYMBOL_GPL(ring_buffer_size);
4305 
4306 static void
4307 rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
4308 {
4309         rb_head_page_deactivate(cpu_buffer);
4310 
4311         cpu_buffer->head_page
4312                 = list_entry(cpu_buffer->pages, struct buffer_page, list);
4313         local_set(&cpu_buffer->head_page->write, 0);
4314         local_set(&cpu_buffer->head_page->entries, 0);
4315         local_set(&cpu_buffer->head_page->page->commit, 0);
4316 
4317         cpu_buffer->head_page->read = 0;
4318 
4319         cpu_buffer->tail_page = cpu_buffer->head_page;
4320         cpu_buffer->commit_page = cpu_buffer->head_page;
4321 
4322         INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
4323         INIT_LIST_HEAD(&cpu_buffer->new_pages);
4324         local_set(&cpu_buffer->reader_page->write, 0);
4325         local_set(&cpu_buffer->reader_page->entries, 0);
4326         local_set(&cpu_buffer->reader_page->page->commit, 0);
4327         cpu_buffer->reader_page->read = 0;
4328 
4329         local_set(&cpu_buffer->entries_bytes, 0);
4330         local_set(&cpu_buffer->overrun, 0);
4331         local_set(&cpu_buffer->commit_overrun, 0);
4332         local_set(&cpu_buffer->dropped_events, 0);
4333         local_set(&cpu_buffer->entries, 0);
4334         local_set(&cpu_buffer->committing, 0);
4335         local_set(&cpu_buffer->commits, 0);
4336         cpu_buffer->read = 0;
4337         cpu_buffer->read_bytes = 0;
4338 
4339         cpu_buffer->write_stamp = 0;
4340         cpu_buffer->read_stamp = 0;
4341 
4342         cpu_buffer->lost_events = 0;
4343         cpu_buffer->last_overrun = 0;
4344 
4345         rb_head_page_activate(cpu_buffer);
4346 }
4347 
4348 /**
4349  * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
4350  * @buffer: The ring buffer to reset a per cpu buffer of
4351  * @cpu: The CPU buffer to be reset
4352  */
4353 void ring_buffer_reset_cpu(struct ring_buffer *buffer, int cpu)
4354 {
4355         struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4356         unsigned long flags;
4357 
4358         if (!cpumask_test_cpu(cpu, buffer->cpumask))
4359                 return;
4360 
4361         atomic_inc(&buffer->resize_disabled);
4362         atomic_inc(&cpu_buffer->record_disabled);
4363 
4364         /* Make sure all commits have finished */
4365         synchronize_sched();
4366 
4367         raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4368 
4369         if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
4370                 goto out;
4371 
4372         arch_spin_lock(&cpu_buffer->lock);
4373 
4374         rb_reset_cpu(cpu_buffer);
4375 
4376         arch_spin_unlock(&cpu_buffer->lock);
4377 
4378  out:
4379         raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4380 
4381         atomic_dec(&cpu_buffer->record_disabled);
4382         atomic_dec(&buffer->resize_disabled);
4383 }
4384 EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
4385 
4386 /**
4387  * ring_buffer_reset - reset a ring buffer
4388  * @buffer: The ring buffer to reset all cpu buffers
4389  */
4390 void ring_buffer_reset(struct ring_buffer *buffer)
4391 {
4392         int cpu;
4393 
4394         for_each_buffer_cpu(buffer, cpu)
4395                 ring_buffer_reset_cpu(buffer, cpu);
4396 }
4397 EXPORT_SYMBOL_GPL(ring_buffer_reset);
4398 
4399 /**
4400  * rind_buffer_empty - is the ring buffer empty?
4401  * @buffer: The ring buffer to test
4402  */
4403 bool ring_buffer_empty(struct ring_buffer *buffer)
4404 {
4405         struct ring_buffer_per_cpu *cpu_buffer;
4406         unsigned long flags;
4407         bool dolock;
4408         int cpu;
4409         int ret;
4410 
4411         /* yes this is racy, but if you don't like the race, lock the buffer */
4412         for_each_buffer_cpu(buffer, cpu) {
4413                 cpu_buffer = buffer->buffers[cpu];
4414                 local_irq_save(flags);
4415                 dolock = rb_reader_lock(cpu_buffer);
4416                 ret = rb_per_cpu_empty(cpu_buffer);
4417                 rb_reader_unlock(cpu_buffer, dolock);
4418                 local_irq_restore(flags);
4419 
4420                 if (!ret)
4421                         return false;
4422         }
4423 
4424         return true;
4425 }
4426 EXPORT_SYMBOL_GPL(ring_buffer_empty);
4427 
4428 /**
4429  * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
4430  * @buffer: The ring buffer
4431  * @cpu: The CPU buffer to test
4432  */
4433 bool ring_buffer_empty_cpu(struct ring_buffer *buffer, int cpu)
4434 {
4435         struct ring_buffer_per_cpu *cpu_buffer;
4436         unsigned long flags;
4437         bool dolock;
4438         int ret;
4439 
4440         if (!cpumask_test_cpu(cpu, buffer->cpumask))
4441                 return true;
4442 
4443         cpu_buffer = buffer->buffers[cpu];
4444         local_irq_save(flags);
4445         dolock = rb_reader_lock(cpu_buffer);
4446         ret = rb_per_cpu_empty(cpu_buffer);
4447         rb_reader_unlock(cpu_buffer, dolock);
4448         local_irq_restore(flags);
4449 
4450         return ret;
4451 }
4452 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
4453 
4454 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
4455 /**
4456  * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
4457  * @buffer_a: One buffer to swap with
4458  * @buffer_b: The other buffer to swap with
4459  *
4460  * This function is useful for tracers that want to take a "snapshot"
4461  * of a CPU buffer and has another back up buffer lying around.
4462  * it is expected that the tracer handles the cpu buffer not being
4463  * used at the moment.
4464  */
4465 int ring_buffer_swap_cpu(struct ring_buffer *buffer_a,
4466                          struct ring_buffer *buffer_b, int cpu)
4467 {
4468         struct ring_buffer_per_cpu *cpu_buffer_a;
4469         struct ring_buffer_per_cpu *cpu_buffer_b;
4470         int ret = -EINVAL;
4471 
4472         if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
4473             !cpumask_test_cpu(cpu, buffer_b->cpumask))
4474                 goto out;
4475 
4476         cpu_buffer_a = buffer_a->buffers[cpu];
4477         cpu_buffer_b = buffer_b->buffers[cpu];
4478 
4479         /* At least make sure the two buffers are somewhat the same */
4480         if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
4481                 goto out;
4482 
4483         ret = -EAGAIN;
4484 
4485         if (atomic_read(&buffer_a->record_disabled))
4486                 goto out;
4487 
4488         if (atomic_read(&buffer_b->record_disabled))
4489                 goto out;
4490 
4491         if (atomic_read(&cpu_buffer_a->record_disabled))
4492                 goto out;
4493 
4494         if (atomic_read(&cpu_buffer_b->record_disabled))
4495                 goto out;
4496 
4497         /*
4498          * We can't do a synchronize_sched here because this
4499          * function can be called in atomic context.
4500          * Normally this will be called from the same CPU as cpu.
4501          * If not it's up to the caller to protect this.
4502          */
4503         atomic_inc(&cpu_buffer_a->record_disabled);
4504         atomic_inc(&cpu_buffer_b->record_disabled);
4505 
4506         ret = -EBUSY;
4507         if (local_read(&cpu_buffer_a->committing))
4508                 goto out_dec;
4509         if (local_read(&cpu_buffer_b->committing))
4510                 goto out_dec;
4511 
4512         buffer_a->buffers[cpu] = cpu_buffer_b;
4513         buffer_b->buffers[cpu] = cpu_buffer_a;
4514 
4515         cpu_buffer_b->buffer = buffer_a;
4516         cpu_buffer_a->buffer = buffer_b;
4517 
4518         ret = 0;
4519 
4520 out_dec:
4521         atomic_dec(&cpu_buffer_a->record_disabled);
4522         atomic_dec(&cpu_buffer_b->record_disabled);
4523 out:
4524         return ret;
4525 }
4526 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
4527 #endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
4528 
4529 /**
4530  * ring_buffer_alloc_read_page - allocate a page to read from buffer
4531  * @buffer: the buffer to allocate for.
4532  * @cpu: the cpu buffer to allocate.
4533  *
4534  * This function is used in conjunction with ring_buffer_read_page.
4535  * When reading a full page from the ring buffer, these functions
4536  * can be used to speed up the process. The calling function should
4537  * allocate a few pages first with this function. Then when it
4538  * needs to get pages from the ring buffer, it passes the result
4539  * of this function into ring_buffer_read_page, which will swap
4540  * the page that was allocated, with the read page of the buffer.
4541  *
4542  * Returns:
4543  *  The page allocated, or ERR_PTR
4544  */
4545 void *ring_buffer_alloc_read_page(struct ring_buffer *buffer, int cpu)
4546 {
4547         struct ring_buffer_per_cpu *cpu_buffer;
4548         struct buffer_data_page *bpage = NULL;
4549         unsigned long flags;
4550         struct page *page;
4551 
4552         if (!cpumask_test_cpu(cpu, buffer->cpumask))
4553                 return ERR_PTR(-ENODEV);
4554 
4555         cpu_buffer = buffer->buffers[cpu];
4556         local_irq_save(flags);
4557         arch_spin_lock(&cpu_buffer->lock);
4558 
4559         if (cpu_buffer->free_page) {
4560                 bpage = cpu_buffer->free_page;
4561                 cpu_buffer->free_page = NULL;
4562         }
4563 
4564         arch_spin_unlock(&cpu_buffer->lock);
4565         local_irq_restore(flags);
4566 
4567         if (bpage)
4568                 goto out;
4569 
4570         page = alloc_pages_node(cpu_to_node(cpu),
4571                                 GFP_KERNEL | __GFP_NORETRY, 0);
4572         if (!page)
4573                 return ERR_PTR(-ENOMEM);
4574 
4575         bpage = page_address(page);
4576 
4577  out:
4578         rb_init_page(bpage);
4579 
4580         return bpage;
4581 }
4582 EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
4583 
4584 /**
4585  * ring_buffer_free_read_page - free an allocated read page
4586  * @buffer: the buffer the page was allocate for
4587  * @cpu: the cpu buffer the page came from
4588  * @data: the page to free
4589  *
4590  * Free a page allocated from ring_buffer_alloc_read_page.
4591  */
4592 void ring_buffer_free_read_page(struct ring_buffer *buffer, int cpu, void *data)
4593 {
4594         struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4595         struct buffer_data_page *bpage = data;
4596         struct page *page = virt_to_page(bpage);
4597         unsigned long flags;
4598 
4599         /* If the page is still in use someplace else, we can't reuse it */
4600         if (page_ref_count(page) > 1)
4601                 goto out;
4602 
4603         local_irq_save(flags);
4604         arch_spin_lock(&cpu_buffer->lock);
4605 
4606         if (!cpu_buffer->free_page) {
4607                 cpu_buffer->free_page = bpage;
4608                 bpage = NULL;
4609         }
4610 
4611         arch_spin_unlock(&cpu_buffer->lock);
4612         local_irq_restore(flags);
4613 
4614  out:
4615         free_page((unsigned long)bpage);
4616 }
4617 EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
4618 
4619 /**
4620  * ring_buffer_read_page - extract a page from the ring buffer
4621  * @buffer: buffer to extract from
4622  * @data_page: the page to use allocated from ring_buffer_alloc_read_page
4623  * @len: amount to extract
4624  * @cpu: the cpu of the buffer to extract
4625  * @full: should the extraction only happen when the page is full.
4626  *
4627  * This function will pull out a page from the ring buffer and consume it.
4628  * @data_page must be the address of the variable that was returned
4629  * from ring_buffer_alloc_read_page. This is because the page might be used
4630  * to swap with a page in the ring buffer.
4631  *
4632  * for example:
4633  *      rpage = ring_buffer_alloc_read_page(buffer, cpu);
4634  *      if (IS_ERR(rpage))
4635  *              return PTR_ERR(rpage);
4636  *      ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
4637  *      if (ret >= 0)
4638  *              process_page(rpage, ret);
4639  *
4640  * When @full is set, the function will not return true unless
4641  * the writer is off the reader page.
4642  *
4643  * Note: it is up to the calling functions to handle sleeps and wakeups.
4644  *  The ring buffer can be used anywhere in the kernel and can not
4645  *  blindly call wake_up. The layer that uses the ring buffer must be
4646  *  responsible for that.
4647  *
4648  * Returns:
4649  *  >=0 if data has been transferred, returns the offset of consumed data.
4650  *  <0 if no data has been transferred.
4651  */
4652 int ring_buffer_read_page(struct ring_buffer *buffer,
4653                           void **data_page, size_t len, int cpu, int full)
4654 {
4655         struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4656         struct ring_buffer_event *event;
4657         struct buffer_data_page *bpage;
4658         struct buffer_page *reader;
4659         unsigned long missed_events;
4660         unsigned long flags;
4661         unsigned int commit;
4662         unsigned int read;
4663         u64 save_timestamp;
4664         int ret = -1;
4665 
4666         if (!cpumask_test_cpu(cpu, buffer->cpumask))
4667                 goto out;
4668 
4669         /*
4670          * If len is not big enough to hold the page header, then
4671          * we can not copy anything.
4672          */
4673         if (len <= BUF_PAGE_HDR_SIZE)
4674                 goto out;
4675 
4676         len -= BUF_PAGE_HDR_SIZE;
4677 
4678         if (!data_page)
4679                 goto out;
4680 
4681         bpage = *data_page;
4682         if (!bpage)
4683                 goto out;
4684 
4685         raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4686 
4687         reader = rb_get_reader_page(cpu_buffer);
4688         if (!reader)
4689                 goto out_unlock;
4690 
4691         event = rb_reader_event(cpu_buffer);
4692 
4693         read = reader->read;
4694         commit = rb_page_commit(reader);
4695 
4696         /* Check if any events were dropped */
4697         missed_events = cpu_buffer->lost_events;
4698 
4699         /*
4700          * If this page has been partially read or
4701          * if len is not big enough to read the rest of the page or
4702          * a writer is still on the page, then
4703          * we must copy the data from the page to the buffer.
4704          * Otherwise, we can simply swap the page with the one passed in.
4705          */
4706         if (read || (len < (commit - read)) ||
4707             cpu_buffer->reader_page == cpu_buffer->commit_page) {
4708                 struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
4709                 unsigned int rpos = read;
4710                 unsigned int pos = 0;
4711                 unsigned int size;
4712 
4713                 if (full)
4714                         goto out_unlock;
4715 
4716                 if (len > (commit - read))
4717                         len = (commit - read);
4718 
4719                 /* Always keep the time extend and data together */
4720                 size = rb_event_ts_length(event);
4721 
4722                 if (len < size)
4723                         goto out_unlock;
4724 
4725                 /* save the current timestamp, since the user will need it */
4726                 save_timestamp = cpu_buffer->read_stamp;
4727 
4728                 /* Need to copy one event at a time */
4729                 do {
4730                         /* We need the size of one event, because
4731                          * rb_advance_reader only advances by one event,
4732                          * whereas rb_event_ts_length may include the size of
4733                          * one or two events.
4734                          * We have already ensured there's enough space if this
4735                          * is a time extend. */
4736                         size = rb_event_length(event);
4737                         memcpy(bpage->data + pos, rpage->data + rpos, size);
4738 
4739                         len -= size;
4740 
4741                         rb_advance_reader(cpu_buffer);
4742                         rpos = reader->read;
4743                         pos += size;
4744 
4745                         if (rpos >= commit)
4746                                 break;
4747 
4748                         event = rb_reader_event(cpu_buffer);
4749                         /* Always keep the time extend and data together */
4750                         size = rb_event_ts_length(event);
4751                 } while (len >= size);
4752 
4753                 /* update bpage */
4754                 local_set(&bpage->commit, pos);
4755                 bpage->time_stamp = save_timestamp;
4756 
4757                 /* we copied everything to the beginning */
4758                 read = 0;
4759         } else {
4760                 /* update the entry counter */
4761                 cpu_buffer->read += rb_page_entries(reader);
4762                 cpu_buffer->read_bytes += BUF_PAGE_SIZE;
4763 
4764                 /* swap the pages */
4765                 rb_init_page(bpage);
4766                 bpage = reader->page;
4767                 reader->page = *data_page;
4768                 local_set(&reader->write, 0);
4769                 local_set(&reader->entries, 0);
4770                 reader->read = 0;
4771                 *data_page = bpage;
4772 
4773                 /*
4774                  * Use the real_end for the data size,
4775                  * This gives us a chance to store the lost events
4776                  * on the page.
4777                  */
4778                 if (reader->real_end)
4779                         local_set(&bpage->commit, reader->real_end);
4780         }
4781         ret = read;
4782 
4783         cpu_buffer->lost_events = 0;
4784 
4785         commit = local_read(&bpage->commit);
4786         /*
4787          * Set a flag in the commit field if we lost events
4788          */
4789         if (missed_events) {
4790                 /* If there is room at the end of the page to save the
4791                  * missed events, then record it there.
4792                  */
4793                 if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
4794                         memcpy(&bpage->data[commit], &missed_events,
4795                                sizeof(missed_events));
4796                         local_add(RB_MISSED_STORED, &bpage->commit);
4797                         commit += sizeof(missed_events);
4798                 }
4799                 local_add(RB_MISSED_EVENTS, &bpage->commit);
4800         }
4801 
4802         /*
4803          * This page may be off to user land. Zero it out here.
4804          */
4805         if (commit < BUF_PAGE_SIZE)
4806                 memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
4807 
4808  out_unlock:
4809         raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4810 
4811  out:
4812         return ret;
4813 }
4814 EXPORT_SYMBOL_GPL(ring_buffer_read_page);
4815 
4816 /*
4817  * We only allocate new buffers, never free them if the CPU goes down.
4818  * If we were to free the buffer, then the user would lose any trace that was in
4819  * the buffer.
4820  */
4821 int trace_rb_cpu_prepare(unsigned int cpu, struct hlist_node *node)
4822 {
4823         struct ring_buffer *buffer;
4824         long nr_pages_same;
4825         int cpu_i;
4826         unsigned long nr_pages;
4827 
4828         buffer = container_of(node, struct ring_buffer, node);
4829         if (cpumask_test_cpu(cpu, buffer->cpumask))
4830                 return 0;
4831 
4832         nr_pages = 0;
4833         nr_pages_same = 1;
4834         /* check if all cpu sizes are same */
4835         for_each_buffer_cpu(buffer, cpu_i) {
4836                 /* fill in the size from first enabled cpu */
4837                 if (nr_pages == 0)
4838                         nr_pages = buffer->buffers[cpu_i]->nr_pages;
4839                 if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
4840                         nr_pages_same = 0;
4841                         break;
4842                 }
4843         }
4844         /* allocate minimum pages, user can later expand it */
4845         if (!nr_pages_same)
4846                 nr_pages = 2;
4847         buffer->buffers[cpu] =
4848                 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
4849         if (!buffer->buffers[cpu]) {
4850                 WARN(1, "failed to allocate ring buffer on CPU %u\n",
4851                      cpu);
4852                 return -ENOMEM;
4853         }
4854         smp_wmb();
4855         cpumask_set_cpu(cpu, buffer->cpumask);
4856         return 0;
4857 }
4858 
4859 #ifdef CONFIG_RING_BUFFER_STARTUP_TEST
4860 /*
4861  * This is a basic integrity check of the ring buffer.
4862  * Late in the boot cycle this test will run when configured in.
4863  * It will kick off a thread per CPU that will go into a loop
4864  * writing to the per cpu ring buffer various sizes of data.
4865  * Some of the data will be large items, some small.
4866  *
4867  * Another thread is created that goes into a spin, sending out
4868  * IPIs to the other CPUs to also write into the ring buffer.
4869  * this is to test the nesting ability of the buffer.
4870  *
4871  * Basic stats are recorded and reported. If something in the
4872  * ring buffer should happen that's not expected, a big warning
4873  * is displayed and all ring buffers are disabled.
4874  */
4875 static struct task_struct *rb_threads[NR_CPUS] __initdata;
4876 
4877 struct rb_test_data {
4878         struct ring_buffer      *buffer;
4879         unsigned long           events;
4880         unsigned long           bytes_written;
4881         unsigned long           bytes_alloc;
4882         unsigned long           bytes_dropped;
4883         unsigned long           events_nested;
4884         unsigned long           bytes_written_nested;
4885         unsigned long           bytes_alloc_nested;
4886         unsigned long           bytes_dropped_nested;
4887         int                     min_size_nested;
4888         int                     max_size_nested;
4889         int                     max_size;
4890         int                     min_size;
4891         int                     cpu;
4892         int                     cnt;
4893 };
4894 
4895 static struct rb_test_data rb_data[NR_CPUS] __initdata;
4896 
4897 /* 1 meg per cpu */
4898 #define RB_TEST_BUFFER_SIZE     1048576
4899 
4900 static char rb_string[] __initdata =
4901         "abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
4902         "?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
4903         "!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
4904 
4905 static bool rb_test_started __initdata;
4906 
4907 struct rb_item {
4908         int size;
4909         char str[];
4910 };
4911 
4912 static __init int rb_write_something(struct rb_test_data *data, bool nested)
4913 {
4914         struct ring_buffer_event *event;
4915         struct rb_item *item;
4916         bool started;
4917         int event_len;
4918         int size;
4919         int len;
4920         int cnt;
4921 
4922         /* Have nested writes different that what is written */
4923         cnt = data->cnt + (nested ? 27 : 0);
4924 
4925         /* Multiply cnt by ~e, to make some unique increment */
4926         size = (data->cnt * 68 / 25) % (sizeof(rb_string) - 1);
4927 
4928         len = size + sizeof(struct rb_item);
4929 
4930         started = rb_test_started;
4931         /* read rb_test_started before checking buffer enabled */
4932         smp_rmb();
4933 
4934         event = ring_buffer_lock_reserve(data->buffer, len);
4935         if (!event) {
4936                 /* Ignore dropped events before test starts. */
4937                 if (started) {
4938                         if (nested)
4939                                 data->bytes_dropped += len;
4940                         else
4941                                 data->bytes_dropped_nested += len;
4942                 }
4943                 return len;
4944         }
4945 
4946         event_len = ring_buffer_event_length(event);
4947 
4948         if (RB_WARN_ON(data->buffer, event_len < len))
4949                 goto out;
4950 
4951         item = ring_buffer_event_data(event);
4952         item->size = size;
4953         memcpy(item->str, rb_string, size);
4954 
4955         if (nested) {
4956                 data->bytes_alloc_nested += event_len;
4957                 data->bytes_written_nested += len;
4958                 data->events_nested++;
4959                 if (!data->min_size_nested || len < data->min_size_nested)
4960                         data->min_size_nested = len;
4961                 if (len > data->max_size_nested)
4962                         data->max_size_nested = len;
4963         } else {
4964                 data->bytes_alloc += event_len;
4965                 data->bytes_written += len;
4966                 data->events++;
4967                 if (!data->min_size || len < data->min_size)
4968                         data->max_size = len;
4969                 if (len > data->max_size)
4970                         data->max_size = len;
4971         }
4972 
4973  out:
4974         ring_buffer_unlock_commit(data->buffer, event);
4975 
4976         return 0;
4977 }
4978 
4979 static __init int rb_test(void *arg)
4980 {
4981         struct rb_test_data *data = arg;
4982 
4983         while (!kthread_should_stop()) {
4984                 rb_write_something(data, false);
4985                 data->cnt++;
4986 
4987                 set_current_state(TASK_INTERRUPTIBLE);
4988                 /* Now sleep between a min of 100-300us and a max of 1ms */
4989                 usleep_range(((data->cnt % 3) + 1) * 100, 1000);
4990         }
4991 
4992         return 0;
4993 }
4994 
4995 static __init void rb_ipi(void *ignore)
4996 {
4997         struct rb_test_data *data;
4998         int cpu = smp_processor_id();
4999 
5000         data = &rb_data[cpu];
5001         rb_write_something(data, true);
5002 }
5003 
5004 static __init int rb_hammer_test(void *arg)
5005 {
5006         while (!kthread_should_stop()) {
5007 
5008                 /* Send an IPI to all cpus to write data! */
5009                 smp_call_function(rb_ipi, NULL, 1);
5010                 /* No sleep, but for non preempt, let others run */
5011                 schedule();
5012         }
5013 
5014         return 0;
5015 }
5016 
5017 static __init int test_ringbuffer(void)
5018 {
5019         struct task_struct *rb_hammer;
5020         struct ring_buffer *buffer;
5021         int cpu;
5022         int ret = 0;
5023 
5024         pr_info("Running ring buffer tests...\n");
5025 
5026         buffer = ring_buffer_alloc(RB_TEST_BUFFER_SIZE, RB_FL_OVERWRITE);
5027         if (WARN_ON(!buffer))
5028                 return 0;
5029 
5030         /* Disable buffer so that threads can't write to it yet */
5031         ring_buffer_record_off(buffer);
5032 
5033         for_each_online_cpu(cpu) {
5034                 rb_data[cpu].buffer = buffer;
5035                 rb_data[cpu].cpu = cpu;
5036                 rb_data[cpu].cnt = cpu;
5037                 rb_threads[cpu] = kthread_create(rb_test, &rb_data[cpu],
5038                                                  "rbtester/%d", cpu);
5039                 if (WARN_ON(IS_ERR(rb_threads[cpu]))) {
5040                         pr_cont("FAILED\n");
5041                         ret = PTR_ERR(rb_threads[cpu]);
5042                         goto out_free;
5043                 }
5044 
5045                 kthread_bind(rb_threads[cpu], cpu);
5046                 wake_up_process(rb_threads[cpu]);
5047         }
5048 
5049         /* Now create the rb hammer! */
5050         rb_hammer = kthread_run(rb_hammer_test, NULL, "rbhammer");
5051         if (WARN_ON(IS_ERR(rb_hammer))) {
5052                 pr_cont("FAILED\n");
5053                 ret = PTR_ERR(rb_hammer);
5054                 goto out_free;
5055         }
5056 
5057         ring_buffer_record_on(buffer);
5058         /*
5059          * Show buffer is enabled before setting rb_test_started.
5060          * Yes there's a small race window where events could be
5061          * dropped and the thread wont catch it. But when a ring
5062          * buffer gets enabled, there will always be some kind of
5063          * delay before other CPUs see it. Thus, we don't care about
5064          * those dropped events. We care about events dropped after
5065          * the threads see that the buffer is active.
5066          */
5067         smp_wmb();
5068         rb_test_started = true;
5069 
5070         set_current_state(TASK_INTERRUPTIBLE);
5071         /* Just run for 10 seconds */;
5072         schedule_timeout(10 * HZ);
5073 
5074         kthread_stop(rb_hammer);
5075 
5076  out_free:
5077         for_each_online_cpu(cpu) {
5078                 if (!rb_threads[cpu])
5079                         break;
5080                 kthread_stop(rb_threads[cpu]);
5081         }
5082         if (ret) {
5083                 ring_buffer_free(buffer);
5084                 return ret;
5085         }
5086 
5087         /* Report! */
5088         pr_info("finished\n");
5089         for_each_online_cpu(cpu) {
5090                 struct ring_buffer_event *event;
5091                 struct rb_test_data *data = &rb_data[cpu];
5092                 struct rb_item *item;
5093                 unsigned long total_events;
5094                 unsigned long total_dropped;
5095                 unsigned long total_written;
5096                 unsigned long total_alloc;
5097                 unsigned long total_read = 0;
5098                 unsigned long total_size = 0;
5099                 unsigned long total_len = 0;
5100                 unsigned long total_lost = 0;
5101                 unsigned long lost;
5102                 int big_event_size;
5103                 int small_event_size;
5104 
5105                 ret = -1;
5106 
5107                 total_events = data->events + data->events_nested;
5108                 total_written = data->bytes_written + data->bytes_written_nested;
5109                 total_alloc = data->bytes_alloc + data->bytes_alloc_nested;
5110                 total_dropped = data->bytes_dropped + data->bytes_dropped_nested;
5111 
5112                 big_event_size = data->max_size + data->max_size_nested;
5113                 small_event_size = data->min_size + data->min_size_nested;
5114 
5115                 pr_info("CPU %d:\n", cpu);
5116                 pr_info("              events:    %ld\n", total_events);
5117                 pr_info("       dropped bytes:    %ld\n", total_dropped);
5118                 pr_info("       alloced bytes:    %ld\n", total_alloc);
5119                 pr_info("       written bytes:    %ld\n", total_written);
5120                 pr_info("       biggest event:    %d\n", big_event_size);
5121                 pr_info("      smallest event:    %d\n", small_event_size);
5122 
5123                 if (RB_WARN_ON(buffer, total_dropped))
5124                         break;
5125 
5126                 ret = 0;
5127 
5128                 while ((event = ring_buffer_consume(buffer, cpu, NULL, &lost))) {
5129                         total_lost += lost;
5130                         item = ring_buffer_event_data(event);
5131                         total_len += ring_buffer_event_length(event);
5132                         total_size += item->size + sizeof(struct rb_item);
5133                         if (memcmp(&item->str[0], rb_string, item->size) != 0) {
5134                                 pr_info("FAILED!\n");
5135                                 pr_info("buffer had: %.*s\n", item->size, item->str);
5136                                 pr_info("expected:   %.*s\n", item->size, rb_string);
5137                                 RB_WARN_ON(buffer, 1);
5138                                 ret = -1;
5139                                 break;
5140                         }
5141                         total_read++;
5142                 }
5143                 if (ret)
5144                         break;
5145 
5146                 ret = -1;
5147 
5148                 pr_info("         read events:   %ld\n", total_read);
5149                 pr_info("         lost events:   %ld\n", total_lost);
5150                 pr_info("        total events:   %ld\n", total_lost + total_read);
5151                 pr_info("  recorded len bytes:   %ld\n", total_len);
5152                 pr_info(" recorded size bytes:   %ld\n", total_size);
5153                 if (total_lost)
5154                         pr_info(" With dropped events, record len and size may not match\n"
5155                                 " alloced and written from above\n");
5156                 if (!total_lost) {
5157                         if (RB_WARN_ON(buffer, total_len != total_alloc ||
5158                                        total_size != total_written))
5159                                 break;
5160                 }
5161                 if (RB_WARN_ON(buffer, total_lost + total_read != total_events))
5162                         break;
5163 
5164                 ret = 0;
5165         }
5166         if (!ret)
5167                 pr_info("Ring buffer PASSED!\n");
5168 
5169         ring_buffer_free(buffer);
5170         return 0;
5171 }
5172 
5173 late_initcall(test_ringbuffer);
5174 #endif /* CONFIG_RING_BUFFER_STARTUP_TEST */
5175 

~ [ 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