Linux Audio

Check our new training course

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