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