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